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 |
|---|---|---|---|
Arm order recognition in multi-armed bandit problem with laser chaos time series
By exploiting ultrafast and irregular time series generated by lasers with delayed feedback, we have previously demonstrated a scalable algorithm to solve multi-armed bandit (MAB) problems utilizing the time-division multiplexing of laser chaos time series. Although the algorithm detects the arm with the highest reward expectation, the correct recognition of the order of arms in terms of reward expectations is not achievable. Here, we present an algorithm where the degree of exploration is adaptively controlled based on confidence intervals that represent the estimation accuracy of reward expectations. We have demonstrated numerically that our approach did improve arm order recognition accuracy significantly, along with reduced dependence on reward environments, and the total reward is almost maintained compared with conventional MAB methods. This study applies to sectors where the order information is critical, such as efficient allocation of resources in information and communications technology.
www.nature.com/scientificreports/ there may be situations where compromises must be made, i.e., other channels will be selected. Now it is obvious that particular channel performance ranking information would be useful when considering non-best channels. Conversely, when there are no other users, a player (the single user) can simultaneously utilize top-ranking options to accelerate the communication ability, similar to the channel bonding in local area networks 10 . The purpose of this study is to accurately recognize the order of the expected rewards of different arms using a chaotic laser time series and to minimize the reduction of accumulated rewards due to too detailed exploration.
Principles
Definition and assumption. We consider a MAB problem in which a player selects one of K slot machines, where K = 2 M and M is a natural number. The K slot machines are distinguished by identities numbered from 0 to K − 1 , which are also represented in M-bit binary code given by S 1 S 2 . . . S M with S i ∈ {0, 1} ( i = 1, . . . , M ). For example, when K = 8 (or M = 3) , the slot machines are numbered by S 1 S 2 S 3 = {000, 001, . . . , 110, 111} . In this study, we assume that µ i = µ j if i = j , and we define the k-th max and k-th argmax operators as max k {} and arg max k {} . The variables used in the study are defined as described below: • X i (n) : Obtained reward from arm i at time step n independent at each time step. x i (n) is observed value. We estimate the arm order of reward expectations by calculating the sample mean of the accumulated reward at each time step. Specifically, the sample means of rewards obtained from arm i by time step n is calculated as follows: In each time step n, we estimated the arm j := arg max k iμ i (n) as the k-th best arm.
Time-division multiplexing of laser chaos. The proposed method is based on the MAB algorithm reported in 2018 6 . This method consists of the following steps: [STEP 1] decision making for each bit of the slot machines, [STEP 2] playing the selected slot machine, and [STEP 3] updating the threshold values.
[STEP 1] Decision for each bit of the slot machine. First, the chaotic signal s(t 1 ) measured at t = t 1 is compared to a threshold value denoted as TH 1 . If s(t 1 ) ≥ TH 1 , then bit S 1 is assigned 1. Otherwise, S 1 is assigned 0. To determine the value of S k (k = 2, . . . , M) , the chaotic signal s(t k ) measured at t = t k (> t k−1 ) is compared to a threshold value denoted as TH k,S 1 ...S k−1 . If s(t k ) ≥ TH k,S 1 ...S k−1 , then bit S k is assigned 1. Otherwise, S k is assigned 0. After this process, a slot machine with the number represented in a binary code S 1 . . . S M is selected.
[STEP 2] Slot machine play. Play the selected slot machine.
[STEP 3] Threshold values adjustment. If the selected slot machine yields a reward, then the threshold values are adjusted in a way that the same decision will be more likely to be selected. For example, if S 1 is assigned 0 and the player gets a reward, then TH 1 should be increased because doing so increases the likelihood of getting S 1 = 0 again. All of the other threshold values involved in determining the decision (i.e. TH 2,S 1 , . . . , TH M,S 1 ...S M−1 ) are updated in the same manner. If the selected slot machine does not yield a reward, then the threshold values are adjusted to make the same decision less likely to take place. For example, if S 1 is assigned 1 and the player does not get a reward, then TH 1 should be increased because of the decreased likelihood of getting S 1 = 1 . Again, all of the other threshold values involved in determining the decision (i.e. TH 2,S 1 , . . . , TH M,S 1 ...S M−1 ) are updated in the same manner.
Arm order recognition algorithm with confidence intervals. Confidence intervals. An overview of our proposed algorithm is shown in Fig. 1a. For each threshold value TH j,b 1 ...b j−1 ( j ∈ {1, . . . , M} , b 1 , . . . , b j−1 ∈ {0, 1} ) and z ∈ {0, 1} , the following values P (z; n) and C(z; n) are calculated: represents a subset of machine arms. If machine i can be selected when the signal s(t j ) is more than www.nature.com/scientificreports/ machine i can be selected when the signal s(t j ) is less than or equal to TH j,b 1 ...b j−1 , then i is included in I j,b 1 ...b j−1 (0) . Otherwise, i is not included in I j,b 1 ...b j−1 (0) . For example, in the case of an eight-armed bandit problem ( Fig. 1b): represents the sample means of rewards obtained from machines in I j,b 1 ...b j−1 (z) . C j,b 1 ...b j−1 (z; n) represents the confidence interval width of the estimated value P j,b 1 ...b j−1 (z; n) . The lower C(z; n), the higher the estimation accuracy. Parameter γ indicates the degree of exploration : a higher γ means that more exploration is needed to reach a given confidence interval width.
Coarseness/fineness of exploration adjustments by confidence intervals. At each threshold TH j,b 1 ...b j−1 , if the two intervals are overlapped, we suppose there is a likelihood of a change in the order relationship between P (0; n) and P (1; n) ; that is, the order of P (0; n) and P (1; n) is not known yet. Therefore, the exploration process should be executed more carefully. Hence, the threshold value should be closer to 0, which is a balanced situation, or we should perform further exploration, so that the threshold adjustment becomes finer. Conversely, if the two intervals are not overlapped, then we suppose a low likelihood of a wrong estimate of the order relationship between P (0; n) and P (1; n) . Hence, we should continue exploration more coarsely so that the threshold adjustment will be accelerated (Fig. 1c).
Results
Experimental settings. We have evaluated the performance of the methods for two cases: a four-armed bandit and an eight-armed bandit. First, the reward probability of each arm is assumed to follow the Bernoulli distribution: www.nature.com/scientificreports/ following conditions: In this experiment, a variety of assignments of reward probabilities ν satisfying the above conditions were prepared, and the performance was evaluated under every reward environment ν . We have defined the reward, regret, and correct order rate (COR) as metrics to quantitatively evaluate the performance of the method.
where n denotes number of time steps, t i (n) is the number of selections of arm i up to time step n, and l m represents the number of measurements in one reward environment ν . For the accuracy of arm order recognition, we considered the estimation accuracy of the top four arms regardless of the total number of arms. We prepared all 144 reward environments ν (all combinations satisfying the above conditions and max i =j |µ i − µ j | = 0.3 ) for the four-armed bandit problems and 100 randomly selected reward environments for the eight-armed bandit problems. The performances of four methods were compared: RoundRobin (all arms are selected in order at each time step), UCB1 (method for maximizing the total rewards proposed in 2002 11 ), Chaos (previous method using the laser chaos time series 6 , only finding the best arm, not recognizing the order), and Chaos-CI (proposed method using laser chaos time series and with confidence intervals). The details of UCB1 used in the present study are described in the Methods section. The purpose of this study is to extend the existing Chaos method to recognize the arm order. We should consider the trade-off between order recognition and reward maximization. As introduced above, RoundRobin and UCB1 were considered to examine quantitative performance analysis. RoundRobin systematically accomplishes the order recognition whereas UCB1 is known to achieve O(log n) regret at time n. We consider that these are appropriate and contrasting representative methods in the literature to examine the trade-off and the essential interest of the present study. Meanwhile, comparison to other bandit algorithms such as Thompson sampling 12 or arm elimination 13 is expected to trigger stimulating future discussions, leading to further improvement of the proposed Chaos-CI algorithm.
Evaluation under one reward environment. The curves in Fig. 2a and b show the time evolutions of regret(n) and COR(n) , respectively, over l m = 12, 000 measurements under specific reward environments ν = (µ 0 , . . . , µ K−1 ) . Specifically, columns (i) and (ii) pertain to the four-armed bandit problems defined by ν = (0.9, 0.8, 0.7, 0.6) x (l) a (l) (s) (s) , . The curves were colour coded for an easy method comparison. In the arm order recognition, Chaos-CI and RoundRobin presented high accuracy in the early time step. In terms of total reward, Chaos and UCB1 achieved the greatest rewards. The postconvergence behavior of Chaos and Chaos-CI is not necessarily the same. The reason is that the parameters and that determine the scale of threshold change are fixed values in Chaos, whereas they change adaptively according to past reward information in Chaos-CI.
Evaluation of the whole reward environments. Figure 3a summarizes the relationship between total rewards and order estimation accuracy: x-axis represents the normalized reward reward † (n) , whereas y-axis represents the COR COR(n) . Here, a normalized reward is defined as follows: Each plot in the graph indicates reward † (n) and COR(n) at time step n = 10,000 under one reward environment ν: Figure 3b shows the time evolution of the average value of each metric over the whole ensemble of reward environments from n = 1 to n = 10,000: Compared to UCB1, Chaos-CI can recognize the arm order faster. On the other hand, Chaos-CI can get more rewards than RoundRobin.
, (1 ≤ n ≤ 10,000). (reward † ν (10,000), COR ν (10,000)) . The more the scatter plot is at the top of the graph, the higher the order estimation accuracy is, and the more the scatter plot is at the right, the greater the obtained reward is. (b) Time evolution of the average value of each metric over the whole ensemble of reward environments ( 1 ≤ n ≤ 10,000).
Discussion
Difficulty of maximizing rewards and arm order recognition. The results of the numerical simulations on the four-armed and eight-armed bandit problems show similar trends: there is a trade-off between the maximized total rewards and arm order recognition. As RoundRobin selects all arms equally, we always achieve a perfect COR at a time step n = 10,000 for any given reward environment. However, we cannot maximize rewards because regret linearly increases with time. On the contrary, in Chaos, we achieved normalized rewards of almost unity at the time step of n = 10,000 with respect to many types of reward environments. However, we can observe inferior performances regarding the arm order recognition accuracy because the arm selection is greatly biased to the best arm. In terms of the COR, the COR on RoundRobin and Chaos-CI (proposed method) quickly converged to unity. In terms of the total rewards, Chaos (previous method) and UCB1 are more active in using the exploitation principle to obtain greater rewards. The proposed method, Chaos-CI, achieves an outstanding performance on the arm order recognition and reward.
Number of arm selections. Figure 4a www.nature.com/scientificreports/ in a linear order. Therefore, the arm order recognition accuracy is faster than UCB1. Although the selections of non-top arms in the linear order cause regret to increase in a linear order, the slope of the linear-order regret is significantly decreased compared with that of RoundRobin by selecting better arms more often or by prioritizing the search (i.e. T [1] (n) > · · · > T [K] (n)). Figs. 3 and 4, the performances of Chaos are very different depending on reward environments ν 1 and ν 2 . This finding is clearly linked with the arm selection number T i (n) . In reward environment ν 1 , all T i (n) evolve in a linear order, but in reward environment ν 2 , T i (n) (i = i * ) is approximately 100 at time step n = 50,000 . Thus, the performance of Chaos heavily depends on the given reward environment. Table 1 summarizes the sample variance of metrics over 100 reward environments in an eight-armed bandit. Our proposed algorithm (Chaos-CI) depends on the order of the arms. If the difference between the average reward of each branch is large, we consider that the ranking estimation is easy. However, from Fig. 3(a) and Table 1, we observe that the proposed method always estimates the rankings with high accuracy regardless of the order of the arms. This estimation accuracy outperforms the performance of UCB1, which is an algorithm that does not depend on the order of the arms. In terms of obtained rewards, Chaos-CI has a larger variance than UCB1 and Chaos but is more stable than RoundRobin.
Environment dependency. As shown in
In the experiments, the expected reward µ i is limited so that the difference in the estimated difficulty does not vary drastically from problem to problem because the larger the difference in the expected reward value of each arm, the easier the problem becomes to solve. Meanwhile, if the difference in reward expectation for each arm becomes even smaller (specifically, smaller than 0.1), the correct order recognition in its exact sense will be significantly challenging. At the same time though, such a case means that there are not significant reward differences regardless of which arm is pulled. Hence, we consider that the evaluation method or the definition of correct order recognition may need revision. We expect these points to form the basis of interesting future studies.
On the other hand, if the reward distribution is more diverse, for example, [0.95, 0.9, 0.6, 0.5], UCB1, which aims to maximize the cumulative reward, will stop selecting the lower-reward-probability arms at an early exploration phase, leading to the degradation of the accuracy of rank recognition. Conversely, since the proposed method adjusts the thresholds based on the confidence intervals in all branches, it is expected that the rank recognition accuracy will not be degraded.
Conclusions
In this study, we have examined ultrafast decision making with laser chaos time series in reinforcement learning (e.g. MAB) and set a goal to recognize the arm order of reward expectations by expanding the previous method, that is, time-division multiplexing of laser chaos recordings. In the proposed method, we have introduced exploration-degree adjustments based on confidence intervals of estimated rewards. The results of the numerical simulations based on experimental time series show that the selection number of each arm increases linearly, leading to a high and rapid order recognition accuracy. Furthermore, arms with higher reward expectations are selected more frequently; hence, the slope of regret is reduced, although the selection number of an arm still linearly increases. Compared with UCB1 and Chaos, Chaos-CI (proposed method) is less dependent on the reward environment, indicating its potential significance in terms of robustness to environmental changes. In other words, Chaos-CI can make more accurate and stable estimates of arm order. Meanwhile, expressing the accuracy of rank estimation in terms of earned rewards in a single metric is an interesting, important, and challenging problem. We plan to explore this in our future research. Such an order recognition is useful in applications, such as channel selection and resource allocation in information and communications technology, where compromise actions or intelligent arbitrations are expected.
Methods
Optical system. The device used was a distributed feedback semiconductor laser mounted on a butterfly package with optical fibre pigtails (NTT Electronics, KELD1C5GAAA). The injection current of the semiconductor laser was set to 58.5 mA (5.37I th ), where the lasing threshold I th was 10.9 mA. The relaxation oscillation frequency of the laser was 6.5 GHz, and its temperature was maintained at 294.83 K. The optical output power was 13.2 mW. The laser was connected to a variable fibred reflector through a fibre coupler, where a fraction of light was reflected back to the laser, generating high-frequency chaotic oscillations of optical intensity 3,14,15 . The length of the fibre between the laser and reflector was 4.55 m, corresponding to a feedback delay time (round trip) of 43.8 ns. Polarization maintaining fibres were used for all of the optical fibre components. The optical signal was detected by a photodetector (New Focus, 1474-A, 38 GHz bandwidth) and sampled using a digital oscilloscope (Tektronics, DPO73304D, 33 GHz bandwidth, 100 GSample/s, eight-bit vertical resolution). The RF spectrum of the laser was measured by an RF spectrum analyzer (Agilent, N9010A-544, 44 GHz bandwidth). www.nature.com/scientificreports/ USB1 algorithm. In UCB1 11 , we select the arm j that maximize the score based on where X j (n) is the expected average reward obtained from arm j, T j (n) is the number of times machine j is played so far, and n is the total numbers of plays so far.
Details of the time-division multiplexing algorithm.
Parameters setting. In the experiments, we set the parameters in Algorithm 1 as follows: α = 0.99 , = 1 , = 0.1 . These are the same values as the previous experiment 6 . The signal s(τ ) is represented by an 8-bit integer type: −128 ≤ s(τ ) < 128.
Convergence of Algorithm 1 based on uniform distribution. This discussion on the convergence concerns only the two-armed bandit, while the random sequences are uniformly distributed and independent each timesomething that does not concern chaotic time sequences. We assume that K = 2 and the time series used for comparison with thresholds follows a uniform distribution of [−1/2, 1/2] at an arbitrary time. We define the value of threshold TH 1 at the beginning of time step n as w(n). The time evolution of w(n) can be represented as The expectation of w(n) is represented as follows.
Because we assume that s(t) follows a uniform distribution, if max{n�, n�} < 1/2, In this case, one of the arms will be selected intensively as time passes.
The above discussion shows that the convergence and performance of Algorithm 1 depend on learning rate α , exploration degree (�, �) , and reward environment (µ 0 , µ 1 ).
Dependency for scale parameter of confidence intervals. Figure 5 shows the influence of the parameter of confidence intervals. γ is a parameter related to the width of confidence intervals. We can see that the correct order rate becomes higher and the obtained rewards smaller as γ becomes smaller, and vice versa. When γ = √ 2 , the reward and the correct order rate are relatively high; we use this value for γ in Chaos-CI described in the main text.
Convergence of the proposed method based on uniform distribution. As described above, we have found that the performance of the algorithm proposed is heavily dependent on parameters (�, �) . Therefore, in the proposed method, exploration-degree adjustments based on confidence intervals are added to Algorithm 1: if the exploration itself is not sufficient, then thresholds are set close to 0 and values of (�, �) decrease, so thresholds are less likely to diverge, which leads to improved accuracy. If exploration is applied sufficiently, then the values of (�, �) increase, so the thresholds are more likely to diverge, which leads to an intensive selection of a better arm and slow increase of regret.
Data availability
The datasets generated during the current study are available from the corresponding author on reasonable request. | 4,964.4 | 2020-05-26T00:00:00.000 | [
"Physics",
"Engineering",
"Computer Science"
] |
Improvement of Emergency Communication Systems Using Drones in 5G and Beyond for Safety Applications
Drones are used for public safety missions because of their communication capabilities, unmanned mission, flexible deployment, and low cost. Recently, drone-assisted emergency communication systems in disasters have been developed where instead of a single large drone, flying ad hoc networks (FANETs) are proposed through clustering. Although cluster size has an impact on the proposed system's performance, no method is provided to effectively regulate cluster size. In this paper, optimum cluster size is obtained through two distinct meta-heuristic optimization algorithms - the Cuckoo Search Algorithm (CUCO) and the Particle Swarm Algorithm (PSO). Flowcharts and algorithms of CUCO and PSO are provided. A presentation of an analytical investigation based on the Markov chain model is provided. To further validate the analytical study, simulation results are presented. Simulation shows the improvement in terms of throughput and packet dropping rate (PDR).
INTRODUCTION
Unmanned aerial vehicles (UAVs), often referred to as drones, have garnered attention in recent times owing to their versatility, independence, and wide range of applicable industries [1][2][3][4].Indeed, a multitude of applications, including those in the telecommunications, military, monitoring and surveillance, search and rescue operations, medical supply delivery, and disaster management, have been seen as being significantly facilitated by drones [5][6][7][8].Over the course of the next ten years, wireless networks that support 5G and 6G are anticipated to be crucial in supporting the fast growing and pervasive use of drones in a broad range of applications [9][10][11][12][13][14].The seamless, high-definition, instantaneous, always-on connection that customers need is the aim of the 6G.6G is expected to meet the high demands of connected devices and automation systems, such as drones and driverless automobiles, in terms of energy economy, latency, data throughput, and dependability.A complete architecture with integrated terrestrial and non-terrestrial networks is part of this [15][16][17][18].Drones are essential in many different scenarios and use cases, some of which may go beyond 5G and 6G.Drones are used in a variety of applications, including remote construction, real-time surveillance, package delivery, and media production.The use of drones has significantly increased.
Rearranging the network in the event of a complex issue that arises in communication systems after natural catastrophes, such earthquakes and floods, takes a lot of time, and the most important thing is to keep people safe.In a situation like this, swift and effective search and rescue operations are desperately required.The first 72 hours of any crisis are crucial for prompt action, and effective search and rescue operations are the only way to meet this need.Nevertheless, a snag in publicity and communication will hinder their efforts.Drone-assisted emergency networks during catastrophes have been devised recently [1], where clustering is suggested to create flying ad hoc networks (FANETs).Ad hoc networks composed of many UAVs are referred to as UAVs ad hoc networks, or flying ad hoc networks, or FANETs for short.The usage of FANETs is seen in Fig. 1.A cluster head (CH) will be present in each cluster, and it will be linked to the emergency communication vehicle (ECV).Through the ECV, CH will be in charge of facilitating communication amongst cluster members (CMs) both inside and outside of the cluster.Each and every kind of communication that occurs within, outside, and between clusters enters via the CH.We demonstrated that the suggested approach performs better than current techniques in our earlier work [1].Cluster size does, however, have an impact on system performance.Performance is dependent on cluster size since factors like as collision probability, channel congestion, and packet loss are all influenced by the number of drones in the cluster.Performance drastically suffers if the cluster size is too big since there will be a lot of packet collisions when the drone count is too high.However, due to a lack of drones, or cluster members (CMs), a tiny cluster was unable to use the radio resources that were already in place.There is no technique to effectively regulate cluster size in the research [1].In this research, we use meta-heuristic techniques to optimize cluster size and hence increase system performance.
Natural intelligence, or artificial intelligence, algorithms have been widely used as search and optimization techniques in a variety of fields lately, including science, engineering, and business [17][18][19].The process of determining the best course of action under certain restrictions for a given goal or objective is known as optimization.Scientists have proposed a novel idea that has been validated via optimization.The goal of optimization is to always get the best results.The research, solution form, and acceptable tolerance are the main points of emphasis for the strongest interpretation.In an effort to address the difficulties encountered in the past, several optimization techniques have been created and used to various disciplines [20].It is usual practice to use mathematical or classical methodologies in the formulation of optimization issues.Because of these techniques' drawbacks-namely, their inelasticity and the need to identify them using mathematical functions-scientists are drawn to develop high-performance, general-purpose alternatives that draw inspiration from natural occurrences.Particle Swarm Optimization (PSO) and Cuckoo Search Algorithm (CUCO) are widely employed in optimization issues [20][21][22][23][24].
The contributions of this study is summarized as follows: drone-assisted emergency communication systems in disasters is presented where clustering is used.The performance of the proposed system is improved through two meta-heuristic algorithms -CUCO and PSO.The cluster size is optimized by algorithms.The Markov chain model is used for an analytical investigation.The analytical analyses are backed up by simulation results, which are also presented.A comparison with the previous study [1], CUCO, and PSO is presented.Simulation shows the improvement in terms of throughput and packet dropping rate (PDR).
The remainder of the paper is structured as follows: Section II outlines the meta-heuristic optimization algorithms.Section III presents the performance analysis.Simulation results are discussed in Section IV.Lastly, Section V provides the conclusions.
A. Cuckoo Search Algorithm (CUCO)
Herd intelligence is the primary focus of the CUCO algorithm.Drones are shown functioning in ad hoc networks.The goal of the CUCO algorithm is to allow drones to communicate with one another socially.The search terminates after the genetic algorithm has finished the required number of generations.Each cluster size (CS) optimizes its size based on historical data to achieve optimal performance.It is the best-positioned drones in ad hoc networks that the CUCO algorithm primarily focuses on estimating.with a randomly determined cluster size will improve their position relative to previous iterations until the objective is met.The CUCO method has been successfully used to a variety of optimization problems [24].In the first algorithm, we see the CUCO method applied to FANETs. Figure 2 shows the flowchart of CUCO algorithm.
In essence, the following phases make up the algorithm: i.An arbitrary beginning CS is used to form a starting swarm.
ii.All drone values are exchanged inside FANETs.
iii.Each drone has a local best of the current generation (pbest).There are as many best in a pack as there are drones.iv.In modern ad hoc networks, the global best (gbest) is derived from the local best.v. Following is an update to the CS.Here, Qid is position and CSid is cluster size values, while and values are random generated numbers.The value of inertial weight is w and z1, z2 are scaling factors.
vi. Steps 2, 3, 4, and 5 should be repeated until the condition for termination is met. .
B. Particle Swarm Optimization (PSO)
PSO is essentially a herd intelligence-based algorithm.FANET-operating drones have been seen.The PSO algorithm relies on drones exchanging social information with one another.The genetic algorithm's generation count determines the search strategy.Using its prior expertise, each CS modifies its position to find the optimal spot on the course.The PSO method is primarily focused on approximating the drones' positions inside the FANET to the drones that have the best FANET positions.The condition in which this approach CS occurs is arbitrary, and often the drones in the herd are in a better position than they were in their prior moves.This procedure is repeated until the goal is obtained.Numerous optimization issues have seen the effective use of the PSO algorithm [25].The PSO method for FANETs is presented in method 2. Figure 3 displays the PSO flowchart.
The following phases make up the algorithm in its entirety: i.A startup herd is created with randomly chosen starting CS.
ii.All of the FANET's drones' conformance values are computed.
iii.The local best (pbest) of the current generation is assigned to each drone.The number of drones is equal to the number of the best in the herd.iv.In the current FANET, the local bests are used to choose the global best (gbest).
Here, Qid position and id V CS values, while 1 rand and 2 rand values are numbers that are generated randomly.w is the inertial weight value and z1, z2 are scaling factors.
III. PERFORMANCE ANALYSIS
Because UAVs can move in three dimensions (3D), obstacles in their route may be easily avoided.A 3D space with movement is shown in Figure 4.For the length of the time T, it is expected that the height h will not change.(x(t), y(t), h) is the location of the UAV represented in terms of time-varying x and y coordinates, x(t) and y(t).The exact positions of the UAV are determined by launch and recovery sites, often known as pre-and post-mission locations.Let (xd, yd, h) represent the destination and (xs, ys, h) the starting point.The symbol d will represent the distance between the starting and destination points.One way to express the minimal .As a result, there exists at least one trajectory that connects the origin and the destination.
There must be one CH per cluster.There are NCL number of CHs as a consequence.The average count of CMs inside a cluster may be expressed as follows [1] 1.
At every given slot, there is a probability that a CM will transmit a packet, which can be given as .
ϕ represents the probability of an idle channel, which may be expressed as 1 (1 ) .
is the probability that a minimum of one CM is broadcasting on the channel during a certain slot.Since ( - 1) CMs are vying for the channel, may be expressed as .
The probability of a transmitted packet colliding is represented by ζ.A collision will happen if any of the remaining ( -2) CMs sends a packet within the same time period.ζ may be expressed as 2 1 (1 ) .
The probability that the current packet delivery on the channel will be successful is represented by η, which may be expressed as With a Poisson process and an arrival rate of ϖa, we may compute the probability of a packet arriving, denoted by σ as where βɛ represents the estimated time of a UAV in each Markov state.βɛ can be given as (1 ) , where the length of a packet transmission with a collision is known as βC, the period of a slot is known as βslot, and the interval of a successful packet transmission is known as βS.βS and βC can be given as , , DIF ay C S del (15) where the size of the packet is denoted by L, the length of the MAC and PHY headers is denoted by Lh, and the data transmission rate is denoted by Rd.The duration of the ACK, or βACK, and the propagation delay is denoted by βdelay.
Assume that μ is the normalized system throughput, which is determined by dividing the average length of a transmitted payload by the average slot time.For the i th cluster, the μ can be given as .
IV. SIMULATION RESULTS
This section assesses the effects of various meta-heuristic optimization strategies in FANETs.A comparative analysis is provided between the suggested system [1] and meta-heuristic optimization techniques.With MATLAB, the simulation results are achieved.Fig. 4 shows the throughput for 50 drones with a cluster size of 5.As the number of drones increases, throughput increases up to a certain point, beyond which it decreases further as more packets compete for the same channel and more collisions occur.Meta-heuristic optimization methods always have throughput that exceeds that of the proposed system [1].When there are fewer drones, the CUCO outperforms the PSO in terms of throughput.PSO algorithm throughput is greater than CUCO when the number of drones is large.Compared to the suggested system, meta-heuristic optimization techniques are more efficient.
The PDR vs the total number of drones is shown in Fig. 5.After the allotted number of attempts, a packet will be destroyed in the event that a transmission attempt is unsuccessful.CUCO has a lower PDR than PSO when there are fewer drones present.A higher drone count results in a lower PDR for PSO.The PDR of the meta-heuristic optimization techniques is consistently less than that of the suggested system.Consequently, it is evident that metaheuristic optimization techniques enhance communication efficiency by increasing throughput and lowering PDR.
V. CONCLUSION
After disaster public safety and security is the most important thing.In this study, we optimize cluster size in FANETs using meta-heuristic optimization techniques to get the best throughput performance for emergency communication systems using drones in disaster.To establish the link between the parameters, an analytical research based on Markov chains is drawn.The results of the simulation are shown.A comparison is shown between PSO and CUCO.It is clear that meta-heuristic optimization techniques improve communication reliability and efficiency.The throughput is highest for PSO when the number of drones is greater than 25.The PDR is also lowest for PSO until the number of drones is 45.It is apparent that PSO can be a better algorithm.
Drones in a herd
Table 1 .
Values of the parameters used in the simulation | 3,138.8 | 2023-11-13T00:00:00.000 | [
"Engineering",
"Environmental Science",
"Computer Science"
] |
C ○ Owned by the authors, published by EDP Sciences, 2013 Statistics for trajectometry
A trajectometer is made of layers of charged particle detectors which measure successive positions along the trajectories; it is generally immersed in a magnetic field, so the curvature of the trajectory provides a measurement of the momentum. A method to perform a progressive fitting of the trajectory (Kalman Filter formalism), incorporating the measurements one after one, with an optimal account for the perturbations (multiple scattering, energy loss), is described with some indications for practical implementations in realistic detector layouts. Useful byproducts of the method and tests of validity are discussed. The procedure appears to be a combination ad libitum of elementary operations on vectors and matrices of fixed dimension (the number of parameters needed to define the trajectory), affording very flexible strategies, including a coupling of the pattern recognition of tracks with the fit of the trajectory, and combination with calorimetric or timing measurements. Extension to non-gaussian errors is discussed.Once the trajectories of an event are independently reconstructed, they may be extrapolated back to the region of production of the particles (target, or zone of intersection of the beams in a collider) and associated to one or several vertices (primary interaction, and possible secondary interactions or decays): a fast and flexible method is described to perform these operations and improve the geometrical reconstruction, hence the kinematical one, by the constraint of a common origin; additional constraints may be added. Here again, the elementary steps consist in linear operations on vector and matrices of fixed dimension, allowing the user to easily proceed by successive trials and to optimize the strategy.
What is a "trajectometer"?
Most of the particle detectors include layers which aim at recognizing and reconstructing as precisely as possible the trajectories of the charged particles produced in the main interaction or in a secondary vertex, usually curved by a magnetic field to provide an information on the momentum.These layers are thin (in terms of radiation lengths) to avoid disturbing the movement.They are generally located close to the interaction point (inner part of the detector for an implementation around a collider, or in first position within a fixed target experiment), before the calorimetric components.However some elements may be set at remote positions to provide a "lever arm" effect for a precise measurement of the curvature: this is the case of external muon detectors, which give both a signature of muons and a relatively precise measurement of their position.A schematic layout of a trajectometer is shown in Fig. 1, in the case on planar layers; for a detector installed around the intersection region of a collider, a better description is obtained by replacing the planes by cylinders.from the interaction region.The solid black lines are layers of material, the dotted black lines are measurement layers.The measurements are the blue circles (with a radius proportional to the uncertainty).
What is known and what is searched for ?
The following properties of the detector are supposed to be known (or determined at a previous stage from calibration data): • the equation of propagation, which determines the shape of the trajectories.In practice, it is defined by the magnetic field map within the trajectometer.
• the nature and the precision of the measurements (drift time, amplitude on signals in strips or pads, etc).In simple cases they just give one coordinate or a fixed combination of coordinates.They may also depend on the direction of incidence on the detection layer.In practice, what is needed is to express the measured quantity as a function of the local parameters of the trajectory (position and direction).The distribution of errors is generally gaussian, but possible exceptions should be evaluated and accounted for.
• the properties of each particle (momentum, direction) at its origin.In practice, we want to obtain a backward extrapolation to the interaction region (the interior of the beam pipe for a collider), or to a secondary vertex of production.A magnetically curved trajectory may be described, when crossing a given reference surface (for example, for a given value of coordinate x), by 5 parameters: two for the position of the intersection with the reference surface (e.g.y, z for the example above) and 3 for the momentum vector (or the momentum and two angles).We suppose here that the mass is known or irrelevant; if not, several reconstructions may be needed, for different mass hypotheses.
• in some cases: a forward extrapolation towards external parts of the detector: e.g. the entry point into a calorimeter of into external muon chambers.
• the position and the composition of the primary vertex and secondary vertices within the detector, if any.This includes making one or several association of particles, and to use the constraint of a common origin to improve the reconstruction at the origin.Here again a procedure using iterative trials may be needed.
The problem of the noise
If the noise processes have negligible effects, we can choose a set of "initial" parameters p 0 (position, direction, momentum) which gives a deterministic prediction of the expected measurements in layer k: mk = F k (p 0 ).Because the measurement errors are independent we can write a χ 2 to be minimized: where m k is the actual measurement, with an error σ k .If needed (significantly non-gaussian errors), this may be replaced by a log-likelihood which is also a sum of independent terms.In most detectors the errors are small, so starting from a first approximation of the trajectory, F k may be replaced by a linear function of the parameters and minimizing the χ 2 consists in solving a linear system of n p equations, where n p is the number of parameters.An iteration may be needed, but in general, the computation is fast.The situation is more delicate if the errors induced by the noise are comparable to the measurement errors, or larger : a perturbation on the particle affects all measurements coming afterwards, so the measurements are no longer independent.The χ 2 should now be written as: where W is the inverse of the total covariance matrix, including the measurement errors and the noise induced errors, with non vanishing covariance terms: the measurements m j and m k are correlated through the perturbations occurring before both of them.This requires heavier computations; moreover, to make an optimal forward extrapolation we need to evaluate another covariance matrix, where m j and m k are correlated through the perturbations occurring after them.
In the following sections we apply to trajectometry an alternative method better suited to estimate the state of a dynamic system affected by random perturbations: the Kalman Filter (KF).The presentation differs slightly from the traditional one, but it is mathematically equivalent.Here we use systematically a weight matrix and weighted mean formalism, which has the advantage of being more "compact", and hopefully more user friendly, in the sense that it relies on various combinations of very few intuitively understandable elementary operations on matrices and vectors.We describe with more or less details the operations within some frameworks (for example for some choices of parameters to define the trajectory), but we cannot give an exhaustive review of all possible implementations: the tools need to be adapted to a given context, after a clear understanding of the detector and the possible and useful approximations.In any case, the choices should be dictated by the gain in terms of the precision on physical quantities of interest for a further analysis.
2 The principle of the Kalman Filter (progressive fitting): a simple unidimensional example
The simplest model of measurement of a "noisy" process
We consider here a point with a random motion on a line.Its position x(t) is measured without bias at times 1, 2, 3, • • • , with independent measurement errors of variance σ 2 .The displacements between two measurements are 0 in average and independent, with the same variance τ 2 (for example, this is a brownian motion, as illustrated in Fig. 2).The measurement errors are independent of the displacements.
Our problem is: if we have n successive measurements X k of the true positions x k , what is the combination of the X k giving the best estimator of the initial position x 1 , or the final one x n ?and why not, the best estimator of any intermediate position x k ?The solution is simple if τ 2 is negligible (at any time: the average of the measurements), or if τ 2 is very large compared to σ 2 (the best for x k is just X k , because the other measurements are too much disturbed).It is less clear if τ 2 is comparable to σ 2 , or, more precisely, if nτ 2 (the total variance of the displacements) is comparable to σ 2 : in that case, the cumulated displacements and the measurement errors cannot be disentangled.A first solution we may think about is the following: the difference between X k and x 1 is the sum of (k − 1) independent displacements and one measurement error, so its variance is V k = (k − 1)τ 2 + σ 2 .We could make the weighted average of the X k , with weights equal to 1/V k .This would be actually the best linear estimator, if the X k − x 1 were independent random variables; but this it not true in our problem, because they include all displacements from the beginning: X k and X l both depend of the steps before k and l.As a consequence, if l > k, cov(X k , X l ) = kτ 2 .
To build the best estimator in a standard way, we have to account for the (n × n) covariance matrix C of the X k .A linear combination a k X k is an unbiased estimator of x 1 if a k = 1 and its variance is minimal if a j a k C jk is minimal within the above constraint.That is, we have to solve a linear system of n equations; the solution may be written as a weighted mean of the X k , with weights w k = j (C −1 ) jk X j .The complexity of the problem grows more than linearly with n.Moreover, if we want to evaluate the final position, or an intermediate one, we have to build another covariance matrix and then solve a different linear system.
Fortunately, there is another way to obtain the same results, with a number of operations proportional to n through the Kalman Filter methodology [ 1,4] (progressive fitting [2,3]).
Tools of the progressive fitting
The fundamental idea is to incorporate the measurements one after one in the algorithm, in such a way that independent random variables are combined at any stage.The procedure is illustrated on Fig. 3, which may help to understand the simple operations behind the mathematical formulae.
Figure 3.
Principle of the forward filter, applied to the same dataset as in Fig. 2 (the black points are the measurements X k of the successive positions x k ).The blue solid line at time k represents X1→k,k , best estimator of x k using the first k measurements (see text below), which is combined with the next measurement X k+1 to give X1→k+1,k+1 , best estimator of x k+1 , and so on.
Let us consider the following elementary steps: where ε 1 is the measurement error on point 1 (variance σ 2 ) and η 1 the displacement from time 1 to time 2 (variance τ 2 ).Because ε 1 and η 1 are independent, we see that • By definition, X 2 = x 2 + ε 2 is another measurement of x 2 , with variance σ 2 .The crucial point is that ε 2 is independent of both ε 1 and η 1 , so X 1 and X 2 are two independent measurements of x 2 .
• The "best" linear combination of two independent measurements (that is, with the lowest variance) is their weighted mean with weights equal to the inverse of their variance.Here, this means that the "best" combination of X 1 and X 2 to estimate x 2 is: (where the notation Xl→m,k means: best estimator of x k using measurements l to m).In the case of gaussian errors, the steps described above may be interpreted as operations on a likelihood function L: including only the measurement X 1 of x 1 , L is a gaussian function; accounting for the random displacement consists in a convolution with another gaussian function; combining two independent measurements is a multiplication of the corresponding gaussian functions.Of course, maximizing L gives the same results as above.We will see later that this formulation can be extended to more complex cases.
The remarkable feature of this algorithm is that it can be iterated to include a further measurement (for convenience we replace σ1→2,2 by σ): • X1→2,2 found above is a measurement of x 2 with variance σ2 , so it is a measurement of x 3 with variance σ2 + τ 2 • X 3 is another measurement of x 3 , with variance σ 2 , and X 3 − x 3 = ε 3 is independent of all random variables used to build X1→2,2 .
• so we obtain the best linear combination of X 1 , X 2 , X 3 to estimate x 3 : Let us note that this may be written in a simpler way using the formalism of the weight (inverse of the variance) to express the combination as a weighted mean: The variance is 1/( w1→2,2 + w 3 ), in other terms: the weight of the combined estimator is just the sum of the weights of the two independent estimators.We can in the same way incorporate the fourth measurement, and so on, and build X1→k,k sucessively for all values of k.This is the "forward filter", which gives eventually the best estimator of the final position.It is clear that we can obtain the best estimator of the initial position through a similar formalism, starting from the last measurement and incorporating the other ones in reverse order.This is the "backward filter", giving, with our notations, Xk→n,k for any k.This is convenient if we have in mind a particle detector, where we are mainly interested to the initial parameters.
We can also build a χ 2 min associated to each step of the procedure, in the usual way: • combining X 1 and X 2 to estimate x 2 , we have As usual, in the gaussian case, χ 2 min = −2 ln(L max ), where L is the likelihood, omitting a constant normalization factor, and it follows the law of χ 2 with N − 1 degrees of freedom, N being the number of measurements included.
The principle of one step of the filter is illustrated in Fig. 4. bottom: minimum of χ 2 ), the dotted line accounts for the next random displacement.The green line represents the next measurement, and the blue one is the combination of these two independent estimators, which is the input for the next step.
Useful byproducts
We can build further tools with very few more efforts: it is easy to obtain the best estimator of any intermediate position using all measurements, that is, to build the "smoother" in the KF terminology (optimal interpolation), by an adequate combination of the forward and backward filters.The forward filter gives X1→k,k as an estimator of x k with a variance that we call σ 2 f , and the backward one gives Xk+1→n,k with a variance σ b 2 (including a term τ 2 to go from Xk+1→n,k+1 to Xk+1→n,k ) ; these estimators include independent measurement errors; the key point is that they also include independent displacement errors (the η j with 1 ≤ j ≤ k for the forward one, k < j ≤ n for the backward one).So they can be simply combined as a weighted mean, defining The same result is obtained by combining X1→k−1,p and Xk→n,k with their proper variances (we just need to include X k once and only once).We can also write an exclusive interpolation (without the measurement X k ) by combining X1→k−1,k and Xk+1→n,k : this will be useful for a discrimination of 03001-p.7 possible outliers, as discussed below.
Up to now we have assumed a perfect situation: both the displacements and the measurement errors have the expected distribution.In practice, in particle detectors, there may be deviations due to rare phenomena affecting the trajectory (e.g. a hard scattering), bad measurements or bad assignments of points to a given trajectory (especially if many particles cross the same detector element).It is useful to build tools to detect abnormal deviations ("outliers") and to define a strategy to get rid of them, as far as possible.Within our very simple model we can define two kinds of tests which can be extended to realistic situations: • for a given time k, the difference X1→k,k − Xk+1→n,k between the forward and the backward filters has a predicted standard deviation b is smaller or comparable to τ 2 , displacements which are large compared to τ may be detected; if not, the test can only discriminate very large deviations.If the distribution of the measurement and displacement errors are gaussian, a probability of χ 2 may be used as a discriminator.
• for time k, the measurement X k and the exclusive interpolation (of variance σ 2 e ) are independent, therefore the variance of their difference is σ 2 e + σ 2 .Here again, if σ 2 e is not too large compared to σ 2 , outliers (abnormal measurements) may be detected.In the gaussian case, the probability of χ 2 gives a good discriminator.
Finally, let us remark that a large sample of clean measurements may be used to perform a calibration of the errors: if σ and τ are not known a priori, the distribution of the residuals mentioned above provide estimators for them, and possibly some hints on the distribution of errors.
Comments
The Kalman Filter was originally suited to dynamical problems like following the trajectory of engines from successive observations.In that case, the forward filter is the most natural tool: it can give in real time an updated estimation of the present state (position, direction, velocity).In applications to particle trajectometry, computations are not done in real time: even if some algorithms are implemented online, their input is a set of measurements coming after the particle has escaped the detector.Moreover the main purpose of the reconstruction of trajectories is to provide the best estimations at starting point (ideally, the vertex where this particle is originating from), so the backward filter is the basic tool.However, extrapolations to external detectors and interpolations may be needed, and discrimination of outliers is quite useful.
The number of operations to build the complete set of filters (forward, backward and smoother) is proportional to n if no outliers are removed.However, if a point k is to be removed after being compared to the interpolation, the forward filter has to be reevaluated at all points after k, and the backward filter at all points before k; the smoother has to be redone everywhere.
More complex examples
In these examples, we want to go progressively to the description of a real detector.In particular, we do not consider measurements labeled by times, but measurements of one or several coordinates as functions of one coordinate describing the successive layers of the detector.To simplify some expressions we will use the following notations for matrices A, B and vectors q:
Straight line planar trajectory (2 parametres, linear model)
In this example (illustrated in Fig. 5) the trajectory may be described as y = ax+b, and the coordinate y is measured as The noise consists in a random scattering (variation of the slope a) with a variance ρ2 k at each measurement layer x k 2 .All measurement and scattering errors are independent .The parameters to be fitted are a and b; we call p the vector (b, a), or, more generally the vector of local parameters (y, dy/dx) at any point of the trajectory.Let C be the (2 × 2) covariance matrix of an estimator, and W its inverse (the weight matrix).If all errors are gaussian, the log-likelihood is a quadratic function of p: where p is the best estimator.In the general case, we will build the "best" linear estimator (using all measurements Y k up to a given position) through linear combinations using the matrices C and/or W at different stages of a progressive fitting procedure.Before that, we will try to analyze this problem of estimating (a, b) with the standard method, by computing the variance/covariance of the measurements.We call ε k the error on Y k and ζ k the variation of slope at x k .Then: The best linear estimator is obtained by minimizing the global χ 2 : that is, we have to invert the (n × n) non-diagonal covariance matrix of the measurements.
Let us now describe one step of a progressive procedure; for convenience we will use local parameters (the value of y at x k and the slope a).For the moment we suppose that we have built the best estimator p1→k,k (matrices C1→k,k , W1→k,k ) using 1→k,k , and we want to build p1→k+1,k+1 .We have to perform 3 operations: • account for the scattering at x k , by evaluating C for p1→k,k as an estimator of the parameters after the scattering.In our model, we just need to account for an additional error on a, so we compute: The value of χ2 is not modified.
• propagate the estimator, going to the local parameters at x k+1 : we have y k+1 = y k + a(x k+1 − x k ), and the slope a is not modified.We write this simple transformation in a matrix formalism: • combine p1→k,k+1 with the independent information given by Y k+1 .The combined χ 2 is: We introduce now the 2-vector of local measurement P k+1 = (A, Y k+1 ) and its weight matrix Actually A is not measured, but the expression of χ 2 does not depend on it; we can set it to 0 by convention.With these definitions we have to minimize: The solution may be written as: ) which is an extension of the weighted mean found in the simplest model (here the weights are matrices).We still have the property of additivity of weights: the weight matrix of the combination is the sum of the weight matrices of the independent estimators.The formalism above may be completely expressed with elementary operations on "atoms" of information (one "atom" = vector of local parameters + weight matrix + minimal χ 2 ), but we have evaded some technical problems for a practical coding of the algorithm, especially: how to begin this recursive procedure.With the first measurement we have not enough information to define both parameters: we have seen that we can manage this situation by taking a singular weight matrix diag(1/σ 2 , 0) and a measurement vector with an arbitrary value for the first element (0 for example): the χ 2 valley may be seen a strip in the (a, b) plane, along the a direction.This "atom" may be propagated to the next position with the formalism above: the valley is still infinite, by now in an oblique direction w.r.t. the local parameters; when combined to a strip in the a direction (local measurement) it results in a finite region, and both C and W are regular.
The handling of the scattering remains to be clarified, because the operation C = C + C scat cannot be done if C is not yet defined (first point).However we can rewrite this operation as which can be performed with one measurement only.
Further comments
The forward filter and the interpolators may be defined in the same way as in the simplest model.The tools to detect measurement outliers or abnormal scatterings may be built using χ 2 differences, with the same principles.
In the previous model the scattering occurred at the positions x k where measurements were done.The formalism may be extended to any configuration: we just need to define the set of positions x k where something happens (measurement or scattering) and to make the corresponding operations on (p, W, χ 2 ) by increasing x (forward filter) or decreasing x (backward filter), with propagation operations in between.In this way forward of backward extrapolations may be done to account for material outside the range of measurements.An important application is the material between the vertex of origin and the first measurement (e.g. the beam pipe), to be accounted for in the analysis of the primary interaction.
A thick scattering region may be described, either as a set of elementary layers handled as above, or globally by computing the (2 × 2) matrix C scat : at first order, the scattering through a thick material is fully described by terms to be added to the covariance matrix of y and a at a given reference position x 0 , including a non-diagonal one to account for their correlation.
Curved planar trajectory (3 parameters)
Here we want to introduce a good approximation of a detector making measurements in a plane xy perpendicular to a magnetic field, for trajectories with a moderate angle w.r.t. the x axis, and a large radius of curvature R compared to the size of the detector.In that case we can describe the trajectory with a linear function of 3 parameters a, b, c: We assume as above that y is measured at k .The 3-vector of local parameters is p = (y, dy/dx, d 2 y/dx 2 ).In this model we can account for scatterings (random variations of dy/dx) and also for both deterministic and random variations of energy, which are expressed as variations of the curvature d 2 y/dx 2 .The formalism of the filters is exactly the same as in the previous model (plus a specific operation to modify the curvature parameter in case of energy loss), with a few technical differences: • The matrix W is regular only when at least 3 points are included in the fit.In practice, it is of rank 1 with one point (the χ 2 valley is a slice in 3D space), 2 with 2 points (the χ 2 valley is a tube).But the same formalism as above may be used: a measurement is represented by a vector (Y k , 0, 0) with weight matrix diag(1/σ 2 , 0, 0) and in the initial vector of parameters undefined values are set to 0.
• In this model the momentum may be estimated from the fit itself so the variance of the scattering angles may be computed without external information; more precisely: the relevant curvature parameter c is proportional to the inverse of the momentum, with a geometrical sign depending on the physical sign of the charge, which is also to be determined.However, the curvature is not supposed to be known at the beginning of the filter.As a consequence, an iteration is needed: for example, the trajectory is fitted first ignoring the scattering, and the curvature found is injected in a second pass; if the curvature is significantly modified, a third pass may be needed.If the measurement range is too short to provide a significant estimation of the curvature, an external information is needed to use the noise formalism in the filters.
Realistic trajectory in space: using a linear approximation
In real detectors, no linear model (as the parabolic one) may represent the trajectories with the accuracy requested by the precision of the measurements.However, if the magnetic field is regular, one can choose initial parameters such that the position and the direction of the trajectory in any measurement layer depends smoothly on them 3 .In these conditions a reference trajectory determined by initial parameters p 0 may be defined as a first approximation, such that the functions F k introduced in Sect.1.3may be replace by a linear expansion: Let us take an example to illustrate a practical application of this method (and possible limitations): a circular trajectory in a (xy) plane (see Fig. 7).
Figure 7.
A planar circular trajectory measured at fixed x.Solid: the reference trajectory; dashed: with a variation of the initial direction ϕ 0 ; dotted: with a variation of the curvature c (changing the initial position y 0 gives a simple translation).The parameters at x 1 (y 1 , ϕ 1 ) depend linearly on small variations of the initial parameters at x 0 .
The initial parameters (at x = 0) are y 0 , ϕ 0 , c = 1/R, where ϕ is the local direction (tan ϕ = dy/dx) and R the radius R with a geometrical sign: by convention we take + for an anticlockwise trajectory; in this model c is constant.With our convention we can write: First we want to evaluate the parameters y 1 , ϕ 1 , c at a fixed abscissa x 1 .The first equation gives two solutions for ϕ 1 (or no one !), and we have to choose one of them.In general it is the nearest one to ϕ 0 (but an actual measurement can correspond to the second solution if it is within the detector...).Once the "right" solution (y 1 , ϕ 1 ) is found, we want compute the derivatives of of y 1 , ϕ 1 w.r.t.y 0 , ϕ 0 , c. Differentiating the first equation we obtain the derivatives of ϕ 1 (with Δx = x 1 − x 0 ): Let us now differentiate the second one: Injecting the expression of dϕ 1 found above in this equation, we can extract after some algebra the derivatives of y 1 : It is interesting to note that when R is large, c is a more convenient parameter than R or R: in that case ϕ 1 −ϕ 0 is small and proportional to c, so all derivatives go to a finite limit when R → ∞; moreover, the value c = 0 (straight line, infinite R) is not a singularity, and the geometrical sign may freely change during a progressive fit.These expressions give us the expression of the 3 × 3 propagation matrix D similar to the 2 × 2 matrix defined in Sect.3.1.We give in parentheses the approximation for weakly curved trajectories (small Δϕ), introducing the length of the arc between the two points = Δϕ/c 1−cos(Δϕ) This formalism should be used with care when the trajectory is close to a real singularity (quasi tangent to the measurement layer, that is cos ϕ 1 0): then the linear approximation is no longer valid, and moreover the actual meaning of the measurement may not be the coordinate y, and depend on the internal structure of the detection layer.In such a situation, it may help to redefine the parameters (e.g.take x at fixed y) and to express specifically the response of the detector under a skimming incidence; if this is not possible, the best solution is to ignore the measurement.
Once a convenient parametrization and a reference trajectory are found, the linear formalism using weight matrices may be applied to the vector δp.The C and W matrices are the same for p and δp.If the deviations from the reference are too large, it may be iteratively redefined until a satisfactory one is found.It is also possible to use different references for different parts of the trajectory; to go from one part to the next one, the parameters and their weight matrix need to be transformed through an operation similar to the propagation described in Sect.3.1 (see below).
Convenient parameters in usual detector configurations
We consider here two main categories of detectors: fixed target experiment or collider.In the first case the detection layers are mainly planes perpendicular to the beam axis (z coordinate) in the forward region, and possibly planes parallel to the beam around the target; in the second one there is a "barrel" part (cylinders of axis along z) and endcaps (planes perpendicular to z).Other configurations are possible, for example with "oblique" layers; this will be discussed later.If there is a magnetic field, we will use S/p to describe the curvature (p is the momentum, S the physical sign).In some cases (e.g.roughly uniform field along z) it is more convenient to use S/p t (p t is the transverse momentum).
Cartesian parameters
When the detectors are planes (e.g. at fixed z), a natural choice is x, y to describe the position within the plane, two slopes (u = dx/dz, v = dy/dz) or two angles to describe the local direction; for example a polar angle θ w.r.t. the z axis, and a azimuthal angle ϕ in projection onto the xy plane.
Cylindrical parameters
If the detection layers are cylinders around the beam axis, cylindrical coordinates (r, Φ, z) are natural parameters for the position: more precisely, the position at fixed r (in a detector layer) is defined by Φ, z (optionally rΦ, z for homogeneity).The direction may be given by θ, ϕ 4 .If the field is uniform and parallel to z axis, the trajectory is a helix of radius R. As in Sect.3.4, we use R with a geometrical sign (+ if the trajectory is anticlockwise in xy projection).Using a point x 0 , y 0 , z 0 on the trajectory and ϕ as a running parameter, the trajectory is defined by:
The "perigee" parameters
It may be useful to summarize the information about the trajectory in one set of intrinsic parameters instead of using an arbitrary reference surface.In the case of quasi-uniform magnetic field along the beam axis (by convention the z axis), we can use the "perigee", point of closest approach to the z axis: if the particle originates from the main vertex, this point will be close to this vertex, so it will give a good approximation of the particle momentum.Another advantage is that a propagation of the trajectory and its error matrix to this point includes most of the material actually crossed by the particle (all material if the perigee is within a vacuum region, e.g. the beam pipe), so if this material is taken into account properly, the perigee parameters may be used in a further step of vertex fitting in a purely geometrical way, without accounting for noise: this will be exploited in Sect.6.
The trajectory is defined by 5 parameters (see Fig. 8 ): the cylindrical coordinates of the perigee (ε, Φ p , z p ), the signed curvature c = 1/R and θ.To avoid discontinuities around the origin when extrapolating a trajectory towards the interaction region, it is convenient to give a geometrical sign to ε: by convention it is positive if the origin O is on the right hand side of the trajectory, and Φ p is defined as ϕ p + π/2.With this convention, we have always x p = ε cos Φ p , y p = ε sin Φ p , and the trajectory may be parametrized as: In the vertex fitting procedure, we need in principle short range extrapolations from the perigee, so we can use the second order approximation in = R(sin ϕ − sin ϕ p ) (distance from perigee in xy projection): The perigee parameters will also be used as a technical tool to compute the derivative matrices needed to propagate the error matrix in cylindrical coordinates (see Appendix).The position of a point along the trajectory is defined by r, Φ; the direction of the tangent is defined by ϕ.In this example, the signed radius R is negative (ϕ decreases with increasing r), and the perigee distance ε is positive.
Propagation of error matrices
In the helix model for trajectories, the analytical computation of the matrix of derivatives D was done in Sect.3.4 for cartesian parameters.Using similar techniques, on can obtain analytical expression for cylindrical parameters, as function of the parameters at the initial and the final point.The computation is developped in Appendix.
If the magnetic field is not perfectly uniform, the trajectory has to be propagated with a precision better than the measurements, so a numerical computation (or a perturbative expansion) may be needed.However, the derivatives may be taken from the analytical expressions, because they give a sufficient approximation for the propagation of errors.
Local change of parametrization
To follow the disposition of the detector layers, it may be convenient to modify the parametrization at a certain point of the trajectory.For example, with the helix model in cartesian coordinates, we use parameters (x, y, θ, ϕ, c) at fixed z in the forward region, and (x, z, θ, ϕ, c) at fixed y in the lateral region.Here using the same notation x for a parameter with a different meaning may be a source of confusion: the transformation of the error matrix should account for a non trivial transformation on the position parameters: an elementary variation δy of y in a plane z = z 0 , at given values of x, θ, ϕ, c, is a translation which results in a displacement of the intersection of the trajectory with a plane at fixed y; using the notation a| b for "a at fixed b", and noting u the unit vector along the trajectory, the variations of the coordinates x| y , z| y are: For the same reason, if the curvature is not negligible, a variation of y| z will affect the direction (actually, only ϕ) at fixed y; if we call δ the displacement inxy projection, we have from the helix model: The jacobian matrix of the transformation from (x, y, θ, ϕ, c)| z to (x, z, θ, ϕ, c)| y is then: The covariance and the weight matrix are modified as For any other change of local parameters, a similar study has to be done to define the jacobian matrix.
Indirect measurements of parameters ("oblique" projection)
Up to now we have represented a layer of by a simple surface (e.g. a plane of wires).The quantity actually measured by a detector is supposed to depend on the position of the intersection of the trajectory with this surface, but it may also depend on the direction of incidence.Let us take two examples: 03001-p.17 • In a plane z = z 0 , we mesure the distance of closest approach of the trajectory to a wire at x = x 0 (see Fig. 9, left).Using the parameters x and a = dx/dz, and assuming that the curvature is negligible at this scale, this means that we measure d = |x − x 0 |/ √ 1 + a 2 , with a precision σ.Within a good approximation, we can take the reference value a ref of the slope, and consider that we measure ref with a precision σ x = σ 1 + a 2 ref (at this level there may be an ambiguity if the extrapolation provided by the filter is not precise enough).
• The detector surface is not exactly perpendicular to z axis (see Fig. 9, right).For example, we measure a coordinate ξ in a plane inclined by α on the xy plane, intersecting the plane z = z 0 at x = x 0 (ξ = 0 at the intersection).We obtain x − x 0 = (cos α + a ref sin α)ξ.Here again we apply to the measurement a factor depending on the local direction.
The real situation may be more complex.For example, in a drift chamber, the relation between the measured time and the local parameters depends on the position (close to the wire or far away).Or we may have to consider in a barrel detector (with cylindrical parameters) detector elements which are planar.In any case, the prescription is to write the local parameter to be measured as a linear function of the quantity which is actually measured, with coefficients depending on the local direction of the trajectory.
Composite measurements
Some detectors (e.g.chambers with tilted wires) provide a measurement of a "composite" coordinate, e.g. a quantity u = αx + βy at fixed z, measured with a precision σ.The formalism of "atoms" introduced in Sect.3.1 is very convenient to account for such measurement u m : it is equivalent to a vector P = (x m , y m , • • • ), where x m , y m are any values such that u m = αx m + βy m , the other components being arbitrary, with a weight matrix of rank 1 (written here with 5 parameters): More generally, let us suppose that we measure in a detector surface a set of that can be expressed locally as a linear combination of the δp: The errors on the measurements of u 1 , u 2 • • • u n may be independent or not.Let us call C U their covariance matrix, and W U = C −1 U their weight matrix.We can introduce in the filter formalism an "atom" made with δp m (any values compatible with the n measurements) and a weight matrix W m = M T W U M of rank n.The result of the weighted means does not depend on the arbitrary choices made to build p m .
Exogenous measurements
We call "exogenous" a measurement coming from a detector which does not belong to the trajectometer, or an information coming from an element of the trajectometer, but of a different nature.In the first category we can include a calorimetric measurement of the energy (if a matching can be established): this may be useful to compute an initial curvature parameter when starting the backward filter (but the ambiguity on the sign has to be solved).In the second category we may have a timing information, or an evaluation of the ionization rate, that can constrain the momentum, or solve the mass ambiguity.When written as a linearizable function of the local parameters, these measurements can be handled in the same way as the composite measurements above.
Comments on practical implementation
A big advantage of the Kalman Filter formalism is to rely on linear operations on vectors and matrices of fixed dimension (number of parameters needed to describe the trajectory), whatever the number of measurements and noise sources.The implementation is computationally efficient if these operations are explicitely coded, without calling functions from a general matrix package.Moreover useful approximations may result in sparse matrices, reducing even more the computations needed.As a consequence, such procedures could even be introduced in high level triggers.
Coupling the pattern recognition to the track fit
In the previous section, we have supposed that the measurements to be affected to a given trajectory were unambiguously defined in a previous step of pattern recognition.In practice, for a complex event including many particles, this preliminary task is far from being easy, and in most cases it cannot be achieved without any ambiguity.The progressive fitting procedure can help to resolve these ambiguities, for example using the probabilities of χ 2 .For a better discriminating power, it may also be used within the pattern recognition procedure itself, to perform a progressive collection of points along the trajectory.The basic procedure [ 5,6] is as follows: • build tentative "segments" using points from a few layers, with loose criteria of compatibility; in this step ambiguities are freely accepted.
• apply a forward and/or backward filter to these segments.
• extrapolate to the next and/or previous layer and try to add a measurement found in this layer, and apply a χ 2 criterion to accept or reject this measurement; at this level ambiguities are still accepted (and possibly extended if several measurements are compatible with the extrapolation).
• iterate the procedure.In principle the χ 2 is more and more selective with more and more points included.
• at the end, resolve the remaining ambiguities if any (or keep some of them open for a final analysis).
In any case, the strategy should be adapted to the context on the following points: chosing the best starting region, tuning the criteria at each step, defining tolerance for missing points, using approximations in the filter (for example: ignoring the noise, assuming a small curvature), etc.In some cases, an external measurement (e.g. from a calorimeter or a muon chamber) may be provide a good starting segment; if the first layers are very precise (as in usual "vertex detectors"), it can be used to define clean segments to be extrapolated forwards, because at this level there are few parasitic tracks produced in the material, and the trajectories are quasi straight lines.Hybrid strategies may also be efficient; there is no general rule on this subject.
Beyond the gaussian approximation
The measurement errors and the perturbations on the trajectory are never exactly gaussian.Some deviations from "gaussianity" are not worrying, because the convolution of different errors tend to 03001-p.19 "gaussianize" the combination.For example, let us imagine a series of hodoscopes which just provide an interval (x 1 , x 2 ) for a coordinate x at different position in y along a straight trajectory in xy plane, described by parameters a, b: in the absence of multiple scattering, each measurement gives slice in the a, b plane, and the global information is a polygon which is more or less extended depending on the position of the trajectory, while the gaussian model gives an ellipse with a shape depending only on the coordinates y k of the hodoscopes; on the contrary, accounting for the multiple scattering results in a smoothing of the distribution of errors.Modern detectors are generally not hodoscopes, and the distribution of errors is often smooth and nearly gaussian.In the case of precise measurements, the non-gaussianity is wiped out by the "noise" along the trajectory.
More serious is the problem of errors with long tails, especially in the energy loss of electrons or positrons, which may be large even through a moderate amount of matter.These tails are propagated throughout the fitting procedure, so that the fitted values do not follow a gaussian distribution, and their variance is underestimated in the gaussian model.We have described above tools to detect abnormal deviations, but we want to go further and try to use explicitly the shape of the error distribution in the case where it is known, or predictable from the parameters of the trajectory.In practice we have to find a reasonable compromise between an ideal procedure (complete description and propagation of the errors), which will appear to be extremely heavy with several parameters, and the available computing power; we also want to have an idea of what we can gain with respect to the gaussian procedure, which is very fast.
The ideal procedure
The fitting procedure is still a forward or backward chain of basic operations (measurement, noise, propagation) along the trajectory, but now the "atom" of information is a density function F(p) in the space of parameters, which express the likelihood of the subset of measurements included from the beginning of the chain.The previous considerations on the independence of the errors are still valid, so the mathematical transformations of F corresponding to the basic operations are: • measurement: combination of independent informations, that is a product: where m is the expression of the local measurement as a function of the local parameters, and f the distribution of the error on m.
• noise: addition of independent errors, that is a convolution: • propagation: going from one layer to the next one consists in a transformation of the local parameters, that is a composition: where P is the transformation from the local parameters in the initial layer to the local parameters in the final one, and J(P) its jacobian determinant.
None of these operations can be performed in a reasonable computing time in a multidimensional space (5 parameters in the standard implementation), without an adequate parametrization of F, f meas and g noise .
The Gaussian Sum Filter
One practical solution is to replace all functions involved in the different steps by a sum of gaussian functions [8].The main advantage is that such functions are defined by a small set of values (the mean p 0 and the weight matrix W); both their product and their convolution are gaussian, and the mean value and the weight matrix of the result have simple expressions.We summarize here the algebra of gaussian functions in a N-dimensional space: and 2 ) −1 In the linear approximation, the propagation may be expressed as in Sect.3: when going from p i to p f , with a jacobian matrix D i→ f = ∂p f /∂p i , we obtain: If we can approximate the measurement and the noise density functions as linear combinations of normalized gaussians, with positive coefficients: we obtain at each step of the procedure F as a combination of gaussian terms, which is automatically positive in the whole space of parameters.Of course, the main problem is that after n steps including each a sum with m i coefficients, F is expressed as a sum of m i terms, so the complexity may be too high if the detector has many layers.This can be partly cured by reducing the number of terms after each step, for example, suppressing the terms with low coefficients, or grouping similar terms into one.The strategy should be tuned for a given detector configuration.
To illustrate the method and the possible gain, we come back to our simplest model with one parameter (Sect.2.1), with one difference: the displacement η between two measurements is no longer gaussian.We adopt here an asymmetric superposition of gaussian functions, with mean value 0: 3 ) /(a 1 + a 2 + a 3 ) In the following, we take for the triplets (a, μ, τ): (10,-1,0.3),(3,0,3) and (1,10,10), which give τ(η) = 4.122; the measurements are gaussian with variance 1.We perform a series of trials of 6 measurements with 5 intermediate displacements; we apply to each sample the standard gaussian filter and the gaussian sum filter (keeping all 3 5 terms), to find an estimator of the initial position.The gaussian sum may be used through its mean value and its standard deviation.Alternatively, one can search for its maximum and the deviations giving a decrease of 1/2 for its logarithm; in this example, there is no significant difference between the two methods.Note that the estimated error on the gaussian sum depends on the actual configuration of the displacements, while the standard filter gives always the same value.
In Fig. 10 we summarize the results on this example: the gaussian sum gives a slightly narrower distribution of errors.More important, it provides an estimation of the variance for each realization, which extends over a large range: computing the "pulls" (deviation/error) with this estimation gives the right spread for any value of the variance.In brief, the average error is not greatly improved, 03001-p.22 but we have a distinction between more or less precise evaluations, with a reliable error for every configuration.
Comments
The main application of a fitting procedure extended beyond the gaussian approximation is the reconstruction of electrons/positrons, accounting for the tail in the distribution of energy loss.We can try to understand intuitively what can be gained.The trajectory is measured over a given segment: if a large energy loss occurs close to the end of this segment, it has no significant effect, whatever the procedure; if it occurs close to the beginning, both the gaussian and the beyond-to-gaussian methods will suffer the same bias on the energy.There may be a significant difference if the large loss occurs in the central region of the segment: the beyond-to-gaussian backward filter includes a tail towards lower curvature (larger energy), so is has more flexibility to modify the curvature when including the points before the large loss, and hence to obtain a better evaluation of the initial energy.
In principle, the formalism may be used outside the measurement range, for example, when including a calorimetric measurement: it can be transformed through an extrapolation to the trajectometer, accounting for the material in between, to give a non-gaussian distribution for the curvature at the beginning of the backward filter ( even if was roughly gaussian within the calorimeter).Similarly, the backward extrapolation to the vertex region may be beyond-to-gaussian, including the material crossed before the trajectometer.The non-gaussian features can be introduced in subsequent kinematical reconstructions; in practice, this may be difficult to implement, especially if the trajectometry and the kinematics are handled in independent modules, in the spirit of "hidden boxes" in an Object Oriented framework.
The vertex procedure (Sect.6)may be coupled to the track fitting to improve the reconstruction.For example, if an electron/positron is supposed to come from a given vertex, the position of this vertex can be used as a "virtual measurement" constraining the initial part of the trajectory and improving the reconstruction of the energy.But this is not possible if one wants to decide whether this electron/positron comes from the main interaction or from a secondary decay; in any case, such a decision is more ambiguous than for a heavy particle.
Fitting a vertex
Once the trajectories have fitted, we have for each one a 5-vector of parameters p i (intersection with the initial surface) with their weight matrix W i .Assuming that a given sample of trajectories comes from the same vertex of interaction, we want to reconstruct this vertex (and possibly check this assumption).This way be done at two levels: • find the best estimator of the 3 coordinates of the vertex (and evaluate errors on them).This may also provide a criterion of quality, e.g. a χ 2 providing a probability for the hypothesis of convergence; if possible, we also want to define a criterion for each individual particle to belong to the vertex.Hereafter we call "simple vertex fit" such a procedure.
• exploit the fact that the trajectories come from this point to improve their reconstruction, that is, add to each trajectory a virtual measurement given by the other ones.This is interesting in view of the kinematical reconstruction of the event.All trajectories are improved, and particularly those measured over a short range, because their parameters (especially the curvature) are poorly defined: an additional point may give a very useful information; but, of course, the criterion to decide whether such a trajectory should be attached to the common vertex is loose.This "full vertex fit" is a priori a complex procedure, because we want to fit 3N + 3 parameters: the coordinates of the vertex and 3 quantities to define the initial state of N particles (e.g.p x , p y , p z or better 1/p and two angles).We will see how the problem can be simplified using a linearization.
The simple vertex fit
The procedure is conceptually simple.From the initial parameters one can deduce the parameters and their error matrix on any surface: this defines a "tube" of probability around the trajectory.When extrapolating the trajectory backwards from the initial surface to the region of the vertex, the errors on the position increase, so the tube gets broader, but over a short range it may be considered as a cylinder, as illustrated in Fig. 11.In other terms, if the position of the vertex is approximately known, each trajectory provides an information on the vertex coordinates that may be summarized in a position with a weight matrix of rank 2 (the position may be arbitray chosen along the axis of the tube): as in Sect.3, combining these informations amounts to make their weighted mean.If at least two non parallel tubes are combined, the degeneracy of the position is removed.
Figure 11.Principle of the "simple" vertex fit.Each trajectory fitted to measurements (black points) provides an initial position with its 2 × 2 error matrix, which is extrapolated backwards to the vertex region (dotted lines).The errors increase with the range of the extrapolation, but in the region of interest (dashed lines) they can be represented by a cylinder, that is, an arbitrary position along a straight line and a constant error matrix on two coordinates (e.g.here x, y at fixed z).Finding the vertex consists in making the weighted mean of these tubes, in the sense defined in Sect.3.1.
We describe now the mathematical procedure.Let us suppose that the parameters are x, y at fixed z, and three more for the direction and the curvature.First, an approximate position (x 0 , y 0 , z 0 ) of the vertex is found from the intersections of the extrapolated trajectories in xz and yz projections.Then, the error matrix C i of each trajectory is propagated to z 0 , using the matrix of derivatives D i = ∂p 0 /∂p i ; actually, we just need the 2 × 5 submatrix D i of the derivatives of x, y (at fixed z = z 0 ) w.r.t.p i to compute the 2 × 2 error matrix C i0 = D i C i D T i and W i0 = C −1 i0 .If we approximate locally the trajectory as x = x 0 + a x (z − z 0 ), y = y 0 + a y (z − z 0 ), we can describe the probability of presence of the vertex at v = (x, y, z) by saying that the 2-vector (x − a x (z − z 0 ), y − a y (z − z 0 )) = u − (z − z 0 )a has a mean position u 0 with a weight matrix W i0 .In the gaussian approximation, this gives a density of probablitity exp(−W i0 [u − (z − z 0 )a − u 0 ] /2), that is a 3 × 3 weight matrix for v (writing W for W i0 in the right hand side): Let v i0 = (u i0 , z 0 ) be the mean position of the extrapolation of the trajectory i.Its weight matrix W iv has rank 2, but if the extrapolations of trajectories at z 0 are not all parallel, the weighted mean of the extrapolated positions of the trajectories is defined as ( i W iv ) −1 i W iv v i0 : this is an estimator of the vertex position (the best one in the gaussian case), with an error matrix ( i W iv ) −1 .
The full vertex fit as a "hierarchical" fit
This procedure ( [3,7,9]) aims to fit 3N + 3 parameters (V = (x v , y v , z v ) and q i = (1/p i , θ vi , ϕ vi ) for each of the N particles) to a set of N 5-fold measurements, e.g.p k = (x i , y i , 1/p i , θ i , ϕ i ) at a fixed value of z for particle i, with a weight matrix W k .The equation of propagation expresses p k as a function of V and q k , and we see that the parameters to be fitted do not play the same role: for any k, p i depends on V, but not on q j if j i.So we can distinguish 3 global parameters and N individual subsets of 3 parameters, which are related to only one measurement: we call "hierarchical" such a fit.
If we want to perform the fit by minimizing a function F (χ 2 or negative log-likelihood), we can write it as a sum over the particles (measured independently): So, if we want to use a minimizing package, we can use an embedded structure of minimizers, where the function of 3N + 3 parameters to be minimized is itself computed at each step through N calls to a minimizer of a function of 3 parameters.In this procedure, the correlations between all parameters are taken into account to find the best path towards the minimum.We can imagine another solution, using an iterative alternate procedure: fit V with fixed q k , then fit each q k with fixed V, and so on.In practice, if the parameters are correlated, the convergence may be very slow 5 .It may be accelerated if F is quasi quadratic around its minimum: in this favourable case, the differences between the parameters at step i and their final values decrease exponentially with i, so after a certain number of steps, one can evaluate a good approximation of the limits, restart the alternate fit, redo the computation of limits, and restart the overall procedure if needed.An advantage of this method is to offer a better control of the convergence, and a way to remove tracks during the iteration, while the first one uses the minimizer package as a "black box" where the internal strategy cannot be modified.
Whatever the strategy, the computing time grows rapidly with N. If the particles produced in the initial interaction may produce secondary vertices (decays or interactions in the material), one has to make trials to choose to best association of tracks to vertices, and the computation may become very heavy.Fortunately, in most detectors, the p k depend linearly on the variations of V and q k within a few standard deviations around the central values: with this approximation, minimizing a global χ 2 depending on 3N + 3 parameters will result in a sparse linear system of equations, that can be solved through a number of operations proportional to N, and moreover adding a removing a track to/from a vertex will be simple.
Linearization of the problem
Let V 0 be an approximate position of the vertex, and q i0 approximate parameters of track i at the vertex (e.g. the ones obtained by a backward extrapolation of p i to z 0 ).In the linear approximation we differentiate the propagation from V, q i to p i : where D i and E i are (5 × 3) matrices of derivatives 6 .So, if each track was fitted individualy as (p f i , W i ) at the initial point, we can rewrite χ 2 = i W i p i − p f i as: where we have introduced the 5-vector of deviations of the individual fits from the predictions of the first approximation at the vertex: ).Using the image of "tubes" in the space of parameters , we can say that the fit of the trajectory i defines a tube of rank 5 in the 6D space (V, q i ), and we have to combine N such tubes in a (3N + 3)-D space.The χ 2 is here approximated by a quadratic function of the parameters, so the minimum is given by a linear system of 3N + 3 equations on δV and the δq i .This system may be split in N + 1 blocks of 3 equations; the first block contains terms for all parameters: the N following ones contain each δV and only one of the δq i : The last blocks of equations give expressions of the δq i as a function of δV: which can be injected into the first one to obtain an equation giving δV: Then each δq i follows from δV.
The left hand side of the linear system written above may be expressed with a sparse matrix W of 3 × 3 blocks, where only the first line, the first raw and the diagonal are non-zero, easily solved by a substitution method: The global covariance matrix on the 3N + 3 parameters is the inverse of the global weight matrix W, that is, written by 3 × 3 blocks: This provides an error matrix on the position of the vertex and re-evaluated error matrices on the indivdual particles.In addition, it should be noted that the procedure introduces a correlation between the different particles: this means that, in principle, this correlation should be taken into account when evaluating the uncertainty on physical quantities built with several particles of a vertex, like the total momentum and energy, equivalent masses, etc.
The system has a unique solution for N > 1, provided the trajectories are not all parallel at the vertex.The total number of operations needed to solve the system grows proportionally to N if N is large.In the gaussian approximation, the minimum of χ 2 , as found with the solution of the linear system, follows a law of χ 2 with 5N−(3N+3) = 2N−3 degrees of freedom.The associated probability gives a criterion to decide whether the set of N tracks are actually compatible with the hypothesis of a common origin.
Flexibility and iterative procedures
In many cases, the bundling of tracks in vertices is ambiguous, either because there are short lived particles giving a secondary vertex close to the main one, or because some secondary particles from remote decays (e.g.K 0 or Λ) are not obviously distinguished.In these conditions, it may be useful to add or remove a track from a vertex to perform a new trial.Let us suppose that we have fitted a vertex with N tracks, giving χ 2 min .We take the fitted parameters as new starting values; then the χ 2 including a (N + 1) th track may be written as: Using A, B i , C i , T i , U as defined above, and introducing A N+1 = D T N+1 W N+1 D N+1 and U N+1 = D T N+1 W N+1 Δp N+1 the minimization gives: Injecting in the first equation the expressions of δq i found in the second and in the third one, we obtain an equation which gives δV: EPJ Web of Conferences hence the δq i .The matrices A − N+1 i=1 B i C −1 i B T i and C −1 i B T i were already computed in the fit with N tracks, so the amount of computation needed is much less than doing a fit with N + 1 tracks ab initio.
The same algorithm can be applied to remove a track from a vertex fit: it is equivalent to add this track with a negative weight matrix −W i .Note also that all operations (multiplications, inversions) involve always 3 × 5 and 3 × 3 matrices and may be coded explicitly in a very efficient way, avoiding any use of indexed arrays.
The tools described above can be inserted in an iterative vertex building.In most detectors the trajectories around the vertex may be described by smooth functions, so the linear approximation is quite adequate, especially when using the "perigee" parameters (no iteration is needed for a given set of tracks).With quality criteria based for example on the probability of χ 2 , a strategy may be defined to determine the best repartition of the tracks in vertices; this strategy may be driven by physical considerations, for example finding the decay of a heavy flavour particle, starting from a "seed" (a large p t lepton, a combination identified by equivalent masses, etc).
Beam profile
The simplest case of constraint for the main vertex is the beam profile, which may be considered as a particular trajectory entering the vertex, except that its parameters should not be re-evaluated in the procedure.If the z axis is chosen along the beam, the lateral profile is usually summarized by the lateral standard deviations σ x and σ y , and the constraint may be expressed by just adding a term (x v /σ x ) 2 + y v /σ y ) 2 to the χ 2 .In the "simple vertex" procedure (Sect.6.1),we just need to add a trajectory with x 0 = y 0 = a x = a y = 0 and a diagonal weight matrix (1/σ 2 x , 1/σ 2 y ).In the "full vertex" fit within the linear approximation (Sect.6.3), the additional term results in adding the diagonal matrix x , 1/σ 2 y , 0) to A and −W b V 0 to U.This formalism may be applied to both fixed target experiments and colliders, but it improves the results only if σ x and σ y are comparable to, or smaller than the errors on extrapolated tracks: in practice, this is true in colliders, where the beam constraint is very useful.
Kinematical and geometrical constraints
A typical example is the reconstruction of the so called V 0 's: remote decay of a neutral object (γ, K 0 , Λ, etc) in two charged particles.The constraint consists in the value of the equivalent mass of the pair with given mass assumptions for the charged particles i and j.The fit may be performed by removing one of the free parameters and replacing it by a function of the other ones; in the case of γ → e + e − , we can force p i and p j to be parallel by making a fit with 7 parameters instead of 9: x v , y v , z v , a common value for θ and ϕ, and the curvatures c i and c j .More generally, the constraint may be expressed as F(p i , p j ) = 0 and handled through the generic method of Lagrange multipliers (see below).
As a result of the fit, we obtain estimators for the trajectory of the neutral particle (a straight line), which can be inserted in a primary or a nearby secondary vertex, through one of the procedures described above: the introduction of 4-vectors instead of 5-vectors in the formalism is straightforward.Conversely, we can introduce an additional constraint in the remote vertex fit, if the neutral particle comes from a previously fitted vertex with a given error matrix.
General procedure with Lagrange multipliers in the linear approximation
The constraint may be applied as a further step after the "standard" fit (with the pure constraint of convergence).Let us call now p f the global vector of values fitted without the constraint F(p) = 0. We can suppose that the constraint is nearly satisfied by p f , that is, F can be linearly expanded in the neighbourhood of p f : where ∇F is the gradient of F at p f .On the other hand, the χ 2 may be approximated by a quadratic expansion around p f : χ 2 = χ 2 min + W p − p f The Lagrange multiplier method consists in finding the minimum of: when varying both the components of p and λ.This is a quadratic function, so the solution can be obtained easily in two steps: • canceling the gradient of G w.r.t.p, which provides p as a linear function of λ: • canceling the derivative w.r.t.λ and introducing the previous expression to solve a linear equation in λ: This value of λ gives p through the first equation.
This procedure is easily extended to the case of several simultaneous constraints We obtain p − p f as a linear function of the λ k and a linear system of equations on the λ k , where the expressions of p−p f may be injected to eliminate them.The values of the λ k obtained from this linear system provide the wanted solution for p.
Appendix: derivative matrix for propagation in the helix model
With the notations defined in Sect.3.5, the trajectory is defined through a running parameter ϕ: We define the signed curvature c = 1/R.To simplify expressions we use t = cot θ as parameter instead of θ.We want to obtain analytical expression for the derivatives of the non-constant parameters (i.e.other than t, c) on a surface, as functions of the parameters on another one.
Planes perpendicular to the magnetic field
The initial parameters are x, y, ϕ, t, c at fixed z 0 .To obtain the parameters at z = z 1 , we can write, with Δz = z 1 − z 0 and = R Δϕ = Δz/t (length of the arc in xy projection): hence the derivatives: The expressions for x 1 , y 1 and their derivatives follow immediately from ϕ 1 , using the notation ΔA = A 1 − A 0 for any quantity A depending on the local parameters: In the approximation of weak curvature (|Δϕ| 1) we obtain "quasi straight line" approximations for the last two: The parameters are now y, z, ϕ, t, c at fixed x; we want to go from x 0 to x 1 .As previously, we first determine ϕ 1 and its derivatives, using similar notations (assuming that the right solution ϕ 1 of the trigonometric equation is chosen; usually it is the closest one to ϕ 0 ); some of the computations are the same as in Sect.
Propagation between cylinders
In the following, we will use as trajectory parameters at fixed r: Φ, ξ and R (or c = 1/R) in xy projection, and z, t = cot θ to complete the description in 3D space.The coordinates of the center C of the projection onto the xy plane may be expressed from any point (r, Φ, ϕ) on the trajectory as: x c = r cos Φ − R sin ϕ (1) y c = r sin Φ + R cos ϕ (2) so the polar coordinates of C are r c , Φ c such that: where we define ξ = ϕ − Φ (deviation from the radial direction).
As intermediate parameters, we use also r c defined above and the z coordinate of the perigee z p = z − R t (ϕ − ϕ p ).Note that with our convention on the geometrical sign S of the curvature (the sign of R), we have Φ c = ϕ p + S π/2.To simplify some expressions we introduce for any point on the trajectory ψ = ϕ − ϕ p (rotation fom the perigee to this point).Introducing ρ c = S r c , we can rewrite (1) and ( 2 The notations used in this Section are illustrated in Fig. 12.We want to obtain the derivatives of the parameters at r = r 1 with respect to the parameters at r = r 0 : to do so we will compute the derivatives of the intermediate parameters (r c , Φ c , z p ) with respect to the initial ones and to the final ones, and use the inversion and the multiplication of jacobian matrices.For convenience, we compute derivatives w.r.t.R instead of c in the intermediate steps.First we consider the transformation from (Φ 0 , ξ 0 , R, z 0 , t) to (ρ c , Φ c , R, z p , t).From the expression of ρ Some approximations may be applied to the first and last derivative of z 1 if the impact parameter |R − r c | is small.
EPJFigure 1 .
Figure 1.Schematic layout of a trajectometer.The green dashed lines are the trajectories of charged particles coming
Figure 2 .
Figure 2. Measurements (blue points) of the position of a point moving randomly on a line.Here a brownian motion is simulated (solid line), and the variance of the displacement between two measurements is τ 2 = 0.25.The solid bars represents measurement errors with σ 2 = 1; the dotted ones include the contribution of the displacements from the beginning, i.e. for point k it is √ kτ 2 + σ 2 .
Figure 4 .
Figure 4.The estimator including all measurements up to a given point is represented in black (top: density of probability;
Figure 5 .
Figure 5.A planar trajectory made of straight line segments between measurement planes, where the slope is randomly modified (simple model for the multiple scattering on particles).The black points represent measurements.
Figure 6 .
Figure 6.The operations of one step k → k + 1 of the filter (as in Fig.4) applied to a 2-parameter model (position and slope of a trajectory).The ellipses are the contours associated to one standard deviation around the central value.Left: solid: the "initial" estimator p1→k,k ; dashed: including the scattering.Right: black,dotted: propagated to point k + 1; green, dash-dotted: measurement P k+1 at point k + 1 (position only → vertical strip); blue: combination.
Figure 8 .
Figure 8. Left: the trajectory in 3D-space (helix of axis along z).Solid: the measured portion M 0 M 1 , dashed: the extrapolation to the perigee P. Right: The projection onto the xy (circle).The position of a point along the trajectory is defined by r, Φ; the direction of the tangent is defined by ϕ.In this example, the signed radius R is negative (ϕ decreases with increasing r), and the perigee distance ε is positive.
Figure 9 .
Figure 9.Examples of "oblique" measurements.The trajectory (blue line) has a slope a = dx/dz.In both cases, x m represents the coordinate effectively measured in the plane z = z 0 through a raw measurement in the detector.Left: the raw measurement is the distance to a line parallel to y axis (e.g. a wire).Right: the raw measurement ξ is taken in a detector inclined by α (thick black line); we see on the figure that x p − x 0 = ξ cos α and x m − x p = a ξ sin α .
Figure 10 .
Figure 10.Standard Filter vs Gaussian Sum Filter(GSF) in the simple 1D model (with measurement error = 1.Top left: the distribution of the random displacements between the measurements.Top right: error of the GSF estimator (solid) vs the Standard one.Bottom left: the distribution of pulls for the GSF (global).Bottom right: the pulls for GSF vs the estimated uncertainty.
Figure 12 .
Figure 12.Parameters used to express the propagation of cylindrical parameters at fixed r. | 18,103.4 | 2013-07-01T00:00:00.000 | [
"Physics"
] |
The Importance of Dialectal Variation in Kerala Curriculum Framework
ORCID0000-0001-67280228 Abstract Purpose: From prior research, language variation is observed to beneficially influence the field of education. Following this hypothesis, the study verifies the importance of dialectal variations in a language, specifically in Malayalam. The study strives to answer the need for linguistic equality and how this can be achieved through the curriculum. Approach/Methodology/Design: A mixed method approach was adopted using questionnaire and personal interviews. Data was collected from University students between the age group 20-30. The material of the study involved different lexical items. The data was analyzed by accounting the number of occurrences and its percentage. Pivot chart was tabulated of the percentage of dialectal variations lexical items against each participant in different category. Findings: The study revealed the lack of awareness of dialectal variations that existed in the selected lexical items. This neglect provides an evidence of the progressing decline in language lexicon that is detrimental to language growth and preservation of vocabulary. The study illustrates how this can be rectified through the curriculum by incorporating dialectal variations in the textbooks. Practical Implications: The study will contribute positively to understanding the importance of incorporating dialectal variations to preserve the existing language lexicon by accommodating the non-standard variation. This step ensuring the equality of regional elements would help in an effective and successful learning of language. Originality/value: This study takes into consideration the regional variations that exist in Malayalam language spoken in Kerala. The study provides a base for further research into mapping dialectology.
Introduction
Over the years, significant importance has been given to documenting variations in language use. The variations within a language whether phonological, morphological, lexical or syntactical are addressed in sociolinguistics. In cases where the internal variations within a language allow the language to be mutually intelligible, the languages are said to be dialects of the particular language (Chambers & Trudgill, 1980). In this scenario, one particular dialect assumes the status of standard language. Often times, the emergence of the Standard Language (SL) or the Language of Instruction (LOI) depends on factors that are historical, political, cultural, ideological or pedagogical. Either way, the dialects provide us with linguistic diversity. Dialectal variation should not be confused with slang that are used by particular group of people. Though slang is vocabulary driven and well-defined in terms of social and regional boundaries, it is an informal form of speech and is closely associated in being part of informal register rather than dialects of the language (Crystal, 1995).
Internal variations can be observed at different levels: phonological, morphological, lexical and syntactic. There have been several views regarding the categorization of these variations. One accepted view is the categorization based on social class or geographical/regional differences (Holmes, 2001). On this basis, dialects are classified as social or regional dialects. The study of dialects on account of above said regional differences is dialectology. Dialectology comprises the mapping of linguistics sub-regions. Accordingly, one can present a dialect map. In Kerala, dialect variation can be dominantly observed via geographical differences though there are elements of social differences.
Theoretically establishing a definite categorization between the dialects is difficult (Wardhaugh, 2006). However, Romaine (2002) summarizes the difference between the two by stating that the regional dialects tell where we are from while social dialects tell who we are. Despite the complexities in categorizing variations, a certain variety becomes the standard language with other varieties as non-standard dialects. The topic of how one variety of language gets regarded as the standard language is of prime importance in the field of sociolinguistic (Cook, 2003;Cheshire, 2005 andLabov, 2003). Majorly this involves a study of government planning and policy changes.
Gradually with the use of a standard variety, a norm began to be observed with regard to the phonological, morphological rules and lexical words. This norm covers decades of print, media and education. At present, when discussing the importance of variations within a language, the focus is not to bring about a change in the standard language; rather it is the official education of the variation. The paper thus tries to validate the importance and need of an official education of variation that exists between the language dialects.
Regional variations are confined to a particular space and often do not find their way across, unless through migration. In situations where migration does happen, the variations are introduced to the subsequent place but they get subdued under the variations that already exist in that particular space. The result is the slow disappearance of the particular variation. This brings us to the question of how much of these variations have disappeared over the years. Though it is not a new scenario, it is one that must be addressed immediately. This does not diminish the importance of study on language extinction rather it emphasizes the need to incorporate dialectal variations as an additional subject of study.
Literature Review
Prior research has also highlighted the advantages of incorporating student's variety (dialect) in education. According to Tegegne (2015), the student's native dialect helps in effective and successful learning. He discusses two hypotheses on language variations that have influenced the field of education; the Deficit hypothesis and the Differences hypothesis. The former advocates for the eradication of the use of dialects in favor of the standard dialect while in the latter, non-standard dialects are seen as different ways of expressing the same idea. This advocates for the use of the non-standard varieties for educational purposes. In a study conducted by Solano-Flores and Li in 2006; it was observed that students performed better when they were administered tests in the local dialect than the standard dialect of the language Haitian-Creole (Tegegne, 2015).
The worst causality among the dialect disappearance is the slow fading away of lexical words. Lexical repertoire is crucial for any language. At present, language teaching educators emphasize the importance to incorporate more vocabulary in teaching and learning. Many studies have associated enhancing the dimensions of lexical repertoire as a source of enriching vocabulary and treasuring language. Though over the years, language word formation continues increasing with numerous new coined words or loaned words, the decrease in words that once belonged to the lexical repertoire of the language is on the rise. Matras (2010) discusses the same in light of the Romani dialects in Britain. With the replacement of English lexicon across the language, Romani lexicon slowly stated depleting. Subsequently, Romani language retreated with increasing use of English words. This is especially true in cases of language that has a number of dialects with an evolving standard or norm language.
In Kerala, a state that has 14 districts it can be said that there are as many dialects of Malayalam as the number of districts. Broadly classified, there are three major regional dialects-south, central, north. There are also differences in dialect along social lines with respect to caste and religion. Prior studies in the field of dialect studies of Malayalam have predominately been done on social lines of religion, caste and tribe. These studies (Subramoniam, 1974;Bhattacharya, 1976;Gopinathan, 1975) provide a comprehensive picture of the linguistic variation present in the state ("Language variation and external influence," n.d.). The dialect survey by V.I. Subramoniam in 1974 identifies twelve dialect areas; South Travancore, Central Travancore, West Vempanad, North Travancore, Cochin, South Malabar, South Eastern Palghar, North Western Phalgat, Central Malabar, Wayanad, North Malabar and Kasaragod. His survey identified these areas through the analysis of Malayalam spoken by Ezhavas and Tiyyas. However, since these studies are individualistic in nature and often deal with one certain community or caste, it is not systematically done from the dialectal point of view.
According to Ethnologue, the regional dialects are; Central Kerala, Kasaragod, Kayavar, Malabar, Malayalam, Mappila, Nagarai-Malayalam, Namboodiri, Nasrani, Nayar, North Kerala, South Kerala, Pulaya (Malayalam, n.d.). These regional dialects are dependent on regional, community and caste lines. The Mappila dialect (spoken by the Mappila Muslim community on Kerala, predominately in Malabar region) differs very significantly from the literary Malayalam when compared to the other dialects in the state. Among the dialectal regions in Kerala, the central Kerala dialect (used in Kottayam district) shows the closest affinity to the written Malayalam SL. The dialectal variations with respect to differences in the pronunciation are mostly colloquial and do not find their way into the formal written format. On the other hand, vocabulary of dialectal variation can be depicted in written format. Officially, Kerala government has signaled the variation in dialects at the district level.
Language Education in India and Kerala
Following the independence of India in 1947, the attention given to education became a major concern to the government of India and the state. From the very first National Education Policy, NEP 1968 and the later 1986 policy, the focus has been to incorporate the cultural and geographical diversity in the nation's education system. The latest National Curriculum Framework (NCF) 2005 guidelines on language, upholds the multilingual 1 character of Indian society. The curriculum emphasizes the need to implement the threelanguage formula 2 to promote multilingual proficiency and national harmony. Focus should be on the recognition of children's home language(s) or mother tongue(s) as the best medium of instruction (these include tribal languages).
In Kerala, in the quest to implement the three-language formula, the option of learning other foreign languages or Indian languages affected the study of mother tongue (Kerala Curriculum Framework (KCF) 2007, p.43). Kerala Curriculum formulators attempt to address this issue, realizing the extent to which the neglect of the mother tongue has been overlooked for a long time. The KCF states that Kerala is one of the only states in India that has not made the learning of mother tongue mandatory to complete one's school education. To remedy this, the new curriculum highlighted the study of mother tongue as the learner's right. The curriculum posits that all learners must be given the opportunity to learn the mother tongue in all stages of schooling. Presently, mother tongue teaching involves just the standard Malayalam, not taking into consideration the elements of differences between them. As rightly said by policy makers, a standard language is necessary for the proper function of government and educational systems. Therefore, this warrantees a uniform norm in official documents, newspaper and textbook formulation. However, the suppression of dialectal variation has been significant in the study of language. The KCF 2007 states that, "the importance accorded to the standard variety of language affects the mother tongue adversely. The domestic language and the dialectical variant of the mother tongue used by the child should be recognized. Language should not be a stumbling block in the construction of knowledge and self-expression" (pp.44-45). Accordingly, at the pre-school, KCF advocates the need to have a framework that has the scope to adapt according to the regional diversities (KCF 2007, pp.74).
There is also a special mention in the KCF 2007 regarding the language variation present in tribal language with respect to the standard language in textbook. The curriculum states the need for "providing learning environment in tribal areas to use their local dialects in standards 1 st and 2 nd and then shift to formal language" (pp. 76) in the section of learning environment. Keeping in mind the variations within the language, KCF 2007 points out the need to include language variations: "The textbooks may be prepared at the district level, considering the regional elements" (pp. 42). This is far from being done.
Also, the need for such a change is relevant on account of the present homogenized nature of Malayalam language present in the textbook. Though the NCF 2005 posits the need to formulate a curriculum that deals with children from different ethnic, social and cultural background, it has not been reflected in the KCF 2007. The latter ignores the diversity that exist in the variations and attempts to homogenize the curriculum with the 'Standard Malayalam' that's visible in all official transactions. What goes unheard is the dialectal variations and the mention of the language of tribal communities in the textbook. In case of the latter, steps are being taken by the school educators and teachers to help in the easy synthesis of standard Malayalam with tribal language.
The process to the formation of textbook at different districts level is not an easy task. Also, an attempt to involve all linguistic stylistics and variations that exist in each dialect is not probable. Such an attempt would only seem to hamper the proficiency of mother tongue. This thus involves a thorough examination of variations that can be deemed to be implemented through written language. Accepting the fact that such a move would take time, the scope of immediate changes seems far-fetched. However, by setting aside the phonological, morphological and syntactic variations (as these often do not go beyond the boundaries of comprehension and speaking into writing), importance can be given to the lexical variations that exist in the language. Such a move would only strength the language and continue to enrich the lexical resources of the language.
Objectives
The following objectives were formulated and a study was conducted to address the same.
1. Is there an awareness of dialectal variations specifically with respect to the lexicon of the language? 2. What is the scope of the above said variations?
Methodology and Procedures
The study involved both quantitative and qualitative approach. The quantitative analysis helps in generalizing the data based on statistics while the qualitative gathers in-depth information and uses words to describe the data. This mixed-method approach ensures the validity and reliability of the research study conducted.
A pilot study was designed to check the feasibility of conducting the study. Initially, the focus was given to assessing people from different regional dialects, however it was found to be not feasible as the same size would be exhaustive. Accordingly, it was decided that the participants would be from different educational backgrounds. This decision was found to be favourable over including people from different regional dialects and characterization on social lines. A list of target words (14 words) was created, taken from textbooks of upper primary and secondary classes (Kerala State Syllabus) for the questionnaire. For uniformity and to homogenize the list, all words belonged to the 'noun' category. The words included the names of fruits, vegetables and fish. For example, 'tapioca', 'papaya', 'bitter gourd', 'pomegranate', 'pineapple', 'sardine' and 'anchovy'. In the questionnaire, the participants were asked to write the first word-the word you know, the one spoken at home or the one they use commonly (could be English). In the second line they were asked to write all the words they know of the particular lexical item in the language. As dialectal variation can also be due to additional factors such as language contact and lexical borrowings, the words selected for the questionnaire were verified to be not affected by these factors.
The prepared questionnaire was given to 16 participants of the University of Hyderabad (UoH) based on convenience sampling and to homogenize the participants' exposure and environment. The participants of the study belonged to varying ages, between the ages 20-30. Out of the 16 participants, 9 were female participants and 7 were male participants. The participants of the study were grouped into three different categories. The data was collected and analysis was done by categorizing the words they wrote. The phonetic transcription of each word was done to account for the phonological variations. However, the phonological variation isn't taken into account as the study focused on the lexicon of the language. Frequency of each word was noted with respect to the four categories mentioned above.
Results and Discussion
The data of the questionnaire is provided in the appendix B. For the purpose of discussion, the different dialectal variation of the lexicon is provided with the percentage of use in each category across participants in Table 1. From the questionnaire data, the percentage of dialectal variation words of the lexical item is plotted against each participant. (Figure 1).
Figure 1. Pivot chart of percentage of dialectal variation lexical items against each participant in different category
From the questionnaire data (Appendix B), it is evident that the category 3 participants knew more of the dialectal variation lexical items when analysed against category 1 and category 2 participants ( Figure 1). The data reveals the increasing use of English words. For example, the use of the word 'passion fruit', 'pineapple' and 'pomegranate' in place of its Malayalam counterparts. In fact, the fruit 'passion fruit' has a Malayalam word-mu:si:liŋa:. But this usage was not known to any one of the participants. This fruit also has other Malayalam words, mu:solikyə and vallina:raŋa. For the word 'pineapple', though the Malayalam words kaitaccakka and kanna:raccakka/ anna:raccakka were given as answers, 6 participants provided pineapple as the first word. Only one participant knew the word 'pritticcakka'. Also, nobody knew the Malayalam 'parakkiccakka'. The same can be said of the word, 'papaya', 'sardine' and 'guava'. Nobody knew the words, kappara, kapparaykka: (for papaya), cuvappuratnam (for sardine), poyyappazam, koyya:kka (for guava).
The only word that almost all the participants were aware of the presence of linguistic variation was 'tapioca'. This can be credited to the fact that tapioca curry is one among the main dish of Kerala. In case of the word 'papaya', the frequency of Malayalam dialectal variation words was less. Only one participant knew the word karu:tta. Also, in case of the word 'anchovy', the nine dialectal variation words had a skewed distribution. Five of the dialectal variation words were known only by 1 participant in category 2 and 3. The same can be said to be the case for the dialectal variation words of 'Pink Perch fish'. Three of the words, sakkara, ma:ŋako:ra, lis was also known only by 1 participant. These are some of the examples that were observed from the pilot study. Overall, there were very few participants who knew almost all the dialectal variation words of the lexical item. There were only three dialectal variation words that were known by three participants of which, one was 'passion fruit' for which the English word was used.
In the first category, one participant was found to know more dialectal variation lexicons when compared to the other participants in the same category. On detailed inspection it was seen that the participant had travelled to different places inside Kerala (different districts) on account of parent's occupation. This explains the participant's better awareness of the different lexicons as the participant was exposed to the different lexicons in his social environment. This shows how the lexicon variations have presently just become regional lexical items and is slowly disappearing from the Malayalam lexicon. The results also showed that the lexical variations were not due to different word formation processes. Rather, the lexical items were a result of the different dialect areas and are arbitrary.
With even one generation of speakers not speaking the words, it can be rightly said that the language has lost specific words existing in the original lexical reservoir of language. As discussed earlier with the example of the Romani dialects (Matras, 2010), the language continued to have a special linguistic repertoire but in comparison a depleted repertoire on account of the increasing use of English language. This is why the KCF 2007 move to create opportunities for learning mother tongue is significant. Making Malayalam mandatory would also go a long way in warranting the survival of Malayalam words to the next generation.
Additionally, from the results it should be noted that certain participants who spent a greater number of years at the University of Hyderabad had a much better knowledge of the different lexicons. This can be explained with regards to the exposure to other dialects, personal interaction, media and age group. Social media, movies and certain YouTube channels are trying to highlight and incorporate the dialectal variations in speaking. Though initial attempts were to serve humour, at present it serves as a medium to assert one's identity. This can be observed in several Malayalam movies where actors and actress take up speaking with distinct pronunciation and vocabulary to highlight where the characters are from 3 . The same can be said of many YouTube channels 4 too. One can also find the different dialects in the literary genre. The 19 th century novelists O. Chandu Menon and C. V. Raman Pillai made use of these dialectal variations through the use of different social, historical and culture aspects. These steps have gone a long way in revitalizing the dialectal variations and its use. Though however extensive this attempt has been, it does not entail a clear awareness of the terms. Especially to those who are not partakers of social media or movies. This is also true in case of rural areas where access to these is limited.
Conclusion and Suggestion
In light of the analysis of the study, it is evident that depletion in lexicon repertoire is serious in the long run for language existence. One important remedy is the official inclusion of dialectal variation in the textbooks. This inclusion can be in the form of footnotes, so that a student who comes across the particular word can be aware of the existence of its Malayalam counterparts. From detailed interviews, it was observed that some of the participants had no knowledge of the existence of particular words. Hence, an official inclusion of words accommodates the need to expose learners of mother tongue to the exhaustive lexical repertoire of the language. Such a move, I believe would go a long way in educating the present and upcoming learners of the many lexical variations existing in the dialects of the language alongside the already present word in textbooks. This would do away with the disproportionate representation of lexical variation and ensure linguistic equality between the dialects.
The new National Education Policy 2020 asserts the importance of mother tongue and the mandatory inclusion of the mother tongue in curriculum (NEP 2020, p. 13). However, not much has been said on the inclusion of regional dialects. In India, a land where many languages have several dialects it is only reasonable to account for regional variations in the textbooks.
Furthermore, future studies incorporating factors such as place of education (rural/urban), education of parents and school syllabus into the design and analysis of the study would strength the result and shed more light on the scope of linguistic variation among dialects. Also, it would be favorable to include more target words for an extensive analysis. One necessary category that needs further research is the study of kinship terms. As kinship terms are bound with both regional and social differences, this category is an excellent source in mapping a dialect geography that comprises regional, community and caste differences on a large scale. Such a step would help to enforce the richness, beauty and diversity of the language for the future generations and an effective growth of dialectal variations. Subsequent studies can be done on other languages to analyze the presence of dialectal variation and to ascertain the importance of the same. | 5,367.6 | 2020-11-03T00:00:00.000 | [
"Linguistics",
"Education"
] |
Solar wind and geomagnetism: toward a standard classification of geomagnetic activity from 1868 to 2009
We examined solar activity with a large series of geomagnetic data from 1868 to 2009. We have revisited the geomagnetic activity classification scheme of Legrand and Simon (1989) and improve their scheme by lowering the minimum Aa index value for shock and recurrent activity from 40 to 20 nT. This improved scheme allows us to clearly classify about 80 % of the geomagnetic activity in this time period instead of only 60 % for the previous Legrand and Si- mon classification.
Introduction
Geomagnetic activity may be defined as the magnetosphere's response to the transitory variation of solar activity. Simon (1985, 1989) and Simon and Legrand (1989) show the existence of two categories of magnetic perturbations: the first is organized into recurrent features and the second forms a range of intense and short events distributed at random. These authors divide geomagnetic activity into four classes: quiet, recurrent, shock, and fluctuating Richardson et al., 2000;Richardson and Cane, 2002;.
To determine the classes, Legrand and Simon (1989) and Simon and Legrand (1989) built a diagram similar to Bartels 27-days rotation using the geomagnetic index Aa from 1868 to 1977. This diagram, named a pixel diagram, represents the geomagnetic data as a function of solar activity for each solar rotation (27 days) and gives an overview of the geoeffectiveness of solar events.
Of these four classes of geomagnetic activity, only three classes (quiet activity, recurrent activity and shock activity) are clearly selected and defined. The fourth (fluctuating) contains all data which does not fit into the other three categories.
In light of the strong correlation between the Aa index and solar wind established by Svalgaard (1977) and the fact that 91.5 % of solar activity is reflected by solar wind speed Simon, 1985, 1989), we decided to try to further refine fluctuating activity to better understand its origins. We find that it is appropriate to subdivide the fluctuating activity into three new classes including Coronal Mass Ejection (CME) manifestations and moderate solar wind effects (Gopalswamy et al., 2003;Ramesh, 2010).
Data
The times of sudden storm commencements (SSC), which are rapid increases in the magnetic field observed at ground, and the number of CMEs per days are taken from http: //isgi.latmos.ipsl.fr/ and http://cdaw.gsfc.nasa.gov/CME list, respectively. The solar wind speed and the international sunspot number (SSN) are obtained from http://omniweb. gsfc.nasa.gov/form/dx1.html. The SSN data are used to determine solar cycle phases.
Data analysis
To classify geomagnetic activity, Legrand and Simon (1989) use the following criteria: (1) SSC times are used to determine when shock events contribute to geomagnetic activity, (2) 91.5 % of solar activity is reflected by solar wind speed, (3) the strong correlation between Aa (Mayaud, 1971(Mayaud, , 1973(Mayaud, , 1980 and the solar wind speed (Svalgaard, 1977). They divided the observations into four classes of geomagnetic activity: Shock activity (SA), quiet activity (QA), recurrent activity (RA) and fluctuating activity (FA), illustrated in Fig. 1a. These four classes are fully described in Legrand and Simon (1989) and . It is, however, important to note that Legrand and Simon (1985) only classified 43 % of SSCs as shock events, including only those on days with Aa > 40 nT. The remainders were classified as fluctuating activity (V ≥ 450 km s −1 ). Similarly, they include in RA only times when Aa is greater than 40 nT.
To refine the fluctuating activity , as suggested in introduction, we use a 142 year-long Bartels diagram of Aa indices, SSC reports and two new classes with the following criteria: (4a) Corotating Moderate Activity. This category of geomagnetic activity is defined as corotating, stable solar wind stream producing moderate geomagnetic effects, Aa between 20 nT and 40 nT. We identify these events in the pixel diagram ( Fig. 1) by the recurrence from on solar rotation to the next of yellow and green areas with no SSCs.
(4b) Cloud Shock Activity. These events are shocks which cause only a moderate increase of activity level. We select only pixels within three days after SSCs with the colours yellow, green in the pixel diagram (Bartels diagram), and so the Aa indices are between 20 nT and 40 nT.
(4c) Unclear Activity. This transient activity is composed of days not included in any other classes. Its annual level is obtained after subtracting Cloud Shock Activity and Corotating Moderate Activity annual levels from the fluctuating activity defined by Legrand and Simon (1989). Thus, we have increased from four to six the number of geomagnetic activity classes. Figure 1a is an example of pixel diagram for year 1994 which illustrates the five clear activity classes: quiet, recurrent, shock (defined by Legrand and Simon, 1989), cloud shock and corotating (defined in this paper).
In the next paragraphs we adopted the following notations: (QA) for Quiet Activity, (SA) for Shock Activity, (RA) for Recurrent Activity, (CSA) Cloud Shock Activity and (CMA) Corotating Moderate Wind Activity. The Unclear Activity due to the oscillation of the neutral sheet will be noted (UA).
For each year, the geomagnetic activity level characterized by the Aa indices can be expressed now by the sum of the six classes of geomagnetic activity: Geomagnetic activity=QA+SA+RA+CSA+CMA+UA (1) By the past with Simon and Legrand (1989), the geomagnetic activity was the sum of four classes of activity: Geomagnetic activity = QA + SA + RA + FA; (2) with FA = CSA + CMA + UA. For each class we determine the yearly level of activity as the sum of all the daily Aa of the days of the class. For example the QA activity level of a year is given by: N : number of quiet days in the year.
Results
For the 142 years covered by this study, we divided the fluctuating activity into three news classes (corotating activity, cloud shock activity and unclear transient activity) to obtain six classes of geomagnetic activity. Figure 2 plots the yearly levels of the different classes of geomagnetic activity and the sunspot number from 1868 to 2009. Figure 2a shows that quiet magnetic days are more prevalent from the end of the 1800s to ∼1950 and at the end of the sunspot cycle 23 (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009) and that the distribution is relatively flat during the second half of the 1900s. Figure 2b superimposes the recurrent activity , the corotating moderate activity and the sunspot number. The higher magnetic activity (RA) dominated the CMA activity during most of the solar cycles from 1868 to 2009. Figure 2c deals with shock activity,cloud shock activity and the sunspot number long time evolution. The lowest shock activity levels were observed during cycles 15 (1913-1923) and 23 (1996-2009) while cloud shock activity reaches its highest level in the same time periods. Figure 2d superimposes the long time variations of fluctuating activity and unclear activity. The difference in these two curves is due to the removal of CSA and CMA from fluctuating activity, which will be discussed later. Figure 2c superimposes the yearly sunspot number on the percentages of shock activity and cloud shock activity from 1868 to 2009. We can see a close link between sunspot number and both activity levels time evolution. Figure 2b shows that the largest percentage of recurrent activity occurs after the maximum in the sunspot number. This figure also shows that significant corotating activity is observed near sunspot maximum but that the largest percentage occurs in the declining phase. Figure 3a and b superimposes the sunspot number on the percentage of shock events and recurrent events, respectively. Figure 3a indicates that the highest contribution of shocks occurs near solar maximum. Figure 3b shows that the most important contributions from high-speed solar wind streams are observed in the declining phase of sunspot cycle.
Discussion and conclusion
We investigate geomagnetic activity and its solar wind origins from 1868 to 2009 (142 years). We define three news classes of geomagnetic activity from the fluctuating activity class of Legrand and Simon (1989). Thus we now define six classes of geomagnetic activity, quiet activity (QA), recurrent activity (RA), shock activity (SA), cloud shock activity (CSA), corotating moderate activity (CMA) and unclear activity (UA). CSA, CMA and UA are extracted from the fluctuating activity previously defined by Legrand and Simon Table 3. Comparison between the geomagnetic classification (Legrand and Simon criterion) and the classification after the refinement of the fluctuating activity. Legrand and New classification Simon (1989) 1989). The new classification allows us to assign about 80 % of the days into specific geomagnetic actitivy classes (as opposed to the FA and UA classes which are composed of days which fit not other group) in comparison to 60 % assigned with the previous scheme . From Fig. 2c, it is shown the strong correlation between cloud shock activity level and sunspot number variations. The concordance between the cloud shock activity and the shock activity time evolutions throughout the many sunspot cycles suggests that both classes have the same mechanism. The different levels of Aa index probably result from differences in shock strength due to differences in the CMEs on the Sun which generate these shocks. We also point out a good correlation (≥0.79) between the yearly level of SA and CSA for cycle 11 to cycle 23 as summarized in Table 2. Thus SA and CSA are different parts of a distribution formed from the same type of event and could be combined into one class. Figure 3a shows the sum of shock activity and cloud shock activity levels from 1868 to 2009. This figure points out a fairly constant in shock event levels was present for all the Legrand and Simon (1989), summarizes their classification scheme and the variation of each class with the sunspot cycle. The recurrent events reach their peak a few years before the sunspot minimum and sometimes near the sunspot maximum. Figure 2b shows that corotating moderate activity which reaches its peak near the sunspot maximum or few years before sunspot minimum (sunspot declining phase) is in concordance with recurrent events occurrence summarized in Table 1. Sunspot maximum and the declining phase are dominated by recurrent high wind stream flowing from coronal holes. In addition, the yearly level of RA and CMA present significant correlation coefficient (≥0.70) for each solar cycle since cycle 11 as shown in Table 2. Thus it appears that CMA and RA are also generated by the same mechanism. These results express the necessity to combine CMA and RA to keep one class as we did for SA and CSA. Figure 3b gives an overview of the recurrent events (CMA plus RA) and the sunspot number from 1868 to 2009. The recurrent events are likely governed by the distribution of coronal hole high-speed solar wind which occurs most often at low-latitudes in the declining phase of the sunspot cycle. The Table 3 summarizes our study and compares it with the results of Legrand and Simon (1989) for the period 1868-2009. We defined four classes of geomagnetic activity but change the criteria from the classifications of . These new criteria allow us to positively classify 80 % (as opposed to 60 % previously). Thus we defined: (1) quiet-days, those days with an Aa index of geomagnetic activity below 20 nT; (2) recurrent activity, the days under the control of recurrent high wind stream with an Aa index above 20 nT without SSC; (3) shock activity, which we attribute to CMEs. We select only those days with non-recurrent SSCs and days within three days after the SSC which show enhanced Aa value above 20 nT; (4) unclear activity. This transient activity is formed by the days which do not fit in the three other classes.
The present refinement in geomagnetic activity classification constitutes a starting point towards a standard classification of solar activity using especially solar physics and solar phenomenon before their geoeffectiveness. | 2,762.2 | 2012-02-01T00:00:00.000 | [
"Geology"
] |
Hypocoercivity and Fast Reaction Limit for Linear Reaction Networks with Kinetic Transport
The long time behavior of a model for a first order, weakly reversible chemical reaction network is considered, where the movement of the reacting species is described by kinetic transport. The reactions are triggered by collisions with a nonmoving background with constant temperature, determining the post-reactional equilibrium velocity distributions. Species with different particle masses are considered, with a strong separation between two groups of light and heavy particles. As an approximation, the heavy species are modeled as nonmoving. Under the assumption of at least one moving species, long time convergence is proven by hypocoercivity methods for the cases of positions in a flat torus and in whole space. In the former case the result is exponential convergence to a spatially constant equilibrium, and in the latter it is algebraic decay to zero, at the same rate as solutions of parabolic equations. This is no surprise since it is also shown that the macroscopic (or reaction dominated) behavior is governed by the diffusion equation.
Introduction
We consider N chemical species S 1 , . . . , S N with different particle masses moving in a periodic box or in whole space. The interaction with a stationary background with constant temperature T triggers first order chemical reactions with reaction rates independent of the velocity of the incoming particle. The velocity of the outgoing particle is sampled from a Maxwellian distribution with parameters taken from the background, i.e. mean velocity zero and temperature T . The resulting reaction network is assumed to be connected and weakly reversible, meaning that for each reaction S i → S j there exists a reaction path S j → · · · → S i . These assumptions lead to a system of N linear kinetic transport equations for the phase space number densities of the reacting species. We shall make the additional assumption that the species can be split into two groups of light and heavy particles, where the particle masses are of comparable size within each group but strongly disparate between the groups. A corresponding nondimensionalization of the equations, assuming at least one light species, will suggest a simplified model, where the heavy particles do not move. As a result we consider a system of kinetic equations (for the light species) coupled, via the reaction terms, to a system of ordinary differential equations (pointwise in position space, for the heavy species).
The construction of equilibrium solutions is straightforward. In equilibrium, the position densities are constant, and the velocity distributions of the light particles are Maxwellians. The position densities are complex balanced equilibria of the reaction network. Existence and uniqueness for given total mass are standard results of the theory of chemical reaction networks.
Our main results are exponential convergence to equilibrium in the case of the periodic box and algebraic decay to zero in whole space. In both situations the rates and constants are computable. Although general results for Markov processes imply that relative entropies are nonincreasing [10], the decay result is not obvious, since the entropy dissipation is not coercive relative to the equilibrium. We employ the abstract L 2 -hypocoercivity method of [6,7] and its extension to whole space problems [2]. The main difficulty is the proof of microscopic coercivity, meaning here that the reaction terms without the transport produce exponential convergence to a local equilibrium, where the total number density of all species might still depend on position and time. Two alternative proofs are presented. In the first one, relaxation in velocity space is separated from relaxation to chemical equilibrium and known results for the latter [8] could be used. The second proof extends the proof in [8] by introducing reaction paths in species-velocity space. For completeness and comparability we fully present both proofs, showing that the second proof never gives a worse result.
The second result is a macroscopic or fast-reaction limit. For length scales large compared to the mean free path between reaction events and for the corresponding diffusion time scales, the system is in local equilibrium and the total number density solves the heat equation.
Systematic approaches to hypocoercivity have been started in [15,18], where Lyapunov functions based on modified H 1 -norms are constructed. More recently, an approach without smoothness assumptions on initial data, motivated by [11], has been developed in [6,7], see [5,14] for overviews. Recently the latter approach has been extended to the analysis of algebraic decay rates in whole space problems [2]. Hypocoercivity for systems of kinetic equations coupled by linearized collision terms has been shown in [4]. For a nonlinear system modeling a second order pair generation-recombination reaction, both hypocoercivity and the fast reaction limit have been analyzed in [17].
This work can be seen as an extension of the corresponding result for linear reaction diffusion models [8], which has recently been extended to general mass action kinetics [9], bringing the theory for reaction diffusion models close to the best results on the global attractor conjecture [13] for ODE models without transport [3].
Many extensions of the present results are desirable. Besides the inclusion of collision effects and of second order reactions, questions of energy and momentum balance pose significant challenges, where a trade-off between mathematical manageability and modeling precision has to be found. One goal is the rigorous justification of the derivation of reaction diffusion systems from kinetic models as an extension of results for linear cases [1].
Finally, we describe the structure of the rest of this article. In the following section the kinetic model is formulated including a dimensional analysis and the reduction to a system with partially nonmoving species. The formal macroscopic limit is presented and our main results on the long term behavior of solutions and on the rigorous justification of the macroscopic limit are formulated. In Sect. 3 our main technical result on 'microscopic coercivity' is proven, i.e. a spectral gap for the reaction operator. Sections 4 and 5 are concerned with the proofs of our main results on long time behaviour and, respectively, on the rigorous macroscopic limit.
The Model-Main Results
We denote the chemical species by S 1 , . . . , S N and the reaction constant for the reaction S i → S j by k ji ≥ 0, i, j = 1, . . . , N , where k ji = 0 means that the reaction does not occur. More completely, also including velocities v ∈ R d , we assume that the jump (S i , v) → (S j , v ) occurs with rate constant k ji M j (v ), as described above independent of the incoming velocity, where the Maxwellian distribution is given by with the Boltzmann constant k B , the constant given background temperature T , and the particle masses m 1 ≤ · · · ≤ m N of the respective species In the following, integrations with respect to x will be written over , where = [0, L] d for the periodic box and = R d for whole space.
The phase space distributions satisfy the evolution system where the left hand side describes free transport and the right hand side the chemical reactions with position densities where we will sometimes also use the notation ρ f , j to avoid ambiguity. We assume that there are N l ≥ 1 light species S 1 , . . . , S N l and N − N l heavy species S N l +1 , . . . , S N . The separation of the two groups is expressed in the assumption In a nondimensionalization we introduce as reference velocity the thermal velocity v th := k B T m N l of the heaviest light species S N l . As reference time t we choose an average value of k −1 i j , i, j = 1, . . . , N . The reference length is given by the equations (1), (2) look the same, but with In particular we have θ i ≥ 1, i = 1, . . . , N l , for the light particles and in the distributional sense as μ → 0. In this limit it is consistent to also look for solutions, where the heavy particles are nonmoving, i.e.
Therefore, for the rest of this work we shall consider the system with N l ≥ 1 and with M i given by (3), subject to initial conditions with initial data satisfying f I ,i ∈ L 1 and write the system (4), (5) in the equivalent form with σ i = 1 for i ≤ N l and σ i = 0 otherwise. We shall consider initial value problems with The system (7) conserves the total number of particles: The total position density and therefore
Local and Global Equilibria
is called a local equilibrium for (7), if it balances the reactions, i.e., if The set of all local equilibria can be described in terms of properties of the directed graph with nodes S 1 , . . . , S N and edges S i → S j , when k ji > 0. Roughly speaking, there is a simple characterization of local equilibria, if the graph has enough edges.
Assumption A1
The directed graph corresponding to the reaction network is connected and weakly reversible, which means that for each pair An example is given in Fig. 1. Note that the path from j to i is in general not unique. For the following a fixed choice of a path of minimal length P i j is used for each pair (i, j). This also means that paths are not self-intersecting in the sense that each reaction S i p−1 → S i p appears only once.
Lemma 2 Let Assumption A1 hold. Then every local equilibrium is of the form
Proof A first consequence of Assumption A1 is that for each i ∈ {1, . . . , N l } there exists at least one k ji > 0 and at least one k i j > 0. This implies that for a local equilibrium Now it is a standard result of reaction network theory (see, e.g., [8,12]), in our simple case of first order reactions related to the Perron-Frobenius theorem, that the connectedness and weak reversibility imply that there is a one-dimensional solution space spanned by (η 1 , . . . , η N ), where all components have the same sign. In the language of reaction network theory these are complex balanced equilibria.
A global equilibrium is a local equilibrium, which is also a steady state solution of (4), (5) compatible with conservation of total mass. Since at least one equation has a transport term, the function ρ from Lemma 2 has to be constant for a global equilibrium. In the case = R d we expect dispersion and consequential decay to zero. Therefore a nontrivial global equilibrium is only defined for = T d by
Microscopic Coercivity-Convergence to Equilibrium
We write the system (7) in the abstract form with the transport operator T and the reaction operator L defined as Lemma 2 characterizes the nullspace of the reaction operator. A projection to this nullspace is given by It is easily seen that is a projection and that f , g = f , g , which implies that is orthogonal. Since the application of the projection involves integration with respect to v and summation over all species, we do not only have L = 0, but the mass conservation property of the collision operator can be written as L = 0.
Considering the quadratic relative entropy with respect to the local equilibrium F suggests the introduction of the weighted L 2 -space H with the scalar product and with the induced norm · . In the case = T d , the members of H are periodic with respect to x. Note that, for i > N l , we have , and it has to be understood that Our main technical result, which will be proved in the following section, is coercivity of the reaction operator with respect to its null space. This property will be called microscopic coercivity. (15) is the orthogonal projection to the nullspace of the reaction operator L : H → H defined in (14). Furthermore there exists a constant λ m > 0 such that
Lemma 3 Let Assumption A1 hold. Then defined by
This lemma is one of the main tools in the proofs of our results on the long time behavior, presented in Sect. 4: Theorem 4 Let Assumption A1 hold and let = T d . Then there exist constants C, λ > 0 such that for every f I ∈ H and with f ∞ given by (12), the solution f of (7), (8) satisfies Theorem 5 Let Assumption A1 hold and let = R d . Then for every f I ∈ H ∩ L 1 (dv dx) there exists a constant C > 0 such that the solution f of (7), (8) satisfies
Macroscopic (Fast Reaction) Limit
We introduce a diffusive macroscopic rescaling x → x/ε, t → t/ε 2 with 0 < ε 1. Note, however, that in the case = T d we still consider a fixed ε-independent torus after the rescaling. The abstract form (13) of our system becomes with a now ε-dependent solution f ε . Assuming convergence to f 0 as ε → 0, the formal limit of the equation implies that f 0 is a local equilibrium, i.e.
. It remains to determine ρ 0 . The rescaled microscopic part R ε : with the formal limit Finally, we also need the conservation law and observe that the diffusive scaling is consistent, since the diffusive macroscopic limit identity holds, which is easily checked, since maps to a vector of centered Maxwellians, T provides a factor v, and the second application of involves an integration with respect to v. The property (20) will also be essential in the proof of decay to equilibrium in Sect. 4 and it guarantees the necessary solvability condition T f 0 = 0 for (19). Substituting its solution into the limiting conservation law should provide the missing information on ρ 0 : where L denotes the restriction of L to (1 − )H. In order to translate the abstract result, we first note that Since application of involves integration with respect to v as first step, the identity . The solution of (19) is then given by A straightforward computation gives
Thus, (21) is equivalent to the diffusion equation
The following result, providing a rigorous justification of the formal asymptotics, will be proved in Sect. 5. It also relies on the microscopic coercivity result Lemma 3.
Theorem 6 Let Assumption A1 and f I ∈ H hold (with
Then the solution f ε of (8), (17)
Microscopic Coercivity (Proof of Lemma 3)
We shall give two different proofs of the coercivity result. Both are inspired by the proof of the corresponding result in [8]. The first approach is based on a splitting between the species and velocity spaces, where for the former the result of [8] can be used directly. In the second approach the reaction paths of Assumption A1 are extended to paths in the larger species-velocity space. In both cases the coercivity constant λ m can in principle be computed explicitly. Since the computations are rather based on algorithms than on explicit formulas, a comparison of the results for both approaches would be difficult.
For the following computations we introduce U i := f i η i M i , V i := g i η i M i and rewrite the reaction operator as where the primes indicate evaluation at v . This implies (11) is used in the second term on the right hand side and the change of variables (i, v) ↔ ( j, v ) in the third: This shows that L can only be expected to be symmetric in the case of detailed balance, i.e. when k i j η j = k ji η i for all i, j = 1, . . . , N . It also shows negative semi-definiteness of L: First proof -separation of species and velocity contributions: The strategy is to split the dissipation term into contributions measuring the deviation from Maxwellian velocity distributions on the one hand, and from reaction equilibria of the position densities on the other hand: The norm of the microscopic part of the distribution is split correspondingly: On the one hand the connectedness of the reaction network implies which allows to relate the first terms on the right hand sides of (24) and (25). On the other hand [8, eqn. (2.15)] could be used for the second terms. For comparability of the results of the two proofs we include the derivation of this second inequality. We start with the relation which is easily verified by adding and subtracting ρ in the parenthesis on the right hand side, expanding the square, and using that ρ is the total density. For each pair (i, j) we use Assumption A1 and choose a path of minimal length P i j from j to i, which implies by an application of the Cauchy-Schwarz inequality. With the definition In the second inequality above we have used that each pair (i p , i p−1 ) occurs only once in a reaction path of minimal length. This concludes the proof of microscopic coercivity with λ m = min{γ 1 , γ 2 }.
Second proof -species-velocity space paths: Starting from the representation (23) and using that the path from j to i is not self-intersecting we have With the Cauchy-Schwarz inequality on R P i j this can be estimated further by As indicated at the beginning of this section, the strategy in these estimates was to extend the path i 0 , . . . , i P i j in the species space {1, . . . , N } to all possible paths of the form As the next step we observe that Combining this with (27) completes the alternative proof of Lemma 3 with λ m = γ 2 , as defined in (26). The result of the second proof is always at least as good as that of the first.
Hypocoercivity
Quantitative results on the decay to equilibrium will be shown by employing the hypocoercivity approach of [7] with modifications introduced in [2]. In the case of the torus, = T d , we assume w.l.o.g. ρ ∞ = M = 0, which can always be achieved by a redefinition of the solution. Thus, in both cases = T d and = R d we expect f → 0 as t → ∞. The functional f → f 2 can then be understood as relative entropy and a natural candidate for a Lyapunov function. However, by the obvious antisymmetry of the transport operator T, d dt which is nonpositive as expected, but vanishes on the set of local equilibria, i.e. it lacks the definiteness required for a Lyapunov function. In [7] a Lyapunov function, or modified entropy, H[ f ] has been proposed, which has the form and with a small parameter δ > 0 to be determined later. In [7,Lemma 1] it has been shown that the operator norm of A is bounded by 1 2 and H[ f ] is equivalent to f 2 for δ < 1. For solutions f of (13), its time derivative is given by From the definition of A it is clear that the property A = A holds. On the other hand, the conservation of total mass by the collision operator is equivalent to L = 0, i.e. also projects to the nullspace of L * . As a consequence, the last term above vanishes: In [7] this property is the consequence of the assumption that L is symmetric, which does not hold here, as noted in the previous section. The first term on the right hand side of (30) controls the microscopic part (1 − ) f of the distribution function. The second term is responsible for the macroscopic part: Lemma 7 With the above notation, where D = i≤N l η i θ i , · 2 = · L 2 ( ) , and g = ρ g F is given by
Proof
The property g = g is obvious from its definition. We use the abbreviation L := (T ) * T and compute Therefore, using again the property A = A, A straightforward computation shows Lg = −D x ρ g F, completing the proof.
As a consequence, the first two terms on the right hand side of (30) provide the desired definiteness, since obviously AT f , f ≥ 0 and AT f , f = 0 ⇒ ρ g = 0 ⇒ f = 0. However, the remaining terms still need to be controlled. We start by using the diffusive macroscopic limit property (20), implying In [7, Lemma 1] it has been shown that the operator norm of TA is bounded by 1 implying, together with the above, Lemma 8 With the above notation, Proof With g as introduced in Lemma 7 and with A * = T (1 + L) −1 = T(1 + L) −1 we have The estimate and an application of Lemma 7 complete the proof.
Lemma 9 With the above notation, Proof We use (remembering L = L(1 − ) and, from the preceding proof, A * f = T g) Lemma 7, and the boundedness of L: Collecting the results of Lemmas 3, 7, 8, and 9 , we obtain Thus, for This shows that H[ f ] is a Lyapunov function. It remains to obtain the decay rates.
in the sense of distributions. The limiting equations are equivalent to the distributional formulation of the heat equation (22) for ρ 0 with the initial condition ρ 0 (t = 0) = R d f I dv .
The uniqueness of the solution of this problem implies the weak convergence of f ε to ρ 0 F. This completes the proof of Theorem 6. | 5,210.8 | 2019-01-24T00:00:00.000 | [
"Physics"
] |
Does college level the playing field? Socioeconomic gaps in the earnings of similar graduates: evidence from South Korea
The socioeconomic gap in participation at university is an enduring policy issue in South Korea, as in many other countries. However, less attention has been paid to the socioeconomic gap in the outcomes from tertiary education. This paper addresses this gap in the literature, using the Korean Education and Employment Panel (KEEP) data to investigate the extent to which the wages of Korean graduates who attended similar higher education institutions vary by socioeconomic background. The results show that a degree appears to largely level the playing field, in terms of earnings, between male graduates from poor and rich backgrounds. For females, by contrast, family background is still a strong predictor of earnings, even after allowing for institution attended and discipline of degree. Further, the wage premium for 2-year and 4-year college degrees also varies by family background. Four-year college degrees, contrary to popular belief, do not always attract a higher wage premium than 2-year college degrees, particularly for men from poorer family backgrounds.
Introduction
Traditionally, education has been highly valued in the Republic of Korea (hereafter South Korea). Most people are convinced that going to university, particularly one of the most prestigious universities, will guarantee good career prospects for any student (Cho, 2016). Participation in higher education is high, and approximately 70% of high school students in South Korea go to university (OECD, 2016). Certainly, up to the 1970s, individuals who went to university in South Korea (about 7% of the cohort at that time), regardless of their secondary school enrolment; approximately 99.7% of students attend. After graduating from upper secondary schools, students can choose whether or not to attend a college, and if choosing to attend, they can select either 4-year academic colleges or 2-year polytechnic colleges depending on their career goals and indeed their scores on a College Scholastic Ability Test (CSAT 1 ). On average, 4-year colleges are more academically selective.
In 2019, 430 colleges and universities offered either a 4-year bachelor's degree or a 2-year associate degree. Specifically, there were 191 4-year colleges, 137 2-year colleges, and 102 other types of college generally offering 2-year options. In 2019, 67.8% of high school graduates enrolled in a degree of some kind. Among students enrolled in tertiary institutions, 66 and 21% of students enrolled in 4-year and 2-year colleges, respectively (Korean Ministry of Education, 2019). The college enrolment rate in South Korea is high by international standards. For example, in the UK, approximately 50% of the cohort enrol in tertiary education, and in the USA, around 40% enrol (NCES, 2019; UK Department for Education, 2019). Participation in tertiary education in South Korea is arguably therefore almost saturated, especially among individuals from socioeconomically advantaged families (Byun & Park, 2017).
Although South Korea has mass higher education, entry is still competitive and on the basis of a national assessment, CSAT. Many students who are not satisfied with their scores on CSAT, held once a year, tend to retake the test and enter college in a later year. In 2018, approximately 20% of students enrolled in colleges had retaken the CSAT at least once (KESS, 2019). Retakes are counter-intuitively related to socioeconomic background. Richer students can afford to do more retakes (and hence have a better chance of achieving the score needed to enter a more prestigious institution) due to the high cost of college admission preparation, particularly private tutoring. For instance, in some high schools in the wealthiest neighbourhoods in Seoul, more than half of high school students retake the exam and enter college in the following years, seeking to attend a more prestigious institution (Jung et al., 2019).
Overall, a late entry into the labour market is a feature of the South Korean system. In addition to the trend of CSAT retakes, all males aged between 18 and 30 in South Korea are required to do military service for approximately 2 years. According to the Ministry of the National Defence, about 85% of males do this between the ages of 20 and 22. Also, many college students take a semester or year off to prepare for English language examinations and various other qualifications. As a result, it takes approximately 12.3 semesters (two semesters per year) to graduate from college in South Korea (14.3 semesters for males and 9.9 semesters for females (Yi, 2016)). Given this, any analysis of graduates' labour market outcomes needs to focus on individuals' earnings at a somewhat older age, namely when individuals have had a chance to have some time in the labour market, normally in their mid-30s.
The return to higher education in South Korea
In South Korea, the wage premium associated with more years of schooling and higher levels of educational attainment has decreased as participation in secondary and higher education has substantially increased (Joo, 2018;Lee, 2011). Although the evidence suggests that there remains a wage premium from a degree in South Korea, there is also increasing heterogeneity in graduates' earnings. Yi and Kim (2016), for instance, found that 2-year and 4-year college graduates earn 5.3 and 14.8% more, respectively, than high school graduates when controlling for work-related characteristics, namely workplace size, firm type, and location. The evidence is mixed, however. Some studies have even suggested that the wage premium from a degree is, in some circumstances, zero in South Korea (e.g., Joo, 2018).
There is also now a sizeable literature exploring the relationship between college prestige and the heterogeneity in graduates' earnings in South Korea (Han et al., 2012;Lee et al., 2018). This literature indicates that the prestige of the 4-year institution that a graduate attended has a strong positive effect on his/her initial job stability and current income (Jung & Lee, 2016). The premium for graduates from Seoul National University, for example, one of the most prestigious institutions in South Korea, is estimated to be around 12% over the earnings of graduates from other less selective 4-year institutions (Han et al., 2012). The literature also suggests bigger quality differences, as measured by various indicators, e.g., reputation, learning environments, and students' outcomes, within the group of 4-year colleges compared to 2-year colleges (e.g., Lee, 2011;Lee et al., 2014;Park, 2014;Yi & Kim, 2016). Generally, the evidence from the existing literature shows that the mean gap in earnings between graduates from 2-year and 4-year colleges may have shrunk or even become zero for graduates from many institutions, bar those from a few top 4-year institutions (Park, 2014). The shrinking earning gap between 2-year and 4-year college graduates is consistent with the greater heterogeneity in 4-year college graduates' earnings (Oh & Chae, 2014). Explaining the sources of such heterogeneity in graduate earnings is important, and one potential explanation is the family background of graduates, an issue we explore in this paper.
The heterogeneity in returns to higher education by gender South Korean labour market data also suggests a high heterogeneity in returns to higher education by gender. In 2019, South Korea had the largest gender wage gap, at 32.5% among the OECD countries (the OECD average was 12.9% (OECD, 2021)). Traditional Korean society has sharply defined gender roles, and as a result, females, even those with a college degree, have long experienced various forms of discrimination in the labour market. The empirical evidence suggests women experience occupational and industrial segregation, lower earnings than males doing similar jobs, and fewer promotion opportunities (Kim & Voos, 2007). The situation is not static, however. Though the gap is still large by international standards, the gender wage gap has gradually decreased since the Equal Employment Opportunity Act enacted in 1988. This Act enabled workers to be legally protected against discrimination on the basis of marital status and pregnancy (Hong, 2011;Monk-Turner & Turner, 2004). For instance, females, on average, earned 68% of male earnings in 2019, which is an improvement on the 1980s when women earned just 40% of male earnings.
One of the significant factors determining the large pay gap between the genders is female career interruption or challenges when returning to work for females after pregnancy and undertaking childcare (Shin, 2011;Kim, 2017). Evidence on the gender wage gap for younger people, i.e., relatively new entrants to the labour market, is therefore mixed. Some studies, for instance, have shown that there is no significant disparity in the earnings and employment rates of males and females for recent graduates in their late 20s, who have not yet experienced any career interruption (Kim, 2017;Kwon, 2018). This is consistent with various statistics that confirm that males and females have similar earnings in their mid-20s right after college graduation but that the gap widens between the genders in older workers, particularly after the age of 30 (Statistics Korea, 2018). Simultaneously, however, much literature has also suggested that more than 80% of females in their 20s have reported experiencing gender inequality in the labour market, and some recent studies have even suggested that taking other characteristics into account, female graduates, on average, earn less than their male counterparts (Choi et al., 2016;Kim & Oh, 2019). Female graduates from more selective 4-year institutions also appear to face even greater earning inequalities in the labour market than those from less-selective 4-year or 2-year colleges (Kim & Oh, 2019). Given the mixed evidence on the extent of the gender wage gap, particularly for young graduates, we model the determinants of earnings separately for males and females.
Socioeconomic gaps in graduates' earnings
This study will contribute to the significant literature from a range of contexts that has suggested a major impact from the family background on graduates' earnings, even after allowing for differences in the higher education institution they attended and the degree subject they studied. Evidence on this issue for the South Korean context will be provided, using a similar methodology to that used in the existing literature (e.g., Britton et al., 2019;Chetty et al., 2017).
Whilst there is considerable evidence of heterogeneity in returns to higher education by subject and institution, there is very little evidence on how graduates' earnings vary by socioeconomic background (Crawford & Erve, 2015). Further, the existing literature suggests contrasting results in different contexts, which is why empirical evidence from an Asian context is so important. In the USA, for example, there is a body of literature that has found significant differences in the return to education by family background, as measured by parental education level and occupational status (e.g., Altonji & Dunn, 1996;Ashenfelter & Rouse, 1997). More recently and using larger-scale administrative data, Chetty et al. (2017) found that graduates from low-income and high-income families had very similar labour market outcomes after allowing for higher education institution attended. The result suggests that family background may not be a strong predictor of graduates' labour market outcomes, once we allow for the specific college attended, and that in the USA, colleges do play a pivotal role in levelling the earnings playing field for students from different socioeconomic backgrounds.
In the UK, by contrast, much of the literature has continued to find significant differences between the earnings of graduates from the lower and higher socioeconomic backgrounds, even after allowing for a rich array of characteristics, namely higher education institution attended and subject discipline (e.g., Britton et al., 2019;Crawford & Erve, 2015). For instance, graduates whose parents were working in a top NS-SEC occupation are more likely themselves to end up working in one of these top jobs and to earn more than graduates whose parents worked in routine occupations (Crawford & Vignoles, 2014;Macmillan et al., 2013). This applies even when comparing graduates who attended the same university. Parental income level also remains a strong predictor of children's subsequent graduate income level in the UK, although the impact of family income on graduates' earnings is halved after allowing for institution attended and subject taken (Britton et al., 2019). This body of literature implies that universities in the UK do not appear to fully level the playing field and eliminate the socioeconomic gap in earnings.
In South Korea, there are many studies showing a strong positive relationship between father's and children's income levels (An & Jeon, 2008;Choi & Min, 2015;Jang, 1999;Yeo, 2008). For example, Choi and Min (2015) found that a son's earnings are, on average, 19% higher if he comes from the top fifth of households in terms of income, as compared to coming from the bottom quintile. There is also evidence suggesting that the likelihood of becoming a graduate and the quality of higher education institution attended is related to family background. Yeo (2008) provided empirical evidence that students from richer families are more likely to attend better quality colleges and as a result tend to have higher wages than graduates from poorer families. To the best of our knowledge however, there are no studies from South Korea that have estimated the extent of the socioeconomic gap in earnings for graduates, after allowing for differences in the quality of university attended and subject studied. This paper seeks to fill this gap.
Data
The data used here are from the Korean Education and Employment Panel (KEEP) survey. The KEEP, first established in 2004, is a longitudinal survey of Korean individuals as well as households. The dataset was designed to examine the relationship between a respondent's household socioeconomic status, their educational achievement, and subsequent labour market outcomes. In its first year, a total of 6000 individuals were selected as a target sample, comprising 2000 middle school seniors and 4000 high school seniors. The selected individuals were then followed up until 2015, with 12 waves of yearly surveys. The sample for the KEEP dataset is stratified into 15 regions (Seoul, six metropolitan cities, and eight provinces), and 4175 middle and high schools across the country were selected through the stratified and multistage sampling methods. Stratification was applied at the school, class, and student level (KRIVET, 2021). Hence, the KEEP was designed to provide a nationally representative sample of Korean middle and high school students. The KEEP is well suited to address our research questions, given that, in comparison to other nationally representative datasets, it includes a richer array of information on Korean youths' educational and labour market experiences, including their high school academic achievement. It also provides information on family background, namely parental income and assets, which is central to our focus.
Sample
In this particular study, the sample was restricted to those who graduated from high school in 2005, who had achieved an undergraduate degree or higher, and who earn at least the minimum legal wage in South Korea as of 2015. Hence, the usable sample is 1229 individuals who attended either 2-year or 4-year colleges and provided usable salary information. This study focuses on the respondents who entered higher education between 2005 and 2009 to allow for the widespread late entry into universities discussed above. In 2015, the targeted individuals were age 28 or 29, and it might be argued that this stage of life is still relatively early to assess the career path of individuals in the South Korean context. With this slight caveat, the data are nonetheless particularly rich in terms of trying to understand the factors that explain wage differences between graduates. Specifically, the data include sufficient information on students' subject of degree and higher education institution, as well as their socioeconomic background and prior academic achievement.
Wages
Individuals' job-related data were collected from the 12 th wave of the KEEP survey in 2015. The KEEP survey does not record each individual's income but instead provides (a) the combined income of an individual and his/her spouse and (b) the spouse's income. Therefore, the spouse's income was subtracted from the combined income to obtain the individual's own income. The data have some other limitations. Specifically, data on years of work experience were not collected. To address the issue, we controlled for the year of graduation since individuals who graduated in the same year will, on average, have a similar length of work experience. Whilst this is not an ideal measure of work experience, it will go some way to control for differences in wages associated with differences in work experience.
The natural log of the individual's monthly wage was used as a measure of their income, and the top 0.5% of observations were excluded as outliers. In addition, individuals earning less than the monthly minimum wage of 1,170,000 KRW (equivalent to £760) in 2015 were considered to be either part-time workers or unemployed, and both were excluded from the sample. We excluded part-time workers because the data were not sufficiently rich to incorporate the reasons why individuals had a part-time job in their late 20s and the extent to which it was a choice, related to family, or other nonwork-related issues. Further, we lacked data on hours of work. Our estimates, therefore, focus on the earning differences among graduates in full-time employment. We acknowledge that this means we are missing the impact of characteristics on the employment decision (including the decision to work parttime), but in the absence of a credible instrumental variable to predict employment, we simply acknowledge that we are providing a partial picture which is focused on full-time employees only. As a robustness check, Tables 8 and 9 of Appendix 3 show the characteristics of those excluded from our analysis due to this data restriction, and there is no obvious systematic relationship between employment status and the other explanatory variables that we used in the model.
Family background
Given the rapid expansion of education opportunities in South Korea, some common timeinvariant proxies for socioeconomic background, e.g., parental education, maybe arguably be less appropriate. In the space of one or two generations, South Korea has gone from having a high level of uneducated workers to a mass higher education society. Hence, parental education is arguably a weaker family background indicator than measures of family income and wealth, recognising that the former in particular may vary over time. We therefore build on a large body of existing literature that has attempted to analyse the impact of family income on the earnings of graduates in various contexts (e.g., Britton et al., 2019;Chetty et al., 2017;Kim, 2014). Hence, monthly family income and family assets are both used as proxies for individuals' family backgrounds in our analysis. Another limitation of our data is that to maximise the quality of our two measures of family background, we have had to take them from subsequent years. To the extent that family income changes over time, this may be an issue, though we expect a high correlation between family incomes over time. The family income data were collected from the 1 st wave of the KEEP in 2004, whereas the family asset data used were collected in the 2 nd wave in 2005 because the latter was measured as a continuous variable in that sweep. The continuous family asset data from the 2 nd wave do, however, have approximately 35% missing values (see Table 5 of Appendix 1). Therefore, we supplement the wave 2 data with categorical family asset data from the 1 st wave, imputing the missing values for the 2 nd wave (see Table 6 and 7 of Appendix 2). Given these limitations of our measures of family background, we were careful to test the robustness of our findings by using the two alternative measures, namely monthly family income and family assets. In any case, there is evidence that big family investments like tuition fees for higher education are not paid for out of income and hence assets might be a more appropriate indicator of family background in this context (Nam & Huang, 2009;Zhan & Sherraden, 2003). The distributions of monthly family income/family assets and their relationship with other indicators of family background, such as the paternal education level and individuals' earnings, are shown in and Figs. 2 and 3 and Tables 10 and 11 of Appendix 4.
College type
Information about individuals' higher education, such as a college type, institution attended, and the subject of degree, was generally taken from data collected in the year they started their higher education. However, data on each individual's final degree was also used to reflect any change of college or subject during the course of his/her studies. It is also possible for individuals attending 2-year colleges to obtain a bachelor's degree equivalent to a 4-year college graduate if they take an additional year of intensive courses. We have opted to classify such individuals as 2-year college graduates, given some evidence that firms do not treat the additional year as equivalent to a 4-year college degree (KCCE, 2019). Individuals who graduated from foreign universities, open universities, and on-line universities, which accounted for less than 1% of the sample, were excluded.
Other controls
Individual's wages are partly determined by their 'ability', broadly defined. Further, the likelihood that an individual enrols in a 4-year or 2-year college is also likely to be determined by his/her ability, in this case his/her academic ability. Hence, it may be that the apparent wage premium for a particular type of college degree may actually be attributable to ability bias, whereby individuals with higher levels of ability select into particular types of college. In our data, we have a measure of earlier academic achievement, prior to entry into higher education, that we can include in our model as a proxy for ability. We are not claiming that this proxy fully measures an individual's productive ability. However, it should provide a measure of academic ability prior to entry into college, which should reduce the ability bias on the coefficient on college type for instance. Specifically, we include a measure of an individual's percentile rank score within his/her high school which indicates how well he/she performed relative to other students in the same school. A lower percentile rank indicates better academic achievement.
We also control in our models for whether the college was attended was located in Seoul. This is partly because institutions in the capital city have historically been viewed as higher prestige. However, we also include it because this variable may also proxy the graduate's current location, given the tendency of graduates to remain where they studied, and salaries in Seoul exceed those in other parts of South Korea.
Analytical model
We estimate the following models on a sample of individuals with earnings at or above the minimum wage in South Korea as of 2015. As discussed above, we use two different measures of family background, e.g., family income and assets, which are measured in different years.
Despite a year gap between the two measures, we argue this is not a major issue due to the permanent nature of family assets. To estimate the effects of monthly family income and family assets separately on individuals' earnings conditional on various educational characteristics, we use the following models: where y i 2015 is log monthly earnings of individual i in 2015 when individuals were age 28/29, P 12004 is monthly family income in 2004, P 22005 is a measure of family assets in 2005, X i is a vector of control variables, and μ i is a normally distributed error term capturing unobserved random factors that influence earnings. The main parameter of interest is β 1 , which measures the relationship between monthly family income/family assets and wages. The results from models that only include family background will be presented first. This provides an indication of the raw correlation between family background and individuals' wages at age 28/29. Additional controls, such as the year of graduation, previous academic achievement, higher education institution attended, and degree subject, are then sequentially added to the model. The change in β 1 indicates the extent to which the various control variables mediate or exacerbate the relationship between family background and individuals' wages. There may of course be interaction effects between family background and different types of higher education institutions. To determine whether this is the case, we also estimate the following models: where Q i is a 0/1 dummy variable indicating whether the individual ever enrolled in a 2year (value 0) or 4-year college (value 1) and both P 1 2004 Á Q i and P 2 2005 Á Q i are interaction terms between monthly family income/family assets and college type. The variable measuring college type records whether the individual ever enrolled in a particular college type over the period 2005-2009, capturing 'late' enrolment into college. The main parameters of interest here are β 1 and β 2 , which capture the combined effect of family background and the interaction effect between family background and college type, on the outcome variable, namely wages.
Estimation and results
The socioeconomic gap in the earnings of graduates Table 1 shows the results of two separate regression models of the log wages of graduates on either the log of monthly family income (equation 1) or the log of family assets (equation 2). Both family background variables are significantly associated with graduates' income, consistent with the existing literature (Choi & Min, 2015;Nam, 2008). When monthly family income and family assets increase by 1% graduates' wages, on average, these increase by 6.5% and 2.9%, respectively. Clearly, graduates' earnings are correlated with their family background, as has been found in almost all countries. The focus of this paper is on the role of different types of a college degree in explaining this positive correlation between family background and graduates' earnings, and we include various measures of the individual's college experience in the model. As discussed earlier, all subsequent models are estimated separately by gender. The tables below show the relationship between family background (as measured by monthly family income in Table 2 and family assets in Table 3) and earnings, conditional on the individual's academic achievement within his/her high school, type of college degree (2-or 4-year), whether the college was located in Seoul, the year of graduation, and the discipline of their degree.
Columns (1) and (4) in Table 2 show the raw correlation between monthly family income and graduates' wages by gender. Monthly family income is positively correlated with the graduate's income for both males and females, but the relationship is statistically significant only for females. Additional controls reduce the correlation between family income and graduate earnings to almost zero for males. For females, by contrast, there remains a positive and statistically significant relationship between family income and graduate earnings after allowing for high school achievement, college type, location, graduation year, and the subject of study.
Similar to the models in Table 2 for monthly family income, columns (1) and (4) in Table 3 indicate the raw correlation between family assets and graduates' wages by gender. With the sample split by gender, sample sizes are reduced, and the correlation between family assets and (1) monthly family income and (2) family assets. Each cell reports the coefficients on family assets and monthly family income from a separate regression, with standard errors in parenthesis (*p<0.10, **p<0.05, and ***p<0.01) graduates' wages for males is statistically insignificant. For females, the magnitude of the coefficients suggests a stronger but statistically insignificant relationship between family assets and graduates' wages. Allowing for college type, whether the college was located in Seoul, high school academic achievement, and the year of graduation in columns (2) and (5) reduces the correlation between family assets and male wages from 1.5 to 0.1%, and the coefficient remains (2) in Table 1) statistically insignificant. The impact of college type on male wages becomes statistically significant, whilst the impact of family assets decreases, and male graduates who attended a 4year college, on average, earn approximately 17% more than those who attended a 2-year college. For females, the coefficient on family assets slightly decreases from 3.4 to 3.0% after controlling for prior academic achievement, institution attended, whether the college was located in Seoul, and the year of graduation. Allowing for the subject of a degree in columns (3) and (6) does not significantly impact the coefficient on family assets for either gender. For females, the coefficients on monthly family income (Table 2) and family assets (Table 3) do not significantly decrease when adding more control variables describing the individual's education. This is not a common finding in the literature on social mobility from other countries. Normally, a strong relationship between family background and graduate earnings tends to be mediated by the inclusion of variables capturing different aspects of the graduate's higher education. This is because a more advantaged family background enables individuals to have higher levels of academic achievement and hence affords access to a more prestigious and higher-earning degree. This is one route by which family background impacts graduates' earnings. Here, by contrast, the results are consistent with females from higher socioeconomic backgrounds being more likely to earn higher wages than those from lower family background, regardless of college attended and degree subject studied. The results suggest that for female graduates, monthly family income (and perhaps the associated social capital and networks) is more important than college and degree subject in determining their labour market opportunities and earnings.
Overall, the results reveal that family income and assets are not strong predictors of graduates' wages, at least for male graduates. In this sense, college in South Korea levels the playing field in terms of male wages. That is, the correlation between family background and male wages is essentially zero after allowing for institution attended and the subject of degree. For females, on the other hand, college does not entirely level the playing field in terms of wages. Monthly family income remains a strong predictor of female graduate wages, even after controlling for college attended and degree subject. As we noted earlier, we are mindful that our estimates are correlational and do not necessarily show a causal relationship between family background and graduate earnings.
The interaction effect between family background and college type
We have a conditional hypothesis, in which the relationship between college type and graduates' earnings depends on the level of family income or assets, for example, because higher income households may be in a position to provide additional support that is complementary to a 4-year college degree and hence increase the return for these students. To put it differently, the wage premium for 2-year and 4-year colleges may also vary by graduates' family background. We therefore also explored whether there was any interaction effect between family background and college type. Four different models (Table 4) were estimated with interaction terms between the two different family background variables and college type, separately by gender.
In column (3) of Table 4, the interaction term between family assets and college type shows a statistically significant correlation with male wages. A mean negative relationship between family assets and male wages, and indeed college type and male wages, is offset by a positive significant interaction between family assets and college type (p<0.10). The results weakly suggest that, on average, attending a 4-year college is negatively associated with wages for males. However, those with greater levels of family assets earn a higher return from attending a 4-year college. This implies that though there is no overall effect of family assets on earnings, there indeed exists a crossover interaction, implying the effect of college type on graduates' earnings varies, depending on the level of family assets. In other words, male graduates with greater family assets, on average, have higher wages when attending a 4-year college, whereas male graduates with lower family assets, on average, earn more when attending a 2-year college. Figure 1 shows how male graduates' income level changes depending on both family assets and different college types.
Some caution is required however, in interpreting the above result since for males, the relationship between family income and wages is statistically insignificant, as are college type and the interaction between college type and family income. This clearly begs the question as to why family assets interact significantly with college type but family income is insignificant when interacted with college type. Family income and assets are of course distinct measures, and it is likely that family assets are more important for education investment decisions. A plausible interpretation may relate to the cost of private tutoring. South Korea is famous for the emphasis families place on education, and the private sector plays a pivotal role in the investments parents make. For instance, more than 60% of high school students participate in private tutoring, and it accounted for roughly 15% of total household expenditure in 2008 (Oh & Kim, 2011;Statistics Korea, 2011). It should also be noted that private tutoring for CSAT preparation often costs more than three million KRW (equivalent to £2000) per month, which is much higher than average monthly household income, which was roughly two million KRW (£1250) in 2010 (Oh, 2013;Statistics Korea, 2011). As such, family assets are likely to be key in bridging the gap between family income and the costs of private tutoring. Given that a body of literature has found that private tutoring is, on average, positively associated with students' academic performance, it may be that family assets, rather than income, better predict students' academic performance and in turn their labour market outcomes. For females, the interaction terms are statistically insignificant. It appears that the value of a 2-year or 4-year degree does not vary by family background for females. The results confirm the findings from the previous tables, namely that for females, higher levels of family income/assets are associated with higher earnings irrespective of college attended or subject studied.
Discussion and conclusion
This paper contributes to the existing literature by providing an estimate of the magnitude of the socioeconomic gap in graduates' earnings in a non-Western context, namely South Korea. In South Korea, much research and policy attention has been paid to the socioeconomic gap in participation at university and indeed to the intergenerational correlation in income. Far less attention has been paid however, to whether a college degree can level the playing field in labour market terms, such that family background is no longer correlated with graduates' earnings. This paper provides new empirical evidence on this issue. We find that family background, measured by either monthly family income or family assets, does not, on average, play a significant role in determining the wages of male graduates, after allowing for institution attended and degree subject. However, our results also suggest that the wage premium for 2-year and 4-year colleges for males does vary by family background. Indeed, males with fewer family assets earn more from a 2-year college degree than from a 4year college degree. Plausible interpretations for this observation include greater quality variation within the 4-year college group and individuals from poorer backgrounds attending, on average, less selective 4-year colleges. In South Korea, since 2000, the number of 4-year colleges has dramatically increased, leading to a large number of less selective institutions, particularly outside of Seoul. As a result, students who would not have been able to attend any type of college in the past are now enrolled in these less selective 4-year colleges. This implies that the heterogeneity in academic achievement levels among 4-year college graduates has also increased, with consequences for their labour market outcomes (Oh & Chae, 2014). More socioeconomically advantaged males tend to access top 4-year college degrees that attract a higher wage premium. The implication is that whilst 4-year college degrees remain a good option for males from more advantaged backgrounds, contrary to popular belief, they do not always attract a higher wage premium than 2-year college degrees for other students.
By contrast, family background is still a strong predictor of female graduate wages, even after controlling for institution attended and subject discipline and even in models that include interactions between college type and family background. In other words, a degree does not level the playing field for female graduates, in terms of their earnings. We noted that South Korea also has the largest gender wage gap of the OECD countries. This may be relevant in terms of understanding why higher education does not eliminate socioeconomic gaps in wages for females. For example, for female graduates, it may be that social networks and social capital are more important in securing a good job, given the evidence of potential labour market discrimination against females as measured by the gender wage gap. A more advantaged family background may be necessary to provide female graduates with better social networks and capital that they need to succeed in the labour market.
Clearly, these findings should be considered in light of some of the data limitations we faced. First, the KEEP data does not include high-quality administrative data on individuals' family income and assets. Instead, respondents were asked to recall the information retrospectively. This may result in measurement error that, if classical in nature, will tend to attenuate the coefficients on the family background variables. Second, we were only able to allow for the type of college attended (2-year or 4-year). We were not able to control for individual institution fixed effects due to sample size limitations and confidentiality issues. Hence, our college type categories include a heterogeneous range of institutions that may hide big differences in the earnings of graduates, depending on the prestige of the institution attended. It is possible that the magnitude of the socioeconomic gaps in the earnings of graduates who went to the exact same institution may differ from the estimates that we provide. Future research could usefully focus on (a) the mechanisms by which family background appears to influence the earnings of female graduates to a greater extent and (b) the differences in labour market outcomes (by family background) for different colleges in South Korea. The latter would require access to large-scale administrative data to generate sample sizes sufficient to compare graduates from individual institutions, along the lines of the large-scale data used by Britton et al. (2019) or Chetty et al. (2017).
In conclusion, simply expanding the educational opportunities and improving access to higher education in the past few decades in South Korea has not necessarily narrowed the socioeconomic gap in the earnings of graduates. In terms of male earnings, it appears that a 4year college degree may only offer a wage premium over a 2-year degree for males from a more socioeconomically advantaged background. These issues should be taken into account when implementing future higher education policies and might support a shift in focus from further expansion of higher education to a greater emphasis on quality improvement, which is indeed in line with recent government policy (Hwang, 2018;KHERI, 2020). Note: this table presents the number of observations with valid information on family assets from the 2 nd wave, imputed based on the 1 st wave data. We used several steps to impute the missing values for the variable of family assets from the 2 nd wave. First, family asset data from the 1 st and 2 nd waves were reviewed and compared. 412 out of 426 missing values for family assets from the 2 nd wave have observations from the 1 st wave (see Appendix Table 5 for further details). The mid-value of each band from the 1 st wave was imputed for the 412 missing values from the 2 nd wave. As a result, we established the new family asset variable, having only 14 missing values | 9,158.8 | 2021-08-13T00:00:00.000 | [
"Economics",
"Education",
"Sociology"
] |
Flow past a Groove
Based on available evidence, it is hypothesized that the net force of friction on a flat solid wall, when fluid flows steadily along it, is reduced by putting one or more grooves in the wall’s surface oriented perpendicular to the mean flow. Among the convincing observations are the existence and history of golf balls which show that golf balls with dimples travel farther than golf balls without dimples. Also there is a laboratory experiment using streak photography of low Reynolds number flow along a straight wall with a square cavity in it, illustrating that the flow jumps right across the cavity’s opening, strongly suggesting that there is no friction of the fluid on the wall in the region of the cavity. One forecast is that if grooves or dimples are made on the inside surface of pipes, the discharged rate of the pipe for fluid flow should become increased.
Introduction
Golf balls have dimples for a good reason, which can be traced back to an accidental discovery in the mid-1800s [1].When small scrapes or nicks occurred in the surface of the ball, observations showed that there was an increase in the distance the ball traveled.By the early 1900s, all golf balls were manufactured with dimples.Is that basically the end of the story?What is the root cause of this phenomenon?Apparently there are available no other practical applications of the idea behind the dimples of present day golf ball?
Suppose that, driven by some unspecified force, steady flow of a constant density fluid moves along a flat rigid wall and then encounters a bump in the wall.The bump's specific shape is not important, but for convenience small slopes will be assumed.Of necessity the fluid next to the bump must rise up the forward face because of the boundary condition that fluid cannot penetrate the solid surface of the bump.Then consider a steady flow along a flat rigid wall that comes to a dip (inverse bump) in the wall.There is no corresponding necessity that the fluid adjacent to the wall follow the surface downward on the front face of the dip.In general, one would not expect such a thing to happen.Note that the upward and downward directions mentioned here have nothing to do with the action of gravity being significant or not.
In searching through the nine fairly standard fluid dynamics books on my shelf for information on flow past a groove, I found only one text with one paragraph on the subject in it: [2].However, on the previous page is a very interesting photograph to go with this paragraph.Streak photography was used to illustrate flow at low Reynolds number along a solid flat wall that had a square cavity in it.The depth of the cavity was comparable to its width.Assume that the flow was from left to right (the text doesn't say but the reference cited undoubtedly does [3]; it is not important for the present purpose).It is seen that the flow along the wall continues right across the opening of the cavity and then goes along the wall again on the other side of the cavity.Although not totally counter-intuitive, such an observation is just a bit surprising.Deductions that arise from this observation are given below.
Groove Opening
To begin with, understanding how the flow can go across the groove's opening needs to be addressed and hopefully understood better.Since the flow appears in the photograph to be steady, the streaks make the streamlines visible.Along a streamline Bernoulli's law is expected to hold approximately: where the speed is greatest the pressure is least.Thus, across the opening the pressure in the flow is lower than just below the flow in the cavity.Although it is seen in the photograph that the flow dips down a small amount at the entrance of the cavity, it is prevented from going further into the groove by the upward pressure gradient holding the flow up.
Strictly speaking Bernoulli's law is not valid when friction is involved.However, across the gap friction within the fluid is less than friction between fluid and solid along the wall because vertical velocity gradients will be less in the groove where there is no non-slip boundary condition within the opening.Also the friction term to be added to Bernoulli's law is small if the curvature of the streamlines is small [4], which it is at the groove's opening according to the streak photograph.
Deductions
An elementary theoretical (analytical) model of flow past a groove appears to be out of reach at present and no such model already exists as far as I can determine.Numerical techniques may be brought into service on the problem in the future, however.Of course further observations are always welcome.In the interval a few important deductions can be put forward based on the existing evidence.
First, it is self-evident from the streak photograph mentioned above that friction between the fluid and the solid wall will not only be small in the region of the groove, compared to that on the flat wall on either side of the groove, but it may actually have the opposite sign, since there appears to be a weak recirculation within the groove.In other words, at the bottom of the groove the friction force probably points upstream!Confirmation by observations of the reduction in friction by the presence of a groove is badly needed.This possible outcome of the effect of a groove on the total friction of the fluid on the wall was not mentioned in the above reference [2].
Second, if one groove reduces the net friction on the wall, even by a tiny fraction, then more grooves should reduce friction by a greater amount.For example, assuming friction reduction is their main purpose, there are about as many dimples on a golf ball as can be fitted in: well over one hundred [1].
Discussion
During translation through the air a golf ball almost always rotates about an axis that passes through the ball, the way the game is normally played.That introduces another force into the problem: the sideways acting Magnus effect.Then confusion can enter because the Magnus effect does have an influence on the distance the ball travels.If a golf ball can be struck without rotation, and if a ball with dimples consistently travels farther than a smooth ball, a stronger case could be made that the dimples reduce the friction between the air and the ball.
While waiting for experimental verification of the two deductions stated above about friction reduction by grooves in a solid boundary, one other application can be hazarded.Dents could be placed on the inside surface of pipes to see if the discharge rate increases.Or circular grooves could be made inside in planes parallel to vertical cross-sections of the pipe, whichever would lead to the greater result or be easier to make or both.
Conclusion
It is concluded that a flat solid wall experiences less friction from a fluid flowing next to it if there are one or more grooves cut into the wall's surface in the direction normal to the mean flow.Laboratory observations strongly suggest that the conclusion is valid but further measurements are needed to confirm it.Assuming the conclusion is right, then it should also be true that a pipe will have a greater discharge rate of fluid flow when grooves are cut into the inside surface. | 1,705.2 | 2015-02-02T00:00:00.000 | [
"Physics"
] |
Innovation Driven Emerging Technology from two Contrary Perspectives : A Case Study of Internet
Internet is a well, organized technological achievement of human being and a rapidly improving medium through time. All the novel technological achievements like web 2.0 or web 3.0 are new epochs of Internet technology and Internet is spreading in multiple dimensions, reforming the paradigm, and innovating the technology in a selfrenewing fashion. In this paper, the technological construction of Internet and the social paradigms are discussed from two contrary perspectives. Either as “problem solvers” or “technical experts”, the characteristics of incumbents of technological positions seems very problematic in terms of their roles in shaping technology. Are they so disinterested and unbiased on creation of technology? Can we reduce their roles as such? How can we make sure that they are neutral? If we put their roles that way, what about freedom of individual decision-making?
Introduction
In 1945, Dr. Vannevar Bush, Director of the Office of Scientific Research and Development, wrote an article about the application of science to warfare.As a coordinator of the activities of leading American scientists, he called for a new relationship between science and scientist.To him, "for many years inventions have extended man's physical powers rather than the powers of his mind."Now, "instruments are at hand which, if properly developed, will give man access to and command over the inherited knowledge of the ages.The perfection of these pacific instruments should be the first objective of our scientists as they emerge from their war work."1Some 40 years later, Langdon Winner proposed a similar perspective to technology.While establishing technological systems, he urged us to reflect upon the potential consequences of such systems.If we cannot interfere at the beginning of designing and developing a right system which means a right world we live in, then, we would not be able to deal with the outcomes."because choices tend to become strongly fixed in material equipment, economic investment, and social habit, the original flexibility vanishes for all practical purposes once the initial commitments are made."2Bush's view of science is optimistic, because he has a very naïve faith in the power of science and scientists.According to this view, science acts in circumstances of scientists' own making and choosing.What we have is our work and development.By the same token, what we prospect is our plans and designs.It is a formation of ordinary science work and thus outcomes are determined.This sort of science view implies not only a positivist and enlightened world paradigm but also a hidden agenda of imposing power from an ivory tower in which scientists live.
Winner has a sort of pessimistic perspective towards the role of technology in the society.If it is possible to determine the outcomes, then there is no possibility to have uncertainty, which is the must of a democratic regime.Winner attempts to avoid undesirable outcomes of technological advances, yet at the expense of freedoms.
Both Winner and Bush are acceptable, in terms of having positive perceptions of science or avoiding negative effects of technology.However, it seems that efforts to control the negative effects of technology or to have positive effects of it to extend the power of mind, as is Bush's attempt, are to threaten or to restrict freedoms of society.Yet, the very nature of science and particularly technological advancements stems from the freedom of individual decision-making.This is a value that fosters the development of societies and individuals.
If, people want technological developments and positive contributions of science to their lives, even if they also want to have their freedoms and individually take their own decisions with their consequences either positive or negative, the question that whether it is possible to have both simultaneously raises another issue: What are the roles of decision makers in shaping the technology?An intense debate over the role of technology in society in late 1960s between Emmanuel Mesthene and John McDermott inspired me to deal such an issue.Mesthene, director of Harvard Program on Technology and Society at that time, argued that technology, neither an alloyed blessing for man nor an unmitigated curse, is a selfcorrecting system.On the other hand, to McDermott, technology had its own politics.Focusing on the nature of contemporary application of technology like in Vietnam War, he defined technology as "systems of rationalized control over larger groups of men, events, and machines by small groups of technically skilled men operating through organized hierarchy."3Though the debate was based upon the different perspectives of rightist and leftist politics, their opposite definitions of decision makers in shaping the technology highlight a fundamental point to my argument: The positions and roles of decision makers in creation of technology.
Innovation Driven Emerging Technology from two Contrary Perspectives: A Case Study of Internet
Emerging Markets Journal | P a g e |88 Our chief interest lies in the question of whether they could be a determinant factor.In order to deal such an ambiguous issue, we are specifically interested in exploring this concern in a new technological form, Internet, whose construct is shaped by a high level of commitment of its pioneers.So, in final analysis, figuring out what is the role of decision makers of technology in shaping the nature of that technology, we argue that if internet, as a newest technological form is so far maintaining both freedom of individual decision making and virtues and advantages of technology, it is possible because of the fact that from the very beginning its pioneers, designers, and architects are eagerly committed to do so.We will investigate the theoretical basis of this idea, and discuss the early history of the Internet in terms of the commitment of its pioneers, decisionmakers and potential decision-makers that affect the development of the Internet.As long as decentralized, interoperable, and open nature of internet technology survives, we assert that this would be possible if only there are committed designers, architects and organizations that can elevate themselves in such a position that they exhibit no drive for commercial or political power. 4per proceeds in the following ways: First, comparing and contrasting two opposite views of decision makers of technology, we derive some specific guidelines that are helpful in determining the characteristics of decision-makers of technological advancements.Then, particularly in the example of internet, we try to reveal the commitment of pioneers of the Internet to open, free and decentralize structure of that new technological form, which otherwise would not be possible.We examine their positions, as well as present efforts in terms of individual decision making and politics of technology. 4Having said that, I have to admit that my view on internet as free and open architecture could be seen a bit naïve by some.For instance, one who sees global economy as a "hegemonic order" and how it deploys the "control utility of network technology" to produce that order or the "universal homogeneous state" would think that "new information and communication technology has not been to free and empower ordinary people but to tighten the screws and make their global economic and political rulers richer and less visible than ever before."Likewise, "insofar as they bolster the already formidable control of capital over the means of power, computer networks are an essentially conservative, not revolutionary, technologyconservative, that is, of the prevailing liberal and capitalist order."See Darin Barney, Prometheus Wired: The Hope for Democracy in the Age of Network Technology, Chicago: University of Chicago Press, 2000, p. 188.
Two Edge Roles of Technology
We can understand the premise of technology's role over society in two opposite ways: As a self-correcting system by Mesthene and as a form of life by McDermott.While the former implies that technology has advantages as well as disadvantages and this does not have to do with our freedom of decision-making since it is almost neutral system, the latter asserts that it is not an arbitrary choice, but an imposition concerning the way we should live with and thereby a tool of suppression of humanity.Since two views are based on and augment certain definitions and characteristics of decision-makers of technology, these two edges about technology's role over society provide us a heuristic tool.
According to Mesthene, technology is "the organization of knowledge for practical purposes."This "organized knowledge" motors social change in creating simultaneously positive and negative effects.In order to understand the impact of technology, we should not isolate either of them and take both at the same time.In fact, what we see as problems or negative effects of technology could be messengers of potential technological advances.Even institutional structures and cultural attitudes of society are subject to that notion: They could offer new opportunities.Yet, since our society depends on individuals and firms, which are looking out new opportunities and they benefit to do so, we cannot realize external benefits embedded in new technologies.These externalities could be either negative or positive.Though positive opportunities are eagerly looked out, there is no way to know what negative externalities are, because "it has not been anybody's explicit business to foresee and anticipate them". 5echnological advances create new opportunities and thereby the alteration of social structures.Negative externalities are because of older structures since they are inadequate to serve new purposes.Individual purposes, thus, without concentrating to what costs we will have at the end, constitute "the institutional fabric" of the society.The negative externalities, that we face is a sort of cost of our individual freedom to pursue our goals whatever their consequences are.In this regard, technology is like a religion: Positive or negative outcomes are not inherent in the technology, yet they depend on "what man will do with technology."6Actually, we cannot attempt to measure or control negative externalities, because these attempts often appear to threaten our freedoms of decision making.If we continue to have positive effects of technology, we should learn to live with negative externalities which would be solved by technology soon.In the long run, that is for sure, technology would maintain general welfare.In this context, Mesthene maintains that since ours is a knowledge society, incumbents of technocratic roles as problem solvers get decision making power.As long as they use "organized knowledge for practical purposes", they would be able to get decision making power.So there is no need to worry about misuse or fraud of authority, because they would be able to take reliable decisions based upon reliable knowledge.
Whereas, McDermott thinks that this view, he calls laissez-innover, is simply to keep positions of those who are in power.The reason is that, technology is a way of "rationalized control over large groups of men, events, and machines by small groups of technically skilled men." 7 The very notion of negative externalities is a "production" for having technology's benefits while avoiding its costs.He thinks that defining technology "the organization of knowledge for practical purposes" is problematic.Attributing to technology "so much flexibility and 'scientific' purity" is taking market as long-term solution for economy.By coining the term laissez-innover, he criticizes the idea that "if the technology or innovation is allowed, will the maximum social good be realized?"8He thinks that concentrating on negative externalities as temporary technical problems creates a ruling technocrats class.Those who got the power as incumbents of technological decision making positions not by their, say, patrimonial characteristics, but by technological skills as problem-solvers are in charge for our own good.They have a bias against ideologies since they are committed to scientifically deal with problems.What it means that we can trust them because of their specific training and professional commitment.To McDermott, this is "an air of mystification around technology's managers."9At this point, he rearticulates the functions of technology decision makers.In fact, they are technical and scientific elites who have highly sophisticated training and education since technology requires doing so.They consist of a ruling class to control masses.So, that enforces the separation between ruling technocrats class and lower classes, a separation enhanced by technological advances and laissezinnover ideology.Besides, the point of "problem solvers" is to assert that technological systems in fact could operate without intervention of human factor.They are resistant to such intervention and do make sure that it is minimal which otherwise would not be "classified" and eventually trusted.In this regard, McDermott argues that technology creates its own politics.Therefore, he lessens the importance of decision makers of technology by calling "technical experts", who make the system rational and efficient, by filtering out the "nonrational" or "nonefficient" elements. 10n fact, both views implicitly propose a role definition of decision maker of technology.When this role is defined as "problem solver", we should be persuaded that this role model does not misuse technology, because otherwise that would not be elevated to such a position.The incumbents of this model are isolated from ideological or political thinking in reshaping technology since they use organized knowledge in a centralized (which means controllable) and institutionalized way.At first glance, it seems that this model has power in terms of determining and reshaping technology because they are elevated and granted to do so.Yet, the very definition of "problem solvers" implies a secondary and complementary function, not a preliminary and determinant factor.Further, the premise of practical purposes is vague.It appears that technology, as a self-correcting system, settle on the practical purposes, not the work force of technology.
On the other hand, when we take our role models as technical experts, we are persuaded that this role model is used to control masses.By acquiring skills and sophisticated education, they are elevated into ruling class whose interests determine technology.Again, this model, yet negatively looking, appears to adore decision makers of technology in terms of determining technology.However, it still praises the organized system and disregards the personal freedom of choice.Thus, it does not make room for individual decision making, reducing the politics of technology into a traditional class conflict base.
Either as "problem solvers" or "technical experts", the characteristics of incumbents of technological positions seem very problematic in terms of their roles in shaping technology.Are they so disinterested and unbiased on creation of technology?Can we reduce their roles as such?How can we make sure that they are neutral?If we put their roles that way, what about freedom of individual decision making?Next section tries to determine specific guidelines in order to deal with these questions.
Useful Guidelines for "Job Specifications" of Technology Decision-Makers
Inspiring two contradictory views outlined above, now, we try to propose two specific guidelines for investigating specifications of technology decision makers, attitudes and positions as well as for having some reference points with which we can evaluate the decision makers of the Internet.
The first guideline refers to "problem solver" as distinctive characteristic of decision makers.This characteristic requires that decision makers place within brackets their personal beliefs and values about the use of technology, including the "bias" that technology has its own politics.Since they
Innovation Driven Emerging Technology from two Contrary Perspectives: A Case Study of Internet
Emerging Markets Journal | P a g e |90 achieve to get their positions through a selective process, they have already been able to think ideologically-unbiased.That is why they are called "problem solvers".Once we take them as problem solvers, then, we accept the notion that there is no misuse of technology.Yet, it does not still abandon our need to define the very purpose of use of technology.Indeed, it needs a certain definition, because it is unclear to determine "practical purposes" on which decision makers build technology.How do we know they are unbiased or "ideology-free"?Being aware of these challenges, Mesthene proposes a new term "institutional innovation".To deal with the new problems because of new technologies and perhaps to make "practical purposes" understandable, he offers to enlarge public decision making.This enlargement has two consequences.First, we should have reliable knowledge and base our decisions on a particular model of society since all decisions are interrelated and thereby affect the whole society.Second consequence is the need of what he calls "institutional innovation".This is to restructure decision making process.Is it a new way to be able to determine the effects of society so that enlarging the positive effects of society?Though it is not clear, the term "institutional innovation" seems not to support that decision makers of technology are only problem solvers.Rather, it undermines individual decision making, by alternating a motor of technological advancements with a static, determined and proposed mechanism that allows allegedly public to join decision making process.This raises an apparent paradox already embedded in the conception of "problem solvers".If we establish a mechanism socalled "institutional innovation", would not it be at the expense of individual decision making?More importantly, taking them as such is reducing their role into a static and non-innovative way, by assuming them reactive not proactive.Thus, it should be clear that such use of the term "problem solvers" implicitly undermines the freedom of individual decision making and minimize the role of decision makers of predetermined set of assessments.
Second guideline refers to "technical experts" as distinctive characteristic of decision makers.This characteristic requires that decision makers are highly specialized workforce employed to make sure that the system works well, which is the domination of a ruling class over masses.Technology and its workforce, in this regard, are not but the agencies of highly centralized and intensive social control.Technology has not only its own politics, but also creates its own working and managing classes.Recalling to return class-based politics, this interpretation of politics of technology seems to ignore the potential contributions of so-called technical experts inherent in their decision making process.Indeed, it is not absolute to argue that the only motivation for technocrat class is to keep their positions intact.Seeing technology as a tool used to control lower classes is, to some extent, ignoring the innovative and entrepreneurship characters of workforce of technology.As a matter of fact, the incentive and motivation for advancement of technology cannot be simply reduced into class conflicts or power relations.This model, to some degree, can work fine, for instance, for explaining the relationship between complex structures of technology and highly skilled and educated workforce in terms of social organization of modern technology, but some of the early motivations that create new technologies, in particular freedom of individual decision making and innovation, make clear that technocrat class do not necessarily exist to control lower classes and thereby their purpose of technological advancements to keep the status quo.Actually, the term technical expert and understanding what underlies this term is far away to explain what motivates the pioneers of technology.Is it reasonable to diminish of keeping class domination, a vaguely proposed claim?At this point, the term selffulfillment that McDermott uses, being only reward of technocrat ruling class is also insufficient in explaining the motivations behind the early history of new technologies.
It appears that, consequently, two contradictory views of politics of technology offer heuristic tools with which we can have some points about the relation between individual decision making and virtues of technology, even though they lack absolute explanations.Simply summarizing, first view is an optimistic one, appraising technocrat class for their knowledge and skills, yet reducing their roles into static and standardized patterns.Although we cherish the positions of decision makers of technology, we, actually, miss their individual freedom on decision making process.On the other hand, second view is a pessimistic one, blaming that class about using technology for their interests, yet ignoring totally thereby missing real motivations embedded in the formation of new technologies.Again, we cherish their positions yet we do not only dislike their misuse of power but also oppose the notion that they are only motivated to control masses and keep status quo.By having these in mind, now, let us look at how the politics of newest form of technology, the Internet, is shaped?
The Politics of A New Technology: Internet
The politics of the Internet in terms of guidelines that we offer above requires a historical overview of this revolutionary information infrastructure.But before this, as a technological form, what characteristics of the Internet and its inventors allow us to criticize two views of technology of politics?What uniqueness of the Internet and its inventors make possible for us a "reverse reading of technology" by arguing that the Internet and particularly its first pioneers do not fit the continuum 2. That the very nature of the Internet is partly, if not mostly, shaped by the characteristics of inventors of the Internet as free-thinkers is apparent.
The first argument is based on the fact that unlike other all mediums before, the Internet allows anyone to communicate instantly with others worldwide.As a network of networks operating "on the basis of multiple implementations of accepted, non-proprietary protocols, standards and interfaces", 11 it transcends national borders and eliminates barriers to the free flow of information.Likewise, there is no central unit in the structure. 12Another difference from other forms of communication is its multi-multi character; actually the Internet allows responsive communication from one-to-one, from one-to-many, and from many-to-one.It permits everyone easily to join itself. 13Rather, "the digitization of information and the ability to transmit it over the telephone network, combined with the decentralized nature of the Internet, mean that the Internet has essentially unlimited capacity to hold information." 14It allows not only users to choose their contents, but also develop their standards and adapt them without changing and affecting the nature of the infrastructure.Thus, current nature of the Internet, open and decentralized, exhibits a unique character amid other forms of communications ever created.This uniqueness of the Internet has shattered not only the domination of traditional forms of communication in terms of freedom of voices, but also the intentions of control over media.According to Newhagen, "the very architecture of the net will work 11 Regardless of Frontiers, Protecting Human Right to Freedom of Expression on the Global Internet, Global Internet Liberty Campaign, http://www.cdt.org/gilc/regardlessoffrontiers.html. 12Having no control unit would be made possible by dividing messages into separate packets each of which had been individually authorized.That was because of the assumption that network would always be seen as unreliable.Yet, to some, it has become its main strength.See Robert Kahn, "Evolution of the Internet", Chapter 11, Revolution in US Information Structure, National Academy Press, www.nap.edu. 13It is so important that, according to Kapor, "future generations will be indebted to this community for the wisdom of building these types of open systems."See Mitchell Kapor, Big Dummy's Guide to the Internet, http://www.umich.edu/~archive/linguistics/bigdummysguidetotheinternet. 14 Regardless of Frontiers, Ibid.against the type of content control these folks (the masters of mass media) have over mass media" 15 In fact, "they have yet to grasp that the Internet can never be merely another profit center in their dreams of empire", because "the Net is built to smash monopolies." 16Even there are some attempts to monopolize the Internet, there will be always "unpaved" portions of cyberspace, "thereby opening the door to a genuine cultural and political renaissance." 17This is so obvious that discussing the freedom of Internet is not an issue: "That seems guaranteed." 18at is the origins of uniqueness of the Internet?What makes possible to have an open, and decentralized information structure?According to Kahn, there are initially two reasons: "Far-sighted investment by the United States government and the active involvement of the research community." 19egarding the government's role in forming and reshaping the structure of the Internet, the supports of United States Government, Defense Advanced Research Project Agency (DARPA) initially, and later the National Science Foundation (NSF), the Department of Education (DOE) and other agencies and departments are obvious.Especially the projects and research focused on new information infrastructure that became a base for the Internet in 1960s and 1970s were not enough attractable for private computer companies.20Government's not only dedications to highly advanced projects but also successful handling of this issue in terms of management and operation have paved the way of the Internet.
Further, some decisions by the FCC have critical impacts on the development of the Internet.For example 1968 Carterfone decision, determining that customers of the AT&T could connect their own equipment to the telephone network so long as the equipment did not in fact harm the functioning of the
Innovation Driven Emerging Technology from two Contrary Perspectives: A Case Study of Internet
Emerging Markets Journal | P a g e |92 network, opened the door to the improvement of the modem.
The second reason of the unique nature of the Internet comes from pioneers of new information infrastructure.
Their commitment to open, decentralized and free cyberspace, as we argue, has not only made their positions exclusive in terms of freedom of decision making, but also had their products, the Internet, an unprecedented effect, breaking the ground where technology challenges human life, threatening its freedom.
Pioneers Inspired to Change the World
According to Rheingold, "the most important parts of the Net began as dreams in the imaginations of a few specific people, who acted on inspiration rather than orders."Neither national defense concerns nor profit motive, but diligent scholars, enthusiastic researchers, and keen teenagers have created the Internet, willing to change the world. 21 there is a need to classify those who pioneered, the very first pioneer of the Internet's pioneers may be Vannevar Bush.Considering a future device called memex, he was first to describe the very nature of the information infrastructure.Memex, according to Bush, is "a device in which an individual stores all his books, records, and communications, and which is mechanized so that it may be consulted with exceeding speed and flexibility."It consists of a desk, which can be operated from a distance, screen, keyboard, and sets of buttons and levers.Its contents would be recorded on microfilms. 22By having this visionary mind, he did not only inspire many, but also had a key role in establishing a team of pioneers (later developed the ARPANET), which initiated many projects that changed information infrastructure, "bringing the government, military and elite academic researchers into a closer embrace than ever before." 23blishing a number of papers related with human and computer interaction in the early 1960s, J. C. R. Licklider is another pioneer.His "Galactic Network" concept proposing a globally interconnected set of nodes through which everyone could quickly access data and programs from any site was very alike of the Internet.He was the first head of the computer research program at DARPA.Though his funding was very limited and computer industry was not ready to the idea of time sharing in machine resources, his vision and persistence made the projects going.He chose most of the first pioneers of information infrastructure project. 24While directing DARPA, he also helped to the formation of Computer Science Departments at many universities. 25 his 1968 paper, "The Computer as a Communication Device," written with Robert Taylor, Licklider defined four principles for human and computer interaction: 1. Communication is defined as an interactive creative process.
2. Response times need to be short to make the "conversation" free and easy.
3. Larger networks form out of smaller regional networks.
4. Communities form out of affinity and common interests. 26fined as "Prophet of the Net" by one scholar, Licklider's vision is seen in the same paper clearly: "The collection of people, hardware, and software -the multi-access computer together with its local community of users -will become a node in a geographically distributed computer network.Let us assume for a moment that such a network has been formed.... Through the network of message processors, therefore, all the large computers can communicate with one another.And through them, all the members of the super-community can communicate -with other people, with programs, with data, or with selected combinations of those resources." 27e creators of networking protocol of TCP/IP, Bob Kahn "I had certain technical ambitions when this project started, but they were all oriented toward highly flexible, dynamic communication for military application, insensitive to differences in technology below the level of the routers.I have been extremely pleased with the robustness of the system and its ability to adapt to new communications technology.One of the main goals of the project was "IP on everything."Whether it is frame relay, ATM, or ISDN, it should always be possible to bring an Internet Protocol up on top of it.We've always been able to get IP to run, so the Internet has satisfied my design criteria.But I didn't have a clue that we would end up with anything like the scale of what we have now, let alone the scale that it's likely to reach by the end of the decade.It seems likely that the Internet will continue to be the environment of choice for the deployment of new protocols and for the linking of diverse systems in the academic, government, and business sectors for the remainder of this decade and well into the next." 28 1980, TCP/IP, a co-production of Kahn and Cerf, was decided that it would be used in the preferred military protocols.Yet, that was not turning point.The turning point was a perfect example of freedom of individual decision making.In Cerf's own words, it is: "In 1988 I made a conscious decision to pursue connection of the Internet to commercial electronic mail carriers.It wasn't clear that this would be acceptable from the standpoint of federal policy, but I thought that it was important to begin exploring the question." 29hn's early goals on this project are remarkable in terms of a perspective of a decision maker of technology, as well.Four basic rules, namely network connectivity, distribution, error recovery and black box design, affected Kahn's thinking: 1.Each distinct network would have to stand on its own and no internal changes could be required to any such network to connect it to the Internet.
2. Communications would be on a best effort basis.If a packet didn't make it to the final destination, it would shortly be retransmitted from the source.Black boxes would be used to connect the networks; these would later be called gateways and routers.
3. There would be no information retained by the gateways about the individual flows of packets passing through them, thereby keeping them simple and avoiding complicated adaptation and recovery from various failure modes.
4. There would be no global control at the operations level." 30milarly, same sort of vision is reflected in the thinking of Tim Bernard Lee who invented WWW in March 1989.His project initially had two main goals: First, like Kahn's design for TCP/IP, WWW hypertext system should have an open architecture, and second, it should be distributed over a communications network.
In his "The World Wide Web: A Very Short Personal History" Lee summarizes his efforts to standardize an "Universal Document Identifier", mentioning how his boss supported his experiments. 31WW was the realization of Lee's dream of creating a "common information space in which we communicate by sharing information".The remaining part of the dream was yet to come.It was web's realistic mirror function that "once the state of our interactions was on line, we could then use computers to help us analyze it, make sense of what we are doing, where we individually fit in, and how we can better work together."Will it happen so? "(It) has yet to happen, but there are signs and plans which make us confident.The great need for information about information, to help us categorize, sort, pay for, own information is driving the design of languages for the web designed for processing by machines, rather than people.The web of human-readable document is being merged with a web of machine-understandable data.The potential of the mixture of humans and machines working together and communicating through the web could be immense" 32 Tracking the dreams and decision making processes of first pioneers, Douglas Engelbart, developer of the graphical user interface, first working hypertext system and first mouse, is also a key figure.Interestingly, his first motivation came from Vannevar Bush's article about his vision for the "memex".Subsequent years, he published a paper called, "Augmenting Human Intellect: A Conceptual Framework", visioning his own information
Innovation Driven Emerging Technology from two Contrary Perspectives: A Case Study of Internet
Emerging Markets Journal | P a g e |94 infrastructure.Next excerpt is notable vision of this inventor: "My professional motivations are strongly oriented toward maximizing the benefit which society might derive from the advancements in the computer field.I might say then that my professional interests are toward the application of automatic informationhandling equipment for helping human society, in the most significant way possible." 33bit different perspective can be seen on Ray Tomlinson, first sender of e-mail.When asked what inspired his invention, he said that "There was no directive to 'go forth and invent e-mail.'"And added: "Mostly because it seemed like a neat idea".Setting out to adapt CYPNET to use SNDMSG to deliver messages to mailboxes on remote machines, through the ARPANET, he feels in a humble way, citing that was "just a minor addition to the protocol." 34garding the outstanding contributions of first pioneers of the Internet to the uniqueness of it, it should be emphasized that these key figures have not only played important role in development of the infrastructure, but also they have been -and still arevery and closely interested in policy and technological changes that could affect the very nature of it.They do so by either participating in government organs or forming their own organizations.The example of former is National Research Council's Computer Science and Telecommunications Board (CSTB) committee chaired by Leonard Kleinrock, one of the pioneers of digital network communications, and helped build the early ARPANET.Being active in policy making with the government, Kleinrock has affected the formation of current framework with his influential 1994 report Realizing the Information Future; The Internet and Beyond.Pursuing the idea that "the nature of the services and styles it (the Internet) can produce is limited only by the imagination of its practitioners", Kleinrock is a typical example of the Internet's first pioneers who choose to keep the dream alive. 35 this regard, it could be said that some organizations and research projects funded by governments (the ARPANET in the United States, the network of the National Physical Laboratory in the United Kingdom, CYCLADES in France, and other networks around the world) have prominent roles in fostering the development of new information infrastructure.Yet, even in these organizations, we see same pattern of thinking rooted in decision making process of individual pioneers.Though McDermott thinks in a different way, hypothesizing that advanced technological institutions are agencies of highly centralized and intensive social control 36 , on the contrary, workforce of these organizations established primarily for military purposes 37 could mostly achieve to determine the early goals and structure of the new information infrastructure in a more independent and free way.Funded by government, these organizations were directed to highly complex research projects.Yet direct control and coercion that would manipulate the researchers were not existent.Thus, they became platforms of freedom of decision making.One of the reasons of this phenomenon may be the management and operation of these organizations.0 Nonetheless, it is difficult to argue that the government funding and military motivations have secondary importance. 41PA has basically pioneered and sponsored three key projects: Advanced Research Project Agency Network (ARPANET), packet radio network, and packet satellite network.According to Kahn, "each of these three networks was individually designed and implemented, but most importantly, the Internet architecture was created to be independent of the detailed design or implementation of any of its constituent networks."He thinks that the success of the Internet lies "on the underlying computer communications technology that had been pioneered in the ARPANET."Further, to him, two important characteristics of the Internet as being no single entity responsible for the overall performance of network and layered host protocols are also aimed in the ARPANET project.42 Another key organization, though its contribution is accordingly limited, is a non-US research center called CERN-the European Laboratory for Particle Physics.When it was founded in 1954, CERN was based upon the idea that atmosphere of freedom -freedom to doubt, freedom to enquire and freedom to discover-is essential for scientific research.Throughout years, CERN has been "the world's largest research laboratory with over 50% of all the active particle physicists in the world taking part in over 120 different research projects.3000 staff members, 420 young students and fellows supported by the Organization and 5000 visiting physicists, engineers, computer experts and scientists specializing in a variety of front-line technologies are collaborating with CERN from 40 countries and 371 scientific institutions."43 Not surprisingly, the researchers of this center have made significant contributions.The WWW project was originally developed by 40 Roberts, Ibid.41 "ARPA wouldn't have happened if what used to be the Soviet Union hadn't shaken a complacent U.S. awake with a tin can in the sky, Sputnik.Wars do wonders for the advancement of technology, and the Cold one was certainly no exception.The way to get a technology advanced is to gather a lot of really smart people under one roof and get them to concentrate on a single project.Of course, that takes some organization and money.Where does that come from?But that's another can of worms -to be opened with relish at a later date.In this case, it was the only body that had a stake in making sure the Net worked -the government."(David Hudson, "Con.txt",Rewired: JOURNAL OF A STRAINED NET, August 9th, 1996.)42 Kahn, Ibid, p. 159.43 Net Valley, http://www.netvalley.com/archives/mirrors/CERN-PR11_94E40thAnni.htm rerchers of CERN in 1990.44 CERN has also played an important role in development of the Internet protocols.First internet protocol was used there during the second phase of the STELLA Satellite Communication Project, from 1981-83, a project inspired by the ARPA IP model.45 Apart their supportive and directive roles through consultation to public and private sectors, the first pioneers of the Internet have also formed critically functioning organizations in establishing standards and forming other components of the Internet.Among the examples of that kind of organizations are Internet Society, IETF, the World Wide Web Consortium (W3C), the Internet Corporation for Assigned Names and Numbers (ICANN), including Commercial Internet Exchange (CIX), aiming to facilitate the exchange of traffic among commercial internet service providers.46 They symbolize the ongoing efforts of the first pioneers of the Internet.Working together, these organizations make sure that the unique nature of the Internet cannot be sacrificed.For example, the W3 Consortium formed by Tim Bernard Lee, the inventor of WWW, as an independent standards making body was to ensure universality of functionality across the industry.Convincing Michael Dertouzos, the head of MIT's Laboratory for Computer Science, Lee established W3C in 1994, in order to "oversee development of common web protocols and promote web interoperability."47 The formation and operation of the W3C is a manifestation of the Internet as most open, decentralized and free medium of all times: Before promoting their standards, W3 staff present a sample code, allowing everybody to raise any concern.Then, they release the standards for the implementation to promote each of their standards.
Conclusion
One scholar points out the historical significance of the Internet: "Instead of a small number of groups having privileged positions as speakers-broadcast networks and powerful newspapers-we are entering an era of communication of the many to the many. ..the nature of the technology itself has opened up a space of much greater democratic possibility." 48n this paper, we argue that this is because of the creative roles of first pioneers in the formation of the very structure of the Internet.Emphasizing their freedom of decision making in establishing the new information infrastructure, we build my argument in to context of two contradictory views of technology of politics.
Innovation Driven Emerging Technology from two Contrary Perspectives: A Case Study of Internet
Emerging Markets Journal | P a g e |96 Either as "problem solvers" or "technical experts", we try to show that the defining characteristics of technology decision makers do not fit in to the frame of first pioneers of the Internet.We never accept the idea that as technical experts or problem solvers, thereby being participators of decision making process, are not interested in shaping the technology.As a matter of fact, they cannot be neutral.In the example of the internet, we believe first inventors of the Internet had their roles very positively.They always had same spirit, making the Internet not only a positive technological contribution to humanity, but also a platform of choice, a mark of respect in terms of freedom of decision making, a spirit which is still alive today.
It should be emphasized that, however, we are not still in a position where we can measure the outcome truly.This is because of the fact that the revolution or evolution whatever it is, is continuing.In a medium which moves speed of light, everything can change at the same speed.Thanks to their ongoing efforts, most of the first pioneers are in charge and still trying to keep "dream" alive.This is a dream that can change the relation between human and technology in a positive way.
Dr. Mehmet Lutfi ARSLAN, Dr. Sadi Evren SEKER and Dr. Cevdet KIZIL P a g e |91| Emerging Markets Journal ranging from Mesthene to McDermott?Answers to these questions proceed in the following set of arguments: 1.The Internet as a technological form has unique characteristics that challenge the notions of technology, which are used by either Mesthene or McDermott.
and Vinton Cerf are living examples of the Internet pioneers.They both worked at DARPA on networking projects.At the beginning, Kahn was working alone.In 1972, he gave a demonstration of a network called ARPANET, connecting 40 different computers at the International Dr. Mehmet Lutfi ARSLAN, Dr. Sadi Evren SEKER and Dr. Cevdet KIZIL P a g e |93| Emerging Markets Journal Computer Communication Conference.That was the first time when the net project gained widely interest.In 1973, Vinton Cerf joined Kahn on this project.They worked on data communications across packet radio networks.They then studied on the development of a standard open-architecture network model, where any computer could communicate with any other.Following statement of Cerf reflects his thoughts about the project: It would be useful to look closely to a few of these organizations.Surely, the most remarkable case is ARPA.ARPA (later became DARPA) was the U.S. Department of Defense's Advanced Research Projects Agency (ARPA).J. C. R. Lick became head of ARPA in 1962 for creating and managing a program for funding research.In 1966, Larry Roberts proposed the ARPANET to ARPA.While the very aim of the organization was militaristic since it was funded by military, Robert's proposal had a more technically focused purpose.It was to explore computer resource sharing and packet switched communications and had nothing to do with nuclear war or survivability."39Regarding the rumor that the Internet was created by military to have survivability of information in a nuclear war depends upon a paper by Paul Baran.Yet, ARPANET started earlier than this paper, based on the36McDermott, Ibid, p. 121.37"Bear in mind that the existing infrastructure was created for something else.It was created for reasons that you wish it hadn't been[military].But without that impetus, the World Wide Web would never have happened.The fax machine would never have come about if it weren't for existing phone lines. | 9,923.2 | 2014-03-05T00:00:00.000 | [
"Computer Science"
] |
Recognition Design of License Plate and Car Type Using Tesseract Ocr and Emgucv
The goal of the research is to design and implement software that can recognize license plates and car types from images. The method used for the research is soft computing using library of EmguCV. There are four phases in creating the software, i.e., input image process, pre-processing, training processing and recognition. Firstly, user enters the car image. Then, the program reads and does pre-processing the image from bitmap form into vector. The next process is training process, which is learning phase in order the system to be able recognize an object (in this case license plate and car type), and in the end is the recognition process itself. The result is data about the car types and the license plates that have been entered. Using simulation, this software successfully recognized license plate by 80.223% accurate and car type 75% accurate.
INTRODUCTION
Nowadays, the amount of private car is increasing, which leads the increasing of traffic level.The vehicle theft is at high occurrence and parking lot for accommodating them in public place becomes lesser.Therefore, the queuing in parking lot is increasing.Based on data on reskrimum.metro.polri.go.id, there were 10.791 cases of vehicle theft in 2006 with national scale, while in the next year it increased into 11.620 cases.From the problem, there is a system that can ease the queuing in accessing toll and in the entrance of parking area to decrease the queuing in the entrance and also the exit.Besides that, it can also decrease the theft level because every vehicle that in and out are recorded and the data is saved to be accessed if it is needed to track missing vehicles.
The method used is recognition pattern with edge detection that is used to find the boundary of certain pattern in image.It is added with the usage of Artificial Neural Network (ANN) that able to learn and solve complicated connection, which is hard to be descripted between the input and output.
METHOD
The method used is divided into two main parts, which are analysis and design.Analysis method includes literature study about character recognition and artificial neural network.While the design method includes preparation requirement (reporting requirement), design, coding and continues with evaluation.
Automatic License plate Recognition (ANPR)
Based on the explanation from the previous research about recognition of license plate, it can be concluded that the detection is needed for life today.ANPR is needed in several fields like: checking the vehicle speed (vehicle with the speed over the limit, the license plate can be recorded automatically); parking area management (parking ticket, missing and vehicle theft, duration of fixed monitoring, audit, hands-free access); traffic management; controlling and monitoring access; control of vehicle obedient on road.
Artificial Neural Network (ANN)
Neural network is network from group of small processing units that is modeled based on human's Artificial Neural Network (ANN) [1].ANN is adaptive system that can change its structure to solve problem based on external or internal information that flows in the network.
Fig. 1: Artificial Neural Network
The understanding model is based on the assumption: (1) Information delivery process occurs on simple element known as neuron.(2) Information connection among nerve through connector (layer in ANN model).( 3) Every connector has weight from the previous calculation result.The calculation is done using multiplication every passing weight.
ANN distributes in several layers with the amount of input that will produce output, between the input and output, there is hidden layer.Every connection between input to output is nodes that have weight.Image training process is done in this process if there is input image so the system will recognize the entered object.Training can be done many times, the more training, the more accurate the recognition system.
Computer Vision
Computer vision is process to obtain, process, analyze and understand the image [2].Generally the processed image is image with high dimension in real world, in purpose to obtain information in number or symbol.
Computer vision system can replace human's duties and make the system more modern.In scientific study, computer vision is the theory of artificial system that extracts information from image.While in technology study, computer vision finds theory and model from construction that is captured by image.
In plate recognizer, the branch of computer vision is used, which is Optical Character Recognizer (OCR), functions to identify number and letter from the plate.
Image Processing
Image processing is on the image definition, and the processes are as follow [3]: (1) Image; image is two dimensional images that are produced from analog image of two dimension, which is continuous into discrete image through sampling process.Analog image is divided into N row and M column then it becomes discrete image.The crossing between row and column is known as pixel.For example is image or discrete point on row N and column M is known as pizel [n,m].(2) Sampling; sampling is process to decide color in certain pixel on image from a continuous image.On sampling process, it is common searched for average color from analog image that then be rounded.Sampling process is also known as digitization.(3) Quantification; occasionally in sampling process, the obtained average color is realized to certain color level.For example if in image there is only 16 levels of grey color, then the obtained average color from sampling process has to be associated to the 16 levels.Association process of average color with certain color level is known as quantification.(4) Noise; noise is image or pixel that interfere the image quality.Noise can be caused by physical disturbance (optic) on acquisition tool or intentionally caused by unsuitable process.For example black or white dot that appears randomly; and it is unwanted in image.The random dot is known as noise salt & pepper.A lot of methods in image processing aim to decrease or relieve noise.
It can be said that image processing is an activity to fix image so it can be interpreted by human or machine (computer).The input is image and the output also in image form but with better quality form the input image.The purpose of image processing is the information that is delivered by input image conveyed and can be processed by human / machine (computer).The conversion of image color from RGB (Red, Green, Blue) becomes gray.An image will be changed into matrix N x M array, then with the changing color so it can be obtained value from every array element that can be said as image element or pixel (picture element).For a digital image, every pixel has its integer value or known as gray level that shows intensity from pixel.
Grayscaling
Grayscale digital image is image, which the value of each pixel is single, only has description with intensity [4].This kind of image is also known as black-and-white image, which specially consists of gray color, with variation from black on the weakest intensity to white on the strongest part.
Grayscale image is different from one bit bitonal on black-and-white image that on the computer image context is image that only has two colors, black and white (also known as bilevel or binary image).Grayscale image has many shades of gray.Grayscale image is also known as monochromatic that shows only one (mono) color (chrome).
Grayscale image is also result of light intensity measurement on every pixel as single tape from electromagnetic spectrum (such as infrared, visible light, ultraviolet, etc.), and in this case grayscale image is on the right monochromatic when the given frequency caught up.However the thing can be synthesized from colorful image.
Conversion from color to grayscale is not difficult, the different weight from color channel effectively represent effect from black-white image with various photography filters on camera.The common strategy is matching the exposure from grayscale image with the colored image exposure.
To convert color, things needed with a grayscale representation of an exposure are firstly we have to obtain values from red, green and blue (RGB) primary encoding with linier intensity from gamma expansion.For the sRGB room, the gamma expansion is defined as: CsRGB is one of three primer gammacompression sRGB in range [0,1] and Clinear is linear intensity value that suitable (in range [0,1]).Then the exposure is calculated as the amount of weight from three linear intensity values.For ITU-R BT.709 primer, like being used in sRGB, the weight of Y = 0,2126 R + 0,7152 G + 0,0722 B gives CIE 1931 exposure.Linear exposure usually needs a gamma that is compressed to be returned to a conventional grayscale representation.By coding the grayscale intensity RGB, each of three primers can be regulated so it can be same with the exposure that will be calculated.For sRGB, the suitable gamma compression is: This is not method that is used to obtain luma on Y' UV and color model that is used on TV standard color as PAL, SECAM and NTSC.The system directly counts gamma that is compressed by luma as linear combination from compressed gamma by primer intensity than using expansion linearization through the use of gamma and compression.On YUV and YIQ model, it used by PAL and NTSC, the luma component (Y') counted as: The coefficient represents the human perception of color, particularly that humans are more sensitive to green and the most sensitive to blue.The model used for HDTV is developed by ATSC using the color coefficient that slightly different, the calculation of luma component described as:
Canny Edge Detection
Canny Edge Detection is an edge detection operator that uses multi-stage algorithm to detect the edges in the image.It was developed by John F. Canny in 1986.Canny also gave a theory of edge detection computational that explained why this technique works.
OCR (Optical Character Recognition)
Optical Character Recognition (OCR) is branch from computer vision.OCR is recognition towards characters whether in letter (big and small) and also number.This technology will enable machine automatically to recognize the character through optic mechanism [5].OCR technology is widely used to convert books and documents into electronic or digital form.OCR enables the user to edit text and recognize motor vehicle.OCR needs setting to read the font specifically, the smart system works with high accuracy especially if the used font is common font.Generally the recognition process through OCR is explained on diagram.Segmentation process is done by separating the object region with the object background, so the image is easily analyzed for object recognition.
The next process is normalization that has two processes, which are: Scaling is function to change size an image, where scaling is a term that tends for maximizing image and shrink for minimize the image.The second process is thinning, morphology process that is used to erase selected foreground pixel from binary image.It is usually used for the process of searching for the bones of an object.The next step is feature extraction.Feature extraction is an image analysis process in identifying the naturethe inherent properties of each character or known as, the features of an object contained in the image.The characteristic is used to describe an object or attribute of an object, then the features that can be used as a character used as recognition process.After the steps above have been done, then OCR is ready to perform recognition stage and will give the output or result of character recognition in digits and letters.
Tesseract OCR
Tesseract is OCR engine open source that firstly developed by HP (Hewlett-Packard) in 1984-1994.At first tesseract was Ryan Smith's object research in HP lab in Bristol.Fig. 4: Tesseract Architecture Tesseract reads the received input in the form of a binary image.Analysis is performed on the connected components to determine which outline the components that will be saved.Outlines gathered together and become blob.Blob organized into lines of text, while the lines and regions analyzed to be fixed pitch and proportional text.Line of text is divided into words by character space.Text with pitch divided each character cell, then the text proportional split into words using fuzzy space.
The first pass separating the words that already exist in the database, and the second pass is a recognition word in image.
Recall and Precision
Recall is the ratio of the number of documents that can be retrieved by a searching process in the relevant image system [4].The formula of the recall is the amount of relevant documents are found divided by the amount of all relevant documents in a collection or can be elaborated by: Where the relevant documents and available documents divided by the number of documents.
Precision is the proportion of the number of documents found and considered relevant for the needs of the information seekers.The formula of precision is the number of relevant documents that are found divided by the number of all documents found, where the relevant documents and derived from documents divided by the number of documents that wish to be identified.
The Used Method
In the process of character identification on the license plate and vehicle type finds many obstacles.The encountered problems are the segmentation process that should be appropriate to determine the license plates of vehicles, tolerance to existing noise in the license plate image, font types and modifications on the vehicle license plate, adequate lighting so that the image can be recognized and position image retrieval.For lighting problem, it will be limited image acquisition, which is done with adequate lighting, so it can be identified clearly and reducing noise.For the font type and plate modification, it will be done character recognition, a process that is commonly used.Problems in training car; so as not to cause a lot of errors, the training will be conducted for three types of cars that will be recognized based on the basic shape and height of the car and classified into sedan, MPV and truck.
Another problem is the problem of segmentation in character, making it difficult to perform manual training to recognize the character.The solving problem is by using OCR from Tesseract contained in the EmguCV library.
Steps of Troubleshooting
Steps that will be done consist of three stages: pre-processing, training process and the process of identification.Broadly speaking, the system will be designed as follows: The first stage is the user inputs car image into the program.After the image is inputted, the program will read and perform pre-processing process, which is the process of image processing from bitmap into an array of 7x7 to the inputted image.After the preprocessing successfully executed then entered the training stage, which is system learning phase in order to recognize an object (in this case license plate and car type).After that the next process will be done recognize process, namely the process of introduction of a system that has previously been trained.After all the process is successful the result will show the introduction of the type of car and license plate of image that has been previously entered.In the diagram above describes the process by which a received image of the car will be done the detection by taking part to be known as the license plate area.After obtaining the license plate area, then it will be done segmentation process towards the car plate, which is the process of separating the image plate from the image.The results of the segmentation plate will be re-segmentation of the license plate characters.After getting the extraction of character, it will be directly included in the license plate character recognition stage of the car's image.In the above diagram is explained that the recognition of the car type is done through sampling and grayleveling process first.After going through the process of sampling and grayleveling, it will be the training process to the image.After the training process of the image is completed, the training results can be saved for reuse.The diagram above is the identification of the type of car flow, to perform recognition on the image of the car, we need to do the sampling and grayscaling process again so that the image can be more easily recognized.After performing the sampling and grayscaling process, it will be continued with noise removal and then will be identified through ANN that has been in training in the previous process.After the recognition by ANN, then the car type of image can be recognized.
Detection and Extraction Plate Number Sampling
The inputted pictures of cars will be sampling.As well as on the theory that the sampling is the process of changing array (matrix) size NxM matrix and each element has its own value.Pictures will be read and then opened with imshow function.The results of these images will be sampling using 7x7 matrixes.
Fig. 9: A Car Original Image
The image that has been through the process of sampling will be converted from RGB to gray or grayscale with level of 28 or 256.
Rectangle Detection
By utilizing the special characteristics of the license plate is rectangular, and then the next step is the rectangle detection.In this rectangle detection process relies on the detection of a line that has been done before as constituent of the rectangle is a collection of lines.It is used rectangle detection to search for a candidate vehicle license license plate because the physical plate form in Indonesia, which supports and is easily recognizable as a rectangle.
Rectangular detecting process is to check towards the two lines are parallel.Then the adjacent lines will be connected to form a parallelogram.Furthermore, the formed parallelogram will be marked as rectangle and normalized angle to 90 degrees.
The results of this rectangle detection generate several candidates as input for the selection process license plate.
Selection of License plate
The selection process of license plate is done by selecting the rectangle that has been detected in the previous stage by using the formula plate ratio = length / width.
The next is after the process of division, it is defined limitation to determine which plate will be selected.The limit is by checking whether the detected rectangle has a character or not.Assuming the license plates have at least 5 characters, then if it is less than 5 characters means it is not included in the license plate region.
Recognition
Character recognition process will read the license plate characters contained in the image by calling the Tesseract functions contained in the EmguCV library.After the character is known, it will be displayed on the output of the textbox.There are two stages, namely: noise filtering and recognition.
Noise Filtering
This process is necessary to fulfill the required parameters for the next process, which is Optical Character Recognizer (OCR).This process is done to fix the license plate so that the object can be inserted in the recognition stage.
This process uses canny edge detection, so that the edges of the numbers and letters character are detected.Furthermore, it is used erosion and dilation process.
Erosion and dilation can be performed simultaneously or even themselves depend on the results we obtained are sufficient to meet the expectations or not.The expected result is each character is subjected to this process forming a valid result.
Erosion and Dilation is used together to unify the separated pixels so ti can clarify the form of letters or numbers and also eliminate noise powder in pixels.
Plate Recognition
Once the noise is eliminated through the process of noise filtering, then it is ready for recognition stage.The license plate recognition stage uses Tesseract OCR engine.(3) Separation of Connected Character.If there are overlapping in the character segmentation result, it will be fixed in this stage.The formed overlap blop will be separated.Candidate for the separation point is found on the concave vertices of polygonal outline approaches and there may be more opposite concave point or line segment.(4) Broken Character Association.If the overlapping character has been successfully separated, check wheterher there are characters that broken or slightly damaged.If there is, the outline of the defective character will be repaired using the best first search.(5) Adaptive Classifier.Featured is a polygonal approach component from a shape outline.In the recognition, the polygon elements are broken down into the shorter sections of equal length, so that the long dimension of the feature vector is eliminated.Some short features are matched with some prototipkal features of training that has been done by Tesseract in .traineddatafile.It makes the clarification process more robust to the lost character.In this stage the character of the plate fitted with the prototype of the characters in the training Tesseract data.When approaching the similarity then the character will be displayed.(6) Recognition Result.If the matched characters are close to the resemblance with the prototype character in the training Tesseract data, then the character will be displayed.The recognition result of license plate characters is as follows.Image, which should be opened, is a digital image that contained the vehicle license plate, and it must be larger than 240x180 pixels because this size can produce accuracy level that is better than Tesseract ocr.
Training Car Type
The training process will be done through artificial neural networks with back propagation method, which after the image through the preprocessing stage the image will be trained, based on the area and height of the car.With some done training, then the system will recognize the type of car based on the digital image input.
The numbers of images that will be used as training sample as many as 16 samples, which consist of 8 cars with background image and 8 cars without background image.Here are the used samples: (1) 80,223%.Automatically then it can be obtained failure rate of this program is approximately 20%.From the analysis, the failure detection caused by several things, namely: the position of the license plate is too far and too close to the plate not detected, the condition of the damaged plate and the effect of poor lighting or too bright.As for the recognition of the car type program has an accuracy rate of 75%.
CONCLUSION
Based on the results of testing to 30 car types that have been implemented in the program, it can be concluded that: This program has reliabilities such as: (1) Time detection and its character recognition of existing numbers and letters on the plates can quickly be recognized by the average time less than 500 milliseconds.(2) All the edges that is considered as square will be detected so that the candidate license plate will not be missed.(3) The program can recognize the car type globally (MPV, sedan and truck).
The program also has its drawbacks include: (1) Unable to detect the modified plate or damaged plate, for example faded plate.(2) Unable to detect all sizes plate, only a certain length and width that can be detected perfectly.Thus the plate recognition system and the car type are not strong enough (robust) in terms of detection, license plate recognition and the type of four-wheeled vehicles.
Fig. 3 :
Fig. 3: Diagram of OCR Process is a library open source cross platform.EmguCV can call functions of the OpenCV library in image processing.EmguCV is compatible with many programming languages, such as C ++, C #, VB, IronPhyton and others.EmguCV can be run by various OS such as Windows, Linux, and MacOS.The advantages of EmguCV are: (1) Cross-Platform: unlike other wrappers that are written in unsafe code.EmguCV writing is written in C#. (2) Compatible with many programming languages: can be used in many programming languages such as C ++, C #, VB and IronPhyton.(3) Compatible with a wide range of Operating Systems: can be used in various kinds of OS such as Windows, Linux and MacOS.(4) There is a drawing class with generic colors and has a depth.(5) Can use the functions of OpenCV.
FinishFig. 6 :
Fig. 6: Diagram of Extraction Flow In Pre-Processing and Character Identification
Fig. 7 :
Fig. 7: Diagram of Pre-Processing Flow and Training Car Type
Fig 11 :
Fig 11: Car Image that has done License plate Selection
Fig. 15 :
Fig. 15: The Recognition Result of License Plate Character (1) The program can detect and recognize the license plate of four-wheeled vehicle with an accuracy of 80.223%.(2) The program can detect and identify the car type with an accuracy of 75%.(3) The program can detect and recognize the license plate number and car type using soft computing with an accuracy of 80%.
( 3 )
The detection of the car type should take a long training process.(4) The recognition of the car type should be the image at the corresponding position.(5) Uncertainty as to the composition of the amount of input, hidden and output layers in obtaining optimal training results.
Table 1 :
Calculation Character Plate | 5,664.4 | 2012-10-31T00:00:00.000 | [
"Computer Science"
] |
Poly(delta-gluconolactone) and Poly(delta-gluconolactone-ε-caprolactone) from delta-Gluconolactone and ε-Caprolactone by Ring-Opening Polymerization
Poly(delta-gluconolactone) (PGL) and poly(delta-gluconolactone-ε-caprolactone) (P(GL-CL)) were synthesized through ringopening polymerization (ROP) and characterized by FT-IR, NMR, XRD, intrinsic viscosity, GPC, DSC, and TGA.The crystallinity of P(GL-CL) with various d-GL/CL ratios (d-GL/CL = 5 : 5, 4 : 6, 3 : 7, 2 : 8, and 1 : 9) was 12.09 to 59.78% while PGL was amorphous. Melting temperature (Tm) of these polymers was 49.8 to 62.0C and decomposition temperature was 282 to 489C depending on the d-GL/CL ratios. In addition, all these polymers were degradable and the degradation rates could be controlled by adjusting d-GL/CL ratios. These results indicated that PGL and P(GL-CL) might be promising novel absorbable materials.
Introduction
Glucose is the most important energy source of living cells and intermediate materials of metabolism, which stored as a polymer, in plants as starch and in animals as glycogen [1]. Beside the function in biology, it is used as medicine, food additives, and chemicals [2]. Because of its huge storage in the celluloses, the glucose and its derivatives had a wide range of applications, such as calcium gluconate and zinc gluconate which are used as sequestrants, acidifiers, curing agents, pickling agents, or leavening agents [3,4]. Although there are extensive researches and applications on the deltagluconolactone (d-GL), the research on its further chemical reaction, derivatives, and synthesis polymers is relatively rare, especially on its ring-opening polymerization (ROP) reaction. This consequence is partly due to the presence of hydroxyl (-OH) on d-GL that is detrimental for ROP, as it leads to deactivation of the polymerization via undesirable side reaction, such as crosslinking-induced gelation [5,6]. The polymerization of d-GL into poly(delta-gluconolactone) (PGL) suggested novel potential applications, like drug delivery system by introducing polar functionalities groups (-OH) as well as the functionalities group of the polymers, which can provide anchor for tissue regeneration through cell attachment, proliferation, control of inflammation, and healing.
PGL might be a good affinity for water while poly(caprolactone) (PCL) is hydrophobic. Considering the similar cyclic structure between d-GL and -caprolactone (CL), copolymerization of d-GL with CL could overcome the drawbacks and induce new functionalities, to yield biobased polymers with possible industrial and medical applications [7,8]. Such polymerization reactions can be accomplished by lipase biocatalysis, and d-GL can be copolymerized withcaprolactone (CL) and -butyrolactone through immobilized lipases to form an aliphatic oligomer (best experiment m/z value 1587) in previous reports [7,8]. The aim of present work is to obtain higher molecular weight polymers through chemical method. To our knowledge, this is the first research in which PGL and P(GL-CL) with high molecular weight were successfully synthesis from d-GL and CL by chemical method. The aim of this work was to discuss synthesis, characterization, and in vitro degradation of PGL and P(GL-CL).
Synthesis PGL and PCL.
In a typical experiment [11,14], a Schlenk tube equipped with a magnetic stirrer was sequentially charged with catalyst (MBTO, 2 equiv), monomers (CL or d-GL, 1000 equiv), and initiator (diethylene glycol, 1 equiv) in n-octane followed by pumping vacuum and high pure dry nitrogen in turn for three times. Subsequently, the temperature was gradually raised to 160 ∘ C and then reaction was conducted at this temperature for assigned time to obtain PGL. PCL was synthesized through the same procedure except that the polymerization temperature was 140 ∘ C. After the reaction, the n-octane was removed in vacuum and then the target product was purified by dissolving in dichloromethane followed by precipitating in ethanol.
Synthesis of P(GL-CL) Copolymers.
All these P(GL-CL) copolymers with various d-GL/CL ratios (d-GL/CL = 5 : 5, 4 : 6, 3 : 7, 2 : 8, 1 : 9) were obtained by one-pot-two-step process (Scheme 1). In short, a Schlenk tube equipped with a magnetic stirrer was charged sequentially with catalyst (MBTO, 2 equiv), initiator (diethylene glycol, 1 equiv), and CL in n-octane followed pumping vacuum and high pure dry nitrogen in turn for three times [15,16]. Subsequently, the temperature was gradually raised to 140 ∘ C and kept for 4 hours and then d-GL was added to the flask and kept for another 4 hours. After that, the reaction temperature increased to 160 ∘ C and was kept for another 4 hours. After completion of this polymerization, the n-octane was removed in vacuum. The target product was purified and obtained by dissolving in dichloromethane, precipitating in ethanol and drying in vacuum to constant weight, sequentially.
In Vitro Degradation in PBS.
An in vitro degradation study of P(GL-CL) was carried out in phosphate buffered saline (PBS). All the samples were shattered into particles (1100-2000 m) and soaked in PBS with a weight-to-volume ratio of 1/30 (g mL −1 ) at 37 ∘ C in a shaking water bath [17]. The incubation time was up to 10 weeks with PBS refreshment every week. The initial weight ( 0 ) of the particles was obtained after the samples were dried in vacuum. The weight ( 1 ) at certain time point was obtained by weighting the Diethylene glycol --- (6) n-Octane TBT Diethylene glycol --- Diethylene glycol --- (10) n-Octane MBTO ---Off-white (11) n-Octane Al complexes Diethylene glycol --- (12) Methylbenzene MBTO Benzyl alcohol Off-white a ---reaction cannot carry out or cross-linking and gelation occur.
specimens after cleaning with water and drying in vacuum.
The weight loss ratio of the samples was calculated using the following: (1)
Fourier Transformed Infrared (FT-IR) of PGL and P(GL-CL).
Fourier transformed infrared (FT-IR) spectra were recorded on a Thermo Nicolet IR-100 spectrometer via sample films coated on KBr crystal.
Intrinsic Viscosity and Gel Permeation Chromatography (GPC) of Polymers.
The intrinsic viscosity measurement of each polymers was carried out in Dimethyl Sulphoxide (DMSO) at 30 ∘ C by Cannon-Ubbelohde viscometer. The intrinsic viscosity values were obtained by a one-point method (the Solomon-Ciuta equation) as follows: where = / 0 , = / 0 − 1, and was polymer concentration in DMSO. And the number average ( ) and polydispersity index (PDI) were measured using Water-2414 equipped with a Waters Styragel HT4 and using Dextran as standards with molecular weights from 1.0 × 10 3 to 2.7 × 10 5 g mol −1 . THF was used as mobile phase at a rate of 1 mL min −1 and the column temperature was 30 ∘ C.
XRD of PGL and P(GL-CL)
. XRD (Philips X'pert Pro MPD) was used to study the crystalline characteristics of the polymers.
Differential Scanning Calorimetry Analysis (DSC) and
Thermogravimetric Analysis (TGA). Differential scanning calorimetry analysis (DSC) was performed on a Perking-Elmer DSC-7 apparatus under nitrogen in the temperature ranging from 0 to 200 ∘ C at a heating rate of 10 ∘ C min −1 . The samples were first heated to 160 ∘ C and then cooled to 30 ∘ C by liquid nitrogen, followed by heating to 200 ∘ C at a heating rate of 10 ∘ C min −1 . The glass transition temperature ( ) and melting point was recorded from the cooling flow and the second heating run, respectively. The thermogravimetric analysis (TGA) was conducted on PerkinElmer Pyris Diamond thermogravimetric/differential thermal analyser (TG/DTA) at 10 ∘ C min −1 from 30 to 600 ∘ C.
ROP of d-GL in Various
Conditions. In this study, different reaction catalysts were investigated for ROP of d-GL, and the results are shown in Table 1. It was found that some general catalysts for polyester polymerization, like TBT, Sn(Oct) 2 , some Al complexes (aluminium isopropoxide, aluminumtriethyl, triisobutylaluminium) and the Rare Earth catalyst Y[N(SiMe 3 ) 2 ] 3 prepared as previous reported [18], failed for ROP of d-GL. When monobutyltin oxide (MBTO) was employed as catalyst and reflexed in n-octane at 150 ∘ C for 8 hours under nitrogen, an off-white polymer (PGL) was precipitated from the solution. The obtained polymer was freely soluble in water and some polar solvents (e.g., THF, DMSO, and CHCl 3 ) but insoluble in alcohol, acetone, and methylbenzene. In addition, various reaction temperature (120, 130, 140, 150, 160, and 170 ∘ C) and reaction time (4, 6, 8, and 12 h) were also investigated in the MBTO catalytic system, summarized in Table 2. It was found that when the temperature was lower than 120 ∘ C, the reaction rate was lower, or the reaction could not proceed. And conversion increasing with rising the polymerization time when temperature was lower than 160 ∘ C. But coloration or decomposition of an aliphatic polyester could be manufactured, when the temperature was higher than 170 ∘ C. Figure 2, the stretching vibration of hydroxyl associated with hydrogen bond shifted from (3468, 3392, 3277, and 3218 cm −1 ) of d-GL to an enlarged peak (3392 cm −1 ) in PGL. In addition, compared to d-GL, the deformation peaks of hydroxyl shifted from 1719 to 1736 cm −1 in PGL. Moreover, the sharp peaks corresponding to the C=O symmetric stretching vibration of ester linkage shifted from 1719 to 1736 cm −1 in PGL, compared with d-GL. Figure 2 showed the FT-IR spectra of copolymers with various d-GL/CL ratios. It was found that the FT-IR spectra of P(GL-CL) compared to PCL showed a broad-enlarged peak assigned to hydroxyl symmetric stretching vibration around 3526 cm −1 . The enlarger hydroxyl peak of P(GL-CL) showed blue-shifted with d-GL/CL ratio increasing, as well as the peaks assigned to C=O symmetric stretching vibration from (1700 to 1750 cm −1 ).
International Journal of Polymer Science 5 Figure 4(a) showed the XRD patterns of the novel polymer PGL was laigh and diffuse, compared to its monomer d-GL. In Figure 4(b), obvious crystallinity peak of PCL appeared around 21.5 and 24 ∘ , and the intensity of P(GL-CL) decreased markedly as d-GL/CL ratio increasing. The degrees of crystallinity of P(GL-CL) copolymers (d-GL/CL = 5 : 5, 4 : 6, 3 : 7, 2 : 8, 1 : 9, and 0 : 10) were 12.09, 28.01, 33.42, 38.79, 40.51, and 59.78% (Table 3) obtained by X'pert HighScore software, respectively. Figure 5(a), DSC data of PGL showed that no melting point could be observed during heating flow run and the glass transition peak appeared at 21.4 ∘ C in the cooling flow. In Figure 5(b), compared to melting point peaks of PCL appearing around 62.0 ∘ C, the DSC curves of P(GL-CL) copolymers started to split off into double peaks at slightly lower than melting point of PCL homopolymer, and it became more obvious as d-GL/CL ratio increases. In particular, the DSC curve of the copolymer (d-GL/CL = 5 : 5) showed split-off peak at 49.8 and 55.8 ∘ C and it showed continuous decline at higher temperature. And the first melting point of copolymers (d-GL/CL = 1 : 9, 2 : 8, 3 : 7, 4 : 6, and 5 : 5) were 54.0, 53.2, 53.1, 52.7, and 49.8 ∘ C, respectively.
The Thermal Properties of PGL and P(GL-CL). In
In Figure 5(c), the TGA curses showed initiate and decomposition degradation temperatures of PGL were 222 and 299 ∘ C, respectively, while PCL's initiate and decomposition temperatures were 282 and 347 ∘ C. It can also be seen that a flight of double stages existed in the TGA curves of P(GL-CL) copolymers and it became more blurred as d-GL/CL ratio increasing; notably, the copolymers with d-GL/CL molar ratio of 5 : 5 showed continuous slowly decline. The degradation temperature of P(GL-CL) was widened, for which decomposition degrade temperatures of copolymers (d-GL/CL = 5 : 5, 4 : 6, 3 : 7, 2 : 8, 1 : 9) were 459, 453, 432, 425, and 415 ∘ C, much higher than PCL and PGL, respectively. Figure 6(a). It could be found that all the samples, except PCL, had a quick weight loss in the initial 2 weeks, followed by a slightly slow weight loss and PGL was easily dissolved in PBS. The weight loss of polymers (d-GL/CL = 0 : 10, 1 : 9, 2 : 8, 3 : 7, 4 : 6, and 5 : 5) were 0.4, 9.1, 19.7, 23.7, 33.6, and 56.2 wt% in the first week and 6.8, 24.9, 35.7, 45.9, 52.3, and 73.5 wt% at the end of the experiment, respectively. With d-GL content increasing in the main chain, the degradation of P(GL-CL) was significantly accelerated. And the degradation rate of copolymers gradually decreased as soaking time increasing; finally the rate of weight loss of copolymers was consistent with the PCL. Figure 6(b) expressed the pH changes of polymers (d-GL/CL = 0 : 10, 1 : 9, 2 : 8, 3 : 7, 4 : 6, 5 : 5) were 7.2, 5.9, 3.8, 2.4, and 2.1 at the first week, respectively, during soaking and the pH of PGL dissolved in PBS was 1.6. And the neutral pH of copolymers was achieved gradually at the end of the experiment.
Discussion
In this work, the d-GL was successfully catalyzed by MBTO into polymer without hydroxyl protection and deprotection and copolymerization with CL through one-pot-two-step process. But it was failed to polymerize d-GL into PGL through melt polymerization and by some general catalysts which could easily catalyst CL into PCL [14,19,20]. Noticed common polarity solvents, such as THF, DMSO, and CHCl 3 , caused crosslinking and gelation in the ROP process of d-GL into PGL ( Table 1). The reason why P(GL-CL) copolymer could not synthesis through one-pot-one-steps process might be that the coloration or decomposition of polymer was induced by the polar dissolvability of CL. However, in the one-pot-two-step process, most of CL was catalyzed into PCL before d-GL was added to the reaction system, so it could be avoiding some side reaction of d-GL induced by the polar dissolvability of CL. Considering the effect of temperature (Table 2), the polymerization temperature of PGL was up to 130 to 170 ∘ C, because the melting point of d-GL was 151 to 155 ∘ C. When the temperature was lower than 120 ∘ C, the reaction rate was lower, or the reaction could not proceed. And coloration or decomposition of an aliphatic polyester could be manufactured [21], when the temperature was higher than 170 ∘ C. Thus, the favorable reaction temperature of PGL might be 150 to 160 ∘ C, and the reaction system should be limited to reflex in n-octane due to the boiling point of noctane was 125 ∘ C. The favorable polymerization time might be 4-6 h in consideration of catalyst d-GL into PGL.
In Figure 1(a), it is obvious to observe that all bands ascribed to hydroxyl, for stretching vibration at 3392 cm −1 and for deformation vibration at 1226 cm −1 , shifted to enlarged peaks in PGL, compared with d-GL, indicating that associated hydroxyl existed in PGL's structure after ROP. And the C=O vibration peak of PGL was band reshifted from 1719 to 1736 cm −1 , indicating C=O group was substituted with electron donating in ROP process (Scheme 2). From Figure 1(a), peak at = 4.0 (ddd, = 2.35, 2.07, 2.18 Hz, 4H) was observed as a steamed bun and tripeak, indicating the hydroxyl attached to three kinds of carbons (CH, CHH, CHH), and it disappeared in D 2 O due to hydroxyl exchange with solvent formed HDO [22], and it is easily observed PGL slow degradation into glucose acid in Figure 2(b). All these results are of FT-IR and 1 H NMR; it is most likely the desired product obtained was successful, and hydroxyl was protected in the whole syntheses process [6]. Along with d-GL/CL ratio increasing in P(GL-CL), the absorption peaks assigned hydroxyl of P(GL-CL) became more obvious and shifted to lower wavenumbers, indicating that hydroxyl was protected and undesirable reaction did not occur in the whole reaction. The same blue-shift phenomenon was observed at the peaks assigned C=O symmetric stretching vibration of P(GL-CL) and the GPC curves of all these P(GL-CL) were observed to be monomodal, demonstrating that the resulting products are a copolymer in nature not a mixture of PCL and PGL [23], so d-GL was copolymerized with CL successfully.
The mechanistic of ROP (Scheme 2) was hypothesized according to previously disclosed publication [18,[24][25][26], which follows typical coordination-insertion mechanism. At first, MBTO formed active species with initiator; then the active species ring-opening d-GL in the chain propagation step and dormant chain were formed by isolating MBTO segment in the chain growing process. As discussed in abovementioned, it was the carbonyl oxygen rather than ether oxygen of d-GL that firstly coordinates with active species during the coordination polymerization, because C=O group substituted with electron donating in FT-IR spectrum of PGL (Figure 1(a)) [26]. And the possibility of primary hydroxyl of d-GL participating in the reaction process could not be excluded because the 1 H NMR (Figure 2(a)) showed much clutter from 3.48 to 3.43 ppm. This hypothesis can be further confirmed by the experiment (Table 1, Entry (10)), which PGL also could be obtain in the reaction system without initiator. In that case, the primary hydroxyl of d-GL might be replacing hydroxyl of initiator in the reaction system. But in D 2 O, PGL could be slowly degrading to glucose acid, indicating that some undesirable reaction did not occur in ROP of d-GL.
The copolymerization of d-GL with CL underwent a similar manner of ROP (Scheme 3). When CL monomer was consumed to some extent and PCL segment was long enough, d-GL could competitively replace CL to insert into the polymer chain, leading to block copolymer poly(GL-CL). In the PCL or PGL dominant component, the other less content unit as a defect existed in chain structure. This hypothesis block chain could be proved by TGA of P(GL-CL), which showed two kinds of stages.
From results of GPC ( Figure 3, Table 3), the PDI of polymers was ranged from 1.16 to 1.41, indicating a narrow molecular weight distribution of the PGL. Compared to PCL, of polymers was shown to be decreasing as the ratio of d-GL/CL increasing, which might be related to the fact that reactivity of d-GL's six-membered ring structure was relatively lower than CL's seven-membered structure [26]. All P(GL-CL)'s GPC curves were observed to be monomodal, demonstrating the resulting products were a copolymer instead of mixtures of PCL and PGL [23]. The DSC data ( Figure 5(a)) of PGL showed that no melting point ( ) was found in heating flow and XRD curves (Figure 4) of PGL was laigh and diffuse, indicating PGL belongs to amorphous structure. Likewise, the XRD curves ( Figure 4) of P(GL-CL) became more laigh and diffuse as d-GL content increasing, showing lower crystallinity of P(GL-CL). And DSC curves of P(GL-CL) split-off more obvious as d-GL/CL ratio increasing at the slightly lower than pure PCL (62.0 ∘ C) and showed continuous decline at higher temperature. These results could be caused by semicrystalline PCL that cannot form a eutectic with amorphous PGL and both homopolymer chain segments were long enough in block chain structure, which are two representatives of melting generated by each segment.
In Figure 5(c), below 222 ∘ C, the PGL loses ca.7% of its weight, suggested PGL crosslinking at high temperatures and low humilities was accompanied by a loss of water. And the backbone thermal degradation temperature of PGL was between 222 and 299 ∘ C. However, the decomposition temperature of P(GL-CL)s was higher than both PCL and PGL. The reason might be that endothermic of PGL segment thermal degradation at first and the PCL segment was protected from backbone thermal degradation. This phenomenon could be explained by the block copolymer structure, which segment with longer PGL start thermal degradation at first and thermal degradation endothermic process protected the shorter one, even leading a slowly continuous decline in TGA curves of P(GL-CL) (d-GL/CL = 5 : 5). The TGA results showed that both PGL and P(GL-CL) had good thermal stability.
Because of pendant hydroxyl of d-GL, P(GL-CL) with more PGL segment could be continually hydrolyzed, resulting in quick hydrolysis and degradation. And the PGL segment in P(GL-CL) also reduced the molecular chain regularity, and this reinforces the short range hydration during the incubation in the PBS and enhanced the hydrophilic. As a result, the higher content of PGL segment in P(GL-CL) copolymers showed more degradation than that of PCL. The degradation of P(GL-CL) in the present research showed 6.8 to 73.5 wt% in the 10-week soaking in PBS with deferent d-GL/CL molar ratios. So degradation rates of polyesters could be further adjusted based on the d-GL/CL proportion. This would great improve the degradation of the PCL and extend its application.
Conclusions
PGL and P(GL-CL) have been successfully synthesized without hydroxyl protection and deprotection. The obtained PGL was amorphous, had a good thermal stability and hydrophily, and could degrade into glucose acid. Because the polar hydroxyl was introduced into P(GL-CL) copolymers, the hydrolytic degradation rate is significantly changed. Moreover, the eutectic cannot have form in P(GL-CL), resulting in the fact that the thermal properties have a tremendous change. Therefore, PGL and P(GL-CL) may be novel absorbable biomaterials because they have tunable biodegradability, as well as good physical and chemical properties, being worthy of being further evaluated in vivo and in vitro. | 4,911.2 | 2018-02-06T00:00:00.000 | [
"Materials Science",
"Chemistry"
] |
Study of jet shapes in inclusive jet production in pp collisions at √s = 7 TeV using the ATLAS detector
Jet shapes have been measured in inclusive jet production in proton-proton collisions at ffiffiffi s p ¼ 7 TeV using 3 pb (cid:1) 1 of data recorded by the ATLAS experiment at the LHC. Jets are reconstructed using the anti-k t algorithm with transverse momentum 30 GeV < p T < 600 GeV and rapidity in the region j y j < 2 : 8 . The data are corrected for detector effects and compared to several leading-order QCD matrix elements plus parton shower Monte Carlo predictions, including different sets of parameters tuned to model fragmentation processes and underlying event contributions in the final state. The measured jets become narrower with increasing jet transverse momentum and the jet shapes present a moderate jet rapidity dependence. Within QCD, the data test a variety of perturbative and nonperturbative effects. In particular, the data show sensitivity to the details of the parton shower, fragmentation, and underlying event models in the Monte Carlo generators. For an appropriate choice of the parameters used in these models, the data are well described.
I. INTRODUCTION
The study of the jet shapes [1] in proton-proton collisions provides information about the details of the partonto-jet fragmentation process, leading to collimated flows of particles in the final state. The internal structure of sufficiently energetic jets is mainly dictated by the emission of multiple gluons from the primary parton, calculable in perturbative QCD (pQCD) [2]. The shape of the jet depends on the type of partons (quark or gluon) that give rise to jets in the final state [3], and is also sensitive to non-perturbative fragmentation effects and underlying event (UE) contributions from the interaction between proton remnants. A proper modeling of the soft contributions is crucial for the understanding of jet production in hadron-hadron collisions and for the comparison of the jet cross section measurements with pQCD theoretical predictions [4,5]. In addition, jet shape related observables have been recently proposed [6] to search for new physics in event topologies with highly boosted particles in the final state decaying into multiple jets of particles.
Jet shape measurements have previously been performed in pp [7], e ± p [8], and e + e − [9] collisions. In this paper, measurements of differential and integrated jet shapes in proton-proton collisions at √ s = 7 TeV are presented for the first time. The study uses data collected by the ATLAS experiment corresponding to 3 pb −1 of total integrated luminosity. The measurements are corrected for detector effects and compared to several Monte Carlo (MC) predictions based on pQCD leading-order (LO) matrix elements plus parton showers, and including different phenomenological models to describe fragmentation processes and UE contributions.
The paper is organised as follows. The detector is described in the next section. Section 3 discusses the simulations used in the measurements, while Section 4 and Section 5 provide details on jet reconstruction and event selection, respectively. Jet shape observables are defined in Section 6. The procedure used to correct the measurements for detector effects is explained in Section 7, and the study of systematic uncertainties is discussed in Section 8. The jet shape measurements are presented in Section 9. Finally, Section 10 is devoted to summary and conclusions.
II. EXPERIMENTAL SETUP
The ATLAS detector [10] covers nearly the entire solid angle around the collision point with layers of tracking detectors, calorimeters, and muon chambers. For the measurements presented in this paper, the tracking system and calorimeters are of particular importance.
The ATLAS inner detector has full coverage in φ [11] and covers the pseudorapidity range |η| < 2.5. It consists of a silicon pixel detector, a silicon microstrip detector and a transition radiation tracker, all immersed in a 2 Tesla magnetic field. High granularity liquid-argon (LAr) electromagnetic sampling calorimeters cover the pseudorapidity range |η| < 3.2. The hadronic calorimetry in the range |η| < 1.7 is provided by a scintillator-tile calorimeter, which is separated into a large barrel and two smaller extended barrel cylinders, one on either side of the central barrel. In the end-caps (|η| > 1.5), LAr hadronic calorimeters match the outer |η| limits of the end-cap electromagnetic calorimeters. The LAr forward calorimeters provide both electromagnetic and hadronic energy measurements, and they extend the coverage to |η| < 4.9.
The trigger system uses three consecutive trigger levels to select events. The Level-1 (L1) trigger is based on custom-built hardware to process the incoming data with a fixed latency of 2.5 µs. This is the only trigger level used in this analysis. The events studied here are selected either by the system of minimum-bias trigger scintillators (MBTS) or by the calorimeter trigger. The MBTS detector [12] consists of 32 scintillator counters of thickness 2 cm organized in two disks. The disks are installed on the inner face of the end-cap calorimeter cryostats at z = ±356 cm, such that the disk surface is perpendicular to the beam direction. This leads to a coverage of 2.09 < |η| < 3.84. The jet trigger is based on the selection of jets according to their transverse energy, E T . The L1 jet reconstruction uses the so called jet elements, which are made of electromagnetic and hadronic cells grouped together with a granularity of ∆φ × ∆η = 0.2 × 0.2 for |η| < 3.2. The jet finding is based on a sliding window algorithm with steps of one jet element, and the jet E T is computed in a window of configurable size around the jet.
III. MONTE CARLO SIMULATION
Monte Carlo simulated samples are used to determine and correct for detector effects, and to estimate part of the systematic uncertainties on the measured jet shapes. Samples of inclusive jet events in proton-proton collisions at √ s = 7 TeV are produced using both PYTHIA 6.4.21 [13] and HERWIG++ 2.4.2 [14] event generators. These MC programs implement LO pQCD matrix elements for 2 → 2 processes plus parton shower in the leading logarithmic approximation, and the string [15] and cluster [16] models for fragmentation into hadrons, respectively. In the case of PYTHIA, different MC samples with slightly different parton shower and UE modeling in the final state are considered. The samples are generated using three tuned sets of parameters denoted as ATLAS-MC09 [17], DW [18], and Perugia2010 [19]. In addition, a special PYTHIA-Perugia2010 sample without UE contributions is generated. Finally, inclusive jet samples are also produced using the ALPGEN 2.13 [20] event generator interfaced with HERWIG 6.5 [21] and JIMMY 3.41 [22] to model the UE contributions. HERWIG++ and PYTHIA-MC09 samples are generated with MRST2007LO * [23] parton density functions (PDFs) inside the proton, PYTHIA-Perugia2010 and PYTHIA-DW with CTEQ5L [24] PDFs, and ALPGEN with CTEQ61L [25] PDFs.
The MC generated samples are passed through a full simulation [26] of the ATLAS detector and trigger, based on GEANT4 [27]. The Quark Gluon String Precompound (QGSP) model [28] is used for the fragmentation of the nucleus, and the Bertini cascade (BERT) model [29] for the description of the interactions of the hadrons in the medium of the nucleus. Test-beam measurements for single pions have shown that these simulation settings best describe the response and resolution in the barrel [30] and end-cap [31] calorimeters. The simulated events are then reconstructed and analyzed with the same analysis chain as for the data, and the same trigger and event selection criteria.
IV. JET RECONSTRUCTION
Jets are defined using the anti-k t jet algorithm [32] with distance parameter (in y −φ space) R = 0.6, and the energy depositions in calorimeter clusters as input in both data and MC events. Topological clusters [5] are built around seed calorimeter cells with |E cell | > 4σ, where σ is defined as the RMS of the cell energy noise distribution, to which all directly neighboring cells are added. Further neighbors of neighbors are iteratively added for all cells with signals above a secondary threshold |E cell | > 2σ, and the clusters are set massless. In addition, in the simulated events jets are also defined at the particle level [33] using as input all the final state particles from the MC generation.
The anti-k t algorithm constructs, for each input object (either energy cluster or particle) i, the quantities d ij and d iB as follows: where k ti is the transverse momentum of object i with respect to the beam direction, φ i its azimuthal angle, and y i its rapidity. A list containing all the d ij and d iB values is compiled. If the smallest entry is a d ij , objects i and j are combined (their four-vectors are added) and the list is updated. If the smallest entry is a d iB , this object is considered a complete "jet" and is removed from the list. As defined above, d ij is a distance measure between two objects, and d iB is a similar distance between the object and the beam. Thus the variable R is a resolution parameter which sets the relative distance at which jets are resolved from each other as compared to the beam. The anti-k t algorithm is theoretically well-motivated [32] and produces geometrically well-defined ("cone-like") jets. According to MC simulation, the measured jet angular variables, y and φ, are reconstructed with a resolution of better than 0.05 units, which improves as the jet transverse momentum, p T , increases. The measured jet p T is corrected to the particle level scale [5] using an average correction, computed as a function of jet transverse momentum and pseudorapidity, and extracted from MC simulation.
V. EVENT SELECTION
The data were collected during the first LHC run at √ s = 7 TeV with the ATLAS tracking detectors, calorimeters and magnets operating at nominal conditions. Events are selected online using different L1 trigger configurations in such a way that, in the kinematic range for the jets considered in this study (see below), the trigger selection is fully efficient and does not introduce any significant bias in the measured jet shapes. Table 1 presents the trigger configurations employed in each p T region and the corresponding integrated luminosity. The unprescaled trigger thresholds were increased with time to keep pace with the LHC instantaneous luminosity evolution. For jet p T smaller than 60 GeV, the data are selected using the signals from the MBTS detectors on either side of the interaction point. Only events in which the MBTS recorded one or more counters above threshold on at least one side are retained. For larger p T , the events are selected using either MBTS or L1 calorimeter based triggers (see Section 2) with a minimum transverse energy threshold at the electromagnetic scale [34] that varies between 5 GeV (L1 5) and 55 GeV (L1 55), depending on when the data were collected and the p T range considered (see Table 1).
The events are required to have one and only one reconstructed primary vertex with a z-position within 10 cm of the origin of the coordinate system, which suppresses pile-up contributions from multiple proton-proton interactions in the same bunch crossing, beam-related backgrounds and cosmic rays. In this analysis, events are required to have at least one jet with corrected transverse momentum p T > 30 GeV and rapidity |y| < 2.8. This corresponds approximately to the kinematic region, in the absolute four momentum transfer squared Q 2 -Bjorken-x plane, of 10 3 GeV 2 < Q 2 < 4 × 10 5 GeV 2 and 6 × 10 −4 < x < 2 × 10 −2 . Additional quality criteria are applied to ensure that jets are not produced by noisy calorimeter cells, and to avoid problematic detector regions.
VI. JET SHAPE DEFINITION
The internal structure of the jet is studied in terms of the differential and integrated jet shapes, as reconstructed using the uncorrected energy clusters in the calorimeter associated with the jet. The differential jet shape ρ(r) as a function of the distance r = ∆y 2 + ∆φ 2 to the jet axis is defined as the average fraction of the jet p T that lies inside an annulus of inner radius r − ∆r/2 and outer radius r + ∆r/2 around the jet axis: where p T (r 1 , r 2 ) denotes the summed p T of the clusters in the annulus between radius r 1 and r 2 , N jet is the number of jets, and R = 0.6 and ∆r = 0.1 are used. The points from the differential jet shape at different r values are correlated since, by definition, R 0 ρ(r) ∆r = 1. Alternatively, the integrated jet shape Ψ(r) is defined as the average fraction of the jet p T that lies inside a cone of radius r concentric with the jet cone: where, by definition, Ψ(r = R) = 1, and the points at different r values are correlated. The same definitions apply to simulated calorimeter clusters and final-state particles in the MC generated events to define differential and integrated jet shapes at the calorimeter and particle levels, respectively. The jet shape measurements are performed in different regions of jet p T and |y|, and a minimum of 100 jets in data are required in each region to limit the statistical fluctuations on the measured values.
VII. CORRECTION FOR DETECTOR EFFECTS
The measured differential and integrated jet shapes, as determined by using calorimeter topological clusters, are corrected for detector effects back to the particle level. This is done using MC simulated events and a bin-by-bin correction procedure that also accounts for the efficiency of the selection criteria and of the jet reconstruction in the calorimeter. PYTHIA-Perugia2010 provides a reasonable description of the measured jet shapes in all regions of jet p T and |y|, and is therefore used to compute the correction factors. Here, the method is described in detail for the differential case. A similar procedure is employed to correct independently the integrated measurements. The correction factors U (r, p T , |y|) are computed separately in each jet p T and |y| region. They are defined as the ratio between the jet shapes at the particle level ρ(r) par mc , obtained using particle-level jets in the kinematic range under consideration, and the reconstructed jet shapes at the calorimeter level ρ(r) cal mc , after the selection criteria are applied and using calorimeter-level jets in the given p T and |y| range. The correction factors U (r, p T , |y|) = ρ(r) par mc /ρ(r) cal mc present a moderate p T and |y| dependence and vary between 0.95 and 1.1 as r increases. For the integrated jet shapes, the correction factors differ from unity by less than 5%. The corrected jet shape measurements in each p T and |y| region are computed by multiplying bin-by-bin the measured uncorrected jet shapes in data by the corresponding correction factors.
VIII. SYSTEMATIC UNCERTAINTIES
A detailed study of systematic uncertainties on the measured differential and integrated jet shapes has been performed. The impact on the differential measurements is described here in detail.
• The absolute energy scale of the individual clusters belonging to the jet is varied in the data according to studies using isolated tracks [5], which parametrize the uncertainty on the calorimeter cluster energy as a function of p T and η of the cluster. This introduces a systematic uncertainty on the measured differential jet shapes that varies between 3% to 15% as r increases and constitutes the dominant systematic uncertainty in this analysis.
• The systematic uncertainty on the measured jet shapes arising from the details of the model used to simulate calorimeter showers in the MC events is studied. A different simulated sample is considered, where the FRITIOF [35] plus BERT showering model is employed instead of the QGSP plus BERT model. FRITOF+BERT provides the second best description of the test-beam results [30] after QGSP+BERT. This introduces an uncertainty on the measured differential jet shapes that varies between 1% to 4%, and is approximately independent of p T and |y|.
• The measured jet p T is varied by 2% to 8%, depending on p T and |y|, to account for the remaining uncertainty on the absolute jet energy scale [5], after removing contributions already accounted for and related to the energy of the single clusters and the calorimeter shower modeling, as discussed above. This introduces an uncertainty of about 3% to 5% in the measured differential jet shapes.
• The 14% uncertainty on the jet energy resolution [5] translates into a smaller than 2% effect on the measured differential jet shapes.
• The correction factors are recomputed using HERWIG++, which implements different parton shower, fragmentation and UE models than PYTHIA, and compared to PYTHIA-Perugia2010. In addition, the correction factors are also computed using ALPGEN and PYTHIA-DW for p T < 110 GeV, where these MC samples provide a reasonable description of the uncorrected shapes in the data. The results from HERWIG++ encompass the variations obtained using all the above generators and are conservatively adopted in all p T and |y| ranges to compute systematic uncertainties on the differential jet shapes. These uncertainties increase between 2% and 10% with increasing r.
• An additional 1% uncertainty on the differential measurements is included to account for deviations from unity (non-closure) in the bin-by-bin correction procedure when applied to a statistically independent MC sample.
• No significant dependence on instantaneous luminosity is observed in the measured jet shapes, indicating that residual pile-up contributions are negligible after selecting events with only one reconstructed primary vertex.
• It was verified that the presence of small dead calorimeter regions in the data does not affect the measured jet shapes.
The different systematic uncertainties are added in quadrature to the statistical uncertainty to obtain the final result. The total uncertainty for differential jet shapes decreases with increasing p T and varies typically between 3% and 10% (10% and 20%) at r = 0.05 (r = 0.55). The total uncertainty is dominated by the systematic uncertainty, except at very large p T where the measurements are still statistically limited. In the case of the integrated measurements, the total systematic uncertainty varies between 10% and 2% (4% and 1%) at r = 0.1 (r = 0.3) as p T increases, and vanishes as r approaches the edge of the jet cone. Finally, the jet shape analysis is also performed using either tracks from the inner detector inside the jet cone, as reconstructed using topological clusters; or calorimeter towers of fixed size 0.1 × 0.1 (y − φ space) instead of topological clusters as input to the jet reconstruction algorithm. For the former, the measurements are limited to jets with |y| < 1.9, as dictated by the tracking coverage and the chosen size of the jet. After the data are corrected back to particle level, the results from these alternative analyses are consistent with the nominal results, with maximum deviations in the differential measurements of about 2% (5%) at r=0.05 (r=0.55), well within the quoted systematic uncertainties.
IX. RESULTS
The measurements presented in this article refer to differential and integrated jet shapes, ρ(r) and Ψ(r), corrected at the particle level and obtained for anti-k t jets with distance parameter R = 0.6 in the region |y| < 2.8 and 30 GeV < p T < 600 GeV. The measurements are presented in separate bins of p T and |y|. Tabulated values of the results are available in the Appendix and in Ref. [36].
Figures 1 to 3 show the measured differential jet shapes as a function of r in different p T ranges. The dominant peak at small r indicates that the majority of the jet momentum is concentrated close to the jet axis. At low p T , more than 80% of the transverse momentum is contained within a cone of radius r = 0.3 around the jet direction. This fraction increases up to 95% at very high p T , showing that jets become narrower as p T increases. This is also observed in Fig. 4, where the measured 1 − Ψ(0.3), the fraction of the jet transverse momentum outside a fixed radius r = 0.3, decreases as a function of p T .
The data are compared to predictions from HERWIG++, ALPGEN, PYTHIA-Perugia2010, and PYTHIA-MC09 in Fig. 1 to Fig. 4(a); and to predictions from PYTHIA-DW and PYTHIA-Perugia2010 with and without UE contributions in Fig. 4 (b). The jet shapes predicted by PYTHIA-Perugia2010 provide a reasonable description of the data, while HERWIG++ predicts broader jets than the data at low and very high p T . The PYTHIA-DW predictions are in between PYTHIA-Perugia2010 and HERWIG++ at low p T and produce jets which are slightly narrower at high p T . ALPGEN is similar to PYTHIA-Perugia2010 at low p T , but produces jets significantly narrower than the data at high p T . PYTHIA-MC09 tends to produce narrower jets than the data in the whole kinematic range under study. The latter may be attributed to an inadequate modeling of the soft gluon radiation and UE contributions in PYTHIA-MC09 samples, in agreement with previous observations of the particle flow activity in the final state [12]. Finally, Fig. 4 (b) shows that PYTHIA-Perugia2010 without UE contributions predicts jets much narrower than the data at low p T . This confirms the sensitivity of jet shape observables in the region p T < 160 GeV to a proper description of the UE activity in the final state.
The dependence on |y| is shown in Fig. 5, where the measured jet shapes are presented separately in five different jet rapidity regions and different p T bins, for jets with p T < 400 GeV. At high p T , the measured 1 − Ψ(0.3) shape presents a mild |y| dependence, indicating that the jets become slightly narrower in the forward regions. This tendency is observed also in the various MC samples. Similarly, Figs. 6 and 7 present the measured 1 − Ψ(0.3) as a function of p T in the different |y| regions compared to PYTHIA-Perugia2010 predictions. The result of χ 2 tests to the data in Fig. 7 with respect to the predictions from the different MC generators are reported in Table 7, for each of the five rapidity regions. Here the different sources of systematic uncertainty are considered independent and fully correlated across p T bins (see Appendix). As already discussed, PYTHIA-Perugia2010 provides the best overall description of the data, while PYTHIA-Perugia2010 without UE contributions and ALPGEN show the largest discrepancies.
Finally, and only for illustration, the typical shapes of quark-and gluon-initiated jets, as determined using events generated with PYTHIA-Perugia2010, are also shown in Figs. 6 and 7. For this purpose, MC events are selected with at least two particle-level jets with p T > 30 GeV and |y| < 2.8 in the final state. The two leading jets in this dijet sample are classified as quark-initiated or gluon-initiated jets by matching (in y − φ space) their direction with one of the outgoing partons from the QCD 2 → 2 hard process. At low p T the measured jet shapes are similar to those from gluon-initiated jets, as expected from the dominance of hard processes with gluons in the final state. At high p T , where the impact of the UE contributions becomes smaller (see Fig. 4(b)), the observed trend with p T in the data is mainly attributed to a changing quark-and gluon-jet mixture in the final state, convoluted with perturbative QCD effects related to the running of the strong coupling.
X. SUMMARY AND CONCLUSIONS
In summary, jet shapes have been measured in inclusive jet production in proton-proton collisions at √ s = 7 TeV using 3 pb −1 of data recorded by the ATLAS experiment at the LHC. Jets are reconstructed using the anti-k t algorithm with distance parameter R = 0.6 in the kinematic region 30 GeV < p T < 600 GeV and |y| < 2.8. The data are corrected for detector effects and compared to different leading-order matrix elements plus parton shower MC predictions. The measured jets become narrower as the jet transverse momentum and rapidity increase, although with a rather mild rapidity dependence. The data are reasonably well described by PYTHIA-Perugia2010. HERWIG++ predicts jets slightly broader than the data, whereas ALPGEN interfaced with HERWIG and JIMMY, PYTHIA-DW, and PYTHIA-MC09 all predict jets narrower than the data. Within QCD, the data show sensitivity to a variety of perturbative and non-perturbative effects. The results reported in this paper indicate the potential of jet shape measurements at the LHC to constrain the current phenomenological models for soft gluon radiation, UE activity, and non-perturbative fragmentation processes in the final state. Data for differential and integrated measurements are collected in Tables 2 to 6, which include a detailed description of the contributions from the different sources of systematic uncertainty, as discussed in Section 8.
A χ 2 test is performed to the data points in Tables 5 and 6 with respect to a given MC prediction, separately in each rapidity region. The systematic uncertainties are considered independent and fully correlated across p T bins, and the test is carried out according to the formula where d j is the measured data point j, mc j (s) is the corresponding MC prediction, ands denotes the vector of standard deviations, s i , for the different independent sources of systematic uncertainty. For each rapidity region considered, the sums above run over the total number of data points in p T and five independent sources of systematic uncertainty, and the χ 2 is minimized with respect tos. Correlations among systematic uncertainties are taken into account in mc j (s). The χ 2 results for the different MC predictions are collected in Table 7, and indicate that PYTHIA-Perugia2010 provides the overall best description of the data. ρ(r) (0 < |y| < 2.8) 30 GeV < p T < 40 GeV r ρ ± (stat.) ± (syst.) cluster e-scale shower model jet e-scale resolution correction non-closure 0. 05 III: The measured differential jet shape, ρ(r), as a function of r in different pT regions, for jets with |y| < 2.8 and 210 GeV < pT < 600 GeV (see Fig. 3). The contributions from the different sources of systematic uncertainty are listed separately. The measured integrated jet shape, 1 − Ψ(r = 0.3), as a function of pT , for jets with |y| < 2.8 and 30 GeV < pT < 600 GeV (see Fig. 4). The contributions from the different sources of systematic uncertainty are listed separately. .3), as a function of pT , for jets with 30 GeV < pT < 500 GeV in different jet rapidity regions (see Fig. 7). The contributions from the different sources of systematic uncertainty are listed separately. | 6,277.6 | 2010-12-30T00:00:00.000 | [
"Physics"
] |
Improving fault identification in smart transmission line using machine learning technique
ABSTRACT
INTRODUCTION
A flaw in electrical hardware is represented by a flaw in its electrical circuit, which causes the flow to diverge from its intended path.Mechanical dissatisfaction, wounds, unnecessary internal and external anxieties, and so forth are common causes of faults.The defect impedance is modest, but the deficiency fluxes are rather significant.During the defects, the force stream is diverted to the issue, and the gracefully to the adjoining zone is hampered.The voltages are unbalanced [1].The deficiency must be identified as soon as possible, which is why an internet of things (IoT) based device was created to speed up the process.It will identify the following four major faults and issue a trip signal for hand-off.Over current shortcoming, propensity issue, and sparking line broken, phase failure defects are the four flaws identified by the model.The electric power system is divided into several sections.One of these is the transmission framework, in which power is transmitted from producing stations and substations to customers via transmission lines.The two methodologies may encounter various types of faults, which is sometimes referred to as a "shortcoming".Flaw is essentially defined as a variety of irritating but unavoidable occurrences that can inadvertently upset the steady state of the force framework, which occurs when the framework's protection fizzles at any point.If ISSN: 2252-8792 Int J Appl Power Eng, Vol. 12, No. 4, December 2023: 359-366 360 not properly monitored, there are a variety of power transmission flaws that might result in power blackouts.Among the most significant are i) faults at the force age station, ii) damage to control transmission lines (tree falling on lines), iii) faults at the substations or parts of dissemination subsystem, and iv) lightening.i) Types of transmission line faults Defects in the force framework might be classified as shunt difficulties or arrangement flaws.Single line-to-ground (SLG) defects are the most well-known type of shunt flaw.When one conduit falls to the ground or comes into contact with the unbiased wire, this type of defect occurs.It could also be the result of trees collapsing in a strong wind.As shown in Figure 1, this type could be approached.
The line-to-line (LL) deficiency is the second most common type of shunt flaw.When two transmission lines are short-circuited, this is said to happen.If a large flying creature remains on one transmission line and contacts the other, or if a tree branch falls on the heads of two force transmission lines, As seen in Figure 2, this type could be conversed with.
The double line-to-ground (DLG) deficit in Figure 3 is the third type of shunt inadequacy.This could be the result of a tree falling on two electrical cables, or it could be the result of other factors.The fourth and most serious type of insufficiency is the reasonable three-stage as in Figure 4, which can occur when three electrical lines in diverse buildings come into contact.
PROPOSED SYSTEM
The electric force foundation is deeply weak in the current framework against a wide range of typical and noxious physical events that can have a negative impact on the matrix's overall display and security.Despite the fact that flaw marker innovation has provided a reliable approach to locate perpetual issues, the specialized group and watch groups still need to observe and examine the devices for extended periods of time to identify broken transmission lines.When a deficit in a transmission line occurs, it is usually hidden, unless it is a major problem.
However, over time, these slight flaws might cause harm to the transformer and even endanger human life.It could also cause a fire.In today's India, we don't have a system in place that will alert us when a flaw occurs on a regular basis.The issue is that because we don't have a genuine memory framework, we end up harming the concealed gear and wind up being a hazard to the people around us.Support or checking of the transmission lines is frequently done on a subsequent premise in order to maintain a strategic distance from such incidents to the maximum extent possible.As a result, there is a greater demand for workers.The reality is that the genuine expectation of this isn't met the same number of times numerous times line disappointment could be due to downpour or tree overturning that can't be predicted.For example, in the Western Ghats, transmission lines are generally routed through the backwoods, and in places like Chirapunjee, massive precipitation virtually brings everything to a halt.It's critical to appreciate the gravity and potential consequences of a line failure.
A global system for mobile (GSM) based transmission line defect detecting system to combat them was proposed [2].When the preset edge is crossed, the microcontroller promptly sends a message to the region lineman and the control station, indicating the precise shaft to post area.This helps us to comprehend a framework that is in use.The true goal of identifying deficit gradually and safeguarding the transformer as Int J Appl Power Eng ISSN: 2252-8792 Improving fault identification in smart transmission line using … (Radhika Venkutuswamy) 361 soon as possible has been found out.It's worth noting that transformers aren't cheap.On a regular basis, an 11 KV transformer costs $3000 USD.As a result, we're developing a realistic and quick-response framework to aid in the improvement of wellness.A transformer is an important piece of equipment for power transmission.Constant observation is an important step in extending the life of a transformer.Similarly, a transformer is a pricey asset, and its failure cannot be repaired as rapidly as other sections of the transmission medium.An approach utilizing the global system for mobile (GSM) communication and a microcontroller is presented to monitor the transformer's failure and communicate it to the intended work force as shown in Figure 5.The information and yield voltages (periods) of the transformer are examined using a comparator.In the case that a critical decrease is identified, the authorities are notified via IoT.An approach for detecting spillage and intensity robbery in transmission is proposed.The proposed method is implemented using an ESP8266 and a microcontroller.It is implemented as two subsystems, one for detecting power outages and the other for detecting burglaries.Estimating the voltage difference between power transmission and gathering detects force spilling or line disappointment.In the suggested framework, the burglary is discovered by calculating the total force consumed and comparing it to the total force sent.The electrical officials are advised to distinguish between the limits using IoT [3].
The failure in the electrical cable is identified and located using a remote sensor network.The force transmission line is partitioned in this paper by remote sensor organizations.It occasionally measures the energy distinction and any deviations, which are then communicated to the appropriate authorities.This proposed paradigm is especially useful for identifying imbalanced problems.For power disappointment detection, a counterfeit neural network is used [4].In any event, neuronal organization necessitates preparation, which may or may not be complete.The support vector machine (SVM) is used to prepare data.All studies concerned with force transmission disappointment have been exaggerated.Furthermore, the sensor does not handle the interference caused by nature in detecting the required data.Furthermore, the computerized reasoning technique used necessitates preparing data, which is difficult to obtain in real time, and the framework's yield is solely dependent on the preparation set.We propose a productive arrangement based on the internet of things (IoT) in this research, as well as several techniques of correspondence.
Currently, the electric force framework is extremely vulnerable to a wide range of natural and malicious physical events, which can have a negative impact on the network's overall appearance and dependability.Furthermore, there is an impending need to prepare the deep-seated transmission line foundation with a superior information correspondence organization, which underpins future operational requirements such as continuous checking and control, which are critical for brilliant lattice coordination.Many electric power transmission companies have traditionally relied on circuit markers to identify faulty transmission line portions.Regardless, pinpointing the precise location of these defects continues to be a challenge.Despite the fact that shortfall pointer innovation has provided a reliable approach to detect longterm flaws, specialized groups and watch groups need actually watch and study the gadgets for extended periods of time to identify defective portions of their transmission lines.
Remote sensor-based transmission line inspection addresses some of these concerns, such as continuous auxiliary awareness, faster deficiency confinement, precise flaw conclusion by recognizable proof and separation of electrical issues from mechanical flaws, cost reduction due to condition-based maintenance rather than intermittent support, and so on [5].These applications establish stringent requirements, such as the delivery of a massive amount of extremely reliable data in a short period of time.The success of these applications is dependent on the creation of a practical and dependable organization with a quick response time.The organization must be able to send sensitive data to and from the transmission matrix, such as the current state of the transmission line and control data.This investigation provides a cost-effective approach for planning a continuous data transmission organization.Sensors are positioned in various places of the force organization to monitor the status of the force framework.These sensors can produce a large amount of data by making fine-grained estimates of a variety of physical or electrical boundaries.A basic test to be attended to in order to construct an intelligent dazzling framework is delivering this info to the control community in a cost efficient and convenient manner [6].Because of the vast scope, vast landscape, unprecedented geography, and basic planning requirements, organization layout is an essential aspect of sensor-based transmission line inspection.Mechanical difficulties, cost savings due to condition-based assistance rather than routine maintenance, and so on.Sensor networks have been proposed for a number of applications, including mechanical state handling and dynamic transmission line rating.Sensors are deployed in various parts of the force network to continuously monitor the status of the force framework.
A switched-mode power supply (SMPS) gracefully produces a similar final output at a reduced cost and with more efficiency [7].The yield power is lighter and smaller for a given yield.This is due to the fact that if the frequency of activity is increased, transformer works can be relocated.The setup or field device consists of four major components: a transformer (CT and VT), a GPS modem, and an Arduino, ESP8266 is shown in Figure 5.The CT and VT primaries, which are associated with the line, sense the framework's ISSN: 2252-8792 Int J Appl Power Eng, Vol. 12, No. 4, December 2023: 359-366 362 current and voltage estimations and send the yield to the Arduino's ADC, which converts the sign to a computerized structure that can be produced by the Arduino's CPU [8].
The circuit serves as the setup's most important concern.It has a large number of programming codes stored in the electrically erasable and programmable read-only memory (EEPROM) that enable it to classify issue types based on voltage and current characteristics.The microcontroller considers these attributes in light of the program to see if they fall within the necessary range.When the voltage and current quality are out of range when compared to the reference, it indicates that there is a problem.The GPS also determines the issue separation in comparison to the device using an impedance-based computation, and then sends this information to the modem for transmission.
In short, the Arduino characterizes, computes the shortcoming separation, and transmits the data to the IoT via the sequential correspondence interface (SCI), which serves as an interface between the Arduino and the ESP8266 Wi-Fi.The ESP8266 serves as a bridge between the Arduino sequential correspondence port and the internet of things [9].In the transmission framework, the gadget is placed at the edge of the sectionalized districts, and the area of the shortfall is established in relation to the gadget's location.The IoT's one-of-a-kind nature is used as a location for the device, which might be saved as a Cloud website page in the control room.
HARDWARE IMPLEMENTATION OF PROPOSED SYSTEM
To realize the concept, an Arduino Mega, a transfer potentiometer, an ESP 8266 transformer, a signal, a temperature sensor, and an LCD is needed as shown in Figure 6.The venture is equipped with a fault detecting equipment that uses a large number of switches at each known location to double-check the precision of the equivalent.The voltage drop throughout the transfer is fed into an ADC, which generates precise advanced data that the modified microcontroller displays on a website page.A 16X2 LCD interfaced with the Arduino Mega microcontroller displays the shortfall at what stage and sends an SMS to the EB office.impedance-based and vigorously programmed issue recognition and area framework for overhead force transmission lines.To reduce the amount of time it takes to rectify a problem and protect expensive transformers from harm or robbery, which is common during extended power outages.To improve the efficiency of specialist teams because the time available to uncover flaws would be restricted.
Machine learning algorithm
Support vector machines (SVMs) are a class of supervised learning methods for classification, regression, and detection of outliers is explained below: i) Step 1: Sensors will initially detect variations in transmission line boundaries such as voltage, scope, longitude, and so on.ii) Step 2: The data is sent from the Arduino microcontroller.iii) Step 3: The microcontroller will send individual indications to the Wi-Fi module for transmission based on the detected data.iv) Step 4: Using the Wi-Fi module, the detected boundaries are sent to the administrator.
Hardware implementation
The hardware implementation of the proposed system is shown in Figure 7. Various approaches for dealing with transmission line disappointments and defects may be received.In this section, a few of them are depicted in depth.The control room will be able to see the precise location of the transmission line as a result of this.The proposed framework, the Arduino Mega is introduced with an inbuilt worker that is connected to the transformers whenever there is a problem, such as a short out, which has a yield in the high voltage and low current range, and an open circuit, which has a yield in the low voltage and high current range.The online page can surely be used to examine and control this.The issue of recognizing the deficit in the dispersion lines and the intended implication to the electricity board (EB) is dealt with by identifying transmission line flaws and reporting them to the EB.The task is in charge of the design and production of the power gracefully, Arduino Mega, circuit board, and ESP 8266, GPS modem.The Arduino's electric appropriation network connects to the electrical wires and delivers a message to the IoT about the type of defect in the line.This reduces the amount of time spent locating the problem.The ultimate goal is to continuously monitor the circulation line state and hence the flaws of the dispersion line due to limits such as overvoltage, undervoltage, SLG, and DLG deficits [10].In the case that one of these events occurs, a message will be sent from the planned controlling unit or substation.The goal is to locate the transmission line fault and to inform the worker about the problem [11].Sensors, such as a fire identifier, a spark indicator, and a multi-issue locator circuit, are used to detect the specific problem in the transmission lines.Sensors detect the transmission line's force properties [12].Sending data to the cloud is a good option.This work is limited to the development of a framework that will identify and locate line-to-line and line-to-ground defects in overhead transmission lines where a problem has occurred [13], [14].The internet of things (IoT) is a networked collection of physical objects that may alter the climate or their own borders, collect data, and transfer it to various devices [15].It is the third wave in the evolution of the internet.This innovation will provide instant access to information about the physical environment and the items that make it up, allowing for creative administrations as well as an increase in proficiency and profitability.Recent occurrences, brilliant sensors, communication advancements, and internet conventions have all aided the IoT [16].A web of things (IoT) organization is envisioned in this venture.The future fate of IoT and its outcomes are given a lot of thought.The identified location at Unnamed Road, Eachanari, Coimbatore, Tamil Nadu, India, is shown in Figure 8.
RESULTS AND DISCUSSION
To evaluate the performance of the suggested system using the accuracy, delay, location identification, and fault probability graphs [17].Performance analysis was done based on the occurrence of fault by applying machine learning and without machine learning is shown [18], [19] in Figure 9. Occurrence of faults was reduced on applying machine learning algorithms [20]- [22].Figure 10 potraits the identification of location of fault that helps the electric workers was improved by applying machine learning algorithms.The proposed system incorporates a machine learning technique that improves accuracy by over 90% while lowering delay time is shown in Figure 11 [23]- [25].
CONCLUSION
Using all existing fault indicator technologies and machine learning techniques, create an automatic and effective fault identification and position system for all overhead power transmission network networks.The transmission line between the EB office and the information, as well as the location latitude and longitude, has been opened.The machine learning method aids in improving accuracy and speed.The new method employs a machine learning approach to boost accuracy by more than 90% while reducing delay time.
Figure 5 .
Figure 5. Working of proposed system
Figure 6 .
Figure 6.Hardware interfacing of proposed system
Figure 7 .
Figure 7. Hardware implementation of the proposed system
Figure 9 .Figure 11 .
Figure 9. Performance analysis based on probability fault | 4,047.4 | 2023-12-01T00:00:00.000 | [
"Computer Science"
] |
Ultrafast electron microscopy for probing magnetic dynamics
The spatial features of ultrafast changes in magnetic textures carry detailed information on microscopic couplings and energy transport mechanisms. Electrons excel in imaging such picosecond or shorter processes at nanometer length scales. We review the range of physical interactions that produce ultrafast magnetic contrast with electrons, and specifically highlight the recent emergence of ultrafast Lorentz transmission electron microscopy. From the fundamental processes involved in demagnetization at extremely short timescales to skyrmion-based devices, we show that ultrafast electron imaging will be a vital tool in solving pressing problems in magnetism and magnetic materials where nanoscale inhomogeneity, microscopic field measurement, non-equilibrium behavior or dynamics are involved.
Introduction
Magnetic materials underlie a wide range of technologies, from data storage 1 and power generation 2 to charged particle optics; 3 in fundamental science, magnetism results from spin or orbital correlations in materials, and is often linked to other ordering phenomena 4 such as superconductivity, 5 ferroelectricity, 6 and charge density waves. 7 Understanding magnetism is crucial for both development of new technological applications and disentangling couplings in strongly correlated materials. Both static and dynamic properties govern the functionality of magnetic materials; this functionality typically relies on a rapid switching or transport of the spin state. Though challenging, a wealth of insight is possible through the direct observation of light-or current-driven magnetization dynamics.
For example, tracking angular momentum flow [8][9][10][11][12] in ultrafast demagnetization is crucial for both understanding the process and for speeding it up; demagnetization at interfaces 13 may show interesting spatiotemporal dynamics. 14 Memory devices based on skyrmions-topologically stable, localized quasiparticles-involve movement of nanometer-scale magnetic textures on ultrafast timescales. [15][16][17][18][19][20][21] As ever-smaller skyrmions are produced, [22][23][24][25][26][27][28][29][30] it will be increasingly necessary to image their dynamic behavior. It may be possible to better understand the atomic-scale motion of spin and charge involved in coherent ultrafast magnetism. 31-33 Highly nonuniform dynamics could be induced by a localized change in magnetic properties, such as a laser-driven transient modification of magnetic anisotropy. 34 These examples illustrate that understanding transient magnetic properties in space and time necessitates high-temporal-resolution imaging. Several probes of magnetism have been developed for this purpose.
Photons-in the visible, 35,36 extreme ultraviolet, 37,38 or x-ray 39-41 spectral range-have been employed to image magnetic dynamics for many years. Magnetization-dependent differences in absorption or reflectivity produce image contrast, and when the timing is varied between pulses of photons used for excitation, called the pump, and pulses used for imaging, called the probe, high temporal resolution is possible. 42 Ultrashort optical pulses can initiate sudden modifications on the magnetic order (see Figure 1), such as demagnetization, 36 domain dilation, [43][44][45] and ferromagnetic-to-paramagnetic phase transitions, 46 all observed using optical pump-probe techniques. These sudden changes can result in magnon generation, 47 all-optical magnetic switching, 48 and spin currents, 49 and can be applied toward technological applications such as terahertz emitters. 50 However, ultrafast photon-based nanoscale imaging has resolution limits. For example, spatial resolution in magnetooptical Kerr effect (MOKE) microscopy is typically restricted to the micrometer scale without the aid of a near-field probe. 51 X-ray magnetic dichroism (XMCD), responsible for many discoveries of nanoscale magnetic dynamics, 15,43,[52][53][54][55][56][57][58][59] can achieve a spatial resolution on the order of ten nanometers, employing imaging beamlines either at synchrotrons with X-ray pulses of tens of picoseconds or longer, 60 or with femtosecond resolution at free-electron lasers 14 and using highharmonic sources. 61 In contrast, XMCD photoemission electron microscopy takes advantage of the spatial resolution of electron imaging. 54,[62][63][64][65] Electrons are a powerful alternative probe of magnetism on the nanoscale. Magnetic image contrast using electrons arises through both inelastic and elastic scattering. With low incident electron energy, scanning electron microscopy with polarization analysis (SEMPA) [66][67][68][69][70][71] probes the local surface magnetization by sampling the spin polarization of emitted secondary electrons, and spin-polarized low-energy electron microscopy (SPLEEM) offers magnetic contrast in reflection, 72,73 because the inelastic and elastic electron mean free paths depend on the relative orientation of electron spin and local surface magnetization. Higher incident-energy electrons can excite atomic transitions sensitive to material electron spin polarization, analogous to XMCD; this technique, called electron magnetic chiral dichroism (EMCD), employs crystalline diffraction to select outgoing electron orbital angular momentum states corresponding to different transitions. 74 EMCD offers atomic-resolution magnetic contrast, 75-77 but the high dose necessary for a good signal-to-noise ratio may restrict the usability of this technique for short electron pulses. Magnetic dichroism with electron orbital angular momentum post-selection 78-82 relies on similar mechanisms but has the potential for an adequate signal-to-noise ratio even with femtosecond pulses and kHz repetition rates. Spin-polarized transmission electron microscopy 83,84 may prove capable of both inelastic and elastic magnetic imaging.
Lorentz transmission electron microscopy
The Lorentz force, an elastic interaction, can be employed for magnetic imaging at a wide range of energies. 85 Lorentz electron microscopy describes modes where the immersive objective lens of a transmission electron microscope is turned off to minimize the external magnetic field applied to a magnetic specimen, and imaging contrast is based on the Lorentz force, as shown in Figure 2. Three contrast modes are most commonly used for magnetic imaging in transmission electron microscopy: (1) Fresnel contrast, 86,87 which uses image defocus to transform phase gradients introduced by the Lorentz force onto incident plane wave-like electrons into intensity in an image (Figure 2b and d); (2) differential phase contrast, 88 which uses a focused beam at the specimen and measures the deflection in momentum space due to the Lorentz force with a quadrant or pixelated detector (Figure 2c and e); and (3) holography, 89,90 which uses a biprism to interfere the plane wave passed through the specimen with an off-axis reference wave passed through vacuum. Fresnel contrast is most frequently used for its simplicity, whereas differential phase contrast is more quantitative and increasingly applied as fast pixelated detectors become more common. Holography is the most quantitative of the three contrast modes, but also requires a high transverse coherence, 91 so it is challenging but possible with pulsed electrons. 92,93 With a pulsed electron source 94 through either photoemission 92,95-100 or a beam chopper, [100][101][102][103][104][105] and synchronized current, field, or laser excitation of the specimen, sub-picosecond temporal resolution has become possible in stroboscopic transmission electron microscopy. [106][107][108] This approach has many applications, but some of the first experiments imaged magnetic dynamics with Lorentz microscopy. 101,109 Moreover, using continuous-beam imaging and in situ short-pulsed excitation of a sample, insights into the ultrafast generation and stabilization of quenched magnetic states are possible. [110][111][112][113][114]
Imaging optical pulse-induced magnetic textures
Standard Lorentz microscopy is an excellent tool to investigate magnetic textures induced by femtosecond optical excitation. Imaging the response of magnetic textures to external stimuli has a long tradition. Commonly, variations in temperature 115,116 and temperature gradients, 117 magnetic fields [118][119][120] or electric currents 121,122 are applied on timescales for which the spin system remains close to local thermal equilibrium. In contrast, in situ femtosecond optical excitation results in a far-from-equilibrium spin state and enables the creation of novel metastable magnetic textures, which can be characterized by continuous-beam Lorentz microscopy, as shown in Figure 3.
As a first example, Eggebrecht et al. recently demonstrated the optically induced generation of a dense vortex-antivortex network in a polycrystalline iron thin film deposited on a silicon nitride membrane. 112 Before excitation, the system shows a magnetic ripple texture, induced by the local fluctuations of the magnetic anisotropy direction (see Figure 3a, upper panel). Applying a single intense ultrashort optical pulse raises the temperature of the iron thin film above the Curie temperature. In the center of the optical focal spot, the iron thin film is transiently demagnetized, whereas the underlying silicon nitride membrane remains cold. On a 100 ps timescale, the iron film thermally equilibrates with the membrane and a local ferromagnetic ordering is re-established, but the fast cooling rate (larger than 10 12 K/s) prevents the system from reaching a homogeneous ordering. Instead, nanoscale ferromagnetic domains are formed, which, at their mutual boundaries, are linked by topological vortex and anti-vortex defects. The "frozen" magnetic state, as shown in Figure 3a (lower panel), is stable at room temperature but can be reset to the initial ripple texture by either applying an external magnetic field or lowfluence optical pulses.
Similar results were obtained for optically excited permalloy (Ni 80 Fe 20 high permeability magnetic alloy) microstructures, in which vortex switching and smallscale vortex-antivortex networks can be induced. 113 In this case, the probability of the final magnetic configuration was shown to depend on the strength of the optical excitation and the shape and size of the magnetic structure.
While out-of-plane magnetized samples are generally less suited for Lorentz microscopy, local topological defects such as domain walls or skyrmions can still provide good image contrast with nanoscale resolution. 119 In a recent study, 114 the optical generation of a metastable skyrmion texture in FeGe thin films was probed by in situ Lorentz microscopy. In equilibrium, FeGe shows a series of different magnetic phases depending on sample temperature and applied magnetic field. A Lorentz micrograph of a mixed phase with helical and conical textures stable at low temperatures and intermediate external magnetic fields is shown in Figure 3b (upper panel). Upon laser excitation starting from this phase, the sample temperature rapidly increases, reaching the stability regime of an adjacent skyrmion phase. Fast quenching results in a metastable skyrmion lattice, as shown in Figure 3b (lower panel). Such all-optical writing approaches could be a fundamental ingredient for novel skyrmion-based information storage and transport applications. The dynamical evolution upon femtosecond optical excitation in a related sample system was further elucidated using time-resolved x-ray scattering, demonstrating ultrafast switching and the existence of a transient topological fluctuation state. 123
Ultrafast Lorentz transmission electron microscopy
With stroboscopic electron illumination, ultrafast Lorentz transmission electron microscopy can be applied to image transient changes in magnetic textures and fast dynamics. for spatiotemporal reversibility of the triggered magnetic dynamics, often necessitating a tailored sample design.
Ultrafast Lorentz microscopy has been employed to study several types of dynamics, including domain wall motion under an oscillating magnetic field 109 or laserinduced thermal gradients, 110,124 laser-induced demagnetization, 100,124-126 and current-driven vortex oscillations. 127 In two recent examples, Möller et al. investigated the response of a magnetic vortex to an applied highfrequency current. 127,128 The interplay between the current-induced force on the vortex, consisting both of Oersted-field and spintransfer torque contributions, and the confinement potential within a magnetic nanostructure results in resonant gyration dynamics of the vortex. Driving the vortex gyration by applying 101.5 MHz radiofrequency currents across a 2.1 µm permalloy square, Möller et al. were able to track the vortex position with a 2 nm precision, 127 as shown in Figure 4, and observed the frequency dependence of the response in real space. 128 Prior work had mapped the time-averaged motion using continuous-beam Lorentz microscopy. 121 The temporal resolution available in ultrafast Lorentz microscopy now captures the precise phase of the gyration, and can measure the dissipation rate and transient ring-down behavior of the oscillation after the driving current is rapidly switched off.
As a further example, several groups have recently investigated ultrafast demagnetization, shown in Figure 5. Ultrafast demagnetization was first observed through optical Kerr spectroscopy in 1996, 36 and its underlying microscopic physical mechanisms are still being debated. 31, [129][130][131][132][133] Ultrafast Lorentz microscopy has the potential to a b 1 µm 250 nm provide further experimental insights due to its very high spatial resolution.
Several recent experiments studied demagnetization behavior with ultrafast Lorentz microscopy with picosecondscale 100,124-126 temporal resolution. Rubiano da Silva et al. recently mapped the partial demagnetization of vortices in nanoscale disks of permalloy with 100 nm spatial resolution and 700 fs temporal resolution, 125 shown in Figure 5a-c. Cao et al. reported on demagnetization of permalloy with a grating pattern caused by interference in the exciting laser, resulting in a coherently precessing magnetic grating, 126 shown in Figure 5d. Zhang et al. observed a peculiar sequence of subpicosecond demagnetization, recovery over 3 ps, and decay to a paramagnetic state over 12 ps of the spiral spin texture in Mn-Ni-Ga. 100 Despite this methodical progress, ultrafast Lorentz microscopy still needs to overcome a number of experimental challenges. Since Fresnel contrast is linked to the applied imaging defocus, reaching a sufficient signal-to-noise ratio often requires rather large defoci in the 100 µm range or above, associated with a significant loss in spatial resolution. Therefore, further progress in high-coherence, high-brightness ultrafast electron sources is necessary to reach a spatial resolution on the order of few nanometers.
Conclusion and outlook
As the experiments presented here show, behavior of magnetic materials at picosecond or shorter timescales may strongly deviate from equilibrium. This far-from-equilibrium behavior is observable with Lorentz electron microscopy both via in situ excitation of ultrafast processes that lead to long-lived changes in magnetic textures, and with stroboscopic excitation and imaging of picosecond or faster dynamics. The combination of a nanometer spatial and femtosecond temporal resolution makes ultrafast electron microscopy an excellent tool for imaging magnetic dynamics. Electron microscopes can measure crystal structure, charge states, electric fields and bonding down to the atomic scale, and strain, temperature, phononic and plasmonic band structures on the nanometer scale. The capability to correlate these properties with sub-picosecond movies of magnetic dynamics will enable identification of nanoscale inhomogeneities invisible in bulk measurements, quantification of local fields used to excite dynamics, and direct measurement of coupled behavior like magnetostriction.
Specifically, this tool is wellsuited to help answer a number of open questions in magnetism related to microscopic magnetic energy landscapes, 127,128 angular momentum transfer processes, 134 ultrafast demagnetization at interfaces, skyrmion dynamics, and coherent ultrafast magnetism. The growth of other ultrafast electron techniques, including time-resolved SEMPA 67-71 and potentially SPLEEM, opens the door to complementary surface-sensitive imaging of dynamics. New manipulation techniques, such as irradiation by a high-charge pulse of electrons, 135 may grow in use. There are numerous opportunities for ultrafast electron imaging to unravel key problems in magnetism.
Acknowledgments
We appreciate helpful discussions with Felix Büttner, and thank Markus Münzenberg and Henning Ulrichs for their contributions to some of the works presented here. We thank the UTEM team for work on maintaining the microscope. We gratefully acknowledge funding by the Deutsche Forschungsgemeinschaft (DFG-SPP-1840 "Quantum Dynamics in Tai acknowledges support of a Lichtenberg professorship funded by the VolkswagenStiftung.
Funding
Open Access funding enabled and organized by Projekt DEAL.
Conflict of interest
On behalf of all authors, the corresponding author states that there is no conflict of interest.
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 licence, 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:// creat iveco mmons. org/ licen ses/ by/4. 0/. | 3,571.4 | 2021-08-01T00:00:00.000 | [
"Physics"
] |
Advances in Databases and Information Systems, Moscow 1996
In our days numerous object analysis and design methods appeared. These methods provide users with easy to understand graphical notations for expressing a wide variety of concepts central to the presentation of software requirements. While such techniques are recognized as useful tools, the graphical notations used with these methods are often ambiguous, resulting in diagrams that are easily misinterpreted. The disadvantage of these methods also is a weak support of reusability. On the other hand, the Synthesis method being developed [3, 1] is focusedon the interoperable information system design basing on the reuse the pre-existing information resources (databases,legacy systems, program packages,etc.). The Synthesis method is a top-down, bottom-up iterative process of analysis, design and development. To take an advantage of the existing Object Analysis and Design OAD methods we embed the Synthesis method into one of such methods so that the phases of requirement planning and domain analysis could be developed using the chosen method technique. After that the Synthesis method continues a design with the reuse of pre-existing resources. Therefore a problem of interfacing of an OAD method and the Synthesis method arises. In this paper we present the mapping of the graphical notation of a specific OAD method (OMT) to the Synthesis entities. In order to prevent ambiguities in the graphical notation in OMT we impose some restrictions on OMT using. Also we augment OMT with the ontological specifications needed to resolve contextual differences between application and pre-existing information resources as well as with predicative specifications of object behaviors.
Introduction
The results reported in the paper were obtained in frame of the Synthesis project1 that is focused on various problems of Heterogeneous Interoperable Information Resource Environment (HIRE) specification, design and management.Synthesis mainly attempts semantic interoperation issues intended for the specification and design of interoperable information systems [4].
In the project the Synthesis method for the interoperable information system design is being developed [1].The method is based on reuse of pre-existing information resources (databases, legacy systems, program packages, etc.).The method emphasizes the design of semantically interoperable compositions of the pre-existing information resources in HIRE.Some specific architecture supporting HIRE (such as [8]) is assumed.Resources should be semantically coherent and their compositions should be consistent and meaningful within the application context.The Synthesis method focuses on the semantic interoperation reasoning process that should lead to concretization of specifications of requirements by views over the pre-existing information resources [4].
The Synthesis method is a top-down, bottom-up iterative process of analysis, design and development.The schema sketching the method is shown on the figure 1.
In the method we focus on the design phase and for requirement planning and domain analysis phases we use conventional techniques of object analysis and design.The architectural framework for Synthesis is designed so that the Synthesis method itself should be neutral to existing (OAD) methods [7].We take OMT as one of the most powerful and popular OAD methods.
The results of these initial phases (OMT object diagrams) are interpreted as the Synthesis specifications of the application requirements.
To work with reuse effectively, OAD methods should provide facilities to search for suitable components and check their conformance to requirements.For these purposes we augment OMT with the ontological specifications.
Synthesis language uses a symbolic notation having precise semantics that is required to reason that certain resource is reusable for a given specification.Mapping of a model developed by a conventional OAD method into a Synthesis model gives to this model precise semantics.Thus Synthesis language can play a role of a metamodel [10] with respect to a conventional OAD method.
The paper is organized as follows.Section 2 provides an introduction to OMT and object model notation.Section 3 gives a short summary of the main Synthesis concepts.Section 4 presents the mapping of OMT object model notation to the Synthesis entities.Section 5 provides an augmentation of OMT with the ontological specifications.
The OMT Methodology
The Object Modeling Techniques [9] provides a comprehensive coverage of the object modeling.It includes the analysis modeling, design modeling, and implementation modeling.The purpose of analysis is to model the real world so that it can be understood.The system design stage concentrates exclusively on the overall architecture of the system.The object design stage concentrates on optimizing and refining the object model, ready for translation into a programming language.Three models of the system are developed initially and then refined in all these stages: the object model, the dynamic model, and the functional model.The object model represents the static, structural, "data" aspects of a system.The dynamic model represents the temporal, behavioral, "control" aspects of a system.The functional model represents the transformational, "function" aspects of a system.
The Object Model Notation
The Object Model concepts include objects, classes, attributes, operations, generalization relationship, instantiation relationship, associations, aggregation relationships, qualifiers, abstract classes, class attributes and operations, link attributes and operations, links as classes and modules.
Figure 2 shows the basic OMT object model notation.
Links and associations establish relationships among objects and classes.A link connects two or more objects.An association describes a group of links with common structure and common semantics.Multiplicity specifies how many instances of one class may relate to each instance of another class.Associations are inherently bi-directional.A role is a direction across an association.
The method allows to define attributes of a link.The link attribute is a named data value held by each link in an association.Sometimes it is useful to model an association as class.Each link becomes one instance of the class.
The method allows to define a qualified association that relates two classes and a qualifier.A qualifier is a special attribute that reduces the effective multiplicity of an association.The qualifier distinguishes among the set of objects at the many end of an association.
Aggregation is the "part-whole" or "a-part-of" relationship in which objects representing the components of something are associated with an object representing the entire assembly.It is a tightly coupled form of association with A class attribute describes a value common to an entire class of objects.A class operation is an operation on the class itself.In method they described as attributes and operations with prefix $.
A module is a logical construct for grouping classes, associations, and generalizations.Modules enable to partition an object model into manageable pieces.The same class may be referenced in different modules.In fact, referencing the same class in multiple modules is the mechanism for binding modules together.
The Basic Synthesis Entities
The Synthesis [5] entities include objects, classes, types, generalization/specialization type hierarchies, subclass hierarchies, classification hierarchies made by metatype relationship, specification inheritance, attributes, functions defined by predicative specifications as object calculus definitions, assertions related to values and their attributes, association metaclasses, modules and ontological specifications.
Abstract data types constitute the basis for the type system of the language providing for construction of arbitrary data types.Synthesis supports the multilevel type system that sets a classification relationship on the data types.A type in the model is a value represented by an object.A type as the value may be produced as the result of type expression evaluation used mainly for type inferencing.A class supports a set of objects of a given type that constitutes an extent of the class.Two types are associated with a class.One defines an interface of the class and other defines an interface of object instances of the class.With metaclasses type that defines an interface of the instances of instances of the metaclass is additionally associated.
Object attributes are treated also as classes of association objects establishing a correspondence between a set of objects in an association domain and a set of objects (values) in an association range.Thus a specification of an attribute is considered as a specification of an association class.Treating of an attribute as an association class provides for introduction of association metaclasses setting a classification relationship on attribute classes and establishing properties of association classes and/or their instances.
The Synthesis entities are represented by frames.Figure 3 shows as example the specifications of classes and types.
Mapping the OMT Object Model Notions into the Synthesis Entities
Classes Generally a class of OMT is mapped into a Synthesis class, its own type and a type describing an interface of instances of the class, except the cases which will be described later.In such cases we treat OMT classes as having persistent extensions.
For OMT classes that according to the application requirements should not provide such extensions we introduced special tags as specific descriptions in OMT notation.In such cases we map classes of OMT to types in Synthesis.
Interfacing of Object Analysis and Design Methods with the Method for Interoperable Information Systems Design
Figure 4 shows an example of the mapping for class proposal.Attributes and Operations Each operation should be defined by predicative specifications using object calculus definitions.
Attributes and operations of an OMT class are mapped to the attributes and functions of a type describing the interface of instances of the class.Figure 4 shows an example of the mapping for attributes name, budget and operation accept proposal of class proposal.
Associations We restrict ourselves only with binary associations.Each association may not have a name, but should have the role names (at least one).
An association is mapped to the attribute (or attributes, if two role names are defined).The attribute becomes an instance of an association metaclass reflecting semantics of the OMT association.The attribute name is the name of the role of the association.The attribute type is a type corresponding to opposite class for one-to-one association or a template set defined on a type corresponding to opposite class for one-to-many association.Many-to-many association is considered as two one-to-many associations.Instantiation relationship Instantiation relationship is mapped as the classification relationship in Synthesis and described by the slot in in a class and type specifications.
Each OMT class having as instances other classes should be considered as metaclass in Synthesis.Such class is mapped to the metaclass in Synthesis.Figure 7 shows an example of the mapping the instantiation relationship between classes meta proposal and proposal.
Figure 9 shows an example of the mapping the association as class between budget association and budget sem class.
Synthesis Ontological Specifications
The specifications resulting from the requirement planning and domain analysis phases should be complete for the semantic interoperation reasoning.For this purpose we augment OMT with the ontological specifications needed to resolve contextual differences between application and pre-existing information resources.
An ontological specification is a set of definitions of context-specific knowledge representation primitives for describing of context vocabulary (names of individuals, functions, predicates, attributes, types, classes) in a form that is both human and machine readable.Ontology contains also a set of rules (axioms) associated with the vocabulary or with particular terms [6].Ontologies provide a conceptual framework for talking about an application domain and an implementation framework for problem solving.Ontologies play the role of a coupling interfaces among shared resources providing the basis for them to interoperate.
In Synthesis ontological specifications usually are organized in a separate ontological module coupled with the information resource specification module.The ontological module consists of a collection of name definitions and concept definitions.
A name definition may contain verbal definitions (human and machine readable definitions of concepts in the natural language) and interconcept relationships (such as hypernym/hyponym, positive, negative).Name definitions are defined as objects of a class OntDef.
Concept definitions syntactically look as class (metaclass) definitions containing a metainformation associated with the concept (e.g., attributes, set of rules).
Figure 10 shows the Synthesis example of information resource specification module and ontological module for the class proposal with attribute budget and association metaclass budget sem.
Representation of Ontological Specifications in the OMT Graphical Notation
We augment OMT with ontological analysis.The results of the analysis are represented in the OMT ontological modules having the following form.Name definitions are represented as objects of OMT class OntDef.The name of each such object is the name of the corresponding concept given with prefix "O".The hypernym/hyponym, positive and negative relationships between concepts are represented by links between objects with corresponding names.Relationships between a class and associated types (defining an own interface of the class and an interface of instances of the class) are represented in OMT by associations between class and types (named class section and inst section correspondingly).For metaclasses the inst inst section association is added.Figure 11 shows an example of representation of ontological specification for class proposal and attribute budget and of association metaclass budget sem in the OMT object model notation.
Conclusion
In this paper we define the interfacing of the OAD method (OMT) to the Synthesis method intended for a transition from conventional OAD method to the interoperable information system design.In the Synthesis method we focus on the design with reuse and apply OMT to cover the requirement planning and domain analysis phases.The mapping of the OMT object model notation into the Synthesis entities is given.
To achieve the required completeness of the specifications for the semantic interoperation reasoning we augment the OMT with the ontological specifications as well as with the predicative specifications of functions.
Figure 3 :
Figure 3: An example of the Synthesis specification
Figure 4 :
Figure 4: An example of mapping of the OMT class
Figure 5 Figure 5 :
Figure 5 shows an example of the mapping the association relationship between classes proposal and person, proposal and researcher.
Figure 6
Figure 6 shows an example of the mapping the generalization relationship between classes person and researcher.
Figure 9 :
Figure 9: An example of mapping of the OMT associations as classes
Figure 10 :
Figure 10: An example of the Synthesis ontological specification
Figure 11 :
Figure 11: An example of representation of ontological specifications in OMT | 3,154.2 | 1996-09-01T00:00:00.000 | [
"Computer Science"
] |
Lyophilization for Formulation Optimization of Drug-Loaded Thermoresponsive Polyelectrolyte Complex Nanogels from Functionalized Hyaluronic Acid
The lyophilization of nanogels is practical not only for their long-term conservation but also for adjusting their concentration and dispersant type during reconstitution for different applications. However, lyophilization strategies must be adapted to each kind of nanoformulation in order to minimize aggregation after reconstitution. In this work, the effects of formulation aspects (i.e., charge ratio, polymer concentration, thermoresponsive grafts, polycation type, cryoprotectant type, and concentration) on particle integrity after lyophilization and reconstitution for different types of polyelectrolyte complex nanogels (PEC-NGs) from hyaluronic acid (HA) were investigated. The main objective was to find the best approach for freeze-drying thermoresponsive PEC-NGs from Jeffamine-M-2005-functionalized HA, which has recently been developed as a potential platform for drug delivery. It was found that freeze-drying PEC-NG suspensions prepared at a relatively low polymer concentration of 0.2 g.L−1 with 0.2% (m/v) trehalose as a cryoprotectant allow the homogeneous redispersion of PEC-NGs when concentrated at 1 g.L−1 upon reconstitution in PBS without important aggregation (i.e., average particle size remaining under 350 nm), which could be applied to concentrate curcumin (CUR)-loaded PEC-NGs for optimizing CUR content. The thermoresponsive release of CUR from such concentrated PEC-NGs was also reverified, which showed a minor effect of freeze-drying on the drug release profile.
Introduction
Nanogels (NGs) from polyelectrolyte complexes (PECs) of hyaluronic acid (HA) have been extensively studied in recent years as biomedical platforms with high potential that have the advantages of nanomedicine, green chemistry, and the inherent properties of HA [1]. Meanwhile, thermoresponsive nanocarriers from thermosensitive polymers exhibiting a lower critical solution temperature (LCST), i.e., increased hydrophobicity upon temperature increase, have also received great attention over the last two decades for drug delivery applications owing to their interesting behaviors, notably thermo-triggered drug release which can be controlled spatiotemporally [2,3]. Combining the two above-mentioned concepts, our previous work developed thermoresponsive polyelectrolyte complex nanogels (PEC-NGs) from the complexation between diethylaminoethyl dextran (DEAE-D) or poly-L-lysine (PLL) and hyaluronic acid grafted with Jeffamine M-2005 (M2005), i.e., thermoresponsive poly(ethylene oxide)-co-poly(propylene oxide) side-chains ( Figure 1) [4]. Such PEC-NGs showed promising properties as nanocarriers for drug delivery, namely simple preparation, hyaluronidase-responsive degradation, thermo-controllable encapsulation of curcumin (CUR) as a model hydrophobic drug, and, especially, high stability in physiological saline media (PBS 1X at pH = 7.4) when including PLL as the constituent polycation [4]. In the above-mentioned context, lyophilization seems to be the most appropriate approach for optimizing both the aspects of conservation and formulation of the PEC-NGs. As reported in our previous work, PEC-NGs from HA-M2005 showed that particle size gradually increased over one month of storage in aqueous suspensions, probably due to the hydrophobic aggregation caused by M2005 chains on the particle's surface [4]. Furthermore, the thermosensitivity of PEC-NGs in aqueous media, which can cause particle shrinkage or reswelling upon heating or cooling, respectively, would probably destabilize these PEC-NGs upon temperature variation during their transportation. In addition, aqueous media may favor microbiological contamination, which can cause HA degradation due to bacterial enzymes [6]. Therefore, water removal by freeze-drying has become a classical approach for better long-term conservation of colloidal systems [7]. Drying can also reduce product weight and volume to facilitate their storage and transportation. Moreover, the dried products after lyophilization may be dispersible within a freely adjustable volume of aqueous media to reach the desired concentration before their use [8]. The dispersant type can also be modified in this process, e.g., saline buffers of different physiological pH values for intended in vitro bioassays or in vivo applications.
Despite such practicality, lyophilization usually brings about irreversible particle aggregation due to different types of stress during the freezing phase, including ice crystal formation, cryoconcentration, and dehydration at the microstructure level [9,10]. The risk of destabilization during the drastic temperature decrease upon freezing may be even higher in the case of thermosensitive systems. For this reason, nanoparticles are usually freeze-dried in the presence of cryoprotectants, i.e., chemical agents which can stabilize In the above-mentioned context, lyophilization seems to be the most appropriate approach for optimizing both the aspects of conservation and formulation of the PEC-NGs. As reported in our previous work, PEC-NGs from HA-M2005 showed that particle size gradually increased over one month of storage in aqueous suspensions, probably due to the hydrophobic aggregation caused by M2005 chains on the particle's surface [4]. Furthermore, the thermosensitivity of PEC-NGs in aqueous media, which can cause particle shrinkage or reswelling upon heating or cooling, respectively, would probably destabilize these PEC-NGs upon temperature variation during their transportation. In addition, aqueous media may favor microbiological contamination, which can cause HA degradation due to bacterial enzymes [6]. Therefore, water removal by freeze-drying has become a classical approach for better long-term conservation of colloidal systems [7]. Drying can also reduce product weight and volume to facilitate their storage and transportation. Moreover, the dried products after lyophilization may be dispersible within a freely adjustable volume of aqueous media to reach the desired concentration before their use [8]. The dispersant type can also be modified in this process, e.g., saline buffers of different physiological pH values for intended in vitro bioassays or in vivo applications.
Despite such practicality, lyophilization usually brings about irreversible particle aggregation due to different types of stress during the freezing phase, including ice crystal formation, cryoconcentration, and dehydration at the microstructure level [9,10]. The risk of destabilization during the drastic temperature decrease upon freezing may be even higher in the case of thermosensitive systems. For this reason, nanoparticles are usually freeze-dried in the presence of cryoprotectants, i.e., chemical agents which can stabilize the microstructure of the products during freezing [7]. Despite the importance of these aspects, few works have thoroughly described the lyophilization of HA-based PEC-NGs. The most relevant work of this type, to the best of our knowledge, was carried out by Umerska et al., who studied the effects of PEC composition and cryoprotectants on the particle integrity of PEC-NGs from chitosan (CTS) or protamine (PROT) in complex with HA or chondroitin sulfate after freeze-drying and reconstitution, for which poly(ethylene glycol) (PEG) and/or trehalose (TRE) were chosen as cryoprotectants [11]. In another work, Gheran et al. freeze-dried NGs from HA and CTS with glucose being selected as a cryoprotectant [12]. It can be seen that the lyophilization method should be adapted to each kind of nanoformulation to ensure particle stability but has not always been thoroughly explained and justified. In this context, the present work aims at understanding the impacts of formulation factors on particle integrity after the freeze-drying and reconstitution of PEC-NGs from HA or HA-M2005 in complex with DEAE-D or PLL, focusing on finding the best lyophilization strategy for HA-M2005/PLL PEC-NGs. The thermo-triggered drug release in vitro from the thermoresponsive PEC-NGs was also verified and reevaluated after lyophilization in order to evaluate the impact of this treatment on drug release.
Preparation of Blank PEC-NGs
PEC-NGs of predetermined charge ratios n−/n+ between HA or HA-M2005 as polyanions and DEAE-D or PLL as polycations were prepared as described in our previous work [4]. Briefly, polyanion and polycation solutions at defined polymer concentration (C P = 0.1, 0.2, or 0.5 g.L −1 ) were prepared and filtered through regenerated cellulose 0.45 µm membrane filter units (Sartorius, Göttingen, Germany). To obtain PEC-NG suspensions at a predetermined C P , the polyanion solutions were rapidly dropped (one-shot) into the polycation solutions at the same C P , followed by mixing under mild stirring (200 rpm) at room temperature for 30 min. All samples were prepared, at least, in triplicate.
Curcumin Loading and Quantification in PEC-NGs
CUR-loaded HA-M2005/PLL PEC-NGs were prepared with the same protocol as our previous work [4]: an amount of 5 mg.mL −1 CUR stock solution in acetone was added to the HA-M2005 solution (C P = 0.5 or 0.2 g.L −1 ), preheated at 50 • C, followed by rapid addition of PLL solution of the same C P and temperature. The CUR to total polyelectrolyte mass ratio was 2:5. The mixture was then mildly stirred (200 rpm) at 50 • C for 30 min and, thereafter, left at room temperature for over 12 h for acetone evaporation. In order to remove precipitated CUR in excess and obtain homogenous PEC-NG suspensions, the mixture was centrifuged at 3000× g for 15 min and the supernatant was filtered through polyvinylidene difluoride (PVDF) 0.45 µm membrane filter units (Millipore, Burlington, MA, USA). All samples were prepared, at least, in triplicate and kept away from light to avoid CUR degradation.
The CUR encapsulated in PEC-NGs was quantified as described in our previous work [4]. Briefly, CUR-loaded PEC-NG suspensions were vortexed with HCl 1 N solution (PEC-NG suspension: HCl 1 N = 10:1 v/v), then with ethanol (aqueous phase: ethanol = 1:5 v/v). The mixture was then centrifuged at 3000× g for 15 min. The absorbance at λ = 420 nm of the supernatant was measured using a Cary 100 spectrophotometer (Agilent Technology, Santa Clara, CA, USA) to determine the CUR concentration in the supernatant and then the CUR concentration in the starting PEC-NG suspensions (C CUR/SUS ). CUR loading capacity (LC) and encapsulation efficiency (EE) were estimated using Equations (1) and (2), with C CUR/WATER being CUR solubility in Milli-Q water (0.3 µg.mL −1 ) [4], C P being polymer concentration, and C CUR0 being the concentration of CUR initially added with respect to the aqueous phase.
Particle Characterizations
The mean hydrodynamic diameter (D h ) and polydispersity index (PDI) of PEC-NGs were determined by dynamic light scattering (DLS) technique in back-scatter mode using Zetasizer Ultra (Malvern Panalytical, Worcester, UK) for homogenous PEC-NG suspension samples without further dilution. Zeta potential (ζ) was measured by electrophoretic light scattering (ELS) technique with the same device. The temperature for all measurements was 25 • C with 120 s of temperature stabilization. All samples were measured, at least, in triplicate.
Particle morphology was verified by transmission electron microscopy (TEM). Specimens were prepared by placing a 10 µL droplet of PEC-NG suspension on a formvarcarbon-coated copper grid for 15 min. The droplet was then gently wicked away using a piece of filter paper. A 10 µL droplet of phosphotungstic acid 1% solution (contrast enhancer) was, thereafter, placed on the grid and removed after 30 s. The specimens were then air-dried and observed under an FEI Tecnai 12 BioTwin electron microscope (Phillips, The Netherlands).
Freeze-Drying and Reconstitution of PEC-NGs
In the case of lyophilization with cryoprotectants, the cryoprotectants of interest (sucrose, glucose, trehalose, or PEGs) were dissolved at determined concentrations in PEC-NG suspensions by vortex mixing in 30 s. The PEC-NG suspensions, with or without cryoprotectants, were frozen overnight at −20 • C in a freezer. The samples were, thereafter, freeze-dried using an Alpha 1-2 LD Plus freeze-drier (Martin Christ, Germany) in primary drying mode for 24 h with pressure under 0.1 mbar, and the condenser temperature was −60 • C. To reconstitute PEC-NG suspensions, unless otherwise indicated, the freeze-dried PEC-NGs were dispersed by vortex mixing for 60 s in the same volume of Milli-Q water as before freeze-drying. Only visually homogeneous reconstituted suspensions were then subjected to DLS and/or ELS for particle characterization. For certain studies, different volumes of PBS 1X buffer (pH 7.4) were applied as dispersants during the reconstitution step according to the experiment's purpose.
Curcumin Release In Vitro
CUR release from PEC-NGs was studied using the dialysis method adapted from the literature [13,14]. In brief, 3 mL of PEC-NG suspension was placed in a dialysis tube (D-Tube Dialyzer Maxi, MWCO 12-14 kDa, Millipore, Germany) and then immersed in 21 mL of PBS 1X (pH 7.4), with 0.5% (m/v) Tween 20 to ensure sink conditions and 0.5 mmol.L −1 ascorbic acid to avoid CUR oxidation. The dialysis was realized under mild stirring (100 rpm) at either 37 or 42 • C to investigate the impact of temperature. In order to quantify the cumulative fraction of CUR released, 700 µL of external media was taken out for CUR quantification and replaced by the same volume of fresh media at predetermined time points. CUR quantification was realized by measuring the absorbance at λ = 420 nm with a Cary 100 spectrophotometer (Agilent Technology, Santa Clara, MA, USA) and a standard curve of CUR concentration in the range of 0.01-1 µg.L −1 in the release medium (PBS 1X at pH = 7.4 with 0.5% m/v Tween 20 and 0.5 mmol.L −1 ascorbic acid). The experiments were repeated on, at least, three batches of PEC-NGs.
Results and Discussion
In the present work, the physicochemical characteristics of PEC-NGs after vs. before the lyophilization and reconstitution process were compared to evaluate the efficiency in preserving their colloidal structures. Firstly, HA/DEAE-D PEC-NGs were studied to highlight the effects of basic parameters, such as the polyanion/polycation ratio in PECs (represented by the molar charge ratio n−/n+) and cryoprotectant type as well as cryoprotectant concentration. The optimal conditions observed were then reinvestigated on different PEC-NGs with PLL replacing DEAE-D and/or the thermoresponsive HA-M2005 replacing HA since HA-M2005/PLL constituted the most potent system for drug delivery from our previous work [4]. This revealed the effects of polycation type and the presence of M2005 grafts on the stability of PEC-NGs after freeze-drying. At this point, the impact of C P before freeze-drying was studied to determine its optimal value for the possibility of concentrating CUR-loaded PEC-NGs after lyophilization. Finally, CUR release kinetics in vitro before and after lyophilization was also evaluated.
Effect of Charge Ratio
Lyophilization was realized for the suspensions of HA/DEAE-D PEC-NGs having different n−/n+ ratios (1.25, 2.5, and 5) at C P = 0.5 g.L −1 without cryoprotectant. These n−/n+ values higher than 1 were chosen for this study since they were previously reported to result in stable PEC-NGs containing mostly HA, which is our main polymer of interest [4,15]. However, after lyophilization, only PEC-NGs with n−/n+ = 5 can be redispersed in Milli-Q water to form homogeneously translucent suspensions, while visible large aggregates remained after redispersing the freeze-dried PEC-NGs having n−/n+ of 1.25 or 2.5. The experiment was then reproduced on these two types of PEC-NG in the presence of trehalose as a cryoprotectant at different concentrations, increasing with a 1% (m/v) interval until obtaining a visually homogeneous reconstitution. The characteristics of rehydrated PEC-NGs at the optimal trehalose concentrations as well as their change factor R (ratio of values after reconstitution to those before freeze-drying) are reported in Table 1. Trehalose was preliminarily chosen as the cryoprotectant since it is generally the most effective among cryoprotectants according to the literature [11,16,17]. Further studies regarding cryoprotectant selection are described in Section 3.2. Table 1. Characteristics of HA/DEAE-D PEC-NGs (C P = 0.5 g.L −1 ) reconstituted after freeze-drying at the lowest trehalose concentrations allowing homogeneous reconstitution (n ≥ 3). It can be seen from Table 1 that PEC-NGs with a higher n−/n+ would require less trehalose during lyophilization for homogeneous redispersion in water during the reconstitution, which was confirmed by particle sizes remaining in the colloidal range with relatively low PDIs. This can be explained by the better electrostatic stabilization of these PEC-NGs since an n−/n+ further from 1 would lead to greater surface net charges, namely, zeta potential changed from −30 to −45 mV when n−/n+ is increased from 1.25 to 5 as reported in our earlier work [15]. Indeed, during the freezing of PEC-NG suspensions, the continuous formation of pure ice crystals should lead to a gradual reduction in the volume of the liquid phase, thus, a local increase in particle and counter-ion concentrations, also known as cryoconcentration [10]. This will result in different stresses on PEC-NGs: (i) me-chanical damage or deformation of PEC-NGs caused by growing ice crystals, (ii) structure interference caused by the loss of hydrogen bonds as water molecules are transferred to the ice phase and then sublimated, and (iii) closer approach of particles and weakened electrostatic stabilization by the charge screening effect of counter-ions which facilitate particle aggregation [10,18]. Logically, a higher absolute zeta potential at the beginning would allow a greater electrostatic repulsion to avoid such aggregation and also facilitate disaggregation during the reconstitution, which was also observed by Umerska et al. [11] and Eliyahu et al. [19]. Furthermore, the higher proportion of HA on the NG surface may also contribute to their better stability during freezing since HA can also have cryoprotective effects [1], e.g., by maintaining intra-and inter-molecular hydrogen bonds on the particle surface during dehydration [20]. However, after freeze-drying and reconstitution, we still observed an increase of nearly 40% in the particle size of HA/DEAE-D PEC-NGs at n−/n+ = 2.5 (Table 1), which was also the most interesting n−/n+ for CUR encapsulation reported in our earlier studies [4]. Therefore, different cryoprotectant types and concentrations were further tested, particularly on PEC-NGs at this n−/n+ in an attempt to better preserve their particle size (Section 3.2).
Effects of Cryoprotectant Type and Concentration
HA/DEAE-D PEC-NGs at n−/n+ = 2.5 and C P = 0.5 g.L −1 were freeze-dried with 1% m/v of different cryoprotectants, i.e., trehalose, sucrose, glucose, and PEGs with M n of 2000; 10,000 or 20,000 g.mol −1 , which have been largely studied as cryoprotectants for nanoformulations in the literature [21,22]. In our case, the addition of PEG caused visible aggregation of PEC-NGs after reconstitution (PEG of M n = 20,000 g.mol −1 ) or even before freeze-drying (PEGs of M n of 2000 or 10,000 g.mol −1 ). On the contrary, trehalose, sucrose, and glucose did not cause visible aggregation or changes in the particle size before lyophilization ( Table 2). The resulting lyophilized cakes ( Figure 2A) could be redispersed to establish visually homogeneous suspensions, which were then characterized by DLS to evaluate the change factors of particle size and zeta potential after lyophilization and reconstitution ( Figure 2B). Regarding the visual aspects of freeze-dried cakes (Figure 2A), using disaccharide cryoprotectants (i.e., sucrose and trehalose) also resulted in porous cakes as the one obtained without cryoprotectant, while using glucose led to a collapsed cake. After reconstitution, although the PDI and zeta potential were well-maintained, a significant increase in particle size was always observed with all three cryoprotectants ( Figure 2B). However, it can still be seen that trehalose was the most effective among these cryoprotectants in preserving the particle size of HA/DEAE-D PEC-NGs while glucose was the least effective, which is in accordance with the visual aspect of lyophilized cakes earlier. Then, by increasing the trehalose concentration, we saw that the particle size can be well-preserved (change factor in the range of 1-1.1) with at least 2% m/v trehalose ( Figure 2C). To explain the cryoprotective mechanisms of these cryoprotectants, three theories have been described in the literature: vitrification, water replacement, and water entrapment or preferential exclusion ( Figure 3) [23,24]. was the least effective, which is in accordance with the visual aspect of lyophilized cakes earlier. Then, by increasing the trehalose concentration, we saw that the particle size can be well-preserved (change factor in the range of 1-1.1) with at least 2% m/v trehalose ( Figure 2C). To explain the cryoprotective mechanisms of these cryoprotectants, three theories have been described in the literature: vitrification, water replacement, and water entrapment or preferential exclusion ( Figure 3) [23,24]. was the least effective, which is in accordance with the visual aspect of lyophilized cakes earlier. Then, by increasing the trehalose concentration, we saw that the particle size can be well-preserved (change factor in the range of 1-1.1) with at least 2% m/v trehalose ( Figure 2C). To explain the cryoprotective mechanisms of these cryoprotectants, three theories have been described in the literature: vitrification, water replacement, and water entrapment or preferential exclusion (Figure 3) [23,24]. In the vitrification theory, during cryoconcentration, the highly increased concentration of cryoprotectants can render the liquid phase more viscous and then form a glassy layer around the NGs when water is completely removed. This can reduce the mobility of particles in the liquid phase to avoid their close approach, thus preventing their aggregation, and eventually constitutes a solid glassy matrix supporting the whole system and stabilizing the NGs in a dry state [7]. In this theory, the glass transition temperature (T g ) of cryoprotectants is an important parameter since a higher T g would lead to a higher stability of the glassy matrix to better preserve the products sequestered inside. Disaccharides are, thus, more preferable cryoprotectants compared with monosaccharides since the former normally have a higher T g [10]. Concerning the three saccharide cryoprotectants used in this work, the order of their efficacy in preserving NG particle size as shown in Figure 2B is consistent with their T g found in the literature: T g of glucose (30 • C) < T g of sucrose (58 • C) < T g of trehalose (108 • C) [25]. The low T g of glucose can also lead to a low T g ' (apparent T g of concentrated cryoprotectants with NGs and other solutes), which may explain the collapse of the lyophilized cake in this case [10]. Although cake collapse is only an aesthetic defect of lyophilized products, which usually has no significant impact on their physicochemical stability, it can, however, lead to a longer time needed for the drying as well as the redispersion steps due to low porosity [26,27]. In the water replacement mechanism, saccharide molecules can replace water to maintain the hydrogen bonds on NG surfaces, thus keeping the NGs in a "pseudo-hydrated" state to avoid interference with their structure upon dehydration as well as reduce particle aggregation through hydrophobic interactions [28,29]. Meanwhile, the water entrapment or preferential exclusion theory has been less widely described and is usually applied for protein cryoprotection [24]. According to this theory, sugar molecules like trehalose can self-organize to "entrap" water molecules, thus reducing the hydration rate of the protein molecules and favoring their folding in a more compact and rigid state, which is more stable and can protect them from the damage caused by ice crystals [30]. This may also be the case for the cryoprotection of our NGs. In theories of water replacement and water entrapment, in order to explain the different effects between sucrose and trehalose despite their similar chemical formula, some studies have suggested that the geometrical structure of sucrose, which is different from that of trehalose, may allow for intramolecular hydrogen bonds that limit its capacity to form hydrogen bonds with water as well as other macromolecules and, therefore, results in lower cryoprotection efficiency [31,32].
Regarding PEG, despite its cryoprotectant effect reported for some systems, its presence might also favor agglomeration or aggregation of nanoformulations [22]. Indeed, it has been known that PEG is incompatible with dextran since the presence of both these polymers in the solution can cause phase separation, known as aqueous two-phase systems [10,33], which might explain the formation of large aggregates observed upon adding PEG in HA/DEAE-D PEC-NG suspensions. PEG has also been reported to act as a crowding agent to intensify the complexation between HA and PLL and, hence, favor phase separation [34]. Furthermore, PEG may also bridge nanoparticles to cause flocculation, known as polymer bridging flocculation [7,35]. In the case of PEG having M n of 20,000 g.mol −1 , the resulting high viscosity may also be a factor preventing homogeneous redispersion of lyophilized PEC-NGs.
From the above results, it can be generally seen that PEG is less effective than the three sugars tested for the cryoprotection of PEC-NGs, while trehalose seems to be the most appropriate cryoprotectant. The noticeable cryoprotective effect of trehalose has also been reported for other HA-based systems, such as HA/CTS nanoparticles [22], cisplatin-conjugated HA nanoparticles [36], or micellar NGs from HA [37]. In regards to cryoprotectant concentration, according to the cryoprotecting mechanisms, cryoprotection efficiency should be optimal only at a sufficiently high cryoprotectant concentration, which is around 2% (m/v) trehalose for HA/DEAE-D PEC-NGs in our case ( Figure 2C). From these results, trehalose at concentrations (C TRE ) between 0-2% m/v was selected for later studies.
Effects of M2005 and Polycation Type
HA-M2005/DEAE-D and HA-M2005/PLL PEC-NGs (C P = 0.5 g.L −1 , n−/n+ = 2.5) with C TRE of 0, 1 or 2% m/v were freeze-dried and reconstituted. During their reconstitution, like HA/DEAE-D PEC-NGs in the previous sections, homogeneous redispersion could not be obtained without trehalose but only with C TRE of at least 1%, thus allowing for characterization by DLS (Figure 4).
Effects of M2005 and Polycation Type
HA-M2005/DEAE-D and HA-M2005/PLL PEC-NGs (CP = 0.5 g.L −1 , n−/n+ = 2.5) with CTRE of 0, 1 or 2% m/v were freeze-dried and reconstituted. During their reconstitution, like HA/DEAE-D PEC-NGs in the previous sections, homogeneous redispersion could not be obtained without trehalose but only with CTRE of at least 1%, thus allowing for characterization by DLS (Figure 4). . The results of freshly prepared PEC-NGs (before L-R) were retrieved from our previous work as references [4].
For HA-M2005/DEAE-D PEC-NGs, CTRE = 1% was sufficient to keep their particle size unchanged ( Figure 4A), which means a more efficient preservation compared with HA/DEAE-D PEC-NGs in Section 3.2, which required at least 2% trehalose. Such a more positive result for HA-M2005/DEAE-D PEC-NGs is probably due to the stabilizing effect of the M2005 grafts through their hydrophobic interactions and LCST behaviors [4]. However, for HA-M2005/PLL PEC-NGs, even a trehalose concentration as high as 2% could not prevent a considerable increase (of around 50%) in particle size ( Figure 4B), which may be attributed to the highly hydrophobic nature of PLL in HA-M2005/PLL PECs [4] which facilitates their aggregation under freezing stress. We, therefore, tried a higher CTRE (i.e., 4% m/v) but such a high trehalose concentration seemed to destabilize the PEC-NGs, as this led to the visible aggregation of HA-M2005/DEAE-D PEC-NGs and a significant increase in the particle size of HA-M2005/PLL PEC-NGs (from 262 ± 24 to 380 ± 29 nm). It has been described in the literature that exceeding the optimal cryoprotectant concentration can sometimes destabilize nanoformulations, whose mechanisms have not always been well clarified [7,38,39]. With M2005 grafts, which constitute the amphiphilic and thermoresponsive properties of PEC-NGs, the aforementioned water entrapment effect of trehalose may be comparable to the salting-out effect of counterions, which overfavors hydrophobic interactions and, thus, aggregation. Moreover, in the case of HA-M2005/PLL PEC-NGs at CP = 0.5 g.L −1 and n−/n+ = 2.5, also the most interesting system among different PEC-NGs studied in our previous work [4], CTRE = 1-2% seems to be the most suitable concentration for their freeze-drying, although the efficiency in preserving particle size was not completely optimal. In addition, with those reconstituted HA-M2005/PLL PEC-NGs, we reverified the NG shrinkage upon heating at 50 °C ( Figure S1), which is typical for their thermoresponsive behaviors [4]. This confirmed that the thermoresponsiveness of HA-M2005-based PEC-NGs still remained after lyophilization. . The results of freshly prepared PEC-NGs (before L-R) were retrieved from our previous work as references [4].
For HA-M2005/DEAE-D PEC-NGs, C TRE = 1% was sufficient to keep their particle size unchanged ( Figure 4A), which means a more efficient preservation compared with HA/DEAE-D PEC-NGs in Section 3.2, which required at least 2% trehalose. Such a more positive result for HA-M2005/DEAE-D PEC-NGs is probably due to the stabilizing effect of the M2005 grafts through their hydrophobic interactions and LCST behaviors [4]. However, for HA-M2005/PLL PEC-NGs, even a trehalose concentration as high as 2% could not prevent a considerable increase (of around 50%) in particle size ( Figure 4B), which may be attributed to the highly hydrophobic nature of PLL in HA-M2005/PLL PECs [4] which facilitates their aggregation under freezing stress. We, therefore, tried a higher C TRE (i.e., 4% m/v) but such a high trehalose concentration seemed to destabilize the PEC-NGs, as this led to the visible aggregation of HA-M2005/DEAE-D PEC-NGs and a significant increase in the particle size of HA-M2005/PLL PEC-NGs (from 262 ± 24 to 380 ± 29 nm). It has been described in the literature that exceeding the optimal cryoprotectant concentration can sometimes destabilize nanoformulations, whose mechanisms have not always been well clarified [7,38,39]. With M2005 grafts, which constitute the amphiphilic and thermoresponsive properties of PEC-NGs, the aforementioned water entrapment effect of trehalose may be comparable to the salting-out effect of counterions, which over-favors hydrophobic interactions and, thus, aggregation. Moreover, in the case of HA-M2005/PLL PEC-NGs at C P = 0.5 g.L −1 and n−/n+ = 2.5, also the most interesting system among different PEC-NGs studied in our previous work [4], C TRE = 1-2% seems to be the most suitable concentration for their freeze-drying, although the efficiency in preserving particle size was not completely optimal. In addition, with those reconstituted HA-M2005/PLL PEC-NGs, we reverified the NG shrinkage upon heating at 50 • C ( Figure S1), which is typical for their thermoresponsive behaviors [4]. This confirmed that the thermoresponsiveness of HA-M2005-based PEC-NGs still remained after lyophilization.
Reduction of Polymer Concentration before Freeze-Drying
In an attempt to further reduce the particle size (i.e., to under 350 nm with PDI under 0.3) of HA-M2005/PLL PEC-NGs at n−/n+ = 2.5 after freeze-drying and reconstitution, PEC-NGs were prepared at a lower polymer concentration (i.e., C P of 0.2 g.L −1 ). This was realized upon the presumption that such a lower C P would result in a smaller particle size before lyophilization and also after reconstitution, which may allow for further concentrating PEC-NGs without causing large aggregates. This time, we also tried using PBS as a dispersion medium during PEC-NG reconstitution for their future applications, such as biological studies in vitro or in vivo. As a first attempt, the PBS volume added for reconstitution was equal to 40% of the suspension volume before lyophilization, thus, a final C P of 0.5 g.L −1 . In this case, PEC-NGs could be homogeneously redispersed even without trehalose but still showed a relatively large particle size (around 400 nm), which means a 2.5-fold increase with respect to the initial size of PEC-NGs at C P = 0.2 g.L −1 before freezedrying ( Figure 5A). Otherwise, the presence of trehalose at 1 or 2% m/v before lyophilization (i.e., C TRE of 2.5 or 5% after reconstitution) resulted in less increased particle size (around 250 nm) ( Figure 5A) with good stability for, at least, 7 days ( Figure S2). Therefore, in comparison with PEC-NGs freeze-dried and reconstituted at C P = 0.5 g.L −1 in Section 3.3, PEC-NGs reconstituted after lyophilizing more diluted suspensions (C P = 0.2 g.L −1 ) could show a lower final particle size despite the same final C P during reconstitution.
realized upon the presumption that such a lower CP would result in a smaller particle size before lyophilization and also after reconstitution, which may allow for further concentrating PEC-NGs without causing large aggregates. This time, we also tried using PBS as a dispersion medium during PEC-NG reconstitution for their future applications, such as biological studies in vitro or in vivo. As a first attempt, the PBS volume added for reconstitution was equal to 40% of the suspension volume before lyophilization, thus, a final CP of 0.5 g.L −1 . In this case, PEC-NGs could be homogeneously redispersed even without trehalose but still showed a relatively large particle size (around 400 nm), which means a 2.5-fold increase with respect to the initial size of PEC-NGs at CP = 0.2 g.L −1 before freeze-drying ( Figure 5A). Otherwise, the presence of trehalose at 1 or 2% m/v before lyophilization (i.e., CTRE of 2.5 or 5% after reconstitution) resulted in less increased particle size (around 250 nm) ( Figure 5A) with good stability for, at least, 7 days ( Figure S2). Therefore, in comparison with PEC-NGs freeze-dried and reconstituted at CP = 0.5 g.L −1 in Section 3.3, PEC-NGs reconstituted after lyophilizing more diluted suspensions (CP = 0.2 g.L −1 ) could show a lower final particle size despite the same final CP during reconstitution. After such an observation, we tried further concentrating PEC-NGs by reducing PBS volume during reconstitution to 20% of the initial volume before freeze-drying to reach a final C P of 1 g.L −1 , in the presence of trehalose with final C TRE of 1 or 2.5% m/v. Such a reduction in reconstitution volume seems to slightly increase the final particle size of PEC-NGs (i.e., to around 300 nm) ( Figure 5A), which remained acceptable. The morphologies of PEC-NGs reconstituted at C P = 1 g.L −1 were also verified by TEM ( Figure 5B). It was also noticed that PDI was always relatively stable around 0.2, indicating good uniformity of particle size regardless of the studied C P and C TRE . We also tried further reducing the initial C P to 0.1 g.L −1 but this did not lead to a considerable difference regarding particle size and PDI, which were still around 300 nm and 0.2, respectively ( Figure 5C). In general, these results show that the approach of lyophilization of diluted PEC-NG suspensions can be interesting for concentrating PEC-NGs and, probably, the encapsulated drugs.
Curcumin Encapsulation and Release
In order to verify the possibility of concentrating encapsulated drugs in PEC-NGs after lyophilization, CUR-loaded HA-M2005/PLL PEC-NGs (n−/n+ = 2.5) were prepared at C P of 0.1 or 0.2 g.L −1 , then freeze-dried with C TRE of 0.1 or 0.2% m/v, respectively, and reconstituted with PBS to reach a final C P of 1 g.L −1 and C TRE of 1% as for blank PEC-NGs in the previous section. As can be seen from Figure 6A, compared with suspensions of the same PEC-NG type freshly prepared at C P = 0.5 g.L −1 , the suspension reconstituted at C P = 1 g.L −1 from initial C P = 0.2 g.L −1 was also visually homogeneous but with a darker yellow color, suggesting qualitatively an effective increase in CUR concentration without causing precipitation. After CUR quantification, the results obtained with the suspensions both before lyophilization and after reconstitution from initial C P = 0.2 g.L −1 were also consistent with those obtained from PEC-NGs freshly prepared at C P = 0.5 g.L −1 ( Figure 6B), since LC values were in the same range and the CUR content was, thus, always in good proportion with C P . However, PEC-NGs prepared at C P = 0.1 g.L −1 showed a much lower LC and, therefore, resulted in suspensions with a lower CUR concentration after reconstitution even at the same final C P of 1 g.L −1 ( Figure 6B). The low loading capacity of PEC-NGs at C P = 0.1 g.L −1 may stem from their more porous and looser structure compared with those prepared at C P = 0.2 g.L −1 , suggested by their similar particle sizes ( Figure 5A, C) despite the much lower C P in the former case. From such results, freeze-drying suspensions at C P = 0.2 g.L −1 with 0.2% trehalose can be considered the most appropriate method for lyophilizing HA-M2005/PLL PEC-NGs of n−/n+ = 2.5, which can preserve satisfactorily both their particle size and CUR content. To verify the thermo-triggered release of CUR from HA-M2005/PLL PEC-NGs (n−/n+ = 2.5) as well as the change in such behavior after lyophilization, the release profiles of CUR at 37 °C and 42 °C, i.e., maximum temperature for in vivo local hyperthermia [3,40], were characterized for PEC-NGs freshly prepared at CP = 0.5 g.L −1 and those reconstituted at CP = 1 g.L −1 in PBS after lyophilization at CP = 0.2 g.L −1 with 0.2% trehalose. The cumulative CUR release over time is presented in Figure 7. . Samples at C P = 0.5 g.L −1 were retrieved from our previous work [4] as references.
To verify the thermo-triggered release of CUR from HA-M2005/PLL PEC-NGs (n−/n+ = 2.5) as well as the change in such behavior after lyophilization, the release profiles of CUR at 37 • C and 42 • C, i.e., maximum temperature for in vivo local hyperthermia [3,40], were characterized for PEC-NGs freshly prepared at C P = 0.5 g.L −1 and those reconstituted at C P = 1 g.L −1 in PBS after lyophilization at C P = 0.2 g.L −1 with 0.2% trehalose. The cumulative CUR release over time is presented in Figure 7.
To verify the thermo-triggered release of CUR from HA-M2005/PLL PEC-NGs (n−/n+ = 2.5) as well as the change in such behavior after lyophilization, the release profiles of CUR at 37 °C and 42 °C, i.e., maximum temperature for in vivo local hyperthermia [3,40], were characterized for PEC-NGs freshly prepared at CP = 0.5 g.L −1 and those reconstituted at CP = 1 g.L −1 in PBS after lyophilization at CP = 0.2 g.L −1 with 0.2% trehalose. The cumulative CUR release over time is presented in Figure 7. In the case of PEC-NGs at CP = 0.5 g.L −1 without lyophilization, the release kinetics at 37 °C was relatively fast at the beginning and gradually slower over time, resulting in 90% In the case of PEC-NGs at C P = 0.5 g.L −1 without lyophilization, the release kinetics at 37 • C was relatively fast at the beginning and gradually slower over time, resulting in 90% of the encapsulated CUR being released after 4 days. As these PEC-NGs have a core-shell structure with a dense and hydrophobic core inside a corona-like hydrophilic shell [4], the first rapid phase of release can be attributed to the diffusion of CUR molecules adsorbed on the hydrophilic shell of NGs while the slow release afterward was probably because of CUR molecules entrapped in NG cores, which have a great affinity to CUR through hydrophobic interactions and, thus, decelerated its diffusion. Meanwhile, the release at 42 • C showed more evident biphasic kinetics with the first phase being significantly more rapid than that at 37 • C. The faster release of CUR at 42 • C is in agreement with the thermoresponsive behaviors of these PEC-NGs, as NG shrinkage, when the temperature is increased [4], may cause the expulsion of water as well as CUR molecules from the NG interior. This is in accordance with a recent work by Khodaei et al. on CUR release from thermoresponsive nanoparticles from PEO-PPO-PEO copolymers, which showed that the CUR release rate was increased upon a temperature rise to 45 • C, but started to decrease when the temperature was higher than 45 • C [41]. Through mathematical modeling, the authors suggested that the former acceleration of CUR release is related to particle shrinkage, while the latter deceleration is due to stronger hydrophobic interactions between CUR and the copolymers. The higher release rate of drugs upon temperature increase has also been observed with other thermoresponsive systems from LCST-type polymers, e.g., methotrexate release from polymeric micelles of poly(N-isopropylacrylamideco-acrylamide)-b-poly(n-butylmethacrylate) [40] or release of CUR and doxorubicin from poly(N-isopropylacrylamide)-coated particles [42]. The results in the present work suggest thus the interest of HA-M2005/PLL PEC-NGs for thermo-triggered drug release, which can be applied for spatiotemporally controlling drug release by applying local hyperthermia in vivo [3]. Compared with those results, the release of CUR was not significantly different when PEC-NGs were concentrated at C P = 1 g.L −1 after lyophilization (Figure 7), which shows that the suggested method for freeze-drying and concentrating PEC-NGs upon reconstitution has no considerable effect on drug release as well as its thermoresponsive characteristics. The smaller standard deviations, in this case, were probably due to the higher CUR concentration as the samples were concentrated, allowing for a more precise absorbance measurement.
Concerning the CUR release mechanism from NGs, this can be governed by a complex set of factors, namely simple diffusion, NG swelling, degradation, or disassembly, as well as external stimuli like temperature in the case of thermoresponsive NGs [43]. In order to better predict the release mechanism from the present PEC-NGs, both the release curves at 37 and 42 • C of PEC-NGs of C P = 0.5 g.L −1 were fitted to different mathematical models most widely used for describing drug release, including a zero-order model, a first-order model, a Higuchi model, and a Korsmeyer-Peppas model ( Figure S3) [44]. It was revealed that CUR release at 37 • C follows first-order kinetics, which confirmed that the release rate was not constant but depended on the actual CUR content remaining in the NGs [43,44]. However, none of the models used could fit the release data at 42 • C, probably due to the copresence of two strictly distinct phases in this case which may correspond to two completely different mechanisms. Therefore, more complete modeling studies should be carried out if the release mechanism at 42 • C needs to be better understood.
Conclusions
This work concerns the effects of different formulation factors on the preservation of HA-based PEC-NGs after lyophilization and reconstitution, especially thermoresponsive PEC-NGs from HA-M2005 and PLL. It was revealed that the most practical approach for the lyophilization of such systems is freeze-drying PEC-NG suspensions prepared in a dilute regime (C P = 0.2 g.L −1 ) with 0.2% m/v trehalose as a cryoprotectant. This approach allowed for concentrating PEC-NGs to at least C P = 1 g.L −1 in PBS and, thus, increasing the drug content (i.e., curcumin) in the reconstitution step for future biological studies, while maintaining the particle size in an acceptable range (i.e., under 350 nm). Moreover, such a process of freeze-drying and concentrating PEC-NGs upon reconstitution did not considerably change the drug release profile as well as its thermoresponsiveness, which would be an interesting behavior for their applications in controlled drug delivery. However, further studies need to be performed in order to predict the biorelevance of the reconstituted suspensions before biological studies, namely, osmolality evaluation since the osmolality after concentrating PEC-NGs and cryoprotectants might exceed the physiological limits. In addition, before biological studies, it would be highly relevant to find a suitable technique for sterilizing these systems, e.g., by autoclave since it has been shown to be efficient for both stabilizing and sterilizing NGs from amphiphilic HA derivatives [45,46]. Moreover, for better preserving PEC-NG structure, it would also be interesting to further investigate the technical aspect of the lyophilization process, e.g., the effects of freezing speed, drying time, temperature, pressure, and secondary drying phase (moisture desorption) on the microstructures of these systems. | 9,833.6 | 2023-03-01T00:00:00.000 | [
"Materials Science",
"Medicine"
] |
Existence and Uniqueness of Global Smooth Solutions for Vlasov Maxwell Equations
Global existence of classical solutions to the relativistic Vlasov-Maxwell system, given sufficiently regular initial data, is a long-standing open problem. The aim of this project is to present in details the results of a paper published in 1986 by Robert Glassey and Walter Strauss. In that paper, a sufficient condition for the global existence of a smooth solution to the relativistic Vlasov-Maxwell system is derived. In the following, the resulting theorem is proved by taking initial data 2 0 f C ∈ , 3 0 0 , E B C ∈ . A small data global existence result is presented as well.
Introduction and Preliminaries
A plasma is one of the four states of matter, which is a completely ionized gas.
For this work, we assume the following.The plasma is: at high temperature.
at low density. collisions are unimportant (i.e.collisions between particles and external forces is negligible).
The plasma is at high temperature implies that where N is the total number of charges per unit volume, and is the mean distance between the particles.Definition 1.2.We call the distance at which the coulomb field of a charge in the plasma is screened a Debye length denoted by a and is defined by: ( ) T e N > we have that e N e N i.e.T Ne a a γ < < < we can interpret this inequality as, the mean distance between particles is small with respect to the Debye length.
Generally speaking, a plasma is collision-less when the effective collision frequency ν ω < -that is the frequency of variation of E, B. In this case collision term.f t ∂ > ∂
The Relativistic Vlasov-Maxwell System
It is a kinetic field model for a collision-less plasma, that is a gas of charged particles which is sufficiently hot and dilute in order to ignore collision effect.
Hence the particles are supposed to interact only through electromagnetic forces.
In this work let us assume that the plasma is composed of n different particles, (i.e., ions, electrons) with the corresponding masses m α and e α .According to statistical physics the set of the particles of this species is denoted by a distribution function ( ) which is the probability density to find a particle at a time 0 t > , at a position x with momentum p.Here in the vlasov Maxwell system the motion of the particles is governed by Vlasov's equation; where v α is the relativistic speed of a particle α , c is the speed of light and E and B are electric and magnetic fields respectively and p is momentum.Here where m α is the mass of the particle α .
From this we can observe that v c α < (hence relativistic system).
The electric field ( ) , E t x and the magnetic field ( ) , B t x satisfies the fol- lowing Maxwell equations. ; ; 0 where ρ and j are the densities of charge and current respectively, and hence they can be computed by: Robert T.Glassey and Walter A. Strauss [1], under the title, "In singularity formulation in collision-less plasma could occur at high velocity", they showed existence and uniqueness of a global smooth solutions in 1 C by taking initial data 0 0 , E B in .Here our work has many similarities with their work, the only difference is taking 0 0 , E B from 3 C and 0 f is from 2 0 C .In another work Robert T. Glassey and Walter A. Strauss [2], also showed uniqueness and existence of 1 C solution by taking sufficiently small 2 C initial data.The proof of this result is sketched in the following.
Simone Calogero, [3], investigated global existence for Vlasov-Maxwell equation by modifying the system in which the usual Maxwell systems are replaced by their retarded parts.Sergiu Klainerman and Gigliola Staffilani, [4] showed a new approach to study VM system , that is they showed global existence of unique solution in 3D, under the assumptions of compactly supported particle density by using Fourier transformation of the classical Glassey-Strauss result.
Oliver Glass and Daniel Han-Kwan, [5], explained that existence of classical solutions, from which characteristics are well defined in 2D by using the concept of geometric control condition and strip assumption.Gerhard Rein, [6] investigated the behaviour of classical solutions of the relativistic Vlasov-Maxwell system under small perturbations of the initial data.More recently, Jonathan Luk and Robert Strain 2014, [7] derive a new continuation criterion for the relativistic Vlasov-Maxwell system.But the unconditional global existence in 3D remains an open problem.Organization of the project: Now let us describe how this project is organized.
In chapter one, we state some definitions and terms which are related to Vlasov Maxwell system and we try to show the solutions of inhomogeneous wave equations with initial conditions in 3 and Gronwall's lemma is stated and proved.In chapter two, the main theorem is stated and we see representations of the fields and used boundedness to prove the existence and uniqueness of the solution.To prove the theorem, we use an iterative scheme.We construct sequences, then using representations of the fields we showed that these sequences are bounded in 1 C , and finally we try to show that the sequences are Cauchy sequences in 1 C .In the last chapter, the main theorem is re-stated by changing the hypothesis (taking small initial data conditions) just to show the reader there is at least one case such that the sufficient condition in the main theorem of the latter chapter holds true.In this chapter, only the theorem is stated and the main steps to prove the theorem are described.
Preliminaries
solves the initial value problem: Kirchhoff's states that an explicit formula for u in terms of g and h in three dimensions is: .; 0 in , .; 0, .; ., on Duhamel's principle asserts that this is a solution of where ( ) 0 , u x t is the solution of the homogeneous equation 0
Gronwall's Inequality
In estimating some norm of a solution of a partial differential equation, we are often led to a differential inequality for the norm from which we want to deduce an inequality for the norm itself.Gronwall's inequality allows one to do this.
Roughly speaking, it states that a solution of a differential inequality is bounded by the solution of the corresponding differential equality.There are both linear and non linear versions of Gronwalls's inequality.We state here only the simplest version of the linear inequality that we are going to use.
Lemma 1.4.Gronwall's lemma [9] If : f is continuous and bounded above on each closed interval [ ] 0,T and satisfies for increasing function ( ) a t and positive (integrable) function ( ) differentiating both sides with respect to t and applying 1.4, we have: Integrating both sides and using the increasing property of the functions gives then using 1.9, and the above bound, we have Proof: On account of the property ( ) ( ) . Hence the fundamental theorem of differential and integral calculus yields: ( ) hence the result.
Existence and Uniqueness of Global Smooth Solutions for Vlasov Maxwell Equations
In this chapter, we are going to establish the existence and uniqueness of global smooth solutions for the system in 1.5 under a sufficient condition.To derive the sufficient condition, we shall consider the case of only one species of particles, then at the end we extend the result to the case of a plasma composed The term E v B + × can be represented by K .That is which we call Lorentz force.
Theorem 2.1.[6] Let Then there exists a unique 1 C solution for all t.
To prove this theorem, we are going to use the concept of representation of the fields and their derivatives.The characteristics equations of the system 1.1 are the solutions of: Hence the solution of this system is: , , , , , , , X s t For the next two sections, the reader can consult the material [10] for more details.
Representation of Electric and Magnetic Fields
x in a similar way.Proof: Here, i T is the tangential derivative along the surface of a backward cha- racteristic cone.Now let us replace the usual operators t ∂ and i ∂ by i T and S .From For relativistic Vlasov Maxwell system, the fields satisfy the inhomogeneous wave equation: Hence, substituting 2.5 in to 2.4, we have ( ) By applying Equation (1.8) in chapter one, we have ( where ( ) ( ) x is the solution of the homogeneous wave equation.From this we can easily see that the second term is i S E .Let ( ) Hence, we can re-write the last integral as: ( ) Now let us integrate the last term using integration by parts in y.Hence, by applying lemma 1.5 in chapter one (integration by parts), this expression reduces to; ( ) where x , hence the above integral reduces to: see the computation of this expression at the appendix part of [7].Hence, 2.8 Therefore, substituting 2.10 and ( ) Similarly, by using the inhomogeneous wave equation for the field B, we have and following the same step, we have This proves theorem 2.2 Proof of uniqueness of Theorem 2.1.
To do this, let , , f E B be two different Classical solutions of 2.1 with the same Cauchy data given.Define where From the Vlasov Equation, where , E B and ( ) the characteristics of this equation are the solutions of: , , when f is written as a line integral over such a characteristics curve, we have is bounded , we can write it as: Now adding 2.13 and 2.14, we have Applying Gronwall lemma, we have 0 E B f = = = .This implies the solution is unique.This proves the uniqueness of theorem 2.1.
Representations of Derivatives of Electric and Magnetic Fields
Theorem 2.3.Assume that ( ) exists as in the hypothesis of theorem 2.1.Then the derivatives of the fields can be represented as: , , f Sf S f without explicit arguments are evaluated at ( ) Now using the fact that j T is a perfect j y derivative, integrating the last integral using integration by parts in y, is equal to: Here the last expression is part of ( ) , a w v .Here, the most singular term is the j T term, which appears in the first expression, it is: Simplifying this we get: since the first term depends only on initial data, hence part of ( ) . Now the second term simplifies as: (see an elementary computation of this in the appendix part of [1]).Hence, this expression is the value of ( ) Now first compute the third term, we have: Now the integrand in the first term simplifies as: Now the second integral becomes: ( ) Adding the Equations (2.15), (2.16) and (2.17), we get ( ) We can have the same result for the magnetic field see [1], in this case the singular term is TT B .This completes the proof of 2.3.
Estimation of the Particle Density
To estimate the particle density take .The characteristics of the Vlasov equation are solutions of the , , 0, , , , 0, , , f t x p f X t x p P t x p = and since ( ) , that is f is non-negative and bounded.Now, we claim that ( )
( ) ( ) ( )
, supsup A similar definition can be done for the electric field E.
Now by applying the norm properties above, the expression in 2.3 can be reduced to Again by taking , we can have a similar bound, since ( ) Therefore, again by applying the norm properties above we have, (
Bounds on the Electric and Magnetic Fields
We already proved in theorem 2.2 that the fields can be represented as: ( ) By our hypothesis we have We have that ( ) Similarly, for S E , we use ( ) . Then integrating this by parts in p, we get: By the support hypothesis, the v-gradient factor is bounded (say by T C ). Hence, ( ) A similar estimate holds for B, See ( [1]).Hence Adding Equations (2.24) and (2.25), we have:
Bounds on the Gradient of the Field
Theorem 2.4.[1] Let in the form of theorem 2.3 above as:
∫∫
Here the first term(data term) which is ( ) just depends on the derivative of the initial data.For the second term, From this the most singular term is the TT E term.Hence Here ω is integrated over the unit sphere 2 S and p is over 3 .We break the ξ integral into two integrals, over [ ] . Since the support of f is bounded in p, the kernel ( ) Hence, for any , 0 ) Hence, from expressions 2.28 and 2.29, we have: For the Sf term, let us integrate by parts in p: For the 2 S term, we write which is bounded and the y-integrals are over the ball y x t − ≤ and III satisfy the same bound as II.Now again split ( ) IV , integrating by parts in p, and the resulting kernel is bounded for v and p K B ∇ ≤ , hence we have; IV , we recall in Section 2.1, j T is a perfect y derivative, hence we can integrate by part in y.Since ( ) Combining these results, we get: Now adding 2.30, 2.31 and 2.37, we get: To get the same result for B, we repeat the same process, (see [7]), and This proves theorem 2.5.
Here putting 2.21 into this expression, we get: is bounded, and hence ( ) K t is also bounded.
Using this estimates, we will proof the existence of the solutions for theorem 2.1.
Existence of Solutions
From the hypothesis we have smooth initial data ( ) ( ) ( ) which is a linear equation (for a single unknown) of the form and with initial condition 0 f , where c and 0 f are 2 C functions.Since ( ) The characteristics of 2.38 are the solutions of: , , 0 d and d with initial data ( ) ( ) ( ) ( ) Lemma 2.5.[10].Given that C .Now let us proceed by induction on n to show the solutions are 2 C .
From the representation theorem 2.2, where ( ) 0 , E t x is the solution of the homogeneous wave equation with the same Cauchy data, and Substituting this in to ( ) ( ) , n S E t x , we can integrate by parts in p. From the induction hypothesis ( ) This proves lemma 2.5.Now let us claim that the estimates 2.21 and 2.26 holds uniformly in n for To show this we follow the same process as we did for , f E and B , the only difference is replacing with the superscripts ( ) and, the expression analogous to 2.21 is; and the analogue of 2.26 is; with constants C depending on T. Now iterating 2.43, we have This tells us that the fields Now an analogue of the result of theorem 2.5 is: Substituting 2.45 in to 2.46, we conclude that; as in such a way given in theorem 2.5, and then subtract this expressions.We can write the TT term as written in 2.28-2.30and then estimate it, we have; Similarly TS and ST terms are written as in 2.31 and then estimated as: For the SS term, let us break up in to several pieces as in 2.33.Following the same procedure and using the known bound in 1 C , we conclude that Now by the known bounds, we have Hence by Gronwall's lemma, the sequences ( ) ( ) x s p s converges uniformly on 0 s T ≤ ≤ .Here, on the parameter , , t x p , the convergence is also uniform, After integrating this along the characteristics, we have Subtracting the second from the first and estimating, we get: From 2.55 and the known bound in 1 C , the first term goes to zero uniformly on [ ] 0,T .The second term in the integrand is dominated by f respectively.Therefore ( ) , , E B f will be the unique solution of the system 2.1 for the simplified case of a single species.
To generalize for n species, we need just a little modification.The operator S now depends on α .
. . t x i e S v α α =∂ + ∇
In this case each f α remains bounded.In the representations of the fields and their derivatives, ρ and j are written as for each α separately.Hence, with these simple modifications, we conclude for several species case.
Uniqueness and Existence with Small Initial Data
In the previous chapter, we have seen that the sufficient condition for the existence of a global 1 C solution for the relativistic Vlasov Maxwell's equations was the existence of a continuous function ( ) with supports in { } x k ≤ which satisfy the constraints, If the data satisfy To prove this theorem, the key step is to show that the paths of the particles spread out with time.Since the paths of the particles are given by the equations ( ) Thus we need to prove that the electromagnetic field decays as t → ∞ .Hence, to prove this theorem, let us introduce a weighted L ∞ norm for the field, as was introduced by [11].Therefore, we use the weight ( )( )
The Structure of the Proof
The main structure of the proof is as established in the last chapter.To prove uniqueness, we use the same step as we did in chapter two, and for the existence, the following construction was used.For given functions That is given the ( ) iteration, then we define ( ) ( ) Finally we define ( ) ( ) , n n E B as the solutions of the Maxwell's equations, (
Characteristics
Here the characteristics are curves defined by the solutions of the equations 3.9 and 3.10.Because E and B are 1 C , the solutions exist as
4 ) 5 )∫∂ 1 . 1 . 3 .
The coupled system of Equations (1.1), (1.2),(1.3)and(1.4) is what we call the Vlasov-Maxwell System which is represented as:In this system the Vlasov equation governs the motion of the particles and the interaction of the particles are described by the Maxwell equations.So, the aim of this work is to derive a sufficient condition for the global existence of a smooth solution to the system 1.5 with initial data In the entire work, we are going to use, the partial derivatives with respect to , while any derivative of order k with respect to t or x or p will be denoted by k and so on with the convention o D f f = .A Short Review on the Cauchy Problem for the Vlasov-Maxwell Equations Now let us provide a short review on the Classical Cauchy problem on the Vlasov Maxwell System.
field representation in theorem 2.2, we have
plying 2 S
f , and ( ) , b w v is the term multiplying 2 Sf γ , which we can see it easily.Now let us determine ( ) , d w v and ( ) we must use the fact that the kernel has zero average.
1
, the p-derivative of the difference can be estimated in terms of the x-
(
where e α is the charge of particles of species α .Again n move approximately in straight lines if E and B are small. , we define ( ) ( ) ( ) ( ) , To show the sequences are Cauchy, let us fix two indices m and n.For j = 0, 1, is an upper bound for the 1 C norm, of the field, we thus have "which is the largest momentum up to time t emanating from the support of 0 f α " [2].Hence, ( ) u t is a continuous function of t for * 0 t T ≤ ≤ . | 4,758.6 | 2018-01-09T00:00:00.000 | [
"Mathematics"
] |
A Caputo (discretization) fractional-order model of glucose-insulin interaction: numerical solution and comparisons with experimental data
In this paper, we investigate a (discretization) Caputo fractional glucose-insulin model qualitatively with incommensurate orders that appear in Bergman's minimal model. After intravenous tolerance testing, the model is used to characterize the blood insulin and glucose metabolism. We also prove that the presented model possesses existence, uniqueness, non-negative, and boundedness solution. We also proceed a systematical studies on the stability of the (discretization) Caputo fractional. Comparisons between the results of the fractional-order, the integer order and the measured real data obtained from patients are presented. These comparisons is shown that the presented Caputo fractional order model is better representative of the system than its integer order form. Numerical solutions of the Caputo fractional model are obtained by using the method of Adams-Bashforth-Moulton type to handle the fractional derivatives. Also, numerical simulations of the discretization fractional derivative order model are used to support the analytical results.
In [6], De Gaetano and Arino has intended a model called the dynamical model which couples the two different parts of the "Minimal Model" into one part given byẋ 6 , and c 7 are parameters. In [7], Derouich et al. have been used a version of the minimal model in modified form to introduce parameters related to physical exercise: In [8], Li et al. had reinvestigated the dynamical analysis of the "Minimal Model" in both modelling and physiological aspect to understanding blood glucose regulatory system: The concept of fractional calculus has great importance in many branches and is also important for modelling real world problems . In this paper, we concerned on the discrete version Caputo fractional order of the minimal model (1): with This paper concerned on a analytical studies of a Caputo fractional-order glucose-insulin model (2) and its discretization. The fractional calculus has great importance for modelling real world problems and is also important in many branches. After intravenous tolerance testing, this model used to characterize the metabolism of blood insulin and glucose. We show that the model (2) possesses existence, uniqueness, nonnegative, and boundedness solution. We also prove that the presented model possesses existence, uniqueness, non-negative, and boundedness solution. We also proceed a systematical studies on the stability of the (discretization) Caputo fractional. Comparisons between the results of the fractional-order, the integer order and the measured real data obtained from patients are presented. These comparisons is shown that the presented Caputo fractional order model is better representative of the system than its integer order form. Numerical solutions of the Caputo fractional model are obtained by using the method of Adams-Bashforth-Moulton type to handle the fractional derivatives. Also, numerical simulations of the discretization fractional derivative order model are used to support the analytical results.
Notation and definitions
For ν ∈ R, the fractional derivative D ν t , can represented by Define the Euler-Gamma function as In [12], the Riemann-Liouville definition first introduced in 1847 and is given by In [13], the Caputo definition first introduced in 1967, and is given by Anton Karl Grunwald [37] and Aleksey Vasilievich Letnikov [38], introduced the Grünwald-Letnikov definition over the interval [a, t] with n ∈ N, is the step size, and a binomial .
Definition 2.2:
For β > 0, the function is called the Mittag-Leffler function of β.
Proof: One has
Then, the solutions of the model (2) Proof: As in [21], one consider the function Thus, for all γ > 0, Thus, by choosing γ < min{q 1 − q 4 |t|, q 2 , q 6 − q 3 }, one obtains Following to Lemma 2.3, one obtains Thus, as starting in R 3 + , the model (2) has uniformly bounded solution lies in the region , with
Stability analysis
For the model (2), we assume that Then, the model (2) has only one equilibrium point E • = (x b , 0, z b ) and its Jacobian matrix J(E • ) at E • is given by Also, its characteristic equation 1 (λ) is given by Then Thus, by using [32], E • is asymptotically stable.
Numerical results
In this subsection, the numerical solutions for Caputo fractional order system (2) are simulated by using the method of Adams-Bashforth-Moulton existed in [40]. Values of the parameters, given in Table 1, are taken from [11], in which these values are obtained using a computer program, named "MINMOD". Consider the following: with
As t −→ (n + 1)m, one obtains the corresponding equation of the model (2) with a piecewise constant argument is given as:
Stability analysis
Here, the dynamical behaviours and stability analysis of the Caputo fractional discretized Glucose-Insulin model (4) is investigated here at the equilibrium point . First, we compute the Jacobian matrix J(E • ) of (4) as follows and its characteristic equation is given by (5) Its discriminant is given by From the Jury's criterion [41], From the Jury test, E • is asymptotically stable if 3 − a 2 1 , and d 2 = a 3 a 2 − a 1 a 2 .
Numerical simulations
Taking the parameter values as shown in Table 1 and consider the following discretized fractional order: By calculation, the corresponding eigenvalue is D = −5.8328e−05. Then, system (4) has a free equilibrium point E • = (287, 0, 403.4). By (6) and Proposition 3.1, the solution of (4) converges to E • (see Figures 4-7). Consequently, the insulin and the activity of insulin excitable tissue glucose uptake are increased and the glucose decreased. For these parameter the corresponding eigenvalues are D = −5.8328e−05. Furthermore, glucose, insulin excitable tissue glucose uptake, and insulin concentration versus time for different cases of ν.Then, (6) and Proposition 3.1 are satisfied and then E • is asymptomatically stable. Behaviour of x(t), y(t), and z(t), for different values of ν, showing glucose, activity of insulin excitable tissue glucose uptake and insulin dynamics are shown in Figures 4-6. Also, the behaviour of Glucose, Insulin excitable tissue glucose uptake and Insulin concentration versus time for different cases ν = 1, ν = 0.95 and ν = 0.90 are shown in Figures 7-9. Now, we list some numerical results for the discretized fractional order (7) of IVGTT glucose-insulin interaction
Comparison results
In this section, Adams-Bashforth-Moulton method was employed as a reasonable basis for studying the solution of a fractional-order model of glucose-insulin system (2). We have tuned for the order of fractional derivative which ensures better fit. We compared the fractional-order model to the experimental data obtained based upon the experimental data used in [11], given in Table 2, during primary glucose-insulin interaction. Furthermore, based upon this experimental data, we demonstrate that, fractional order Bergman's minimal model is better representative of the system of glucose and insulin in blood as compared to its integer order version. As in Figure 1, the numerical results of the fractional-order model are closer to the real measured data of the patients more than the results of the integer-order. For ν = 0.95, this fractional order model gives better fit on the experimental data. It is worthy to note that the provision of changing fractions in different ways as well as changing parameters is still there, and by availing this provision, it is possible to get a very close fit. In comparison with its integer order version, the proposed model is superior. The reason is that the increase in the glucose level is less that of the integer order version. The Plasma insulin concentration in (mU/L) is illustrated in Figure 2. As shown in this figure, the proposed model outperforms the integer order version. The initial increase of Plasma insulin concentration for the proposed model is much less than that its integer order version. Simulation results verify the satisfactory performance of the proposed model in comparison with Time Glucose Insulin 0 9 2 1 1 2 350 26 4 287 130 6 251 85 8 240 51 10 216 49 12 211 45 14 205 41 16 196 35 18 192 30 20 172 30 22 163 27 32 142 30 a previous related work. Comparison of average and rms values of absolute difference from the experimental data is given in the Table 3. On the other hand, discretization method was employed as a reasonable basis for studying the solution of a fractional-order model of glucose-insulin system (2). We have utilized the above mentioned model, and tuned for the order of fractional derivative which ensures better fit.
Conclusions
The Caputo fractional-order glucose-insulin model (2) and its discretization system (4) are investigated. We showed that the fractional system (2) possesses existence, uniqueness, non-negative, boundedness solution. We also deduced a detailed analysis on the stability of the model (2) and its discretization system (4). Comparisons between the results of the Caputo fractionalorder (2), the model of integer one and the measured real data obtained from patients are presented. These comparisons is concluded that the presented fractional order model is better representative of the system than its integer order one. Numerical solutions of the model (2) are obtained by using the method of Adams-Bashforth-Moulton type to handle the fractional derivatives. We also obtained the solution of the discretization model (4) and a numerical solution of the system which shows that effect of time on the concentrations x(t), y(t) and z(t). | 2,170 | 2021-01-01T00:00:00.000 | [
"Medicine",
"Mathematics",
"Engineering"
] |
Don’t shoot the messenger! A criminological and computer science perspective on coordinated vulnerability disclosure
In the computer science field coordinated vulnerability disclosure is a well-known practice for finding flaws in IT-systems and patching them. In this practice, a white-hat hacker who finds a vulnerability in an IT-system reports that vulnerability to the system’s owner. The owner will then resolve the problem, after which the vulnerability will be disclosed publicly. This practice generally does not focus on potential offenders or black-hat hackers who would likely exploit the vulnerability instead of reporting it. In this paper, we take an interdisciplinary approach and review the current coordinated vulnerability disclosure practice from both a computer science and criminological perspective. We discuss current issues in this practice that could influence the decision to use coordinated vulnerability disclosure versus exploiting a vulnerability. Based on different motives, a rational choice or cost–benefit analyses of the possible reactions after finding a vulnerability will be discussed. Subsequently, implications for practice and future research suggestions are included.
Introduction
Computer hardware and software products are designed to be as user-friendly as is possible, trading security for usability in some cases (Newman and Clarke 2003;Van Schaik et al. 2017).Consequently, enterprising security researchers and criminal hackers may identify flaws within computer devices in order to make them operate in unintended ways (Jordan and Taylor 1998;Taylor 1999).These flaws are commonly referred to as vulnerabilities, as they enable an attacker to gain access to computer systems and data for malicious use.When an individual identifies a vulnerability, they basically have four options: (1) do nothing about it, (2) report the flaw to the vendor or a related security organization for mediation, (3) report the flaw publicly, (4) keep this information private so that it can be used for attack, either by the person who identified the vulnerability, or by selling the vulnerability to someone else at an underground market.
Public reporting on vulnerabilities has evolved over the last 30 years, reflecting shifts in the dynamics between security organizations and the hacker community.Initially many security researchers tried to shame vendors by disclosing all details as soon as the vulnerability is discovered.Such a move would enable attackers to use the vulnerability to compromise systems before they can be corrected.In the last few years, reporting has tended more towards coordinated disclosure, where a researcher privately contacts a vendor to resolve the vulnerability before going public with his findings.Additionally, there has been an increase in "bug bounties" where a person is paid for vulnerability disclosures by security vendors (NTIA 2016).
The general term that will be used in this article to refer to vulnerability disclosures is coordinated vulnerability disclosure (CVD).In general, CVD is a practice in which a hacker who finds a vulnerability in an IT-system reports that vulnerability to the system's owner.The owner will then resolve the problem, after which the vulnerability can be disclosed publicly.In order to prevent criminal use of the vulnerability, it is key that the hacker does not share or publicly disclose the vulnerability before the problem has been fixed.The details and different CVDforms will be discussed later in this paper.The overarching goal of having a CVD policy is to make IT-systems more secure and prevent the criminal use of vulnerabilities in IT-systems (ISO/IEC 2014;NCSC 2013;NTIA 2016).
The Netherlands is one of the few countries in the world with official guidelines for vulnerability disclosure.In 2013, the Dutch National Cyber Security Centre (NCSC) introduced a guideline for Responsible Disclosure (NCSC 2013).This document provided guidelines for the vulnerability disclosure process both from the researchers as well as organizational point of view.The Dutch Public Prosecutor has officially endorsed this guideline and has taken elements of it as a decision framework for when to prosecute (Public Prosecution Service 2013).Since 2013, there have been many successful CVD-cases, ranging from large disclosures by academic researchers to small disclosures that lead to configurational changes (NCSC 2017).There have been several cases where a discloser even ended up with a job at the vulnerable organization, but also cases with successful prosecution when the discloser went too far (Van't Hof 2016).Last year the US guidelines have been published (Department of Justice 2017), but for the sake of clarity the focus of this paper will be on the Dutch guidelines.
The overarching goal of CVD shows a focus on the victim side and data-breach prevention and other victimization types.This makes sense as the CVD policy originates from the computer science field, which generally focuses on making IT-systems more secure.CVD policies also seem to target so-called white-hat or ethical hackers.Criminological inquiries, however, focus on the offenders engaged in criminal hacks and misuse of vulnerabilities (for a review see Holt and Bossler 2016).
So, what can we learn from a combined computer science and criminological perspective on CVD?What are the key requirements for a successful CVD policy and how do these relate to criminological explanations for criminal hacking?What are the main problems with current CVD policies and how do these relate to ethical and criminal use of vulnerabilities?Will a CVD policy mainly work for white-hat or ethical hackers or can we expect it to help potential offenders to choose the ethical instead of the criminal path?And lastly, which empirical research questions should be addressed to further inform us about these questions?In this paper, we will shed light on these questions from both a computer science and criminological perspective.
Coordinated vulnerability disclosure
The Netherlands was one of the first countries to legally recognize the practice of CVD policies.At the time it was called responsible disclosure.The need for a formal policy on vulnerability disclosure arose as a result of some cases that were reported in Dutch media, in which it was unclear if a hacker acted responsibly or if the hacker crossed a line and acted criminal (Van't Hof 2016).Therefore, in 2013 the NCSC of The Netherlands published guidelines for responsible disclosure policies.Later the term "responsible" has been deemed too loaded; the new term "coordinated" conveys that CVD is a process between two equal participants.Coordinated vulnerability disclosure is now used nationally and internationally.The vulnerability disclosure process is described in the guidelines for disclosure of potential vulnerabilities in products and online services (the ISO/IEC 29147:2014) of the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC), see ISO/IEC (2014).
In order to look at CVD from a criminological perspective, it is first necessary to discuss all aspects of CVD as it arose from computer science.The main goal of an established CVD policy is to invite white-hat hackers to report any vulnerabilities they find in an IT-system to its owner.They should also not discuss the vulnerability with anyone else or disclose it publicly somewhere.In this way, the vulnerability is likely only known to the owner and the discloser, which means that the exploitation risk of that vulnerability is minimized.The owner will then try to mitigate the vulnerability as soon as possible, ideally in consultation with the discloser.After the vulnerability has been fixed, the discloser and owner will decide if and how it should be disclosed to the public (ISO/IEC 2014; NCSC 2013; NTIA 2016).
This policy is beneficial for the IT-systems' owners, as they will learn about their vulnerabilities and potentially improve their security posture.This policy provides some certainty for both parties, especially the disclosers who may have committed a crime by finding the vulnerability.As long as the discloser abides by the policy's terms, the IT-system's owner should generally not report their actions to the police.In this way both parties collaborate in their common goal to improve cybersecurity (NCSC 2013).It should be noted, that currently there is no guarantee that the public prosecutor will not prosecute a discloser for any crimes that have been committed.
Representative information about the type and amount of vulnerabilities that are disclosed by using CVD is not available.Nevertheless, some descriptive information based on recent reports is helpful in understanding the nature of CVD.The NCSC of The Netherlands generally only handles CVD reports about their own infrastructure, central governmental organizations, and private organizations who handle critical infrastructure.Their latest annual report (NCSC 2017) indicates that the large majority of CVDs are about vulnerabilities in websites (78%), like cross-site scripting (32%).Other reports included software vulnerabilities (9%) and configuration errors in hardware and software (3%).
While the NCSC sees a rise in CVDs in comparison to previous years, they see a decline in false positives, i.e. reports that eventually did not include a real vulnerability.The NCSC (2017) argues this reflects a maturation process on the disclosers' side.A survey from the National Telecommunications and Information Administration (NTIA 2016) among security researchers showed that 92% of their respondents disclose vulnerabilities by using CVD.
Bug bounties
Initially CVD programs gave out small rewards for successful disclosures, such as t-shirts, small gadgets or listing the researcher in a hall of fame.Many researchers accept this and use it to boost their reputation.Recent years has seen some professionalization of CVD by offering monetary awards, so called bug bounties (Finifter et al. 2013).Microsoft (Microsoft Bounty Programs https ://techn et.micro soft.com/enus/library/dn425 036.aspx, 2018) and Google (Android Security Rewards Program Rules, https ://www.google.com/about /appse curit y/andro id-rewar ds/, 2018) have programs where researchers may be eligible for up to $250,000 for specific disclosures.At the same time several companies have started that help other companies in setting up CVD and bug bounty programs.HackerOne, a third-party platform for hosting vulnerability disclosure and bug bounty programs, claims to have over 800 active disclosure programs (Hackerone 2017).It should be noted, however, that bug bounties are only a small part of CVD.Most organizations with a CVD policy do not offer monetary rewards.Bug bounty programs seem to assume a financial motive for finding and exploiting vulnerabilities, something that criminological research discussed later in this paper has shown to be only partially true.
Problems with current CVD practices
Although the goal of CVD policies is clear and statistics indicate a positive development of these policies and their users, current policies have some problems that should be discussed in order to understand the possible problems of these policies in preventing crime on both the victim and the offender side.Taking a traditional deterrence approach, problems with the reporting process may influence a person's decision to following CVD guidelines.
The organization's response
Organizations should adopt a CVD policy because they want to increase their security, though this also means that the organization should be able to respond to a reported vulnerability.In addition, organizations without a CVD policy may also receive a vulnerability report.When there is no CVD policy, it is not clear to disclosers how the organization will respond.The expected reaction of such an organization may influence the behavior of a possible discloser: these organizations could (1) respond gratefully and patch the vulnerability as soon as possible, (2) ignore it, (3) deny it, or (4) report to the police.An organization that does not have a CVD policy may, for example, not know how to respond or not understand the vulnerability and could therefore decide to ignore it or deny the vulnerability's existence.They may even misinterpret the intentions of the reporter and report it to the police as a crime.
Even organizations that do have a CVD policy might not have the capacity to handle big vulnerabilities, which may delay the patching process.The longer a vulnerability has not been patched, the higher the risk of rediscovery or that the discloser decides to make it public anyway (Herr et al. 2017).Most CVD policies state how much time they would take before fixing a vulnerability, but that could easily be 6 months.In a response to that, new companies now arise who handle coordinated vulnerability disclosure for small companies (Huang et al. 2016).
Moreover, the goal of having a CVD policy is to keep vulnerabilities private until they are patched.This means, however, that the outside world including the discloser cannot see that an organization is working on a patch.Therefore, it is key that an organization keeps communicating with the discloser about the patching process, which is also what the majority of the researchers in the NTIA (2016) report expect.Nevertheless only 58% received a notification when the vulnerability had been patched.Depending on a person's motive, this could influence the discloser's behavior.
Unclear or unjust rules
In order for a CVD policy to work, both the company and the discloser need to stick to the rules in the policy.An absence of clearly identified rules may lead to a lack of disclosures, as would guidelines that are too strict.For example, deadlines in the policy could force a company to publicly disclose a vulnerability that has not yet been patched, as they do not know how the discloser would respond if they would not.
For the discloser, there is no guarantee that he or she will not be prosecuted under current CVD guidelines (NTIA 2016).An organization without a policy may report it to the police immediately, as could organizations with clear policies if they believe the discloser did not abide by their rules.In The Netherlands, the public prosecutor could also decide to prosecute if they believe a crime has been committed.For most disclosures some form of system-trespassing is necessary, as it is not possible to ask for permission from the system owner.For example, in the survey from the NTIA (2016), researchers indicated that they generally find vulnerabilities in their daily activities, without actively looking for them.In that sense, requiring asking for permission partly defeats the purpose of having a CVD policy.
For some organizations, it is publicly known how they generally handle vulnerability disclosures.First, bug bounty programs are publicly known and some organizations are very open about their CVD policies and they actively encourage the hacker community to test their systems.However, there is a big difference between open and closed communities, even in the same sector.For example, while the Linux community actively encourages people to find vulnerabilities, Microsoft historically tended to prosecute people who disclose vulnerabilities (e.g., Steinmetz 2016;Taylor 1999).Similarly, when looking at the hacker subculture, there is a general tendency to share vulnerabilities within the subculture, but not with others like law enforcement or large commercial companies that are not open source (Taylor 1999).These unclear and sometimes unwritten rules result in a situation in which one person will be prosecuted for the same behavior for which someone else would get an acknowledgement or even a bounty.This could result in the opinion that the rules are not fair or even unjust, which may influence if and how someone discloses a vulnerability.
Public disclosure
When the vulnerability has been patched, or when the deadline as described in the CVD policy has expired, the discloser and the IT-system's owner can decide together to disclose the vulnerability to the public.There are several reasons to do so.First, it could be a way to provide the discloser with some acknowledgement for his or her work and abilities to find this vulnerability.53% of the researchers in the NTIA (2016) report stated that they expect to get some form of acknowledgement, although it should be said that a minority (14%) prefers to remain anonymous.
Another reason to disclose these vulnerabilities is to inform the public about the vulnerability and what should be done to prevent exploitation of the vulnerability.It could be the case that other IT-systems have similar vulnerabilities or patching the vulnerability in software requires an update from users (Department of Justice 2017).The amount of information that a company is willing to share about the vulnerability may, however, be limited.The discovery of the vulnerability may be embarrassing for the company, affect their finances, or reveal too much of the underlying operation.This limits the usability of the disclosed information and may influence a person's decision to report a vulnerability to a party that has not shown openness about vulnerabilities.
In a similar fashion, some recent incidents have shown that governments are sitting on vulnerabilities in order to engage in offensive attacks (Ablon and Bogart 2017).They may have found these vulnerabilities themselves, but it is also very likely that they have bought these vulnerabilities at underground markets for exploits (Fung 2013;Healey 2016).They do not disclose these vulnerabilities, not even to the system owners, which has caused some major damages when these vulnerabilities ended up in the wrong hands.For example, the Wannacry ransomware used the EternalBlue vulnerability, which is said to be discovered by the National Security Agency (NSA) several years ago (Nakashima and Timberg 2017;Titcomb 2017), and was not disclosed until the ShadowBrokers published it.Microsoft patched the vulnerability, but 3 months later many systems were still vulnerable which enabled the large and worldwide damage of the Wannacry ransomware (Newman 2017).This is likely one of the reasons that some parts of the hacker culture have a tendency to share vulnerabilities within the community, but not with others and especially not with governments (Taylor 1999).Additionally, by buying these vulnerabilities at underground markets, governments may send the message that they are not supporting CVD, as they are rewarding criminals who sell their exploits.
Knowledge about CVD among possible offenders
Several of the problems discussed above may influence a person's decision about how to handle a vulnerability.To be able to make a decision a person first needs to know about the possibility to report a vulnerability through CVD, and then must know the policy's rules.From the NTIA (2016) report, it is clear that most people who could be regarded as security researchers know about these policies.As also acknowledged by the NTIA it may very well be the case that their respondents have an interest in CVD or at least already know about it.It is unknown to what extent this can be said for the general population.For the purposes of this work, we will assume that a person with the skills necessary to identify vulnerabilities in the wild knows about the possibility to use CVD.
Motives for CVD reporting
A first step in understanding the criminological side of CVD is to understand the motives for both criminal use of vulnerabilities and using CVD instead.Based on the general idea behind CVD, one could say the main reason to report a vulnerability is to increase cybersecurity.For example, Van't Hof (2016) describes a hacker who has made thousands of CVD reports and who sees it as his "personal mission" (p.226).Even though this particular hacker does not go public after a successful disclosure, in general CVD may also be a way to gain status in the hacker community as most researchers who responded to the NTIA (2016) indicated that they expect some form of acknowledgement for their actions.Experiences from some organizations that have CVD policies and experiences at the National Cyber Security Centre also show that some security researchers specifically ask for recognition so that they can use that to build their CV by showing their skills.
Additionally, vulnerabilities may result from fairly easyto-fix and well-known problems.Reporting that kind of vulnerability may even result from some form of frustration about the system's owner's inability to prevent these well-known vulnerabilities.Lastly, bug bounty programs added an important reason to report a vulnerability: money.Bounties may not be a pivotal drive, as only 15% of the researchers in the NTIA (2016) report indicated they expected a payment.A description of a young hacker by Van't Hof (2016) can be seen as a reflection of the motives above: "I ask whether the cash bounties are important to him.Not really, he tells me.He hacks for the recognition in whatever form that comes.He wants to solve the puzzle and he wants to show other people that he has done so " (p. 215).
The motives to report may not be substantial enough to warrant reporting for some individuals due to the inherent risks involved.The NTIA (2016) shows that the unclear rules and risk of prosecution could be enough to keep individuals from reporting a vulnerability.Additionally, the previously discussed frustration around the communication about a vulnerability is a reason to consider disclosing it publicly for 50% of all researchers in the NTIA (2016) report, and 32% actually disclosed publicly because of unmet timelines.Even though these researchers may not exploit the vulnerability they identify, their public disclosure may help others to do so instead.
Nevertheless, their public disclosure may be the only way to force a company to fix the problem, inform other system administrators who have the same vulnerability, or to warn the users of the affected systems.In short, even with good intentions the decision between keeping a vulnerability private and public disclosure may not always be clear.
Motives for criminal hacking
It is important to note that not reporting a vulnerability, if identified, is not currently criminal.Using that vulnerability to engage in criminal hacks is, however, illegal and viewed as part of the hacking-process.An individual may use a vulnerability to gain access to a system, and then access the data on that system or use its functionality for other criminal purposes (Holt and Bossler 2016;Taylor 1999).Criminological research has indicated some motives for hacking and related behaviors.These motives could shed some light on the reasons why a person would decide to exploit a vulnerability or sell it at an underground market, instead of disclosing it or doing nothing with it (Holt and Bossler 2016).
Three different categories of motives for hacking and related offenses can be informative in understanding offending versus CVD.First, some criminal hacking occurs due to the challenge of breaking into a system, curiosity, a need to learn or understand a system, feelings of addiction, feelings of power, etcetera (e.g., Holt 2007;Voiskounsky and Smyslova 2003;Weulen Kranenbarg 2018;Woo 2003).These intrinsic motives could also account for the desire to identify vulnerabilities without exploiting them.However, after breaking in a person may be curious about the data that is stored on a system and may download that data.This is against the rules of most CVD policies.An example of this is a well-known case described in Van't Hof (2016), where a person hacked into the computer systems of a hospital.While the defendant said he had ethical motives, he also states that his "curiosity drove him to access the server on more than one occasion" (p.183) and he also accessed patient records of specific celebrities.In this case, the court ruled that the defendant had gone too far and his behavior was no longer proportional.
A second motive is related to peer associations and personal ego development.In the criminal hacking community, showing that you broke into a system will give you more social status (e.g., Holt 2007;Nycyk 2010).By extension, identifying an unknown vulnerability and selling that or utilizing it in personal hacks would be a demonstration of serious skill.In the more white-hat community, however, showing that you reported a vulnerability through CVD or legitimate reporting channels may increase an individual's social status (Van't Hof 2016).In fact, there is anecdotal evidence that some hackers have begun to donate bug bounty payments to charities, which helps to elevate an individual's reputation and status (Hackerone 2017).The community that a person is part of could therefore strongly influence a person's actions after finding a vulnerability.
Third, many modern criminal hacks are driven by the desire for monetary gain (e.g., Chan and Wang 2015;Grabosky 2017;Holt and Kilger 2012;Kshetri 2009;Provos et al. 2009;Smith 2015;White 2013).This could have two effects on vulnerability reporting.First, a person could decide to sell a vulnerability in the underground community or, second report vulnerabilities to bug bounty programs to turn a profit.We will now further discuss how these motives may influence the rational choice decision to exploit or disclose a vulnerability and we will discuss some things that may influence this decision in favor of using CVD.
Rational choice theory
One of the oldest criminological frameworks applies the rational choice perspective, where an individual considers the costs and benefits of offending when presented with opportunities to engage in crime.Should the benefits outweigh the costs that person may be more likely to offend (e.g., for a review on cybercrime see Holt and Bossler 2016).Regarding vulnerability disclosure, most researchers just find vulnerabilities during their daily online activities (NTIA 2016).They do not specifically look for them in specific IT-systems.Similarly, both traditional criminal opportunities as well as cybercriminal opportunities generally arise during normal daily activities (Weulen Kranenbarg et al. 2017Kranenbarg et al. , 2018)).
One of the main costs associated with offending is the negative social consequences stemming from detection, such as arrest, prosecution and any resulting punishments (e.g., Pratt et al. 2006).The decision to offend is based on the perceived detection risk and costs relative to the benefits the individual receives.For most cybercrimes, apprehension rates are still very low (e.g., Holt and Bossler 2016;Wall 2007) which may make some individuals more likely to offend in cyberspace.Under current CVD practices, the risk of legal action after disclosing a vulnerability may be an important cost in the cost-benefit analyses for CVD.Additionally, if there are too many rules or if the disclosure process is too time consuming, this may also have a negative effect on this cost-benefit analyses for CVD.
Since the costs may be somewhat high for following CVD processes, individual motives may be an equally important factor in the outcome of vulnerability reporting.Individuals motivated by curiosity and social rewards may be more willing to report a vulnerability if they can receive some sort of additional social rewards for their actions.For example, if a company invites a discloser to help testing a patch for the vulnerability, it may make them feel more integrated into the process and see enough benefit to use CVD.Similarly, a person seeking peer recognition may be more affected by leveraging well-known role models such as regarded white-hat hackers who actively argue for the importance of using CVD instead of exploiting vulnerabilities.
Lastly, with respect to financial motives, some researchers have tried to make a costs-benefit analysis between bug bounty programs and the underground market.Allodi (2017) analyzed a Russian cybercrime forum.The results showed that the prices in the underground forum are the same or higher than in bug bounties or other legitimate markets.Also, a vulnerability could be sold more than once in the underground market, while it generally can only be sold once in the legitimate market.Additionally, in most criminal hacking cultures, working together with governments or large companies is not accepted (Holt 2007;Taylor 1999).Therefore, even if bounty payments are very high, reporting vulnerabilities may be offset by social costs to an individual's reputation.However, in general the costs of possible negative social consequences in combination with some payment seems to make bug bounty programs at least somewhat effective (Ransbotham et al. 2012;Zhao et al. 2015).Additionally, as some governments also buy exploits through underground markets, selling an exploit at those markets may also have a negative impact on a person's reputation.
Conclusions and discussion
The rise of coordinated vulnerability disclosure policies presents a unique challenge for criminological and computer science research as it is not entirely clear what factors affect the decision to handle a vulnerability.A person could decide to do nothing, exploit the vulnerability or sell it at an underground market, disclose the vulnerability publicly, or disclose the vulnerability privately by using CVD.The motives of the individual actor will directly shape their cost-benefit analyses regarding the organizational and criminal justice system responses to such a disclosure.
In light of the issues identified in this analysis, it is clear that there are ways to improve the current CVD policies' structure to increase the likelihood that actors report when they identify a vulnerability.From a situational crime prevention perspective (e.g., Newman and Clarke 2003), there are ways to affect the attackers' decisionmaking calculus in ways that could increase reporting or minimize criminal use.One potential avenue would be to increase awareness of CVD, which would remove excuses for not reporting vulnerabilities through CVD.Without this information, a hacker's knowledge base is limited, thereby rendering their decision-making process substantially bounded.Creating programs that try to teach young hackers about the rules and possibilities around CVD, may increase awareness of the mechanisms and potentially improve the likelihood of reporting.
Additionally, providing a positive form of peer recognition through overt positive acknowledgements from the legal hacking community about successful CVD strategies, a potential offender may see the benefits of using CVD.This could be achieved through actively pushing information about successful CVDs to the general media, so that they can also show the positive and constructive side of hacking instead of only the negative criminal side.Such a strategy could not only increase compliance but also further eliminate hackers' excuses to not report (e.g., Holt and Bossler 2016;Newman and Clarke 2003).Additionally, this may stimulate the debate about the rules of CVD policies and when a discloser has crossed the line.More positive public information about CVD among large companies or governments may also demonstrate the value of reporting vulnerabilities to these organizations, despite the negative image this may have in some parts of hacking culture.
Another option based on situational crime prevention models would be to provide easy access to positive alternatives in the event of identifying a vulnerability to remove offender excuses for not reporting.For example, just as studies that use banners to inform potential system trespassers about the negative consequences of system trespassing (Maimon et al. 2014;Testa et al. 2017;Wilson et al. 2015), clear and eye-catching information about a website's CVD policy could help a person understand there are rules and guidelines to report a vulnerability.Additionally, it would be advisable to keep the threshold for reporting low, to make sure that the potential costs of CVD are as low as possible.This would also call on organizations to respond seriously, act quickly and set a date for making it public, keep the discloser updated, and make sure that their rules are clear and easy to find.Taking such steps would reduce the provocations and excuses of hackers that they have no clue what occurs when a vulnerability is reported.If an organization struggles with the fact that a discloser may have committed a crime in finding a vulnerability, organizing hackathons or other ways of actively inviting hackers to test systems, may partly reduce the chance that a person does something that is against the rules.
With respect to the organization's response, it may be valuable to keep an open communication line with the discloser.During the disclosure process, the discloser can be invited to test possible patching, or perform additional (paid) research for the organization for new products or services.As mentioned before, some organizations even use the disclosure process as a recruitment tool.These follow-ups after the disclosure process may provide disclosers with an interesting challenge or lead to legitimate profession.
It should be noted that these concepts have yet to be empirically tested, as with most situational crime prevention research related to cybercrime (e.g., Holt and Bossler 2016).In order to understand the potential of CVD in preventing cyber-offending some empirical research implications should be discussed.The current empirical work from, for example, the NTIA (2016) cannot tell us to what extent CVD is also being used by people who would otherwise exploit a vulnerability, or how much people actually know about CVD.Examining these issues with both general population samples and groups of IT-professionals would improve our understanding of the awareness of CVD.Additionally, there is no empirical research that directly asked disclosers why they used CVD.This may inform our knowledge of the relationship between individual motives and CVD reporting.Additionally, it would be very informative to see if individual reporting decisions vary based on situational factors specific to an individual, such as the type of vulnerability, organization impacted, motives, potential bounty or acknowledgement, and other related factors.
By addressing these research questions in interdisciplinary research, in the future CVD may be even more effective in achieving its main goal: preventing the exploitation of vulnerabilities in IT-systems.In the future it may not only achieve that goal by making IT-systems more secure in patching vulnerabilities, but also by steering potential offenders in the direction of CVD instead of exploitation. | 7,376.6 | 2018-11-19T00:00:00.000 | [
"Computer Science",
"Law"
] |
Chlorella protothecoides Microalgae as an Alternative Fuel for Tractor Diesel Engines
Biodiesel has attracted a great deal attention recently as an alternative fuel due to increasing fuel prices and the imperative to reduce emissions. Among a wide range of biodiesel resources, microalgae are a promising alternative fuel source because of the high biomass, lipid productivity and environmentally friendliness. Microalgae is also a non-edible food, therefore, there will be no impact on the human food supply chain. In this work, petroleum diesel (PD) and biodiesel from the microalgae Chlorella protothecoides (MCP-B20) blend have been used to examine the performance and the emission of a 25.8 kW agriculture tractor engine. Two engine speeds at maximum power take off (PTO) power and torque have been selected for analysis using analysis of variance (ANOVA). The results showed that there is no significant difference between the engine performance when microalgae biodiesel blend (MCP-B20) and PD were used. However, a significant reduction in CO, CO2 and NO emissions was found when MCP-B20 was used. These outcomes give strong indication that microalgae can be successfully used in tractors as alternative fuel.
Introduction
The increased demand on energy associated with fossil fuel has resulted in problems such as high prices, supply depletion and emissions, which in recent years have been the main factors encouraging an increased focus on alternative fuels. Enhancing energy security is another factor for pushing toward alternative fuels to vary the sources of energy and reduce the demand on importing oil. Dyer & Farm tractors and other agricultural machinery consume a significant amount of fuel and contribute to the gas emissions which cause global warming. Biodiesel is regarded as one of the best alternative fuels for diesel engines [1,2]. Biodiesel is renewable, environmentally friendly, non-toxic, biodegradable and does not require significant modification to existing technology [3]. Biodiesel is the product of reacting triglycerides with an alcohol in the presence of an acid or base catalyst to form a fatty acid ester and glycerol. The process of converting vegetable oil or animal fat to biodiesel is called transesterification [4][5][6][7]. Transesterification has been widely used in biodiesel production to enhance the fuel properties and reduce the viscosity [5,8]. However, using biodiesel as a source of energy can lead to a rise in food prices. At the same time biodiesel availability cannot cover even current transportation fuel demand [9]. The main sectors that consume conventional liquid fuel are transport and agriculture and by 2011-2012 the worldwide demand for diesel fuel will be some 66.90 million tonnes with the anticipated increased consumption in the agricultural sector [10]. Among different biodiesel feedstocks, microalgae were reported as a potential future fuel. Microalgae are microorganisms which have the ability to convert CO 2 and sunlight into useful oil. The oil productivity of microalgae is much greater than that of crop oils. In the same way microalgae have the ability to grow under such conditions that the effect on agriculture is insignificant. The topic of biodiesel production from microalgae has received a great deal of attention recently because it is renewable, environmentally friendly and has the ability to convert CO 2 to oil feedstock for biodiesel production. Microalgae biofuel is non-toxic, contains no sulphur and is highly bio-degradable. After extracting the oil, the leftover material can be used as soil fertilizer or used in ethanol production [4]. Microalgae have high biomass and lipid productivity per unit of area in comparison with crops [11]. Chisti [9] reported that the demand for transportation fuel can only be covered by microalgae as a renewable source. The same amount of biodiesel from microalgae (for 30% w/w oil content) compared to rapeseed or soybean crops requires less land, around 49 to 132 times [12]. Furthermore, microalgae are non-edible and can grow under various conditions in which there is no significant effect of human food supplies [13,14]. One of the major arguments against the production of biodiesel from agricultural crops is that it will result in food shortages and raise food prices. However, the production of biodiesel from microalgae circumvents these arguments. Chlorella species' benefits include fast growth and easy cultivation, which make it an attractive potential energy source, however, it is not commercially viable due to its low lipid content [15]. Haik et al. [16] conducted research studies to understand the effect of using algae fuel on the combustion quality, in-cylinder pressure and heat released. Haik et al. [16] reported that algae oil methyl ester properties are similar to those of diesel, and thus they used algae fuel (both algae oil and algae oil methyl ester) successfully in a single cylinder Ricardo engine. The combustion results demonstrated that algae methyl ester produced a marginally higher heat release rate in comparison to diesel. The heat increment can be attributed to the longer combustion delay period of algae fuel in comparison to diesel. However in the case of algae oil, it was found there is a noticeable reduction in both in-cylinder pressure and heat released compared to diesel. This reduction can be attributed to the difference in the properties of algae oil and algae oil methyl ester. A similar outcome was concluded by our research team (results were not reported here).
In agricultural work, tractors are the major power source in the agricultural field and one of the main fuel consumers in the biodiesel production chain. Dorado et al. [2] tested biodiesel from olive waste cooking oil in a direct-injection engine (Perkins, model AD 3-152A) and satisfactory performance and no statistically significant differences were achieved throughout the test using biodiesel and diesel fuel. However; up to a 26% increase in the brake-specific fuel consumption and less than 8% of the power loss were pointed out in comparison with diesel. Matthew [17] used a diesel tractor (John Deere 3203, John Deere, Moline, IL, USA) to test the performance and emissions when using No. 2 diesel (D2), 20% biodiesel (B20), and 100% biodiesel (B100). The test focused on the variation in specific fuel consumption, power take-off (PTO) torque, PTO power, thermal efficiency and NO x emissions between those different fuels. The results showed that there are no statistically significant differences between diesel and B20. Neel et al. [18] used a 23.9 kW compact utility tractor (John Deere 3203) to compare the PTO power performance, fuel efficiency, and NO x using petroleum diesel, biodiesel blend (B20), and neat biodiesel (B100) at rated PTO speed (540 RPM) and at peak torque load conditions. The results showed that when the tractor was fuelled by D2 or B20, similar performance was found at peak torque. Biodiesel B20 was recommended by [19] as the optimal biodiesel blend due the similar performance to diesel.
In this work, tractor PTO tests were conducted to evaluate the tractor engine performance and emissions using petroleum diesel (PD) and microalgae biodiesel (MCP-B20) under different tractor operation conditions. The performance and the emission map have been drawn for wide open "throttle" (WOT) and half open "throttle" (HOT). Further analyses were conducted using SPSS analytics software for the PTO rated speed and peak power during the WOT test. The analysis of variance (ANOVA) was used to identify the significant differences between the means values of the two fuels at the selected speeds.
Experimental Tractor and Apparatus
A stationary PTO test has been conducted in the agricultural equipment laboratory at the University of Southern Queensland. The test aims to study the performance and emissions of a farm tractor fuelled with 20% Chlorella protothecoides microalgae biodiesel blend (MCP-B20). The experimental results were used for comparison with petroleum diesel (PD). The tractor used in the experiment was a John Deere 4410 e-hydro, as shown in Figure 1. The engine in the JD4410 tractor is a Yanmar 3TNE88, three cylinder water cooled diesel engine with 18.8:1 compression ratio. The engine power is 25.8 kW and the manufacture's estimated PTO power is 21.3 kW. The engine torque at rated speed is 87.9 Nm. In order to apply load on the tractor engine, a PTO dynamometer was used (Figure 1, items 4 and 5). The dynamometer was modified by adding a load cell and digital monitor to measure PTO speed (RPM) and torque (Nm) to calculate the power (kW). The dynamometer was calibrated prior to the tests using five standard weights of 20 kg each. The calibration started by running the monitor of the dynamometer in the calibration mode. Then the zero torque was set. Standard weights of 100 kg were suspended on a one meter arm to apply load on the load cell. At that time the torque of 981 Nm was set on the screen. An optical tachometer was used to check the PTO speed. The ratio of the engine to the PTO speeds was found to be 4.815. The level of accuracy of this load cell is around ±2.5%.
A BEA 460 Bosch gas analyser was used to monitor the tractor exhaust gases for O 2 , CO, CO 2 , NO and Lambda. The gas analyser was connected to a laptop to record and save the data ( Figure 1, item 3). The operating ranges (sensitivity) and resolutions are shown in Table 1. Prior to the test, the device was subjected to maintenance and calibration by the manufacturer and daily standard calibration. To compare the noise level produced from the tractor engine using the different fuels, a Larson Davis SoundTrack TM LxT Sound Level Meter was placed at the operator head position.
Test Fuels and Procedure
Chlorella protothecoides microalgae oil (100%, MCP-O) was obtained from Soley Institute Turkey. The MCP-O was converted to biodiesel through transesterification. The transesterification procedure was conducted by heating the oil to 48 o C. For one litre of MCP-O, an amount of 9 g of NaOH was added to 220 mL of methanol and mixed. The mixture was added to the oil and mixed for 40 minutes. After 10 hours, the oil phase was separated to another flask and centrifuged to remove the glycerin produced. The biodiesel was then washed with 500 mL of water. After one hour, another washing with 500 mL of water was conducted. The mixture of biodiesel and water was centrifuged to remove all the remaining water from the biodiesel. Table 2 represents the properties of the fuels used in this test PD and MCP-B100. The chemical formula of MCP-B100 is C 18.151 H 34.376 O 1.942 as reported by [20,21]. To prepare the MCP-B20 blend, 20% by volume of MCP-B100 was mixed with 80% by volume of PD. The specification of MCP-B20 was calculated from the PD and MCP-B20 depending on their volume in the blend. MCP-B20 have higher cetane number comes from the high percentage of unsaturated FAMEs.
The fuel consumption rate was measured using the measuring cylinder shown in Figure 2. Initially a digital flowmeter was used to measure the mass flow rate of diesel; however the data fluctuated. Thus, a volumetric cylinder was successfully and accurately used in this research work [22]. The measuring cylinder was fitted with a two-way valve at both the top and bottom. The top two-way valve allowed for the fuel returning from the injectors to either enter the cylinder or return to the tank. The bottom two-way valve allowed for the fuel to either be drawn from the cylinder or to be supplied from the tank. In order to obtain a fuel consumption reading, the top and bottom valve were opened simultaneously allowing return fuel to enter the cylinder whilst fuel was being drawn from the bottom.
A separate tank was connected while MCP-B20 fuel was being used. The time was recorded using a stop watch for the fixed volume of fuel consumed. The fuel consumption measurements were repeated three times at each speed to reduce the error. At the end of the test using PD, the system was cleaned and the fuel filter was changed to prevent any contamination from the previous test.
In this work the tractor test results of the comparison between PD and MCP-B20 were presented using the WOT and HOT tests. At WOT the engine speed was fixed at 2700 RPM (PTO speed 550 RPM) at no load using a fuel controller. Then the load was applied on the tractor until a fix reduction in speed was achieved. Subsequently, PTO and engine speed, torque and power, fuel consumption, exhaust gas temperature, noise level and exhaust gas emission were monitored. The same procedure was followed for the HOT test, however, the engine speed without load was fixed on 2000 RPM (PTO speed about 415 RPM). In each test, the tractor engine was warmed up for 30 minutes using PD fuel. In order to have consistent readings, the tests were conducted under the same atmospheric conditions where humidity, atmospheric pressure and temperatures are relatively constant. The dynamometer was carefully calibrated prior to the tests. Each test was repeated three to five times and average data was used and reported in this work. Tables 3 and 4 show the descriptive statistical and ANOVA tests for the tractor test at WOT for two speeds from the performance experiment. The first (point) speed was rated PTO speed 540 RPM (engine speed 2600 RPM). The second (point) speed was at peak PTO torque (engine speed 1500 RPM). In these tables, the F values show the (statistically) significant differences between PD and MCP-B20 fuel for all the parameters in this study.
Tractor Engine Performance
In this work, the tractor PTO test was conducted to evaluate and compare the tractor engine performance using PD and MCP-B20 at WOT and HOT. The input power (IP), torque, brake power (BP), brake specific fuel consumption (BSFC), noise level and exhaust temperature are presented in the following sections. Figures 3 and 4 show the relationship between the tractor engine speeds and the engine gross input power (GIP) and engine brake power (BP) using PD and MCP-B20 for the WOT and HOT tests. The GIP power is the result of lower calorific value multiplied by the fuel flow rate. From Figures 3 and 4, it can be seen that the maximum GIP take place at the engine speed of around 2500 RPM for both fuels at WOT. The maximums GIP are 75.9 and 73.1 kW and the maximums for BP are 15.8 kW for PD and MCP-B20 respectively. Table 3, Figures 3 and 4 show that at the rated PTO speed (2600 RPM engine speed) there are significant reductions in GIP 5.1 kW and BP 0.77 kW when the tractor is fuelled by MCP-B20 in comparison to PD. This reduction is due to the difference in heating value between microalgae fuel and diesel. This finding agrees with Neel et al. [18] and disagrees with Kulkarni et al. [19] who found a significant difference and insignificant difference, respectively, between tractor PTO power between PD and B20 biodiesel at rated PTO speeds. There is no significant difference that can be seen in Table 4 at the point of maximum PTO torque between PD and MCP-B20, which agrees with the insignificant difference found between PD and B20 biodiesel at typical pumping speed and peak PTO torque in Kulkarni et al. [19] and Neel et al. [18] respectively. A similar trend is found for the HOT and WOT test, however the MCP-B20 shows higher GIP and BP at 1,800 RPM. These findings associated with results in Table 4 and Figures 3 and 4 show that the differences between the PD and MCP-B20 for the general case are not significant. The close results between PD and MCP-B20 are due to the higher density for MCP-B20 which reduces the effect of the reduction in the lower calorific value for MCP-B20 than PD. Figure 5 presents the tractor engine torque for the WOT and HOT tests against the tractor engine speeds. The maximum tractor engine torques are found at engine speeds between 1300 RPM and 1500 RPM for both fuels. The maximum difference between the PD and MCP-B20 for the WOT test of 5.4 Nm occurs at rated PTO speed (2600 RPM engine speed). The ANOVA test for this point in Table 3 showed that there is a significant different in tractor engine torque between the two fuels. This finding agrees with Neel et al.'s [18] results for PTO torque between PD and B20 fuel. However Kulkarni, et al. [19] found no significant difference in the engine torque between PD and B20 fuel. At the peak torque point, and all points in Figure 5, the results show comparable outcomes in the WOT test. These results agree with Kulkarni et al.'s [19] and Neel et al.'s [18] findings at 1,800 RPM and the peak torque results, respectively. The HOT test shows that when the tractor is fuelled by MCP-B20, both a reduction and increment can be found when compared with PD. Figure 6 shows that the engine efficiency decreases when engine speeds increase for both fuels. The higher engine efficiency at low engine speeds comes from high load applied which reduced engine speed. This reduction comes from low heat losses at high load [23]. The HOT test shows that MCP-B20 excels in giving higher engine efficiency for the speeds below 1700 RPM. The higher efficiency for MCP-B20 comes from the lower IP in spite of the slightly lower BP to PD. That is related to the better combustion that comes from the extra oxygen in biodiesel [24]. Tables 3 and 4 illustrate that there are no significant differences in tractor engine efficiency between PD and MCP-B20 at the rated PTO speed and for the peak torque. This is in agreement with the finding of Kulkarni et al. [19] and Neel et al. [18] as they found no significant differences in thermal efficiency at rated speed, pumping and peak torque between PD and B20 biodiesel.
Brake Specific Fuel Consumption (BSFC)
The BSFC was calculated by dividing the fuel consumption rate by the BP. Figure 7 illustrates the relationship between the tractor engine speeds (RPM) and engine BSFC (g/kWh) for PD and MCP-B20. However this figure shows an increase in the BSFC when the tractor is fueled by MCP-B20 than the PD at most engine speeds except at 2500 RPM and 900 RPM for the WOT test, lower BSFC is found at 1500 RPM for the HOT test. Statistically, Table 3 and Table 4 show that there are no significant differences in BSFC between the two fuels at rated PTO speed and at the peak torque points. That is due to the high SD found at this parameter.The increase in BSFC is understandable due to the lower calorific values of MCP-B20 compared with that of mineral diesel. However, the density of MPC-20 is higher than that of diesel. Figure 8 shows the tractor exhaust temperatures (°C) against the tractor engine speeds (RPM) for PD and MCP-B20. From the WOT and HOT tests it can be seen that the exhaust temperatures increase when the engine speed increases until reaching the maximum, then the temperatures decline. The maximum exhaust temperatures for the WOT test for both fuels are between the engine speeds of 1700 RPM to 2500 RPM, while for the HOT test the maximum exhaust temperatures are found at the engine speeds between 1300 RPM to 1700 RPM. Tables 3 and 4 show that there are no significant differences between the two fuels at the rated PTO speed and peak torque points in tractor exhaust temperatures. The maximum difference between the two fuels of 59 o C is found at 1300 RPM for the WOT test.
The relationship between tractor engine speeds (RPM) and noise level (db) measured at the tractor operator head position for PD and MCP-B20 are presented in Figure 9 at WOT and HOT. The noise level in this figure mostly originates from the tractor engine. However, there is considerable noise that comes from the PTO dynamometer and its shaft which couldn't be avoided during the test. That makes the noise level at all speeds considerably high. The noise coming from the dynamometer is not constant at all engine speeds, it is affected by tractor vibration which is the main cause for high levels of noise at the speeds of 2500 RPM and 1100 RPM for the WOT test as well as 1900 RPM for the HOT test. At these speeds the tractor vibration is significantly greater than at other speeds. The noise caused by the tractor vibration that contributes to the tractor noise is insignificant when fueled by PD and MCP-B20 as revealed in Tables 3 and 4. The peak noise occurs at the engine speed 2500 RPM for both fuels. At this point there is a considerable load associated with high engine speed that result in the highest noises of about 92 db and 92.1 db for PD and MCP-B20 respectively for the WOT test.
Carbon Monoxide (CO) and Carbon Dioxide (CO 2 )
The comparison of CO emissions from PD and MCP-B20 in Figure 10 shows that MCP-B20 produced a significantly less CO% than PD at most points in the WOT test and at most the engine speeds in the HOT test except 1800 RPM, at which MCP-B20 gives a slightly higher CO%. This result is in a good agreement with the Nabi et al. [1] conclusion that biodiesel from cotton seed oil produces less CO than net diesel. Less CO is an indication of better combustion in MCP-B20 than PD which comes from the extra oxygen in the biodiesel form. This significant reduction in CO% is confirmed by ANOVA in Table 3 and Table 4 that show a significantly high reduction in CO% produced from the tractor engine when fuelled by MCP-B20 compared to PD at the rated PTO speed and peak torque. The maximum differences between the concentration of CO are clearly noticable at the lower engine speeds ranging between 1300 RPM and 1100 RPM for both the WOT and HOT tests, and these results agree Yusaf et al. [25], however, there is a disagreement with the higher results found in CO PPM produced from the engine fuelled by B25 of crude palm oil biodiesel when compared with diesel fuel. Figure 10. The relationship between tractor engine speeds (RPM) and CO (%) for PD and MCP-B20. Figure 11 depicts the effect of tractor engine speeds (RPM) on CO 2 (%) for PD and MCP-B20. It is seen that the CO 2 (%) is relatively steady for the engine speeds below 2500 RPM for the WOT test and below 1700 RPM for the HOT test. The percentages of CO 2 then declined dramatically during both throttle tests for both fuels. The reason for this declination can be due to the lower load applied on the engine at the high speeds and the considerably lower fuel consumption rate makes the engine produce lower CO 2 (%). Figure 10 and Table 4 show that both fuels present very comparable percentages of CO 2 at most engine speeds. On the other hand, Table 3 illustrates that MCP-B20 significantly emits lower CO 2 % than PD at rated PTO speed. This may be due to the reason that, in a complete combustion, the lower carbon to hydrogen ratio in biodiesel results in producing less CO 2 emissions from biodiesel than diesel [24]. The advanced air-fuel ratio with the increasing engine speed has improved the air-fuel mixing process in the combustion chamber which resulted in higher burning gas temperatures. This condition converts more CO emissions to CO 2 emissions with much more CO 2 production.
It is important to mention here that PM, UTHC and CO are known products of incomplete combustion. One of the advantages of using microalgae fuel in diesel engine is to reduce black smoke in the exhaust gas. The formation of soot from exhaust emission is proportional with the change of combustion temperature. However due to unavailability of UTHC and soot measurement, the UTHC and PM were not reported. Figure 11. The relationship between tractor engine speeds (RPM) and CO 2 (%) for PD and MCP-B20.
Oxygen (O 2 ) and Lambda (λ)
The results of the WOT test and HOT test which display the O 2 percentage against the tractor engine speeds are given in Figure 11. In stoichiometric air-fuel ratio and complete air-fuel mixing, the presence of O 2 in the exhaust gases gives an indication of the quality of combustion. It can be seen from Figure 12 that, at high engine speeds, the existence of O 2 in the exhaust gases increase for both fuels and both the WOT and HOT tests. This increase can be due to the higher air-fuel ratio at high engine speeds that come from low loads applied on the engine at high speeds. This result is confirmed by lambda results in Figure 13 which present a comparable pattern between O 2 and lambda. Diesel engines normally run with maximum air flow, while the fuel can be controlled. This shows the difference between WOT and HOT in which the air is constant but the fuel flow is reduced for the HOT test which justifies the lower O 2 percentage for the HOT test compared to the WOT test at the same speeds. Comparable results of O 2 % are found between MCP-B20 and PD as shown in Figure 11 and Table 4. However, Table 3 shows a very high increase in O 2 % when the tractor is fueled by MCP-B20 compared to PD at the rated PTO speed. As the same speed there is no significant difference in the engine efficiency between the two fuels, the reason for the higher O 2 % may be due to the extra Oxygen in the biodiesel structure.
Lambda is the actual air-fuel ratio divided by the stoichiometric air-fuel ratio. When Lambda is higher than one, the engine runs lean, which is normal in diesel engines [22,26]. Lambda's values for the tractor engine when running on MCP-B20 and PD at WOT and HOT tests are graphically presented in Figure 13. Figure 14 shows the effect of the engine speeds on the NO emission measured by (PPM) for the tractor engine fuelled by PD and MCP-B20 for the WOT and HOT tests. A similar trend can be observed for both fuels for the WOT and HOT tests in the emission of NO. It can be seen from Figure 14 that NO emissions decrease when engine speeds increase by reducing the toque applied on the engine. Substantial differences are found between low and high engine speeds for both fuels. The maximum values of 1126 PPM and 1194 PPM at 900 RPM are found to decline to 170 PPM and 161 PPM for PD and MCP-B20 respectively. The increase of nitrogen oxides in diesel engines is due to the sufficient amount of oxygen, in-cylinder temperatures and the residence time of reaction at elevated temperatures. It can be seen from Figure 14 that MCP-B20 emitted slightly higher NO levels than PD at all the engine speeds for the HOT test and at most engine speeds at WOT. Table 3 and Figure 14 show that the increase in NO emission of MCP-B20 at WOT in comparison with PD is [19] and Neel et al. [18] findings as they reported an insignificant difference between the NO x from PD and biodiesel B20.
Conclusions
Microalgae is environmentally friendly, has high biomass and lipid productivity, and is a non-edible oil, thus it can be excellent alternative fuel for diesel engines in agricultural tractors. The results of this work demonstrated that MCP-B20 can be used commercially as fuel for tractors with no modification. The PTO tests for the tractor engine at WOT and HOT show that the tractor power, torque and BSFC at MCP-B20 and PD are close and acceptable. The emission results demonstrate great reduction in CO and CO 2 when MCP-B20 was used, which are very encouraging outcomes. The ANOVA analysis for two points at maximum PTO torque and power shows that there are significant reductions in the values of toque, power, CO and CO 2 emissions of MCP-B20. The next phase of this project will be focusing on using microalgae MCP-B100 in the same tractor. The results will be assessed and compared with PD. | 6,553.4 | 2013-02-06T00:00:00.000 | [
"Engineering",
"Environmental Science"
] |
Online Collision-Induced Unfolding of Therapeutic Monoclonal Antibody Glyco-Variants through Direct Hyphenation of Cation Exchange Chromatography with Native Ion Mobility–Mass Spectrometry
Post-translational modifications (PTMs) not only substantially increase structural heterogeneity of proteins but can also alter the conformation or even biological functions. Monitoring of these PTMs is particularly important for therapeutic products, including monoclonal antibodies (mAbs), since their efficacy and safety may depend on the PTM profile. Innovative analytical strategies should be developed to map these PTMs as well as explore possible induced conformational changes. Cation-exchange chromatography (CEX) coupled with native mass spectrometry has already emerged as a valuable asset for the characterization of mAb charge variants. Nevertheless, questions regarding protein conformation cannot be explored using this approach. Thus, we have combined CEX separation with collision-induced unfolding (CIU) experiments to monitor the unfolding pattern of separated mAbs and thereby pick up subtle conformational differences without impairing the CEX resolution. Using this novel strategy, only four CEX–CIU runs had to be recorded for a complete CIU fingerprint either at the intact mAb level or after enzymatic digestion at the mAb subunit level. As a proof of concept, CEX–CIU was first used for an isobaric mAb mixture to highlight the possibility to acquire individual CIU fingerprints of CEX-separated species without compromising CEX separation performances. CEX–CIU was next successfully applied to conformational characterization of mAb glyco-variants, in order to derive glycoform-specific information on the gas-phase unfolding, and CIU patterns of Fc fragments, revealing increased resistance of sialylated glycoforms against gas-phase unfolding. Altogether, we demonstrated the possibilities and benefits of combining CEX with CIU for in-depth characterization of mAb glycoforms, paving the way for linking conformational changes and resistance to gas-phase unfolding charge variants.
M onoclonal antibodies (mAbs) have played a dominant role in the treatment of various disorders, 1,2 and this class of human therapeutics is still rapidly expanding. Currently, most therapeutic mAbs are IgG1 subclass, but other IgG subclass formats are also being produced (IgG2 or IgG4). 3 Besides a variation in subclass, therapeutic mAbs can contain a plethora of different post-translational modifications (PTMs), including lysine clipping, pyroglutamate formation, glycosylation, oxidation, and deamidation. 4,5 Changes in the PTM profile may influence the conformation and thereby impact biological functions (i.e., efficacy and safety of therapeutic proteins). For instance, mAb charge variants, such as deamidated species and proteoforms with isomerization of asparagine residues, may result in decreased binding affinity and potency. 6 Novel strategies are needed to fully characterize charge-related structural changes to ensure efficacious and safe therapeutic products.
Over the last decades, cation-exchange chromatography (CEX) coupled with ultraviolet (UV) detection has found its way into quality control as a reference technique to monitor charge variant profiles of mAb-based therapeutic products. 7,8 Nevertheless, these robust and routinely used methods often lack the ability to perform in-depth structural characterization without time-consuming fraction collection followed by offline mass spectrometry (MS). Fortunately, recent advances dealing with interfacing CEX and native MS (nMS) enabled the development of methodologies providing online separation, identification, and characterization of charge variants. 9 Most importantly, the replacement of traditional non-volatile salt gradients with low ionic strength pH gradients using (MScompatible) ammonium-based mobile phases boosted the application of CEX−nMS for charge variant analysis. 5,10,11 While these pH gradients already result in sufficient separation efficiency for most proteins, proteins with high isoelectric points (pIs) or heterogeneous proteins consisting of many proteoforms may benefit from a pH gradient accompanied by a (minor) increase in the salt concentration, a so-called saltmediated pH gradient. 10,12 Especially, the latter enabled online identification of mAb proteoforms with altered pI, such as proteoforms with differences in glycosylation, deamidation, and isomerization. 7,13,14 Currently. CEX−nMS applications focus largely on structural characterization of charge variants, while their effect on the protein conformation or gas-phase unfolding pattern is not explored.
To decipher the conformational landscape of mAbs, ion mobility (IM)-based technologies are particularly suitable. 15 The availability of traveling wave IM spectrometry in commercial MS instruments opened many possibilities to study global protein conformation in the gas phase. 16 Unfortunately, IM−MS measurements often have low resolving power for species with related conformations. As a consequence, the collision cross sections (CCSs) obtained from these measurements are often not very informative for intact mAb analysis. 3,17 To circumvent this poor resolution as well as gain deeper insights into the gas-phase behavior after activation, IM-based collision-induced unfolding (CIU) has been proposed as a suitable alternative. So far, CIU has been employed to investigate different properties and modifications of mAbs 3,17−19 and antibody−drug conjugates. 20,21 While these classical CIU approaches face various practical challenges preventing routine use, including manual buffer exchange and time-consuming data acquisition process, online coupling of size exclusion chromatography (SEC) with CIU allowed automation of this workflow tackling these challenges. 22 Here, we present an innovative online CEX−CIU method for the in-depth characterization of different mAb populations in their native state in a fast and straightforward manner. Using this approach, mAb populations were separated according to their pIs, and their specific unfolding patterns were acquired by increasing the collision voltage (CV) in the trap cell prior to IM separation during elution of the selected mAb population. This analytical strategy was developed to record the CIU fingerprints of intact or IdeS-digested mAbs showing the capability of this technique to generate CIU fingerprints of CEX-separated protein populations. Subsequently, CEX−CIU hyphenation was tailored to decipher conformational differences of an equimolar mixture of three nearly isobaric mAbs. Finally, the potential of CEX−CIU for charge variant analysis is illustrated by studying gas-phase unfolding resistance of different remodeled glycan-variants of a reference mAb. ■ EXPERIMENTAL SECTION Materials. Eculizumab (Soliris, Alexion), trastuzumab (Herceptin, Roche), and pembrolizumab (Keytruda, Merck) were obtained from their manufacturers. Glycan remodeling was performed by treatment of trastuzumab with Trans-GLYCIT (Genovis, Lund, Sweden) according to the producer specifications with and without the presence of a fucosidase. For middle-level analysis, enzymatic digestion was performed by incubation of one unit of the IdeS enzyme (FabRICATOR, Genovis, Lund, Sweden) per microgram of mAb at 37°C for 60 min. Prior to further analysis, the samples were buffer exchanged to 50 mM ammonium acetate using 10 kDa Vivaspin 500 filters (Sartorius, Goẗtingen, Germany) with an end concentration of 2 mg/mL. CEX Separation. The CEX measurements were performed using an Acquity UPLC H-class system (Waters, Wilmslow, UK) equipped with a quaternary solvent manager, a sample manager operated at 10°C, a column oven, and a TUV detector. A BioResolve SCX column (2.1 × 100 mm, 3 μm) was used from Waters. The mobile phases were composed of 50 mM ammonium acetate at pH 5.0 (A) and 160 mM ammonium acetate at pH 8.6 (B). For intact trastuzumab, a linear gradient from 50 to 70% B in 10 min was applied. The middle-level trastuzumab separation was achieved by a linear gradient from 30 to 70% B in 10 min. The mixture of mAbs was separated by first maintaining 0% B for 1 min followed by a linear gradient from 30 to 70% B in 9 min. All CEX methods contained a column cleaning step at 100% B for 3 min and a re-equilibration step at 0% B for 7 min. The injected amount of intact mAbs was 20 μg and that of middle-level mAbs was 12 μg. The flow rate was 0.1 mL/min, and the UV wavelengths were 280 and 214 nm.
CEX−ESI−CIU Experiments. CEX was coupled online with a Synapt G2 HDMS mass spectrometer (Waters) operated in positive mode with a capillary voltage of 3 kV and a sample cone voltage of 180 V. The backing pressure of the Z-Spray source was increased to 6 mbar. Desolvation and source temperatures were set to 450 and 90°C, respectively. Desolvation and cone gas flow rates were 750 and 60 L/h, respectively. The cone voltage of the Synapt G2 was lowered to 80 V (for intact analysis) or 60 V (for middle-level analysis). The argon flow rate was set to 2 mL/min. Prior to the IM cell, ions were focused in the helium cell (120 mL/min). In the IM cell, the N 2 flow of 60 mL/min was applied, and the wave height and velocity were 40 V and 800 m/s. Other settings were similar as mentioned for the CEX−nMS analysis. The m/ z range was from 1000 to 10,000. External calibration was performed using cesium iodide (2 mg/mL in 50% isopropanol).
For the CEX−CIU experiments, the CV in the trap cell was stepwise increased from 0 to 150 V (in steps of 10 V). One complete CIU fingerprint was obtained in four CEX runs, where each CEX run contained four IM−MS functions per selected species (0−30, 40−70, 80−110, and 120−150 V). For each of these voltage steps, the number of scans and scan time were 4 and 0.2 min, respectively. The start and end times of the functions were adjusted to the retention time of the species of interest, and between the functions during elution, a window of 0.5 s was kept ensuring effective application and stable CVs.
ESI−CIU Experiments. Intact or middle-level trastuzumab was infused in an electrospray ionization (ESI) source using a syringe pump (KD Scientific, Holliston, MA, USA) with a 250 μL syringe (Hamilton, Bonaduz, Switzerland) at a rate of 3 μL/min. The parameters of the Synapt G2 were as described for CEX−CIU experiments. The CV in the trap cell was manually increased from 0 to 150 V in steps of 10 V, and from each CV, 0.5 min was acquired.
Data Analysis. The CEX−IM−nMS data were processed with MassLynx (v4.1), and data from CIU experiments were analyzed and visualized using CIUSuite 2 (v2.2) 23 or ORIGAMI ANALYSE (v1.2.1.6). 24 The chromatographic resolution was calculated with R s = 1.18 × ((t R2 − t R1 )/(W FWHM1 + W FWHM2 )), where t r is the retention time (min), and W FWHM is the peak width at half-height (min). The drift times obtained with CEX−IM−nMS were converted into CCS values using external calibrants, including β-lactoglobulin and concanavalin A for the middle level and concanavalin A, alcohol dehydrogenase, and pyruvate kinase for the intact level. 25 Arrival time distributions (ATDs) were smoothed with the Savitzky−Golay algorithm with a window length of 5 and polynomial order of 2. The data were interpolated with a factor of 2 on the CV axis. For all CIU analyses, averaged and differential plots, including root-mean-square deviation (RMSD) values, were obtained. Feature detection and CIU50 analysis were performed using standard mode, where the minimum feature length was 2 steps, the feature allowed width was 0.75 ms, no CV gap length within a feature was allowed, drift time spectrum was the centroid at the maximum value for each CV, and transition region padding was 15 V. The univariate feature selection (UFS) plots were generated to highlight diagnostic voltages [i.e., high −log10(p-value) scores] to classify mAb subclasses.
CEX for mAb Charge Variant Separation at Intact and Middle Levels.
According to the separation capabilities of CEX for therapeutic mAbs, we implemented two different CEX−nMS methods to monitor charge variants of a reference mAb (trastuzumab) at both the intact and middle levels. Most efficient separation of charge variants of trastuzumab was achieved using a salt-mediated pH gradient with mobile phases composed of 50 mM ammonium acetate at pH 5.0 (A) and 160 mM ammonium acetate at pH 8.6 (B). Applying these mobile phases with a gradient from 50 to 70% B, we resolved acidic and basic species from the main form of intact trastuzumab ( Figure 1a) with similar separation efficiencies as previously reported. 10,11 The main form of trastuzumab eluted at 13.6 min with a full width at half-maximum (FWHM) of 0.77 min. Besides this form, additional acidic and basic variants were separated, which could be assigned based on previous (bottom-up) data. 10,11,13,26 The acidic species, including deamidated variants, eluted at 11.9 min and were baseline resolved from the main peak (resolution of 1.6). Basic charge variants due to isomerization of Asp residues eluted later at 14.9 min (resolution of 1.2) (Figure 1a; Table S1).
Although analysis of intact mAbs can be very informative, CEX−nMS on mAb subunits upon IdeS digestion 27 (i.e., cleavage of mAb below the hinge region) allowed localization of charge variants at the subunit level along with more precise mass measurements facilitating PTM identification 17 ( Figure 1b). The CEX separation of the Fc and F(ab′) 2 fragments was accomplished using the same mobile phases with a gradient from 30 to 70% B (Figure 1b; Table S1). While the Fc fragment eluted as a single peak at 8.0 min, multiple peaks were observed for the F(ab′) 2 domain. Besides the main F(ab′) 2 peak at 13.9 min, also acidic and basic species were separated at 12.3 and 14.8 min, respectively. The latter indicated that the deamidation and isomerization occur on the F(ab′) 2 domain, which is also consistent with published data. 26,28 The peak width of the Fc peak was smaller, whereas the F(ab′) 2 peak was very similar to intact trastuzumab. This is in line with expectations since low ionic strength mobile phases have been described to focus on the chromatographic peaks, while higher ionic strength results in broader peaks. 8 In our case, the pH gradient starts at low ionic strength (50 mM) and increases in ionic strength with time.
Coupling of CEX and CIU. We then aimed at direct coupling of CEX and CIU to gain information on the conformation of chromatography-resolved protein populations. Similar to our previously described SEC−CIU optimization procedure, 22 all critical CIU parameters were adapted to obtain successful hyphenation of CEX and CIU, including the acquisition time, magnitude, and number of CV steps. The CV in the trap was increased from 0 to 150 V in steps of 10 V. The maximum CV of 150 V was selected to minimize the risk of fragmentation and obtain high-quality data while still covering the most diagnostic CIU energy range of mAbs. Furthermore, a cone voltage of 80 V was sufficient for intact mAb analysis, whereas for the middle level, 60 V was chosen to prevent changes in mAb conformation ( Figure S1). To find the optimal number of CV scans per run and the acquisition time, the CEX peak width (around 0.80 min) as well as the spectral quality was considered. We compared the effect of four CVs per chromatographic peak (each CV was applied for 20 s during which five scans of 4 s were recorded) with six CVs (15 s for each CV during which five scans of 3 s were recorded) ( Figure S2). Even though the latter condition reduced the total analysis time (three 20 min CEX−CIU runs instead of four), noisy CIU features are obtained above 100 V due to a lower signal-to-noise ratio. Therefore, a total of four CEX−CIU runs (0−30, 40−70, 80−110, and 120−150 V) were chosen to record a complete high-quality CIU fingerprint ( Figure 2). Importantly, the lowest voltages of the run covering the 0−30 V range should be recorded at the peak maximum due to low desolvation and trapping efficiencies at these low voltages Analytical Chemistry pubs.acs.org/ac Article together with a low-intensity signal at the beginning of the chromatographic peak ( Figure S3). Compared to the previously reported SEC−CIU method, 22 CEX−CIU required a longer analysis time to record the whole dataset from 0 to 150 V. More precisely, a complete CEX−CIU fingerprint will take 80 min (four times 20 min) compared to 15 min (three times 5 min) for SEC−CIU. However, CEX−CIU supplemented the CIU workflow with a chromatographic separation dimension rather than solely automated sample introduction in the MS instrument as for SEC−CIU, counterbalancing the longer acquisition times with improved separation power. To benchmark our method development, we first applied the optimized CEX−CIU method to the analysis of the reference mAb trastuzumab, resulting in four different states within the whole CIU fingerprint (Figure 2b). The conformational transitions between these four different states were observed at 18.2 ± 0.2, 34 ± 2, and 91 ± 3 V, respectively. One feature of this fingerprint is the coexistence of unfolding states 2 and 3 over a wide voltage range (between 40 and 120 V). This particular CIU signature has already been reported for IgG1 fingerprints of high charge states (27+), 22 being in good agreement with the isotype subclass of trastuzumab. The CIU fingerprints of the CEX-separated Fc and F(ab′) 2 domains were recorded using similar experimental parameters ( Figure S4). The Fc subunit (14+) showed two transitions at 27 ± 2 and 116 ± 3 V ( Figure S4a). To provide more insights into the latter transition, a manual inspection of the mass spectra of the Fc subunit along with the analysis of the corresponding ATD profile was performed ( Figure S5). According to these results, the upper limit of the Fc fingerprint was set to 130 V as further activation led to fragmentation of molecular ions and thus the total loss of the MS signal. As depicted in Figure S5, the most unfolded state (around 12 ms) starts to be populated at 30 V and becomes the most intense conformation in the ATD at 40 V. Beyond 90 V, this conformational state starts to be less populated, favoring the increase of an additional unfolding state with a shorter drift time (around 10 ms). Although further analysis should be carried out to confirm the structural mechanism behind this conformational change, these results could suggest that the Fc subunit of mAbs can undergo compaction under specific experimental conditions. For the F(ab′) 2 part (21+), three transitions were observed giving rise to four progressively unfolding states ( Figure S4b). These transitions are observed at 22.5 ± 0.1, 72 ± 7, and 128 ± 7 V, respectively.
Similar CIU fingerprints of mAbs and mAb subunits have been previously reported, 3,17,22 indicating that the obtained CIU fingerprints of trastuzumab were not affected by the upfront CEX separation. The RMSD value between CEX− ESI−CIU and ESI−CIU at the intact level was 6.6% ( Figure S6a), in line with differences observed in previous studies using the combination of SEC with CIU. 22 The RMSD between the middle-level analyses using CEX−ESI−CIU and ESI−CIU was 6.5% ( Figure S6b). In both cases, the obtained RMSD was similar to the RMSD values between replicates ( Figure S6). The slight differences observed at the intact level between the inter-and intra-RMSD values can stem from the fact that both analyses were not recorded at the same flow rate (4 μL/min for ESI−CIU and 250 μL/min for CEX−ESI−CIU) along with the variation of the IM−MS signal transmission as a consequence of the chromatography peak elution when recording CIU fingerprints in combination with CEX. However, the RMSD values pinpoint that online CEX−CIU affords comparable and consistent CIU results, leading to the conclusion that neither the fractionation of the CIU recording nor the interaction with the CEX stationary phase significantly impairs the quality of the CIU fingerprints of therapeutic mAbs.
Proof of Concept: CEX−CIU for Nearly Isobaric mAb Mixture Analysis. Since the CEX−CIU approach was shown to be well suited to monitor the unfolding mechanism of different chromatography-selected populations, this analytical strategy was applied to a mixture of three therapeutic mAbs at the intact level, as a proof of concept of the methodology. The proposed application can be particularly interesting for biopharmaceutical companies since the simultaneous characterization of individual components contained in an mAb mixture is gaining more attention due to novel promising coformulated mAb products. 29 The selected mAbs belonged to different isotype subclasses exhibiting differences in inter-chain disulfide patterns. Previous studies have shown that the CIU mechanism strongly depends on the inter-chain disulfide pattern of mAbs, and therefore, the CIU fingerprints of these mAbs should show different unfolding energies and number of transitions. Three therapeutic mAbs with relatively similar molecular masses but substantially different pIs (i.e., 6.1 for eculizumab, 7.6 for pembrolizumab, and 9.1 for trastuzumab 30 ) were chosen for the mAb mixture analysis. Thereby, the relatively small difference in molecular masses hampered studying unfolding patterns with classical direct infusion Analytical Chemistry pubs.acs.org/ac Article techniques or applying non-denaturing size-based separations, such as SEC. 22 In this case, CEX−CIU is a suitable alternative, since mAbs can be separated based on pIs prior to the nMS analysis (Table S2). Additionally, this approach allows simultaneous monitoring of minor charge variants within the mAb mixture. The CEX gradient started by maintaining 0% B for 1 min to ensure elution of eculizumab, whereafter a linear increase from 30 to 70% B in 9 min enabled separation of pembrolizumab and trastuzumab. The obtained CEX chromatogram showed three (major) distinct peaks corresponding to the different mAbs (Figure 3a,b), where eculizumab elutes at 1.8 min, pembrolizumab at 8.9 min, and trastuzumab at 15.1 min.
After optimizing the CEX separation, the CIU fingerprints of the separated mAbs were recorded and compared (Figure 3c). Using this approach, four CEX−CIU runs of 20 min (in total, 80 min) were required to obtain complete CEX−CIU fingerprints for the three mAbs. A clear advantage of this approach is the low risk of interferences in the CIU pattern coming from different species since the three mAbs elute at different retention times. To illustrate the detected differences, the CIU50 values were calculated (Figure 3d). In the case of trastuzumab, three different unfolding events were observed, while only two transitions were detected for eculizumab and pembrolizumab. Thereby, a CIU50 value of 95.1 ± 0.2 V was determined for this last transition (CIU50-3), being the sole mAb that undergoes unfolding transition at collision energies higher than 50 V. The RMSD between trastuzumab measured as part of the mAb mixture and measured as single mAb was 6.1%, whereas the RSMDs between the different mAb subclasses within the mixture were ranging from 11.2 to 18.8% ( Figure S9). To strengthen these results, linear discriminant analysis was performed (Figure 3e), showing the ability to distinguish the three mAbs based on their CIU features. This indicates that the CIU patterns of the mAbs are indeed different, and no additional variation was introduced upon mixing the mAbs. Additionally, the UFS plot pinpointed the 120−140 V region as the most diagnostic energy range to perform the inter-subclass comparison (Figure 3f), being in agreement with previous reports. 17,31 After the analysis of these results, the benefit of the CEX−ESI−CIU approach is clearly shown to study a co-formulation mixture of therapeutic mAbs providing a separation of the different components in the first dimension (CEX) and subsequently recording the unfolding pattern of each protein population. The outcome of this approach not only improves the characterization of coformulated mAbs without jeopardizing the chromatographic resolution but also reduces the overall analysis time by acquiring multiple CIU fingerprints at once.
CEX−ESI−CIU To Monitor Differences in Gas-Phase Unfolding of mAb Glyco-Variants.
As previous reports have shown that glycosylation impacts the gas-phase unfolding of mAbs, 18,32 we applied the novel CEX−CIU platform to study the influence of glycosylation, particularly sialyation, on the gas-phase unfolding pattern as a particular type of charge variants.
The different glycoforms of trastuzumab were created using glycan-based enzymatic remodeling (see the Experimental Section for details on the procedure). The heterogeneous pool of glycans present on the initial product (T0) was first partially removed (EndoS2 treatment), leading to the deglycosylated mAb (T1) followed by replacement with a particular glycoform (G2S2 with or without core F) (Figure 4a). The remodeled glycans contained additional negative charges from the sialic acids, which resulted in modification of the protein surface charges. Since we were interested in differences in glycosylation located on the Fc part, the glycoengineered trastuzumab formats were digested with the IdeS enzyme, and Fc and F(ab′) 2 subunits were further analyzed by CEX−CIU. As expected, the Fc fragments with remodeled glycans showed a shift to lower retention times in CEX, specifically 6.0 min for G2S2 and 6.3 min for G2S2F compared to 8.0 min for T0 (Figure 4b). Notably, the Fc fragment carrying G2S2F showed a minor peak around 7.8 min, indicating that the reaction was not complete and a minor amount of the T0 product was still present. However, due to the CEX baseline separation of the different Fc fragments, there was no inference of these species in the region where the CIU fingerprints were recorded. The peaks of the F(ab′) 2 fragment eluted at the same retention time (14.5 min) for all samples and showed similar mass spectra, suggesting that no other modifications are induced on this part during the sample treatment as expected ( Figure S10).
The CIU fingerprints of the Fc and F(ab′) 2 fragments were acquired (Figures 5, S11, and S12). As previously mentioned, the Fc CIU fingerprints were recorded until 130 V to reduce Analytical Chemistry pubs.acs.org/ac Article ion activation and ensure enough IM signal intensity. However, substantial differences in unfolding behavior between glycoforms were observed in the fingerprints of the Fc domain in the 0−130 V range. This clearly highlights the benefit of CIU workflows over conventional IM−MS, which was not able to distinguish the different glycoforms in their ground state ( Figure S13). Comparing the CIU fingerprints of the glycosylated T0 product and the deglycosylated T1 product revealed the same number of transitions. Nevertheless, some differences can be observed regarding the gas-phase kinetic stability of the unfolding features of both subunits. For instance, the most unfolded conformation of T0 (12 ms) is the most populated conformation at 110 V, whereas for T1, this conformation undergoes transition at lower collision energies, and thus, the conformation observed at 10 ms is the most populated state at 110 V (Figure 5b,c). This indicates that deglycosylation of trastuzumab results in an Fc domain more prone to gas-phase unfolding, which is in agreement with previously performed differential scanning calorimetry (DSC) as well as CIU experiments. 18,20,33,34 Regarding the CIU fingerprints of sialylated Fc variants, differential features can also be detected upon inspection of the unfolding patterns. Interestingly, an intermediate unfolding transition can be observed around 52 V for the G2S2 and G2S2F products, which was not observed for T0 or T1 (Figure 5a,c). Furthermore, the most unfolded conformation of those products (state 3) was still detected as the most intense conformation at the highest voltage (130 V). These results clearly pinpoint that the presence of sialic acid glycans confers higher unfolding resistance to the Fc domain of the mAb. According to our results, the presence of core fucoses on these glycans does not influence the gas-phase unfolding mechanism of these products reflected in very similar unfolding patterns. The latter was also confirmed by previously reported DSC data, where afucosylated IgG1 glycoforms did not show different thermal unfolding compared to the fucosylated counterpart. 35,36 Overall, the differences between the glycan-remodeled Fc subunits were highlighted when performing differential analysis between the fingerprints ( Figure S11). Upon pairwise comparison, RMSD values greater than 14% were obtained (excluding the comparison between G2S2 and G2S2F, where almost no differences were observed), while the variation of the technical replicates was lower than 6%. According to the UFS plot, the most diagnostic voltages to differentiate the glycoengineered Fc domains were between 100 and 120 V (Figures S11c and 5b). In addition, the Fc CIU features showed differences in the 20−60 V range. The extracted ATDs at 40 V (Figure 5d) revealed clear-cut differences between the glycoengineered Fc domains compared to T0 and T1, with a different number of coexisting populations along with different relative intensities, allowing the categorization of the Fc species. As above mentioned, the showing that differences in the ATD profile can also be detected in the low-energy range of the CIU fingerprints. The F(ab′) 2 CIU fingerprints showed no differences in resistance to gas-phase unfolding ( Figure S12).
Analytical Chemistry pubs.acs.org/ac
Article presence of fucose does not significantly alter the Fc gas-phase unfolding mechanism and hence the high similarity of G2S2 and G2S2F ATD profiles. The CIU patterns of the F(ab′) 2 subunits revealed the same number of unfolding transitions along with very similar CIU50 values. Upon pairwise comparison of the F(ab′) 2 fingerprints, the RMSD values were close to those obtained upon technical replicates which evidenced the similarities between the CIU fingerprints of those domains, suggesting that the F(ab′) 2 unfolding mechanism was not affected by the Fc glycan scaffold ( Figure S12). The same conclusion was drawn in previous studies based on DSC measurements, demonstrating that the melting temperatures of the Fab and CH3 domains of different N-glycosylated mAbs were not affected by the presence of different glycan moieties. 37 Altogether, the CEX−CIU approach was perfectly adapted to monitor structural modifications of mAbs while simultaneously obtaining information on the unfolding mechanism of the chromatography-selected species. The proposed experimental setup allowed the correlation between the glycan-engineered scaffold and unfolding signatures in the CIU fingerprints, providing valuable information on the influence of local structure alteration (i.e., altered glycan moieties) on the unfolding pattern of the Fc subunit of trastuzumab.
■ CONCLUSIONS
In this study, we aimed at the innovative hyphenation of CEX with CIU experiments to provide thorough characterization of therapeutic mAbs. This experimental setup combined the benefits of CEX separation to afford information about protein heterogeneity together with the CIU analysis to allow the characterization of those protein populations beyond their folded state. After developing this approach at both the intact and middle levels, we showed that the analytical strategy is well suited to record the global conformation and the CIU unfolding pattern of different CEX-separated species without affecting the resolution and chromatography separation capabilities. Moreover, the recorded CEX−CIU fingerprints contained a similar level of information compared to those obtained under classical ESI−CIU conditions.
To show the applicability of our innovative CEX−CIU platform, we focused on a mixture of three therapeutic mAbs close in mass but with different pIs and representatives from different subclasses (with different inter-chain disulfide patterns). The three CIU fingerprints exhibited different unfolding transitions leading to the conclusion that CEX− CIU was able to maintain the key features of the unfolding mechanism of each mAb, thus allowing the differentiation of those proteins. Furthermore, these results showed that the overall CEX−CIU analysis time could be significantly reduced when dealing with therapeutic mAb mixtures. Additionally, CEX−CIU coupling emerged as an appealing alternative in the particular case of isobaric mAb mixtures since these proteins cannot be separated using SEC−CIU. In this case, CEX−CIU affords a baseline-resolved separation according to mAb pIs; hence, CEX−CIU combines the synergy of separation based on pIs and CIU recording.
Besides, applicability of CEX−CIU to obtain glycan-variantspecific information was investigated by studying the gas-phase unfolding of Fc regions carrying different glycan moieties. The Fc CIU patterns showed variations in gas-phase conformational transitions as a result of altered glycosylation, while conventional IM−MS measurements were unable to provide clear conclusions ( Figure S13), highlighting the suitability of the CIU approach to study protein populations with similar CCS values. Based on the CIU patterns, it was shown that deglycosylated Fc fragments were more prone to unfolding events, while remodeled glycoforms showed increased resistance to gas-phase unfolding.
Altogether, the results provided in this study evidence that hyphenation of CEX with CIU has the potential to provide indepth characterization of therapeutic mAbs. First, CEX−CIU affords practical benefits, such as automation of CIU fingerprint recording while maintaining pI separation without any prior sample preparation, leading to an increased throughput. More importantly, CEX−CIU not only enables the identification of protein populations using classical CEX− nMS conditions but also offers the possibility of simultaneous conformational characterization of those chromatographyresolved populations. In future, the application of this approach could be further expanded to low-abundance charge variant analysis, including deamidation and oxidation, either in drug products or in forced degraded samples. Both the level of structural information and the streamlined analysis afforded by the CEX−CIU approach make this technique particularly interesting to be integrated into R&D laboratories performing analysis of mAb-derived proteins.
Influence of cone voltages on middle-level analysis; CEX−CIU fingerprints with four or six IM−MS functions; effect of the order of applied CVs during peak elution; CEX−CIU of middle-level trastuzumab; individual mass spectra and ATDs at different energies of the Fc subunit; comparison of ESI−CIU and CEX− ESI−CIU experiments; CIU50 analysis of multiplexed mAbs (27+); CIU fingerprints and CIU50 analysis of multiplexed mAbs (26+); CIU fingerprints and CIU50 analysis of multiplexed mAbs (28+); mass spectra of F(ab′) 2 fragments before and after glycan remodeling; RMSD plots of Fc fragments with remodeled glycosylation; CEX−CIU fingerprints of F(ab′) 2 domains before and after glycan remodeling; IM−nMS analysis of Fc domains with differences in glycosylation; retention times and relative peak areas of CEX separation; and characteristics of multiplexed mAbs (PDF) | 7,565.6 | 2023-02-15T00:00:00.000 | [
"Chemistry",
"Medicine"
] |
Effect of age on inter and intra-subject variability in acceptable noise level ( ANL ) in listeners with normal hearing
Article history: Received September 28, 2012 Received in revised format 19 November 2012 Accepted 20 November 2012 Available online November 27 2012 Several industrial engineering and psychological studies have shown that noise affects the important aspects of communication for both adults and children. For speech understanding by hearing aids users and language development in children, an accurate hearing is very important. A metric has been developed for measuring an individual's acceptance of noise while listening to speech in quiet. This metric is known as Acceptable Noise Levels (ANLs). Studies have shown large inter-subject variability in acceptance of background noise. An argument has been made that an acceptance of background noise is a “means” that will help to solve the puzzle and monumental problem of hearing aid rejection. Meanwhile, within subject age dependency has not been investigated. This study is conducted to determine if ANL inter and intra-subject variability under music signal depends on age. Twenty subjects participated in the study (average age = 29; SD = 3.7; range 23-35 years). All participants had hearing level not worse than 25dB HL at octave frequencies from 250Hz to 4000Hz. Listeners’ task was to adjust the level of music played in quiet to their most comfortable listening level and then to adjust the level of background noise to the maximum level that they still consider acceptable while listening to music. Further, music is not a speech signal that many researchers have used to determine the significance of age on ANL inter-subject variability. Results of this study supported the findings of others on age dependency, which shows that ANL inter and intrasubject variability is independent of age.
Introduction
Although, human beings are very adaptable creatures, high levels of interfering sounds are always detrimental to their effective communication.In human factors arena, these interfering sounds are commonly referred to as noise.However, the term noise has two general meanings; a narrow meaning and broader one.Noise in its narrow meaning, is a wideband sound consisting of infinite number of components with constant amplitudes and random phases (Adams & Moore, 2009).Such noise is referred to as physical noise.One example of physical noise is thermal noise in which all components have similar amplitudes.Thermal noise is the noise underscoring most physical processes including spontaneous brain activity.The idealized form of thermal noise, in which all components have exactly the same amplitudes, is called white noise.Another example of physical noise is 1/f noise whose components have amplitudes roughly proportional to 1/f, where f is the frequency of a given component.The idealized form of 1/f noise is called pink noise (Gilden, 2001).In its broad meaning, noise is any unwanted sound, including physical noise, regardless of its source.Orthophonic (direct) speaking and listening are the primary communication modes in most educational settings.As a result, Noise Levels (NLs) and Reverberation Times (RTs) of such environments, laboratories, theater rooms, libraries, and administrative offices should be satisfactorily low that speech produced by teachers, presenters, surgeons and responding students, and audience is intelligible to all listeners (Fasanya & Letowski, 2009).Results of Bradley (2007) study indicated that even quite modest amounts of background noise or reverberation could interfere with speech perception and consequently impair educational outcomes.Similarly, people listening to radio, TV, and other media will not understand speech and enjoy the program if the noise levels in their environment are too high.If the background noise interferes with the speech or music, the listener initially increases the signal level but eventually turns the program off if the noise level becomes unacceptably high.This situation happened often among workers performing repetitive task listens to radio to make this task less boring.Many students do read and write while at the same time listening to music in order to mask all other distracting sounds.
A number of studies have been conducted on the significance of signal-to-noise ratio (SNR) to the speech comprehension of normal listeners, as well as those who are hearing-impaired with and without hearing aids (Adams & Moore, 2009;Gengel, 1971;Finitzo-Hieber, & Tillman, 1978;Carhart & Tillman, 1970;Plomp, 1976;Sutter, 1985;Dirks et al., 1982;Schum, 1996;Cooper & Cutts, 1971;Killion, 1997).In these studies findings showed that SNR has a significant effect on speech intelligibility for both hearing impaired listeners and normal listeners.High SNR enhanced better speech intelligibility than lower SNR.About two decades ago a metric was developed to measure person's ability to function in the presence of noise; this metric is known as acceptable noise level (ANL), which is also called acceptable SNR by Nabelek et al. (1991).ANL is the maximum level of background noise listeners are willing and ready to accept without being tired or tensed when listening and following speech in a quiet condition.ANL is calculated as the difference between listener's most comfortable listening level (MCLL) and the maximum background noise level (BNL) the listener could accept (e.g.If MCLL = 30dB and BNL = 12 dB, ANL = 18 dB ).Nabelek et al. (1991) study was to find a lasting solution to the complaint of individuals with hearing loss, and about their inability to comprehend speech when there is background noise.Several studies have been conducted on speech understanding with noise.These studies have used testing of pre and posthearing aid applications to show improved speech comprehension in noise when using hearing aids.The two goals of clinical audiologist are for patients to receive overall benefit from hearing aids and to use their hearing aids, consistently.Based on the statistics of hearing aid use, at least one of these goals is not being met (Allyson, 2006).The effort to determine factors that contribute to a person's success with amplification is perhaps one of the most challenging issues facing audiologists today.
However, several studies have shown that audiometric configuration, degree of hearing loss, type of hearing loss, embedded noise in hearing aids etc. are factors that influence people's willingness to use hearing aids (Franks, & Beckman 1985;Garstecki, & Erler, 1998;Kochkin, 1996;Nabelek, et al., 1991;Crowley, & Nabelek, 1996;Allyson, 2006).The main application of ANL is to predict a person's success with hearing aids.This is also used to assess noise acceptability while listening to speech in various environmental settings.For example, Freyaldenhoven et al. (2006) showed that either acceptance of noise or the listener's willingness to wear hearing aids was significantly affected by the gain compensation of the hearing aid or the changes in the ear mold venting.In addition, the authors demonstrated a significant improvement in masked Speech Recognition Threshold (SRT) with binaural versus monaural amplification.It was shown that there was no improvement in ANL with binaural versus monaural amplification for most listeners.In addition, Harkrider and Smith (2005) reported that the amount of background noise, which listeners were willing to accept while listening to speech in a monotic (monaural) condition correlated with the amount of background noise they were willing to accept while listening to speech in a dichotic (binaural) condition.A study conducted by Nabelek et al. (1991) which used data from 191 listeners with hearing impairment showed that people with low ANLs (≤7.5 dB) accepted a significant amount of background noise, and were found to be successful users of hearing aids.Meanwhile, people with mid to high ANLs (≥7.5 dB) accepted low background noise, and were found to be less likely to be successful users of hearing aids.This indicated that as individual's ANL increases, the possibilities of success with hearing aids decreases.Therefore, it is important to investigate the psychological and physiological processes that reduce ANL.
Moreover, researchers have consistently used speech as the signal to determine individual ANL.As an exception to all other researchers, Kattel et al. (2008) expanded on the concept of ANL metric, where they used music as a signal, believing that research has demonstrated that music defines human personality (Rentfrow & Gosling, 2003).Their findings were similar to other researchers' findings where speech was used as the signal; ANLs for music is independent of the type of background noise used.Other researchers evaluated the effect of music as background noise and found that it has more complex influences than other types of background noise (Gordon-Hickey & Moore, 2007).However, the study further revealed that ANL for music was likely to be better compared with ANL for twelve-talker babble.According to the authors, listeners were more willing to accept music as a background noise than speech babble.The results further revealed that ANL for the music samples were not correlated with preference for the music samples, indicating that ANL for music was not related to music preference.In 1982, Franks conducted a study on hearing aid preference judgment among normal-hearing and hearing-impaired subjects with speech and music signal.The results of the study revealed that with music signals, normal-hearing listeners pay attention to both cutoff frequencies whereas hearing-impaired listeners only took the low-frequency cutoff.Clarity was also rated high for overall music quality along with fullness, in the study conducted by Balfour and Hawkins (1992) to investigate the quality of sound stimulus.In Gordon-Hickey and Moore study, music was used as the background noise, on contrary to its usage in a study conducted by Kattel et al. (2008).Music has rarely been the subject of research in hearing science.Gordon-Hickey and Moore (2007) further documented information about researchers that have studied the use of hearing aid technology and cochlear implant technology (Gfeller et al., 2000;Leal, et al., 2003;and Kong, et al., 2004) for giving musical satisfaction to the hearing impaired.
A characteristic feature of ANL seems to be a large inter-subject variability in acceptance of background noise while listening to speech.According to Rogers et al. (2003) large inter-subject variability has been shown to be relatively independent of gender, hearing sensitivity Nabelek, et al. (2006) and Freyaldenhoven, et al. (2007) and the type of background noise distraction Nabelek, et al. (1991) and primary language of the listener Von Hapsburg and Bahng (2006).Other studies (Nabelek, et al., 1991;Freyaldenhoven & Smiley;Nabelek, et al., 2006) have showed that ANL is independent of age.Nabelek et al. (2004) demonstrated that hearing aids type had no effect on ANLs, but speech-in-noise scores increased with the introduction of amplification.According to Nabelek (2005) acceptance of background noise is a "means" that will help to solve the puzzle and monumental problem of hearing aid rejection.Therefore, to ascertain the dependency of each factor, more than two or three authors' findings will be appropriated for a reasonable conclusion.Fasanya and Letowski (2009) substituted music for speech to determine peoples' ANL and the results proved similar to other findings.
Objectives of the Study
Based on the current research findings on the subject matter, the authors hypothesized that intersubject variability in ANL will depend on age, intra-subject variability, and background noise type (signal type).Thus, the objectives of this study were to investigate the effect of age and background noise type on ANL and intra-subject variability dependency of age.
Participants Selection and Characteristics
The sample included 20 undergraduate and graduate students at a public university located in the eastern region of United States.After providing them with the needed research information including the authors' commitment to their confidentiality, the students volunteered to participate in the study.The ages of the students ranged from 23 to 35, with an average age of 29 and standard deviation (SD) of 3.7.Initial screening showed that all participants had normal hearing defined as not worse than 25 dB HL (hearing level) at octave frequencies in the 250 Hz to 4000 Hz frequency range.Participants were diverse and consisted of ten Africans eight Afro-Americans and two Caribbean.The participants indicated their music genre preference as reggae (9), gospel (8), and hip-hop (3) for their test signal.Using Cochran (1963) sample size determination formula shown in Eq. ( 1), at confidence level of 90% with assumption that the population is large.Therefore, assume p=.5 (maximum variability), the sampling error is calculated to be 18%.(1)
Apparatus and Test Materials
The The positions of all volume controls except for M-audio volume control were fixed.The arrangement of the loudspeakers is shown in Fig. 1.The loudspeakers delivering background noise were positioned at the front (N2) and the back (N1) of the listener about two feet away.Two other signal loudspeakers (RS and LS) were positioned at ± 45 o azimuth and about three feet from the listener center line.The output voltage of the master volume control of the M-audio sound card was calibrated by simultaneous recording of the sound pressure levels in dB SPL at the right ear location of the average head height of the listener using a sound level meter.Pearson Correlation coefficients (r values) were calculated to established reliability of individual responses for the twenty listeners' ANLs and the intra-subjects variability measures.The data for each of the three background noise types were compared.Correlations between ANLs values and the ANL intra-subject variability were found not to be statistically significant under different background noise types.Meanwhile ANLs intra-subject variability coefficient for babble noise was found to be significant with listeners' ANLs for both pink and babble noise, respectively (p = 0.019, and p = 0.04 and r = 0.5168, and r = 0.46185) Table 2 showed that none of the factors compared with age were statistically significant at α = 0.1.Pearson Correlation results further showed high correlation between background noise types (r = 0.9811, between pink and white noise, r = 0.8611 between white and babble noise and r = 0.8973 between pink and white noise); suggesting that ANLs variability did not depend on background noise types.This is also the first time Pearson Correlation coefficient (r value) will be used to determine the dependency of background noise types.
Conclusions and Discussions
Based on results of the analyses, ANLs obtained using pink, white and babble noise were reliable over the short period of three (3) minutes used in the experiment for listeners with normal hearing.
Results were in agreement with other studies like the ones conducted by (Nabelek et al., 1991;Nabelek et al., 2006;and Freyaldhoven & Smiley in press) which found that inter-subject variability does not depend on age, even though hearing aids users were the participants for their study.Likewise, participants' ANL inter-subject variability did not depend on background noise type.These findings confirmed some previous results (Nabelek et al., 1991).Results also showed that ANLs recorded using speech as signal can be replaced with music to determining successful use of hearing aids and for measuring human hearing fatigue.Results further showed that intra-subject variability is age independent.Interchangeability of speech to music for determining background noise and age dependency on ANL is a new finding.
The results of this study are robust leading to conclusions on age dependency in two dimensions.It showed that any variability in ANLs in any form has nothing to do with age.Therefore, the findings of this study may be of interest to manufacturers designing hearing aids, audiologist, hearing fatigue researchers and practitioners fitting hearing aids to patients.However, some limitations were observed during this study.Time duration for participants to complete the three sessions appeared to be too much.This seemed to scare people away; thus contributing to the reduced number of participants (twenty).Future research could include hearing impaired listeners that use hearing aids in a daily basis.Background noise frequency bandwidth may also be considered in a future analysis so that concrete conclusion about background noise dependency can be drawn.
study was conducted in a large Industrial Acoustics Chamber (IAC) Audiometric Booth at the Department of Industrial, Manufacturing and Information Engineering (IMIE) of Morgan State University (MSU) Baltimore, Maryland.The audiometric booth meets noise criteria for uncovered ears (American National Standards Institute, 1991) and equipped with an array of four Studiophile BX5A loudspeakers.Two of the loudspeakers were used to deliver test signals and two others were used to emit background noise.The signals (3 types of music) and noises (3 background noises) were stored on two computers with M-Audio sound cards and Sony Sound Forge 7.0 software.The three music selections used in the study were gospel recording "Leaning on the everlasting arm" by Evangelist Tolu David, hip-hop recording "Can't Hide from Love" by Jay Z and reggae recording "Tuff Gong (Is this love)" by Bob Marley & The Wailers Legend.Each listener used the preferred music recording as the music signal.The three background noise maskers were (a) pink noise (b) white noise and (c) speech babble of 12 voices(Frank and Craig, 1984).All the signals were stored in an IBM computer and normalized to have the same relative average RMS level of -25.0 dB measured at the output of the M-audio sound card.Both the music and the noise were played from IBM computers using Sony Sound Forge 7.0 software and WINAMP software for noise looping.
on age.Meanwhile, intra-subject ANL values for pink noise showed age dependency at 90% confident level (α = 0.1) [F = 2.53, p = 0.0892].Overall statistics analysis showed that either inter-subject variability ANLs or intra-subject variability ANLs depends on age or background noise type with Wilks' Lambda value (0.00057) and (p > 0.05).
Sound pressure level was measured using Metrosonics Chameleon sound level meter (dB SPL).Prior to the beginning of the experiment, calibration was done on sound level meter for accuracy and verified several times during | 3,950.8 | 2013-02-01T00:00:00.000 | [
"Physics"
] |
Numerical Determination of the Secondary Acoustic Radiation Force on a Small Sphere in a Plane Standing Wave Field
Two numerical methods based on the Finite Element Method are presented for calculating the secondary acoustic radiation force between interacting spherical particles. The first model only considers the acoustic waves scattering off a single particle, while the second model includes re-scattering effects between the two interacting spheres. The 2D axisymmetric simplified model combines the Gor’kov potential approach with acoustic simulations to find the interacting forces between two small compressible spheres in an inviscid fluid. The second model is based on 3D simulations of the acoustic field and uses the tensor integral method for direct calculation of the force. The results obtained by both models are compared with analytical equations, showing good agreement between them. The 2D and 3D models take, respectively, seconds and tens of seconds to achieve a convergence error of less than 1%. In comparison with previous models, the numerical methods presented herein can be easily implemented in commercial Finite Element software packages, where surface integrals are available, making it a suitable tool for investigating interparticle forces in acoustic manipulation devices.
Primary radiation force potential
To obtain Equation (8)
Secondary radiation force potential
The derivation of this section follows the steps of Silva and Bruus.
According to Gorkov's potential theory, the acoustic radiation potential of any arbitrary field, except a plane travelling wave, can be obtained as: Moreover, in our case, the total velocity potential is the sum of the velocity potential of the external field and the rescattered field: Here the first term corresponds to the primary radiation potential, the last term is small compared to the second and third terms. Moreover, For the second term of Equation (S8), first ∇ ⋅ ( ) = cos cos ℎ + − = cos cos ℎ − (S12a) such that the second term of Equation (S8) is , | = , cos cos ℎ − (S12b)
Its real part being
Re , = − , cos cos ℎ + (S12c) As the first term of the secondary radiation potential, Equation (S5b), depends on the velocity potential of the external field, which is real, and the real parts of the scattered velocity potential are given by Equations (S11b) and (S12c) The second term of the secondary potential, Equation (S5b), can be calculated by splitting the scattered potential: Equation (S14b), requires the calculation of some auxiliary terms (the gradient of the scattered velocity potential): S15b) and consequently, the first term of Equation (S14b) after simplification and using 2 cos − sin = (S17c) Equation (S17b) can be written as and now Equations (S5b), (S13) and (S18) yield the secondary radiation potential: Note that this force is directly proportional to the monopole scattering coefficient of the scatterer particle.
Secondary radiation force in the polar direction
The derivatives of the different terms containing : and on substitution to Equation (3b) and Equation (10) As all terms contain sin or sin 2 , the above force goes to zero when = 0 However, when = /2, only terms cos , sin 2 or 1 + cos 2 disappear, leaving which is only zero at either the nodes or antinodes, where sin 2 ℎ = 0.
Mesh convergence analysis
Mesh convergence analysis was carried out using a uniform and a non-uniform meshing to assess the convergence speed of the numerical method and verify its robustness. In both cases, we use a scaling parameter mesh_size to define the minimum and maximum discretization steps. For the uniform mesh, the maximum mesh size is given as /mesh_size , while the minimum mesh size as /(2 ⋅ mesh_size). For the non-uniform mesh, the scattering and probe particles are meshed using the above minimum and maximum values, but the other domains (the fluid domain and the PML) are meshed using a coarser mesh, with minimum and maximum size of /(0.6 ⋅ mesh_size) and /(1.2 ⋅ mesh_size), respectively.
As the polystyrene particle in water has a lower contrast compared to the polystyrene particle in air, and consequently its relative secondary radiation force is lower, we chose this former case to analyze the mesh convergence. The results for various cases (node, antinode) and along different directions (radial and z) can be seen in the following Figs. S1-S4. The convergence error is defined as where Fm denotes the secondary radiation force obtained using the m th mesh_size parameter. In all cases, the probe particle was placed 0.35 distance from the scatterer, this distance is approximately halfway between the node and antinode and therefore has a non-zero force. Figure S1. Convergence analysis results when the scattering particle is aligned with the pressure antinode and the probe particle is along the z direction. The convergence error is plotted as a function of the mesh size parameter (top graph) and as a function of the number of degrees of freedom (bottom graph). Figure S2. Convergence analysis results when the scattering particle is aligned with the pressure node and the probe particle is along the z direction. The convergence error is plotted as a function of the mesh size parameter (top graph) and as a function of the number of degrees of freedom (bottom graph). Figure S3. Convergence analysis results when the scattering particle is aligned with the pressure antinode and the probe particle is along the r direction. The convergence error is plotted as a function of the mesh size parameter (top graph) and as a function of the number of degrees of freedom (bottom graph). Figure S4. Convergence analysis results when the scattering particle is aligned with the pressure node and the probe particle is along the r direction. The convergence error is plotted as a function of the mesh size parameter (top graph) and as a function of the number of degrees of freedom (bottom graph).
Except for the last case ( Figure S4), the convergence is fast, and no significant difference between the uniform and non-uniform mesh can be observed, as far as the number of degrees of freedom is concerned. For non-uniform mesh, for all four cases the error is below 1%, when mesh_size > 35, meaning that the particles are meshed between 0.0143 < mesh < 0.0286 and the other domains between 0.0238 < mesh < 0.0476 .
For the uniform mesh, the error is below 1% for all cases when mesh_size > 23, meaning that for all domains the discretization size is 0.0217 < mesh < 0.0435 .
As for the 2D case, increasing the number of degrees of freedom in the same order of magnitude as for the 3D case would require an inefficiently large mesh, we decided to use a quartic discretization in this case compared to the cubic discretization of the 3D case. The quartic discretization allows expansion of a medium-sized mesh into a large number of degrees of freedom using fourth order approximation of the solutions over each mesh element. Moreover, as for the 3D case no significant difference was observed between a uniform and non-uniform mesh, for the 2D investigations we only applied a simple uniform mesh with the same characteristics as before.
These results are summarized in Figures S5 and S6. Figure S5. Convergence analysis results when the scattering particle is aligned with the pressure node and the probe particle is along either the r or z direction. The convergence error is plotted as a function of the mesh size parameter. Figure S6. Convergence analysis results when the scattering particle is aligned with the pressure antinode and the probe particle is along either the r or z direction. The convergence error is plotted as a function of the mesh size parameter.
In both cases an extremely fast convergence can be observed irrespective of the particle position: the error is already less than 0.01% when the mesh scale parameter is 20. For the 2D case, the relationship between the mesh size parameter and the number of degrees of freedom is summarized in Table S1. Although the methods are not applicable for simulations of touching spheres, we show that the separation distance (the surface-to-surface distance) of the two spheres can be arbitrary low. In Figure S7, secondary radiation force results for PS particle in water, where the scatterer is at the antinode can be seen, showing separation distances as low as 0.001λ with successful simulation. Figure S7. Secondary radiation force when the scattering particle is aligned with the pressure antinode and the probe particle is along the z direction. The distance in this case corresponds to surface-to-surface distance of the particles. | 1,992.2 | 2019-06-29T00:00:00.000 | [
"Physics"
] |
Heat Transfer Enhancement in Microchannel Heat Sink Using Nanofluids
1.
Introduction 1.1 Introduction to microchannel heat sink (MCHS)
The microchannel heat sink (MCHS) cooling concept is first proposed by Tuckerman and Pease [1] in 1981.MCHSs are the most common and cost-effective hardware employed for the thermal management of MEMS devices.Tuckerman and Pease [1] pointed out that decreasing liquid cooling channel dimensions to the micron scale will lead to increase the heat transfer rate.They demonstrated experimentally a forty-fold improvement in heatsinking capability in Si-base microchannels anodically bonded to Pyrex cover plates.Since then, intense research on MCHS have been conducted to study the heat transfer and fluid flow characteristics of MCHS.There are two main configurations for the application of microchannel cooling which are direct cooling and indirect cooling.Direct cooling requires a direct contact between the surface to be cooled and the coolant fluid as illustrated in Fig. 1.1a.This scheme reduces the thermal resistance between the surface and the coolant and thus, enhances the cooling effectiveness.However, electrical and chemical compatibility between the coolant and device itself needs to be ensured for this system to work [2].An alternative to the above configuration is the use of a metallic heat sink to conduct the heat away from the device to a coolant which is forced through circular or noncircular grooves in the heat sink.Such an indirect cooling configuration shown in Fig. 1.1b allows for a greater flexibility in coolant selection at the cost of increased thermal resistance between the device and the heat sink due to the heat diffusion resistance in the heat sink itself [3].Microchannels are very fine channels of the width of a normal human hair and are widely used for electronic cooling purposes.In a MCHS, multiple microchannels are stacked together as shown in Fig. 1.1 (b) which can increase the total contact surface area for heat transfer enhancement and reduce the total pressure drop by dividing the flow among many channels.Liquid or gas is used as a coolant to flow through these microchannels.The large surface area of MCHS enables the coolant to take away large amounts of energy per unit time per unit area while maintaining a considerably low device temperature.Using these MCHS, heat fluxes can be dissipated at relatively low surface temperatures.
Cooling performance of MCHS
In order to drive the development of compact and efficient thermal management technology for advanced electronic devices, cooling devices have to be in light-weight, small in size and of high performance.Steinke and Kandlikar [4] presented a comprehensive review of friction factor data in microchannels with liquid flows.They indicated that entrance and exit losses need to be accounted for while presenting overall friction factors losses in microchannels.Most of the data that accounted for friction factor loss show good agreement with the conventional theory.They also provided a new procedure for correcting measured pressure drop to account for inlet and outlet losses.Furthermore, three-dimensional fluid flow and heat transfer phenomena inside heated microchannels were investigated by Toh et al. [5].They solved the steady laminar flow and heat transfer equations using a finite-volume method.It was found that the heat input lowers the frictional losses and viscosity leading to an increase in the temperature of the water, particularly at lower Reynolds numbers.Peng and Peterson [6,7] performed experimental investigations on the pressure drop and convective heat transfer of water flow in rectangular microchannels.They found that the cross sectional aspect ratio had a great influence on the flow friction and convective heat transfer both in laminar and [8] performed experimental and numerical analysis of the effect of axial conduction on the heat transfer in triangular microchannel heat sink.They pointed out that the bulk water and heated wall temperatures did not change linearly along the channel.Inspired by the MCHS idea, new designs and modeling approaches of high performance cooling devices have been proposed, including using nanofluids as coolants in the MCHS study.Nanofluids are produced by dispersing nanometer-scale solid particles into base liquids such as water, ethylene glycol (EG), oils, etc. Lee et al. [9] used 38.4 nm of Al 2 O 3 and 23.6 nm of CuO particles to enhance the thermal conductivity of water and EG.They showed that the enhancement percentage in thermal conductivity was not only a function of concentration and conductivities of the particles material and liquid, but it is also function of particle size and shape.Koo and Kleinstreuer [10] considered nanofluid flow in a representative microchannel, and conduction-convection heat transfer for different base fluids such as water and ethylene glycol with 20 nm CuO-nanoparticles.They come out with several suggestions which are a base fluid of high-Prandtl number such as ethylene glycol and oil should be used, using nanoparticles with high thermal conductivity are more advantageous, and a channel with high aspect ratio is desirable.Chein and Huang [11] analyzed silicon microchannel heat sinks performance using nanofluids with a mixture of pure water and nanoscale Cu particles as coolants with various volume fractions.The MCHS with two specific geometries, one with W ch =100 μm and L ch =300 μm, the other with W ch =57 μm and L ch =365 μm, were examined.They found that the performances were greatly improved for these two specific geometries by using nanofluids as the coolants compared with pure water due to the increase in thermal conductivity of coolant and the nanoparticles thermal dispersion effect.The effectiveness of nanofluids for single-phase and two-phase heat transfer in microchannels is analyzed by Lee and Mudawar [12].They indicated that the higher single-phase heat transfer coefficients are achieved in the entrance region of microchannels with increased nanoparticle concentration.However, the enhancement is weaker in the fully developed region.They suggested that nanoparticles should not be used in two-phase MCHS.This is because once boiling commences, particles begin to deposit into relatively large clusters near the channel exit due to localized evaporation.This clustering phenomenon quickly propagates upstream to fill the entire channel, preventing coolant from entering the heat sink and causing catastrophic failure of the cooling system.Jang and Choi [13] analyzed numerically the cooling performance of MCHS with nanofluids.They showed that the cooling performance of a MCHS with water-base nanofluids containing diamond (1%, 2 nm) at fixed pumping power of 2.25W was enhanced by about 10% compared with that of a MCHS with water.There have been relatively few recent studies on nanofluid flow and heat transfer characteristics as comparing with those of pure fluid [14][15][16].These studies showed that the heat transfer coefficient was greatly enhanced using nanofluid compared with pure fluid although there is a slight increase in pressure drop due to the presence of nanoparticles in MCHS operation.The enhancement depended on Reynolds number, particle volume fraction, and particle size and shape.It should be noted from the above literature review, however, that limited studies are available on nanofluid flow and heat transfer characteristics of rectangular shaped MCHS performance and this has motivated the present study.
The current study mainly focuses on 3D computational simulation of heat transfer and laminar liquid flow characteristics in MCHS.Following this introduction section, the governing equations and numerical model is explained, followed by studies on the geometrical parameters with various MCHS shapes (Section 3), effects of different nanoparticle volume fractions (Section 4), effects of different nanofluids types (Section 5), and effects of nanoparticle in different base fluids (Section 6).
MCHS model
The physical configuration of the MCHS is schematically shown in Fig. 2.1.The heat supplied to the MCHS substrate through a top plate is removed by flowing fluid through a number of 25 microchannels.This article focuses on heat transfer and liquid flow in three different cross sectional shapes of MCHS including rectangular, trapezoidal, and triangular.The dimensions of three different sets for each cross sectional shape of MCHS are given in
Governing equations
In the analysis of the entire domain of MCHS, it is necessary to set up the governing equations (continuity, momentum, and energy).For the specific case of heated flow through microchannels, the governing equations are solved with the following assumptions: 1.Both fluid flow and heat transfer are in steady-state and three-dimensional.2. Fluid is in single phase, incompressible and the flow is laminar.
3. Properties of both fluid and heat sink material are temperature-independent. 4. All the surfaces of MCHS exposed to the surroundings are assumed to be insulated except the top plate where constant heat flux boundary condition, simulating the heat generation from electronic chip, is specified.Thus, the governing equations in dimensionless form which are used in present study for heated MCHS can be written as [17]: Continuity ∆ is the dimensionless pressure, and is the dimensionless temperature.
Given the complexity of these equations, computational methods of solving them are required.The methodology used in the present study to solve these governing equations is described in the following section.
Boundary conditions
Boundary conditions for all boundaries are specified for this simplified computational domain.Fig. 2.1a shows the general MCHS computational model used in this study and Fig. 2.1b shows the cross sectional shapes of channel considered in the study.Heat, supplied to the aluminum substrate through a top plate, is removed by flowing fluid through a number of microchannels.At the entrance of the MCHS assembly (z = 0, from Fig. where A is the channel flow area, P is the channel wet perimeter, k is the channel width, h is the channel height, and is the channel hypotenuse.The Reynolds number in this work was ranged from 100 to 1000.In calculating the velocity, the fluid is assumed to be evenly distributed into all microchannels.The transverse velocities at the inlet are assumed to be zero.On the aluminum substrate, the velocities are zero, and it is assumed to be an adiabatic surface.Boundary conditions at the inlet: At the outlet: At the fluid-solid interface: At the top plate: In Eq. (2.6), U and θ are the dimensionless fluid inlet velocity and dimensionless temperature, respectively, P is the dimensionless pressure at the outlet, n is the direction normal to the wall or the outlet plane, and q w is the heat flux applied at the top plate of the heat sink.The heat flux applied at the top plate was ranged from 100-1000 W/m 2 .
Numerical solution using FVM
The governing conservation equations Eqs. ( 2) -( 4) with the corresponding boundary conditions and equations for solid and fluid phases are simultaneously solved as a single domain conjugate problem using the standard finite volume method (FVM) with a hybrid differencing scheme [17].The standard SIMPLE algorithm is used as the computational algorithm [18].The iterations are continued until the sum of residuals for all computational cells became negligible (less than 10 −7 ) and velocity magnitudes did not change from iteration to iteration.Because of the assumption of constant fluid properties and negligible buoyancy, the mass and momentum equations are not coupled to the energy equation.Therefore, the temperature field is calculated by solving the energy equation after a converged solution for the flow field is obtained by solving the momentum and continuity equations.
Numerical implementation, grid testing and code validation
The distribution of hexahedral cells in the computational domain is determined from a series of tests with different number of cells.For example, for the case of rectangular MCHS, computational cells with 1.9x 10 5 , 2.8 x 10 5 , and 3.4 x 10 5 grids are used to test the grid independence of the solution.The results are shown in Figs.2.2 and 2.3 which present the dimensionless temperature profiles and dimensionless pressure drop across the heat sink versus Reynolds number, respectively.The dimensionless temperature and dimensionless pressure are defined in Eq.3.1(Section 3) and Eq.5.1(Section 5), respectively.It can be seen that almost identical results obtained when 2.8 x 10 5 and 3.4 x 10 5 grids are used.Thus, based on the results shown in these figures, a computational cell with 2.8 x 10 5 grids is employed for all the numerical computations in this study.Similar study was carried out for other MCHS shapes.
Effect of geometrical parameters of various MCHS shapes
This section investigates the effect of geometrical parameters on heat transfer and fluid flow characteristics for different cross sectional shapes of MCHS including rectangular, trapezoidal, and triangular with the dimensions as stated in Table 2.1-2.3,respectively.For this Section, water is used as the working fluid as the main aim is to investigate on the geometrical effects.For Section 4, 5, and 6, various nanofluids are used as the working fluid and only one particular cross sectional shape of MCHS with specified dimensions is considered for each Section.The dimensionless temperature profiles, heat transfer coefficient, friction factor, and thermal resistance which affected by geometrical parameters are discussed and presented.
Temperature profile
In the MCHS operation, the high temperature region should occur at the edge of the MCHS since there is no heat dissipation by fluid convection while the low-temperature region should occur in the region where microchannels are placed, especially at the middle regions of the MCHS due to the high heat transfer coefficient.To address this point, the dimensionless average temperature profiles of rectangular cross-section MCHS for different dimensions of the channel area at Re = 500 are presented in Fig. 3.1.The dimensionless temperatures for each channel can be defined as: Where T f is the fluid temperature, T i is the inlet temperature, and T w is the wall temperature.It can clearly be seen from Fig. 3.1 that the lower temperature profile occurs in the channels [19].For fluid in channels close to the edge of the MCHS, higher temperature profiles are observed due to the high heat transfer from the high-temperature edge of the MCHS.
Heat transfer coefficient
The computed averaged heat transfer coefficient in each channel of the rectangular crosssection MCHS for various hydraulic diameters is illustrated in Fig. 3.2.The magnitude of heat transfer coefficient decreases with the increase in hydraulic diameter and the trends are the same for each cross sectional shape of the MCHS.This is due to the fact that lower pressure drop in larger hydraulic diameter corresponds to lower inlet velocity driven into the MCHS.Furthermore, due to the difference of the channel hydraulic diameter which is caused by the difference of channels' area and perimeter, the averaged heat transfer coefficient in each MCHS is also different under a given inlet velocity.For each type of the MCHS, the middle channel (channel number 14) has the highest averaged heat transfer coefficient value as expected.The averaged heat transfer coefficient value for other channels is seen to decrease depending on their distances from the wall.The averaged heat transfer coefficient distribution for all types of MCHS is almost symmetrical with respect to the centerline of the MCHS.It is shown that the heat transfer coefficient of the MCHS is greatly influenced by the hydraulic diameter of the channel, as the hydraulic diameter decreases, the heat transfer coefficient increases.
The computed average heat transfer coefficient versus Reynolds number for different cross sectional shapes of the MCHS is presented in Fig.
Friction factor
The effect of geometrical parameters on the friction factor, f for different microchannel shapes is discussed in this section.For the current study, the friction factor is calculated using Darcy equation [20]: Where D h is the hydraulic diameter, is the pressure drop, is the density of water, u in is the inlet velocity of water, and L c is the length of channel.Fig. 3.4 shows the friction factor, f, at different W c /H c ratios for rectangular cross-section MCHS.It is clearly observed that f increases with the increase of W c /H c ratio of the channel.Therefore, the flow resistance will increase evidently when the ratio is increased.The reason behind this is that the increase in W c /H c ratio causes a decrease in the flow area and the pressure drop becomes significant.The validity of the present numerical results was also approved by Kandlikar et al. [21] results, where f increases with the increase of W c /H c ratio for rectangular cross-section MCHS.Fig. 3.5 shows f with three geometrical parameters including bottom-to-top width ratio (b/a), height-to-top width ratio (h/a), and length-to-hydraulic diameter ratio (L/D h) , which affect the friction and heat transfer in the trapezoidal cross-section MCHS.Reynolds number, fRe from the present numerical simulation for trapezoidal cross-section MCHS with the experimental work done by Wu and Cheng [22].As noticed, a very good agreement is obtained between the two studies whereby it is shown that fRe rises with the increase of Reynolds number.As can be noticed from Fig. 3.6, fRe increases linearly with the increase of Reynolds number due to the linear dependency of pressure drop on constant inlet velocity for a fixed hydraulic diameter in fully developed laminar region.Thus, fRe is linearly proportional to Reynolds number for fully developed channel flow.
The f with different tip angles of triangular cross-section MCHS is illustrated in Fig. 3.7.It is clearly observed that f increases with the increases of tip angle of the channel.The validity of the present numerical simulation result was also approved by the results of Kandlikar et al. [21].The results show that the flow behavior is very similar for all types of cross sectional shapes of MCHS where the friction factor decreases with the increase of Reynolds number.
In overall, From the Figs.3.4,3.5, and 3.7, it is also apparent that f for rectangular shaped microchannel is most, for trapezoidal shaped microchannel is less, and for triangular shaped microchannel is least.This is due to the triangular shaped microchannel has a smallest area.
For trapezoidal shaped microchannels, the flow area is changed suddenly from larger area to smaller area, thus the pressure drop would be large.Therefore, f value for trapezoidal shaped microchannels is also larger than the triangular shaped microchannels.For the rectangular shaped microchannels, the flow area is the largest which corresponds to the highest value of f.Therefore, f of water flowing in MCHS, having the same width and height but with different cross sectional shapes can be very much different due to the difference of the cross sectional shape and geometrical dimensions of the channels.
Thermal resistance
The performance of the MCHS is commonly presented by the thermal resistance (R th ) which is defined as: Another important parameter in MCHS operation is the pressure drop across the MCHS which relates to the coolant pumping power required.The pumping power is used to drive the coolant in MCHS operation.It is the product of the pressure drop across the heat sink,∆ and volume flow rate, Q: where Q is defined as: For rectangular cross-section MCHS, where N is the number of microchannels, W ch is the width of channel, H ch is the height of channel, and u in is the inlet velocity of water.For trapezoidal cross-section MCHS, where N is the number of microchannels, k is the width of channel, h is the height of channel, and u in,nf is the inlet velocity of water.
For triangular cross-section MCHS, where N is the number of microchannels, k is the width of channel, h is the height of channel, and u in is the inlet velocity of water.The effect of using different cross sectional shapes of MCHS on thermal resistance, R th versus pumping power is shown in Fig. 3.8.The pumping power is determined from the calculated volume flow rate for rectangular, trapezoidal, and triangular cross-section shape using Eq.3.5-Eq.3.7,respectively, and computed pressure drop using Eq.3.4.It is evident from this figure that the R th for rectangular cross-section MCHS is the lowest followed by trapezoidal and triangular cross-section MCHS.Thus, a MCHS with rectangular crosssection is capable for removing high heat flux due to its small difference between maximum wall temperature and inlet temperature at particular pumping power compared with that for MCHS with trapezoidal and triangular cross-sections.
Effect of nanoparticle volume fractions on MCHS performance
Nanofluids are produced from a suspension of nanoparticles with specified volume fraction in a conventional base fluid such as water, have been gaining interest recently due to its potential to greatly outperform traditional thermal transport liquids in MCHS [23].However, the cooling effectiveness of using nanofluid as well as addressing other technical problems, such as agglomeration of the nanoparticles which yield to MCHS clogging due to the high particle volume fraction are vital to practical implementation of the nanofluid-cooled MCHS.In the current study, the effect of using three different particle volume fractions on heat transfer and liquid flow characteristics in MCHS will be discussed.Since the cooling performance of MCHS with various cross sectional shapes is investigated in previous chapter, now a rectangular cross-section MCHS with specific dimensions as shown in Table 4
Thermophysical properties of nanofluid
In this study, aluminum oxide with different particle volume fractions is used as the working fluid.The thermophysical properties required for flow as listed in Table 4.2 are calculated using the following equations [24,25,26,27]: Heat capacity: Thermal conductivity: Where is particle volume fraction, the subscript "nf" refers to nanofluid, "bf" refers to base fluid, and "p" refers to particle.
Temperature profile
The dimensionless temperature distribution at each channel of MCHS at Re = 900 for various particle volume fractions and different heat fluxes is shown in Figs.4.1-4.3.The dimensionless temperature is calculated using Eq.3.1 (Section 3).The presence of nanoparticles has an effect of reducing the temperature as the particle volume fraction of nanofluids increasing due to its higher dynamic viscosity and lower heat capacity compared to pure water.However, these conditions only can be greatly achieved for the case of heat flux = 1000 W/m 2 compared to 100 W/m 2 and 500 W/m 2 .For case 100 W/m 2 and 500 W/m 2 , there is no significant differences between nanofluids and pure water, and the results are inconsistent for temperature distributions.Therefore, the presence of nanoparticles could enhance cooling of MCHS under the extreme heat flux conditions, the same trend was reported by Tsai and Chein [15] and Murshed et al. [28].Where h nf is the nanofluids heat transfer coefficient value and h avg,water is the average heat transfer coefficient of pure water.It is inferred that the nanofluids can enhance the heat transfer of MCHS as its volume fraction increases from 0% to 2.5%.However, nanofluid with 5% volume fraction is not able to enhance the heat transfer or performing almost the same result as pure water.This may due to agglomeration of the nanoparticles.validity of the present results is confirmed with Murshed et al. [28] results.From Fig. 4.4, it is also observed that the middle channel (channel 14) has the highest averaged heat transfer coefficient value as expected.The averaged heat transfer coefficient value for other channels is seen to decrease depending on their distance from the wall.The present results as illustrated in Fig. 5.6 show similar trend with the results obtained by Liu and Garimella [29] and Chein and Chen [19].Since higher temperatures are observed due to the heat transfer from the high-temperature edge of the heat sink, the averaged heat transfer coefficient is lower in channels close to the edge of the heat sink.
Friction factor
The variation of the friction factor for various particle volume fractions is shown in Figs.4.5 and 4.6 for Reynolds number ranged from100-500 and 500-1000, respectively.The friction factor is calculated using Darcy equation [20]: Where D h is the hydraulic diameter, p is the pressure drop, is the density of nanofluid, u in , nf is the inlet velocity of nanofluid, and L c is the length of channel.The results show that the friction factor is similar for all particle volume fractions where the friction factor decreases with the increase of Reynolds number.Based on Figs.4.5 and 4.6, it can be stated that the increment of dynamic viscosity due to the presence of the nanoparticles in water, only appears to give a slight rise in friction factor especially for low Reynolds number range (Fig. 4.5).For high Reynolds number range (Fig. 4.6), the nanofluid flow yields significant effect on friction factor.Fig. 4.7 shows a comparison of the predicted friction factor from the present numerical simulation against previous experimental studies [ 12,30].A very good agreement was obtained for φ = 0% (pure) and φ = 1% alumina, which shows that the friction factor decreases with the increase of Reynolds number.[12] and [30].Where τ nf is the nanofluid wall shear stress and τ avg, water is the average value of pure water wall shear stress.As depicted in Fig. 4.8, the wall shear stress increases with increasing Reynolds number for a specified percentage of fractions.It is clearly seen that for all cases investigated, the presence of nanoparticles substantially increases the wall shear-stress as compared with pure water.In general, for all cases examined, the least wall shear stress occurs at lower value of φ = 1 % while highest value occurs at φ = 5 %.
Thermal resistance
Since the maximum wall temperature (T w, max ) of nanofluid-cooled MCHS is lower than pure water-cooled MCHS, reduction of thermal resistance in nanofluid-cooled MCHS is then expected.If the presence of nanoparticles in coolant did not produce extra pressure drop which relates to the coolant pumping power required, it can then be considered as another benefit of using nanofluid as the coolant in MCHS operation.The volume flow rate, Q for rectangular shaped microchannel is defined as: Where N is the number of microchannels, W ch is the width of channel, H ch is the height of channel, and u in , nf is the inlet velocity of nanofluid.
The cooling performance of the MCHS using nanofluid is by exploring the results for the thermal resistance, R th as defined in Eq.3.3.The variation of thermal resistance versus pumping power is presented in Fig. 4.9.The pumping power is determined from the calculated volume flow rate as defined in Eq.4.8 and computed pressure drop using Eq.3.4 (Section 3).As depicted in Fig. 4.9, the nanofluid is shown to have significantly lower value of R th as its particle volume fraction increases than pure water.For the current geometry under study, the maximum reduction in R th that can be achieved is about 57.6% for nanofluid with particle volume fraction of 5% at 0.0145 W. This great improvement in R th is mainly due to the increase in thermal conductivity and decrease in temperature difference between inlet and maximum wall temperature as the particle volume fraction increases.The additional reduction in R th is clearly due to the thermal dispersion.Since water has the highest R th value , it is not preferable for removing high heat flux.The percent enhancement of MCHS's cooling performance is actually calculated by using: Where R th,nf corresponds to thermal resistance of particular nanofluid and R th,water corresponds to thermal resistance of pure water.
Effect of using different nanofluids on MCHS performance
Based on the results that obtained in Section 4 which focused on the effect of using different particle volume fractions of Al 2 O 3 -H 2 O on heat transfer and liquid flow characteristics in MCHS, it is known that nanofluids have the capability to enhance the heat transfer performance of MCHS with smaller pressure drop penalty.However, due to production of various nanofluids in market, further investigations on using different types of nanofluids in MCHS are necessary because the heat transfer enhancement of MCHS is strongly depend on thermophysical properties of dispersed nanoparticles in base fluid [31].Thus, in this chapter, the effect of using six types of nanofluids on heat transfer and liquid flow characteristics in triangular cross-section MCHS with specified dimensions is comprehensively investigated.The specific dimensions of the triangular MCHS used in this section are given in Table 5.1.A constant particle volume fraction value of 2% is considered for all types of nanofluids studied in this chapter.The thermophysical properties for all types of nanoparticles, base fluid (water), and the nanofluids with particle volume fraction of 2% are listed in Table 5.2.For all types of nanofluids studied, the density, thermal conductivity, and dynamic viscosity appear significantly higher while the specific heat of nanofluids is lower than pure water.For instance, diamond-H 2 O nanofluid has a relative increase of 5.03%, 6.11%, and 5% in the density, thermal conductivity, and dynamic viscosity and has a relative decrease of 5.90% in specific heat, respectively, compared with pure water.In this section, different types of nanofluids which are Al , and TiO 2 -H 2 O are used as working fluids.The thermophysical properties as listed in Table 5.2 including density, heat capacity, thermal conductivity, and viscosity for all types of nanofluids which are involved in the governing equations are calculated using Eqs.4.1-4.4 (Section 4), respectively.
Heat transfer coefficient
The thermal performance of using different types of nanofluids in MCHS is examined by presenting the results of the dimensionless heat transfer coefficient along the length of channel number 1 in Fig. 5.1.It is observed that all types of nanofluids-cooled MCHS could be able to enhance the heat transfer compared with pure water-cooled MCHS.The waterbase nanofluid containing diamond has the highest heat transfer coefficient value followed by SiO 2 , CuO, TiO 2 , Ag and Al 2 O 3 .However, there is no significant difference that can be detected between TiO 2 and CuO due to their small difference in thermal diffusivity.However, there is a very small increment in CuO-H 2 O-cooled MCHS especially at the outlet of the channel compared with that for TiO Since the velocity at the entrance of the channel was kept constant at Re = 600, it can be noticed that the heat transfer coefficient decreases linearly along the length of channel because of the development of thermal boundary layer for all cases.The highest value of heat transfer coefficient occurs near the entrance.The present results show a similar trend with the results obtained by Lee and Mudawar [12].It is also depicted from Fig. 5.1 that the difference in dimensionless heat transfer coefficient between various nanofluids is too small but it follows trend and not too far from reported values.From the results obtained, lowest thermal resistance is expected with diamond-H 2 O nanofluid cooled MCHS due to its lowest temperature distribution.Therefore, the presence of diamond nanoparticle in water could greatly enhance the cooling of MCHS compared with other types of nanofluids.This is due to the highest thermal transport capacity of diamond in nature and diamond particles are often used as filler in mixtures for upgrading the performance of a matrix (composition of mixtures) as reported by Xie et al. [32].Thus, it is reasonable to expect that the addition of diamond nanoparticles in water would lead to heat transfer enhancement.where ∆ is the pressure drop of particular microchannel configuration, ρ is the water density, and is the inlet water velocity.It can obviously be seen that the high pressure region occurs at the entrance and the low pressure region occurs at the outlet of the heat sink.This is due to the needs of high pressure to push the fluid flow along the direction of the microchannels out from the outlet plenum of heat sink.Thus, the pressure decreases linearly along the length of channel for all types of nanofluids studied including pure water.Because of that, there is no apparent difference observed between pure water and nanofluids flows particularly at the outlet region of MCHS.However, compared with pure water cooled case, slightly larger pressure drops for nanofluid-cooled MCHS were found except for Ag-H 2 O nanofluid which has the lowest value of pressure drop among nanofluids types.The water-base nanofluids containing SiO 2 has the highest value of pressure drop and followed by diamond-H 2 O respectively.The pressure drop in water-base nanofluids-cooled MCHS containing Al 2 O 3 and TiO 2 are slightly lower than diamond-H 2 O-cooled MCHS.Since the viscosity of nanofluids is larger than pure water as indicated in Table 5.2, larger pressure drop is expected in nanofluid-cooled MCHS.Another reason that contributes to increase the pressure drop is the deposition of nanoparticle in water is increasing the wall roughness [28].However, Ag-H 2 O-cooled MCHS still has lower value of pressure drop than pure water-cooled MCHS and there is no extra pressure drop in CuO-H 2 O-cooled MCHS as compared with pure water-cooled MCHS.The reason for this result is Ag-H 2 O has the lowest Prandtl number (Pr = 5.6 and followed by CuO-H 2 O (Pr = 6.1).The small difference in pressure drop between other types of nanofluids which can be seen in Fig. 5.2 is due to small differences in their Prandtl numbers and SiO 2 -H 2 O yields the highest pressure drop across the MCHS due to its highest Prandtl number (Pr = 6.8).Therefore, the pressure drop penalty is not only affected by the viscosity of nanofluids but also the Prandtl number of nanofluids which need to be considered in nanofluids selection as a coolant for MCHS.Apart from that, slight increment in pressure drop particularly for Ag-H 2 O and CuO-H 2 O-cooled MCHS which can be considered as one of the benefits of using nanofluid in MCHS operations in order to achieve overall heat transfer enhancement with small pressure drop penalty.
Thermal resistance
The pumping power which is the product of the pressure drop across the heat sink,∆ and volume flow rate, Q, is calculated using Eq.3.4.The volume flow rate, Q for triangular shaped microchannel is defined as: Where N is the number of microchannels, k is the width of channel, h is the height of channel, and u in,nf is the inlet velocity of nanofluid.
The thermal resistance for all types of nanofluids as function of pumping power is shown in than pure water.The maximum reduction in R th is achieved in diamond-H 2 O MCHS is 3.1292 K/W when pumping power is 0.0097207 W. The results indicate that the cooling performance of a MCHS at pumping power = 0.005 W is enhanced by about 2 % for waterbase nanofluids containing diamond, SiO 2 , CuO, and TiO 2 , 1.13% for water-base nanofluids containing Al 2 O 3 , and 1.04% for water-base nanofluids containing Ag respectively, compared with that of the MCHS with pure water.The percent enhancement of MCHS cooling performance is also calculated using Eq.4.9.The great improvement in R th is mainly due to the increase in thermal conductivity and decrease in temperature difference between inlet and maximum wall temperature.The additional reduction in R th is also due to the thermal dispersion of particle volume fractions.
Effect of using different base fluids on MCHS performance
Since base fluid for all types of nanofluids studied in Section 5 was water, it is important and necessary to know the heat transfer and liquid flow characteristics in MCHS using other types of conventional base fluids such as ethylene glycol (EG), glycerin, and oil.By using other types of conventional base fluids instead of using water, the MCHS cooling performance could be enhanced or may not.Thus, considering nanofluid flow in a microchannel, the effect of using different types of base fluids including water, EG, engine oil, and glycerin on heat transfer and liquid flow characteristics in trapezoidal shaped MCHS which was selected for this chapter with specified dimensions is extensively discussed in this chapter.The specified dimensions are given in Table 6.1.
Since the best uniformities in heat transfer coefficient and temperature can be achieved in MCHS using diamond nanoparticle as obtained in Section 5, the diamond nanoparticle is selected and considered for all types of base fluids studied in this chapter.The thermophysical properties for diamond nanoparticle, for all types of base fluids, and the nanofluids with particle volume fraction of 2% formulated are listed in Table 6.2.For all types of nanofluids, the density, thermal conductivity, and dynamic viscosity appear significantly higher while the specific heat of nanofluids is lower than the base fluids.For instance, the EG-diamond nanofluid has a relative increase of 4.28%, 6.12%, and 5% in the density, thermal conductivity, and dynamic viscosity and has a relative decrease of 4.77% in specific heat, respectively, compared with EG base fluid.The thermophysical properties as listed in Table 6.2 including density, heat capacity, thermal conductivity, and viscosity for all types of base fluids for nanoparticles are calculated using Eqs.
Temperature profile
The effect of different types of base fluids on dimensionless temperature variation is presented in Fig. 6.1.It can clearly be seen that the temperature of pure water is always highest among all types of base fluids.By comparing dimensionless temperature distributions between different types of base fluids, glycerin-diamond has the lowest value of temperature while water-diamond has the highest value of temperature and followed by EG-diamond and oil-diamond.Based on results shown in Fig. 6.1, it can be stated that the lowest thermal resistance is expected in the glycerin-diamond fluid cooled MCHS due to its lowest temperature distribution.Therefore, the presence of diamond nanoparticle in glycerin could greatly enhance the cooling of MCHS compared with other types of base fluids studied.This is because of glycerin has the highest dynamic viscosity in nature compared to other base fluids and diamond particle mixed properly in glycerin which contribute to increase the thermal transport capacity of the mixture.
Heat transfer coefficient
The thermal performance of using different types of base fluids in MCHS is examined by plotting results of the dimensionless heat transfer coefficient along the length of channel number 20 as shown in Fig. 6.2.Koo and Kleinstreuer [10] reported that higher Prandtl number base fluids-cooled MCHS could be able to enhance the heat transfer performance of MCHS as compared with water base-cooled MCHS.This was proven by the present results where the glycerin base-cooled MCHS has the highest ℎ value due to the highest Prandtl number value followed by oil, and EG.Kleinstreuer [10].However, there is no significant differences in ℎ for EG, oil, and glycerine but when the plot is zoomed in, there are slight differences in ℎ where glycerin base-cooled MCHS has the highest ℎ value, oil is in between, and EG is the least.For each type of the heat sinks, the middle channel (channel number 14) has the highest ℎ value as expected.The averaged heat transfer coefficient value for other channels is seen to decrease depending on their distances from the wall.The averaged heat transfer coefficient distribution for all types of MCHS is almost symmetrical with respect to the centreline of the heat sink.Fig. 6.4 shows the dimensionless heat transfer coefficient in each channel of various substrate materials for EG-base fluid.The highest ℎ is achieved in copper made microchannels, which has the lowest thermal diffusivity followed by aluminium, steel, and titanium.The increase in ℎ could be attributed to the decreased conduction resistance along the microchannel wall due to much lower thermal diffusivity in copper made microchannels.The averaged heat transfer coefficient of oil and glycerine base-cooled MCHS shows a similar trend as for EG base-cooled MCHS.However, the condition becomes different for low Prandtl number base fluid (water) where the highest ℎ is present in steel made microchannels as depicted in Fig. 6.5.This proves that the thermal diffusivity of substrate material on the heat transfer in micro domains has significant effect for low and high Prandtl number base fluids.Thus, based on the results presented in Fig. 6.5, it can be stated that the heat transfer performance of a low Prandtl number base fluid such as water can be greatly enhanced by using it as a coolant in MCHS of a high thermal diffusivity made substrate material such as steel.
The width to height ratio (W c /H c ) has a significant effect on the Poiseuille number for the rectangular cross-section microchannels.Poiseuille number increases linearly with the increase of W c /H c .
For the trapezoidal cross-section MCHS, the height-to-top width ratio (H/a), the bottomto-top width ratio (b/a), and length-to-hydraulic diameter ratio (L/D h ) are the important design parameters for trapezoidal microchannels.Poiseuille number increases linearly when H/a and L/D h decreases while b/a increases.
The tip angle ( ) has a remarkable effect on the Poiseuille number for the triangular cross-section MCHS.Poiseuille number increases linearly with the increase of .
In order to achieve overall heat transfer enhancement in MCHS, it is envision that rectangular cross-section MCHS with a small hydraulic diameter was of greater benefit to the heat transfer coefficient with a lower penalty in pressure drop and friction factor compared with trapezoidal and triangular cross section MCHS.For nanofluids in MCHS: The cooling performance of MCHS improved well when the particle volume fraction increases.However, the heat transfer becomes poor in larger amount of particle volume fraction of nanofluid flow as the heat transfer could not be enhanced compared with pure water.
The presence of nanoparticles substantially increases the friction factor and wall shear stress.Both are rises with the increase of particle volume fractions.
The nanofluid-cooled MCHS has lower thermal resistance than pure water-cooled MCHS.The thermal resistance decreases as the particle volume fraction increases.
Diamond nanoparticles dispersed in water is preferable since diamond-H 2 O-cooled MCHS has highest heat transfer coefficient value among the others. Ag nanoparticles dispersed in water is recommended as it has lowest pressure drop and wall shear stress among the other nanofluids studied.
In order to achieve overall heat transfer enhancement, a base fluid of high Prandtl number (210-12800) such as glycerin, oil, and ethylene should be used to maximize the merits of adding nanoparticles for fluid flow in MCHS.
Different types of base fluids do not have a significant influence on the friction factor.
The heat transfer for low Prandtl number fluid flow such as water is greatly enhanced in high thermal diffusivity made material such as steel compared with low thermal diffusivity materials.These findings suggest that for low Prandtl number base fluids, the substrate material of high thermal diffusivity is desirable and vice versa in order to achieve a better heat transfer in MCHS.
Fig. 2 .
Fig. 2.1.(a) Schematic diagram of the MCHS (b) Section of the MCHS cross sectional shapes with its dimensions.
Fig. 3 .
6 shows a comparison of predicted Poiseuille number which is the product of friction factor, f and
5
Wall shear stressThe effect of different volume fractions of nanoparticles on the local dimensionless wall shear stress is also investigated in the current study and the results are shown in Fig.4.8.The dimensionless wall shear stress is calculated using:
Fig. 5.2.Dimensionless pressure drop variation along the length of channel No.1.
Fig. 6 . 1 .
Fig. 6.1.Dimensionless temperature profile versus number of channels for pure water and various base fluids.Fig.6.3 shows the dimensionless heat transfer coefficient in each channel of the heat sinks for various base fluids-cooled MCHS.The ℎ for high Prandtl number base fluids is always higher than that for water base fluid.This is due to the fact that high Prandtl number base fluids are experiencing stronger thermal flow development effects as reported by Koo and Kleinstreuer[10].However, there is no significant differences in ℎ for EG, oil, and glycerine but when the plot is zoomed in, there are slight differences in ℎ where glycerin base-cooled MCHS has the highest ℎ value, oil is in between, and EG is the least.For each type of the heat sinks, the middle channel (channel number 14) has the highest ℎ value as expected.The averaged heat transfer coefficient value for other channels is seen to decrease depending on their distances from the wall.The averaged heat transfer coefficient distribution for all types of MCHS is almost symmetrical with respect to the centreline of the heat sink.
Table 2 .
1. Dimensions for three different sets of rectangular cross-section MCHS.
Table 2 .
2. Dimensions for three different sets of trapezoidal cross-section MCHS.
Table 2 .
3. Dimensions for three different sets of triangular cross-section MCHS.
Table 4 .
.1 is considered.Results of interest such as temperature, heat transfer coefficient, friction factor, wall shear stress, and thermal resistance are reported in this section.The forgoing thermophysical properties for the nanoparticle (alumina), base fluid (water), and the alumina-water nanofluid of three different volume fractions formulated are listed in Table4.2.The density, thermal conductivity, and dynamic viscosity of the nanofluid appear significantly increased while the specific heat of nanofluid decreased with the increase of its particle volume fraction compared to pure water.The nanofluid of φ = 5% alumina, for instance, has a relative increase of 14.88%, 15.17%, and 12.46% in the density, thermal conductivity, and dynamic viscosity and has a relative decrease of 14.14% in specific heat, respectively, compared to pure water.1. Dimensions of the rectangular cross-section MCHS. | 10,236.8 | 2012-02-24T00:00:00.000 | [
"Physics"
] |
A switchable light-input, light-output system modelled and constructed in yeast
Background Advances in synthetic biology will require spatio-temporal regulation of biological processes in heterologous host cells. We develop a light-switchable, two-hybrid interaction in yeast, based upon the Arabidopsis proteins PHYTOCHROME A and FAR-RED ELONGATED HYPOCOTYL 1-LIKE. Light input to this regulatory module allows dynamic control of a light-emitting LUCIFERASE reporter gene, which we detect by real-time imaging of yeast colonies on solid media. Results The reversible activation of the phytochrome by red light, and its inactivation by far-red light, is retained. We use this quantitative readout to construct a mathematical model that matches the system's behaviour and predicts the molecular targets for future manipulation. Conclusion Our model, methods and materials together constitute a novel system for a eukaryotic host with the potential to convert a dynamic pattern of light input into a predictable gene expression response. This system could be applied for the regulation of genetic networks - both known and synthetic.
Background
Gene expression systems with both spatial and temporal regulation are key components of engineered and synthetic biological networks. Engineered systems generally use a controlled external stimulus to signal to a specific promoter element, producing a rapid and dose-dependent response [1]. The external stimulus, used at the level of both the whole organism and cell culture, has often been a small, cell permeable molecule, which functions as an activator for the corresponding promoters [2][3][4]. Heat shock gene promoter systems can also be utilised for conditional gene expression using heat or irradiation as the stimulus [5].
The yeast artificial light switchable promoter system proposed by Shimizu-Sato et al. demonstrates many of the advantages of inducible systems, including low background expression, high inducibility, reversibility and dose-dependence [6]. It combines these desirable features with non-toxicity and a lack of pleiotropic and unantici-pated effects which are inherent properties of chemically inducible systems. This system is based on the properties of the plant phytochrome B photoreceptor (PhyB), which reversibly changes its conformation in response to red (λ max = 660 nm) or far-red light (λ max = 730 nm). The farred light absorbing conformer (PhyB Pfr) binds to the phytochrome interacting factor 3 (PIF3) protein, whereas interaction between the red light absorbing conformer (PhyB Pr) and PIF3 is much less efficient [7]. In the proposed system, PhyB and PIF3 are expressed as chimeric proteins, fused to the DNA-binding (GBD) or the transcriptional activator (GAD) domain of the GAL4 transcription factor, respectively, giving a typical two -hybrid interaction assay. The cis component of the system is the lacZ reporter gene controlled by a GAL4-responsive artificial promoter. In darkness, PhyB-GBD binds the promoter, but does not induce transcription. Red light illumination converts PhyB into the Pfr form, therefore facilitating PhyB-PIF3 interaction, which recruits PIF3-GAD to the GAL4-dependent promoter resulting in the activation of transcription. Subsequent far-red light illumination coverts PhyB Pfr to Pr and this is followed by the dissociation of the PhyB-GBD -PIF3-GAD complex and abrogation of transcription. The authors demonstrated the dose-dependent response of the system and the dynamics of photoreversible activation of the lacZ reporter gene, derived from quantitative liquid culture assays.
Recently, another genetically encoded signalling system based on PhyB -PIF3 interaction, with different chimeric proteins, has been successfully used for photoswitching of actin assembly through the Cdc42-WASP-Arp2/3 pathway in E.coli [8].
All phytochromes (PhyA-E) in the model plant Arabidopsis thaliana are capable of light-dependent conformational changes, but interacting proteins have only been investigated for the two most abundant phytochromes (PhyA and PhyB) [7,9,10]. FAR-RED ELONGATED HYPOCOTYL 1 (FHY1) and FHY1 LIKE (FHL) proteins control the nuclear import of PhyA via specific interactions with the Pfr conformer [11,12]. It follows that, besides the PhyB-PIF3 pair, other phytochrome-interacting protein combinations could be employed as the "light sensing" module of the expression system.
Functional phytochrome receptors consist of the apoprotein and the covalently linked chromophore called phytochromobilin. Since the chromophore is not synthesised in yeast, an analogous compound, phycocyanobilin (PCB), purified from cyanobacteria, is added to the media. PCB is taken up readily by yeast cells and is autoligated by phytochrome apoproteins resulting in photochemically functional phytochrome photoreceptors [13][14][15]. When expressed in yeast with PCB, PhyA behaves like other phytochrome receptors: the Pr ↔ Pfr conversion is controlled by red and far-red light [15][16][17].
The light switch described by Shimizu-Sato at al., translates light-dependent protein interactions into transcriptional regulation of a selected gene [6]. Beta-galactosidase is the most widely used reporter gene in yeast; however, the protein has a half-life of more than 20 hours in this system, and it can be detected in vitro only [18]. By comparison, the firefly luciferase has a 1.5 hour half-life in yeast, and luciferase activity (luminescence) can be monitored in real-time and in vivo, which makes this reporter a better tool for monitoring dynamic changes in transcription, as has been elegantly demonstrated recently through the monitoring of the cell-cycle and respiratory oscillations monitoring in agitated liquid yeast culture [19,20].
Our aim was to create and mathematically model an inducible gene expression system, based on the principles described above, but containing novel components that provide more stringent regulation and in vivo real-time detection of transcription in yeast colonies on solid media.
Selection and testing of components for the light inducible expression system
Detection of promoter induction via beta-galactosidase activity is a well characterised method in S. cerevisiae however it requires time-consuming sampling and in vitro analysis. In order to provide a real-time, in vivo detectable reporter in our system, the GAL4-responsive GAL1 promoter was fused to the firefly luciferase gene (GAL1:LUC) (Fig 1). Figure 1B shows the resulting gene circuit in the community standard Systems Biology Graphic Notation (SBGN) [21]. As a constitutive control, the ADH1:LUC (ALCOCHOL DEHYDROGENASE I) construct was prepared and stably integrated into the genome. Yeast colonies prepared as described in Materials and Methods reached a steady state of luminescence 16-18 hr after luciferin was applied (Additional file 1). As expected, ADH1:LUC produced much higher light emission than GAL1:LUC independent of the GBD/GAD fusion proteins expressed (Additional files 1 and 2). Separate set of patches were irradiated with red light (R) or far-red light (FR), or R immediately followed by FR (R/FR), or were kept in darkness. R light induced a rapid increase of luminescence in the case of yeast patches expressing GAL1:LUC, but not in the ADH1:LUC-expressing patches or in patches expressing GAL1:LUC without GBD/GAD fusion proteins (Fig 2 and Additional files 1 and 2). Luminescence reached a maximum 14-16 hr after the R light, followed by a slow decrease. In contrast, FR light alone induced very low levels of luciferase activity, which was essentially the same, when R light treatments were followed by FR light immediately (Fig 2). Since the relative (fold) induction was the highest in yeast cells having the GAL1:LUC reporter and expressing PHYA-GBD and FHY1-GAD (Fig 2A), this set of interacting proteins were used in further experiments. These results demonstrate that (i) appropriate LUC markers can be used to report phytochrome photoconversion and light-induced protein-protein interactions in our system; (ii) LUC enzyme activity is unaffected by light in yeast and (iii) yeast patches grown on solid media and treated with luciferin represent stable and reliable experimental material for luminescence imaging.
Luciferase as a reporter for gene expression in yeast
In order to calculate changes in the rate of transcription from real-time luminescence data, it was necessary to determine the relationship between transcription and enzyme activity.
The copper inducible CUP1 promoter was fused to luciferase gene, expression was induced and CUP1: LUC RNA and luciferase activity were tested over 7 hours time course. Figure 3 shows a 3-4 h delay in the induction of LUC activity relative to LUC mRNA expression. These data contributed to determine the kinetic parameters (mRNA half-life, translation rate) for the Luc reporter model.
An unidentified compound functions as a chromophore Text for this sub-section
Phytochromes are chromoproteins consisting of the apoprotein and a covalently linked, linear tetrapyrrole chromophore, phytochromobilin (PΦB) [22]. In the absence of chromophore, phytochromes cannot absorb light, do not show light dependent conformation changes and, therefore, do not function as photoreceptors. Phytochrome apoproteins are synthesised in the Pr form in plants and after autoligation of PΦB are capable of light absorption and photoconversion into the Pfr conformer.
Light-responsive gene promoter system , or PHYBNT-GBD/GAD-PIF3 (C) fusion protein-pairs were grown in darkness to form patches (merged colonies) for two days at 30°C, treated with 2.5 mM luciferin and transferred to 22°C for 17.5 h. Separate yeast patches were irradiated with single red (R), or far-red (FR) light pulses, or with red pulses immediately followed by far-red pulses (R/FR), or were kept in darkness (Dark). Luminescence values normalised to the pre-pulse levels are shown; time 0 h is the start of the light treatment. The luciferin pretreatment is shown in Additional file 1. E: Selected luminescent images of yeast patches used to obtain data in panel A. I: red light-induced, NI: non-induced dark control, T0: last images before the light pulse. Consecutive luminescent images taken in every two hours are shown. Pictures are displayed in pseudo-colors: red-white or blue-black colors indicate high or low expression levels, respectively.
Cyanobacteria (Synechococcus and Synechocystis sp.) harbour phytochrome-like photoreceptors, which use a chromophore (PCB) with similar structure to that of PΦB [23]. Plant phytochromes binding PCB are fully functional photoreceptors and, because of the relative ease of PCB purification it is generally used as an exogenously added chromophore [24]. However, it was unclear whether yeast cultures contained chromophore-like compounds that could serve as the chromophore for plant phytochromes. To test this, yeast cells with the GAL1: LUC reporter and expressing PHYA-GBD and FHY1-GAD were grown on media lacking PCB and the same light treatments were administered as in Fig 2A. To our surprise, significant R induction was detected in the absence of PCB (Fig 4). The fold-induction was reduced to 30% compared to the results with added PCB (Fig 4 vs. Fig 2A). Moreover, Figure 4 shows that FR light alone, or R followed by FR light also gave qualitatively very similar results to the photoreceptor with PCB. The basal expression level of GAL1:LUC was not affected significantly by the presence or absence of PCB (Additional file 2). These results demonstrate that an unidentified compound naturally present in yeast, can serve as a chromophore for phytochromes expressed in this heterologous system.
Model assumptions and structure
To use our regulatory system for synthetic biology we developed an ordinary differential equation (ODE) model of its function based on kinetic data from the literature and experimentally determined parameter values ( Fig 1A, B and Additional file 3). The model describes all the known phytochrome properties (e.g. photoconversion, dark reversion, sequestration, etc), using yeast phytochrome data to provide a realistic description of the light-switch function (for the detailed model description and structure, see Methods).
In summary, the model assumptions are: 1) Overall concentrations of Phys and PIF3/FHY1/FHL are constant, 2) Before the light impulse all the phytochromes are in the inactive (Pr) form, and sequestered in a slow acting pool. This might be related to their inclusion in SAPs(Sequestrated Areas of Phytochromes)-like structures similar to those observed by microscopy in cytosol [15] (see Methods for more details).
3) The dark reversion rate is the same for the free Phy and the Phy-FHY1/FHL complex. 5) The luciferase -luciferin subsystem is approximated as a steady state before light treatments.
Kinetics of induction of luciferase mRNA and luciferase activ-ity from the CUP1:LUC reporter gene Figure 3 Kinetics of induction of luciferase mRNA and luciferase activity from the CUP1:LUC reporter gene. Yeast cells harboring the CUP1:LUC construct were grown in liquid rich media overnight at 30°C. CUP1:LUC expression was induced by 1.5 mM copper-sulphate (final concentration) at time = 0 and aliquots were harvested hourly. The samples were used to prepare crude protein extracts and to isolate total RNA. Luciferase activity was measured by in vitro assays and luciferase mRNA was determined by qRT-PCR reactions. Switching without PCB Figure 4 Switching without PCB. Yeast cells harboring the GAL1:LUC reporter and expressing PHYA-GBD and GAD-FHY1 fusion proteins were grown in darkness at 30°C for two days on media without PCB. Luciferin and light treatments and imaging were performed as in Fig. 2A. 6) The initial sharp decrease in luminescence following the application of luciferin is due to the diffusion of luciferin from the point of application on the yeast patch into the agar medium. This enables the model to provide a good fit to conceptually similar systems with different interaction partners. For example, the adapted model fits Shimizu-Sato's data with good accuracy using parameter values derived from the literature [6,7,16,25]. This model is simpler, for the main part because the slow LacZ degradation obscures the long-term kinetics (see the model equations in Methods and simulation results in Additional file 4).
To account for the initial difference in cell density for each yeast patch that affects luminescence intensity ( Fig 5) and for the non-uniformity of the solid media, the initial conditions were set for each patch (each experiment) individually, so that each experimental curve is considered for two regimes, "diffusion" and "phytochrome". The former starts from the application of luciferin with initially decreasing luminescence, which approaches an approximate steady state after 17-18 hours. The latter begins from the light treatment and continues to the end ( Fig 6A). The first regime fits separately to the diffusion part of the model and provides the initial luciferin level at the time of light application. Such decomposition of the initial conditions is introduced to describe the temporally changing substrate availability that emerges from the solid culture conditions. Modelling the solid media allows a much wider variety of experimental applications in conditions that are most common in yeast and synthetic biology. The diffusion coefficient in our experiments corresponds to diffusion rate of 3 to 10 mm/h, which is in good agreement with literature data for agar gel [26].
Parameter values for model equations were obtained from fitting the model to all the timeseries data from luciferase imaging (Fig 6A, B), within the parameres ranges derived from the literature ( Table 1). The parameter data for luciferase protein degradation rate was supported by additional experiments using cycloheximide treatment of yeast cultures constitutively expressing the luciferase gene (data not shown). The degradation rate constant is estimated to be 0.2-0.8, which corresponds to a 0.8-3 hour half-life. This is similar to the value measured in yeast and mammalian cell cultures and plants [27][28][29].
Model predictions
The refined model both captures qualitative dynamics and enables a quantitative description of the light switching behaviour. Moreover, it enables us to deduce which parameters would be critical for particular behaviours of the system. By varying these parameters we showed that predicted intermediate state during the photoconversion of Pfr_FHL is crucial to match the slow switching off of the observed LUC expression (Fig 7A). The biochemical nature of this state as well as its experimental measurement is the subject of further experiments. Also, according to the model simulation (Fig 7B, C), shortening of reporter protein half-life does not affect the longevity of reversal of the transcription activation but significantly reduces the intensity of the luminescence. The light switch model also gives several predictions about the long-term system behaviour ( Fig 8A). In particular, we predicted based on experimental data for 50 h, that in the experimental conditions considered the complete removal of the transcriptional activation effect should take a relatively long time (100 hours), and this has been confirmed by experiments (data not shown). Furthermore, with the given dynamics, we can manipulate subsequent applications of R and FR, to achieve a wide range of desirable profiles of transcription activation ( Fig 8B, c, and 8D). Fig 8B illustrates the different types of behaviour of light input, which depend on the interval between R and FR treatments. It is clear from simulations that a small interval (2 min) between R and FR causes the increase of transcription rate and, accordingly, the increase of the luminescence intensity with time. Meanwhile, a longer interval (5 hours) produces the stable base line of input oscillations. On the basis of these simulations one can create a specific protocol of light input, combining the given modes as required, and thus obtaining the "square" shape ( Fig 8C) with two modes of light regime, and "sigmoid" shape ( Fig 8D) with three modes. Thus, the overall system could be used as a tool for the design of experiments with flexible perturbations of the system, for example, by changing the time intervals between.
Discussion
We developed a photo-regulatory genetic switch for yeast cells that combines several desirable properties. In addition to the widely recognised interacting pair PhyB-PIF3 we have tested other possible protein combinations. We found that the PhyA-FHY1 (and PhyA-FHL) pair provides higher induction level with lower background than that of the PhyB-PIF3 pair in our experimental conditions.
Previous experiments were carried out using agitated liquid yeast cultures at 30°C [6]. We used yeast colonies grown on solid media, for the reason that this setup facilitates light treatments, continuous monitoring of luminescence and potentially allows spatial patterning of light input and biological response. Our experimental system represents a reliable, reproducible and simple set up for investigation of dynamic transcription.
Phytochrome photoperception in yeast has previously been reported with addition of exogenous chromophore (PCB) [6,11]. Our system also showed light responsiveness without exogenous chromophore, albeit at a lower level (Fig 4). We propose that phytochromes can employ an unidentified compound from yeast as a chromophore, but the light absorbing efficiency of the constituted receptor is less than that of the holoprotein binding PCB. As a result, R treatment induces a reduced amount of phytochrome Pfr, which results in less efficient induction of transcription. The heterologous chromophore may be specific for some yeast strains, or its weak activating effect could have been difficult to detect using previouslyemployed reporter genes.
We developed a mathematical model that describes the system and fits the experimental data with great accuracy. The model incorporated experimental variability arising from the cultures on solid media, via substrate diffusion that corresponds to observations and sets the initial sub- (Fig 6B), model fits better to the longer intervals (at least 1 hour between R and FR treatments), compared with the shorter treatment intervals, when FR is given immediately or 30 min after R. It can be seen from the time series that there is only a small quantitative difference between immediate and 30-mindelayed FR, while after a 1 hour delay the shape of the response resembles that of a single R treatment (Fig 6B), differing only in amplitude. This qualitative shift in the system behaviour requires further analysis, which may shed light on mechanism of transcription activation by R light and inactivation by FR light in the system. We found that the kinetics of induction were slower in our conditions compared to previously reported experiments [6]. For our tests, yeast patches were grown at 30°C for two days, and then, due to technical issues, the plates were moved to 22°C in the imaging chamber, so the effect of light treatments was investigated at 22°C. We found evidence that the system responds more quickly at 30°C (data not shown), but a complete explanation requires further investigations.
Time course of luciferase luminescence intensity in different light conditions
Our system does not display an instantaneous shutting off of target gene expression. It takes a substantial period of time to completely remove the Luc signal after FR treatment. Modelling suggests that this is not simply due to stability of the Luc reporter (see Fig 7B, C), but rather reflects persistent PhyA activity. It should be noted that far-red exposure does not convert all the active PhyA into Pr form, but by itself produces about 3% of Pfr form [30]. Additionally, we propose a residual physical interaction between the Pr form of PhyA and FHY1/FHL as a possible explanation for the slow kinetics. This was supported with model simulations that correspond to the experimental kinetics. However, the properties of the intermediate state remain to be determined.
Conclusion
The current work initially aimed to create a system to provide well-defined, light-induced perturbations in tran- scription to a genetic oscillatory circuit, to effect entrainment of the oscillations to a rhythmic light regime. The light switchable system presented here meets the requirements for an entrainment tool; moreover the mathematical model will facilitate the design of any desired entrainment mode. Hence, using the light switch with the corresponding model provides a powerful tool for regular perturbing any gene system of interest with a predictable amplitude and period. Moreover, with spatially-patterned light inputs, such as images, the system would allow spatio-temporal regulation, which could facilitate a greater understanding of biological processes in which inter-cellular communication is involved.
Constructs, yeast strains and growth conditions
Plasmids expressing PHYA-GBD, PHYBNT-GBD, GAD-PIF3, GAD-FHY1 and GAD-FHL fusion proteins have been described [6,11,12]. PHYBNT corresponds to an Nterminal fragment of PHYB containing residues 1-621. GAL1, CUP1 and ADH1 promoters containing full 5' untranslated regions and the 3' un-translated region (terminator) of the GAL2 and ADH1 gene were amplified from S. cerevisiae PJ69-4A genomic DNA using the following primers: GAL1 Fwd: 5'-AAAGTCGACATTACCACCATATACATATCC-3' The promoter:luciferase-terminator constructs were assembled in pBluescript SK plasmid using the restriction sites designed for the PCR primers (sites are underlined in the sequences of the primers above). All plasmids were transformed into E.coli by the SEM method and cultured under standard conditions [31]. The GAL2 or the ADH1 terminator was used for the GAL1, CUP1:LUC or the ADH1:LUC construct, respectively. The constructs were verified by sequencing and re-cloned in the integrating plasmid pδ-UB [32]. The final clones were linearized with XhoI and transformed in yeast strain PJ69-4A by standard LiAC/carrier DNA/PEG protocol. Transformants were plated on synthetic dropout media (SD) without uracil SD(-U). Selected strains carrying the GAL1:LUC construct were co-transformed with plasmids pD153 or pGADT7 (Clontech) expressing GBD-or GAD-fusion proteins, respectively [6]. Transformants were selected and maintained on SD(-LW) plates. Preparation of media and transformation of yeast cells was done according to the Clontech Yeast Protocols Handbook (Clontech).
Model simulation and predictions
In vivo luminescence imaging, light treatments 2 ml of selective SD media was inoculated with yeast cells and incubated for 16 hr at 30°C with agitation. 20 μl drops of the cultures were transferred to SD agar plates containing 10 μM PCB, irradiated with far-red light at 70 μmolm -2 s -1 fluence rate for 10 min and incubated for 48 hr at 30°C in darkness. All further manipulations were conducted under green safety light. Yeast cells formed merged colonies (or patches) with 5-8 mm diameter. 20 μl of 2.5 mM luciferin solution was pipetted at the center of each patch and the plates were transferred in the imaging chamber at 22°C. Images were taken every 15 minutes using a liquid nitrogen-cooled CCD camera (Visitron Systems GmbH, Munich, Germany). Luminescence was quantified using the Metamorph software (Molecular Devices, Downingtown, PA). Unless stated otherwise, light treatments were administered 17-18 hr after the application of luciferin. The duration of each light treatment was 10 min and fluence rate of light (independent of wavelength) was 70 μmolm -2 s -1 . Red and far-red light was provided by Snap-Lite LED modules (Quantum Devices, Barneveld, WI).
Induction of CUP1:LUC expression, qRT-PCR and in vitro luciferase assays
Yeast cells carrying the CUP1:LUC construct were inoculated in 10 ml of SD(-U) media and were grown for 16 hr at 30°C with agitation. The starter cultures were diluted to a final volume of 100 ml with fresh SD(-U) media. CUP1:LUC expression was induced by adding CuSO 4 solution to a final concentration of 1.5 mM. Samples were harvested hourly from induced and non-induced cultures. 2 ml or 100 μl of the cultures were pelleted and frozen for RNA quantification or for luciferase assays, respectively. Total RNA was isolated by using the RNeasy Plant Mini Kit (QIAGEN) according to the manufacturer's instructions. cDNA synthesis and qRT-PCR was performed as described [33]. Primers for qRT-PCR were: Luciferase-specific signals were normalised to ACTIN 1 (ACT1) levels for each sample. For in vitro luciferase activity measurements, frozen cell pellets were re-suspended in 100 μl of Cell Culture Lysis Buffer (Promega), and vigorously vortexed. After incubation on ice for 5 min, cell debris was pelleted by centrifugation and the supernatant was used as crude protein extract. 20 μl of protein extracts was mixed with 30 μl of the Steady-Glo Luciferase Assay Reagent (Promega) in the wells of a microtiter plate and luminescence was measured in the TopCount NXT luminometer (Perkin-Elmer) for an hour after the addition of the reagent. Counts during monitoring were averaged and normalized to total protein content of the extracts. Protein concentrations were determined by the Bradford assay [34].
Model description and structure
Our principal model system (Fig 1A, B) includes two chimeric proteins: phytochrome fused to the GAL4 DNAbinding domain (Phy_GBD), in the active (Pfr) and inactive (Pr) forms, and binding protein PIF3 (or FHL/FHY1) fused to the GAL4 activation domain (FHL_GAD). According to existing experimental data, the recombinant phytochromes are quite stable in yeast. Although the light lability of plant PhyA Pfr is well-described, no detectable difference was observed between the stability of the Pfr and Pr forms of oat PhyA over an 80 hour time period in yeast [24]; moreover, no significant decay in the total PhyA and PhyB amounts over 120 hours was reported [35,15]. This provides the basis for assuming that our model proteins are present constitutively, so neither production nor degradation occurs in the model system. Two pools, Pool_Phy and Pool_PIF3, fulfil the mass conservation laws for Phy and PIF3.
In plants, the Pr forms of phytochromes are localized in the cytoplasm in the dark and are translocated to the nucleus in their Pfr form after light absorption [36]. In the yeast system, however, all fusion proteins are constitutively nuclear-localized due to the natural Nuclear Localisation Sequence (NLS) present in the GBD tag or the presence of the SV40 NLS motif fused to the GAD fusion partner. Therefore, in this system the only light-dependent event is the interaction of phytochromes with their corresponding protein partners. Taken together, these details give us reason to locate the interacting proteins and the processes of association and dissociation in the nucleus.
Not instantaneous kinetics of induction (Fig 2) prompted us to suggest the existence of two phytochrome pools: slow and fast. It has been reported that the sequestration of recombinant PhyA into the cytosolic SAPs (sequestered areas of phytochrome) in yeast has no dependence on light [15]. We, therefore, propose the presence of sequestered and free Phy pools (less and more easy to access, respectively) in the nucleus with a reversible interchange occurring between them. We assume that only the free pool is available for binding to its interaction partner, and, thus, the transition between slow (sequestrated) and fast (free) pool is responsible for the shape of initial light response.
It is well known that the phytochrome photoconversion cross-section (σ) for Pr and Pfr forms depends on the wavelength of light. Red (approximately 660 nm) and far red (approximately 730 nm) light are the most effective for Pr → Pfr and Pfr → Pr photoconversions, respectively. Nevertheless, it is evident from the cross-section data that the absorption spectra of the Pr and Pfr forms of Phy significantly overlap [30]. This means that monochromatic light of biologically relevant wavelength (i.e. red) does not convert all the Phy to the Pr or Pfr form, but rather determines the specific distribution ratio of the forms in the total Phy pool. We thus have to account for the activation and inactivation of phytochrome by both red and farred light, so that: Exact values for Ki and Ka for the different wavelengths were adopted from [30]. In the model Pr ↔ Pfr transitions are applied for both associated and free form of the phytochromes.
Dark reversion has been reported for PhyA and PhyB in yeast cultures [15,35,13]. According to these data, only a fraction of the total Pfr pool is subject to dark reversion (20-40% of the total amount) with a half-life of 20-40 min. For simplicity in the current model we assume a single Pfr pool that is dark reversible and has a longer halflife than the range suggested by Hennig et al [13]; however, the model is still in good agreement with the overall kinetics described in the literature [35,15].
We assume that dark reversion of the complex Pfr_PIF3 (Pfr_FHY1, Pfr_FHL occurs with the same rate. Therefore, both the photoconversion and dark reversion processes contribute to dissociation of the transcriptional activation complex. Finally, for the PhyA_FHY1/FHL complexes, we assume the existence of an additional state, Pr_FHY1/FHL, that has the ability to activate transcription to some extent, as it has been previously demonstrated by [11]. Although the reference above corresponds to PhyA, in our experimental conditions PhyB demonstrated the same kinetics (Fig 2C), so we assume the intermediate state for PhyB-PIf3 as well. According to our hypothesis, this complex is produced as an intermediate product of photoconversion of the Pfr_FHY1/FHL complex after FR exposure. Thus, we propose that Pr proteins that have previously been Pfr can interact with FHY1/FHL and activate transcription.
Mass Action kinetics were used to describe complex formation and dissociation, translocation, translation, and degradation. Transcription was described with a Hill function and the reporter enzymatic reaction follows Michaelis-Menten kinetics (see Fig 1B for
( )
The equations (1)-(5) describe changes in concentrations of all the phytochrome components, while (6) and (7) correspond to changes in concentrations of luciferase mRNA and protein, respectively.
Luminescence level is calculated according to the Michaelis-Menten equation: Light emission is measured in terms of Relative Light Units (RLU) per second and this corresponds to the rate of the light emission reaction for the colony [28]. The parameter RLU is a conversion factor that translates the number of moles of luciferin reacted into the RLU measurement by the instrument. This also accounts for the discrepancies in colony sizes (Fig 5), growth rate, and instrument characteristics.
'Diffusion' part of the model
Our experimental setup involves the application of a relatively small amount of luciferin substrate (20 μl) to a yeast patch, growing in a 100-mm diameter plate on an agar gel of 5-7 mm thickness. We assumed that the initial decrease in luminescence level just after luciferin application predominantly resulted from the diffusion of substrate through the gel. This was confirmed by an additional experiment (Additional file 5). Taking into account that the thickness of the gel is much smaller than the diameter of the plate, we assumed that the diffusion of luciferin could be described with the diffusion equation in polar cylindrical coordinates: The particular solution of form was found to fit experimental data with the best accuracy.
Here, S is the cytosolic luciferin concentration, D is the diffusion coefficient, r0 is the effective colony radius, A and B are constants of integration.
PHYB-PIF3 Model Reactions (via Shimizu-Sato's system)
Model for the Shimizu-Sato's system has the similar structure but differs in reporter -LacZ. Model lacks the description of the reporter protein kinetics due to the stability of LacZ protein and the overall relative shortness of the timescale investigated in the paper (2 h) (See Additional file 4 for the data and model simulation).
Estimation of photoconversion rates
For estimation of photoconversion rates we used the data for the photoconversion cross-section of Pr, Pfr andP and Pfr/P ratios at photoequilibrium of type -I phytochrome [30].
Fitting to experimental results
The model was developed in SBTOOLBOX2 for MATLAB and fitted with a particle-swarm optimisation algorithm from the SBPD package in SBTOOLBOX2 [37,38]. Experiments were designed to cover all possible states of the system that have to be addressed in the model. We started with fitting the model to the simple experimental protocol, including dark conditions and red light application with or without the subsequent immediate far-red application ( Fig 6A). Dark experiments taken separately provided us with parameter values for the luciferase system (see Table 1), namely the degradation and translation rates, that were fixed during the following optimization procedure. Light response parameter values were estimated from R and R-FR experiments. For that the model was simultaneously fitted to five sets of ON-OFF experiments, each containing seven experiments: R, dark and five combination of R followed by FR with intervals 0 h, 0.5 h, 1 h, 3 h and 9 h (Fig 6B). Thus, a total of 35 timeseries (each of 210-360 timepoints) were fitted simultaneously. Fitting results demonstrate a good accuracy (see Fig 6A, B) with the root mean square deviation of 1.9*10 -3 .
As we aimed to account for increasing variability arising from solid culture conditions, our model parameters comprise the members which appear specific in each experiment. First of all, this relates to parameters corresponding to 'diffusion' section (D, r0, A and B) as they establish the initial conditions by the time of light treatment. Secondly, parameter RLU that accounts for variability in colony size and growth rate also has to be locally | 7,702.8 | 2009-09-17T00:00:00.000 | [
"Biology"
] |
Assessing the utility of whole-genome amplified serum DNA for array-based high throughput genotyping
Background Whole genome amplification (WGA) offers new possibilities for genome-wide association studies where limited DNA samples have been collected. This study provides a realistic and high-precision assessment of WGA DNA genotyping performance from 20-year old archived serum samples using the Affymetrix Genome-Wide Human SNP Array 6.0 (SNP6.0) platform. Results Whole-genome amplified (WGA) DNA samples from 45 archived serum replicates and 5 fresh sera paired with non-amplified genomic DNA were genotyped in duplicate. All genotyped samples passed the imposed QC thresholds for quantity and quality. In general, WGA serum DNA samples produced low call rates (45.00 +/- 2.69%), although reproducibility for successfully called markers was favorable (concordance = 95.61 +/- 4.39%). Heterozygote dropouts explained the majority (>85% in technical replicates, 50% in paired genomic/serum samples) of discordant results. Genotyping performance on WGA serum DNA samples was improved by implementation of Corrected Robust Linear Model with Maximum Likelihood Classification (CRLMM) algorithm but at the loss of many samples which failed to pass its quality threshold. Poor genotype clustering was evident in the samples that failed the CRLMM confidence threshold. Conclusions We conclude that while it is possible to extract genomic DNA and subsequently perform whole-genome amplification from archived serum samples, WGA serum DNA did not perform well and appeared unsuitable for high-resolution genotyping on these arrays.
Background
Array technologies are designed to rapidly genotype hundreds of thousands of single nucleotide polymorphisms (SNPs) across the genome using a relatively small amount of DNA. Advances in genotyping platforms have allowed cost-effective, whole-genome scans of multiple individuals in large-scale association studies. These studies are aimed at identifying genetic factors affecting many important complex human diseases. DNA samples from carefully characterized populations that are necessary to carry out adequately powered genome-wide association studies (GWAS), however, are often limiting. Collecting a sufficient number of appropriate samples for GWAS can be a complex and expensive collaborative challenge. There are potential alternative sources of genomic DNA for which important medical phenotypes have been recorded but their utility for GWAS has been poorly explored.
Archiving of serum samples is widely practiced in research and clinical domains [1]. Records of phenotype information associated with individual serum samples may be valuable for the GWAS setting. Archived serum samples are, therefore, an attractive and convenient potential source of genomic DNA. As an obstacle to genotyping, limited DNA yield could be overcome by taking advantage of recent whole-genome amplification (WGA) technologies [2][3][4][5][6][7][8].
Multiple displacement amplification (MDA), in particular, is an improved WGA technology known to minimize amplification bias, incomplete genome coverage, and generation of relatively short fragments. MDA utilizes a highly processive ϕ29 DNA polymerase and a mix of random hexamer primers and is capable of amplifying partially degraded and low quantity DNA sources. Reliable MDA-based whole-genome amplification of DNA from serum samples has been demonstrated [9]. Previous studies have successfully utilized WGA DNA from archived sera on a relatively small number of SNP markers using single assay methods [10][11][12]. A recent study by Mead et al showed that WGA serum DNA could be used for custom targeted medium-throughput genotyping [13]. Nevertheless, an unbiased genome wide scan with SNP markers using residual DNA in archived serum specimens will require the use of high throughput array-based genotyping. It is, therefore, necessary to determine whether this is technically feasible with standard technologies that are currently available for GWAS.
In this study, we describe a large and statistically robust study of genotyping WGA serum DNA on a widely-used whole genome SNP-genotyping panel. The Genome-Wide Human SNP Array 6.0 features about 1.8 million genetic markers, including assays for more than 906,600 single nucleotide polymorphisms (SNPs) and 946,600 probes for detection of copy number variation [14]. We further compare the performance of two different genotype-calling algorithms that are compatible with SNP 6.0 genotyping array (Birdseedv.2.0 and CRLMM). Compared to previous studies, the results of our work provide more precise and accurate estimates of genotyping efficiency and error rates using WGA serum DNA. Establishing archived serum samples as a reliable, alternative DNA source may boost the power of large-scale GWAS particularly in cases where available DNA samples from suitable subjects are limited.
Results
Quality control for WGA serum DNA SNP 6.0 genotyping We successfully isolated genomic DNA from 100% of seventy-five, 20-year old 500 uL frozen serum samples using the Qiagen DNA Mini Kit (Qiagen, CA). The yield of genomic DNA ranged from 31.5 ng to 608.4 ng (mean +/ -SD, 202.7 +/-122.3 ng). Purified DNA samples were then whole-genome amplified via multiple displacement amplification (MDA). MDA was the method of choice for WGA in this study because our pilot studies using Affymetrix TG array demonstrated superior genotyping results from MDA-amplified serum DNA compared to samples amplified via other WGA methods (results not shown). Furthermore, Affymetrix recommends the use of Repli-G for their genotyping chemistry. We avoided the use of WGA methods which includes random fragmentation of DNA target because >200 bp fragment lengths must be amplified when using SNP 6.0 genotyping arrays. In this study, WGA of serum DNA samples by MDA resulted in 5500-fold to 65000-fold amplification relative to input DNA. Our yields for WGA DNA ranged from 30800 ng to 125900 ng (mean +/-SD, 18536 +/-11754 ng). To reduce genotyping error associated directly with variation in the WGA [15,16], the 40 top-yielding samples were carried over to the genotyping process. All samples passed to the genotyping step produced relatively uniform quantity and quality of DNA (purification yield of 229 +/-136 ng; whole-genome amplification yield of 84409 +/-6086 ng; moderate to intense amplified PCR bands, and intact 10 kb fraction in 1% agarose gel). Two aliquots were taken from each DNA preparation for WGA and were genotyped as technical replicates. All WGA DNA were successfully run on the SNP 6.0 array, with the exception of 1 technical replicate sample. However, quality control (QC) call rates of WGA serum DNA samples ranged from 49.2% to 95.6% (mean +/-SD, 67.1 +/-10.1%) when analyzed by the standard Birdseed algorithm [17]. WGA serum DNA performance based on the SNP6.0 QC call rates showed poor correlation with yield after DNA purification (r 2 = 0.122), WGA yield (r 2 = 0.164) and intensity of 60 bp PCR-amplified band (r 2 = 0.001).
Call rates
To evaluate SNP 6.0 array genotyping efficiency on WGA serum DNA samples, we measured the proportion of SNPs with missing calls in the genotyped samples. Genomic DNA from peripheral whole blood paired with whole-genome amplified DNA from freshly isolated serum of 5 individuals were run on SNP 6.0 arrays as controls. Genotype data were inferred by implementing the Birdseed v2.0 algorithm at 0.1 confidence threshold in separate clusters of non-amplified and WGA DNA samples (see Additional File 1 and Additional File 2). Percent call rates from WGA serum DNA and genomic DNA are shown in Figure 1. Overall, WGA DNA samples from sera yielded significantly low call rates (mean +/-SD, 45.0 +/-2.7%), which corresponded to 409,366 markers out of approximately 906,600 SNP markers represented on the SNP 6.0 array. In contrast to WGA serum DNA samples, non-amplified DNA samples from peripheral whole blood yielded excellent call rates, ranging from 96.9% to 99.5% (mean +/-SD, 98.0 +/-1.1%). The average call rate of WGA DNA samples from archived sera (mean +/-SD, 44.6 +/2.7%) did not differ from WGA DNA samples from freshly isolated sera (mean +/-SD, 48.1 +/-1.8%) (p = 6 × 10 -5 ). Call rates between technical replicates were correlated (r 2 = 0.71), thus demonstrating that high quality serum DNA samples consistently produce high genotyping call rates compared to samples with poor quality DNA( Figure 2).
Genotype concordance between technical replicates
We further evaluated the quality of SNP 6.0 genotyping performance on WGA serum DNA samples by measuring the repeatability of calls between technical replicates as shown in Figure 3 (see Additional File 3). Reports for concordance were restricted to SNP loci with complete genotype calls on both technical replicates of a given individual. Genotype data between technical replicates gave modest percent concordances, ranging from 77.3 to 99.9 (mean +/-SD, 95.2 +/-4.6) over approximately 340,000 SNP loci. We then determined the percent contribution of allele switch (AA ←→ BB) or heterozygote dropout (AB → AA or BB) to the observed global discordance as shown in Figure 3 (see Additional File 4). In all technical samples, discordance occurred largely due to heterozygote dropout, with an average of 85.9%. Allele switch was less common and contributed to 14.1% of the global discordance.
Genotype concordance between paired samples
In order to assess whether accurate genotypes can be obtained from WGA serum DNA using SNP 6.0 arrays, we examined the fidelity of WGA serum DNA genotypes to their corresponding non-amplified DNA genotypes ( Table 1). Calculation of concordance was restricted to SNPs where calls are made on both samples of a pair. Unlike technical replicates, paired samples showed gross discordances that ranged from 17.77% to 44.25% of the average 282,256 called markers (mean +/-SD, 31.03 +/-12.05%). In all paired samples, the majority of discordant markers were caused by heterozygote dropout (AB→AA or BB), with an average of 50%. Of the observed heterozyogote dropouts, 69.3% were due to conversion of purine to another purine or pyrimidine to another pyrimidine. In contrast, other genotyping errors such as allele switch (AA←→BB) and heterozygote gain (AA or BB → AB) contributed an only average of 16.9% and 33.2%, respectively, to global discordance ( Table 2). The genotype concordance (reproducibility) for identical samples using the SNP6.0 microar-Relative call rates between archived samples and controls
Performance Comparison of Genotype Calling Algorithms
Previous studies have shown that genotype calling algorithms show variable performance in normalization and clustering [18]. We sought to determine whether a different clustering algorithm would improve the repeatability and accuracy of genotype inference when WGA serum DNA samples were used. Genotypes of technical replicates and paired samples were inferred using CRLMM (Corrected Robust Linear Model with Maximum Likelihood Classification) algorithm from separate signal intensity clusters of non-amplified and WGA DNA samples at 0.999 confidence threshold (see Additional File 5, Additional File 6, Additional File 7 and Additional File 8).
Chip Quality Measures as Sample Rejection Determinants
To investigate the basis for sample rejection by CRLMM, we visually assessed the chip quality of WGA samples. Signal intensities from probe A (Theta A) and probe B (Theta B) of SNPs assayed on SNP6.0 platform were plotted to capture the extent of differences between the intensity values AA, AB and BB genotypes. In figure 5, we show that sample performance correlated well with the separation of genotype clusters on the chip of a given sample. Genotyped samples with high confidence values and high call rates, such as genomic DNA (Figure 5a) or a representative successful WGA serum DNA (Figure 5b), showed three distinct clusters, corresponding to the three genotypes. On the other hand, very poor separation of clusters is the hallmark of the poor-performing and rejected WGA DNA samples ( Figure 5c).
Effects of Amplicon Fragment Length on Individual Marker Performance
We next determined whether amplicon fragment length affect genotype concordance and marker genotyping efficiency. As a measure of marker genotyping efficiency, we computed the average confidence values and concordance *Number of markers with calls between paired samples **%Concordance is the proportion of markers with consistent genotype calls in a paired sample.
Discussion
Our data show that DNA from archived serum samples cannot substitute for good quality genomic DNA using the standard technologies employed in this study. In our effort to establish serum as an alternative DNA source for future large-scale GWAS, we have extensively tested the performance of WGA DNA recovered from sera of 40 twenty-year old archived samples on the SNP 6.0 platform. Large standard deviations in DNA yield during the early stage of DNA preparation suggested that DNA from archived serum samples were variably prone to degradation. Selection of WGA DNA samples with uniform yield and quality did little to improve genotyping performance.
Of note, all WGA DNA from serum samples in the study gave call rates of less than half of the total markers assayed on SNP 6.0 platform, far below the call rates of our nonamplified genomic DNA controls (mean +/-SD, 98.03 +/ -1.05%). More importantly, genotypes of WGA DNA from serum samples and their non-amplified DNA coun- terpart showed gross mismatches affecting 80-300,000 markers out of ~430,000 total called markers. For samples in which non-amplified genomic DNA were available, the relative proportion of mismatches as reflected by %discordance was consistent between paired samples (Table 1) and their corresponding technical replicates (see Additional File 3). For example, the best performing sample (INF002) consistently produced the least %discordance in both paired sample and technical replicate compared to other samples. Our data altogether suggest that the quality of DNA that can be extracted from archived serum samples is inherently very poor. This factor alone accounted for the vast majority of genotyping failure in this study.
We further studied possible sources of inaccurate genotypes. Heterozygote dropout was the most common error detected in discordant assays in both technical replicates (85%) and paired samples (50%). This suggests that DNA in archived serum samples were prone to allele bias during WGA. Allele switch in genotypes of technical replicates may have arisen from generation of amplification error during WGA stage although poor genotype calling is the more likely explanation. This accounted for <15% discordances in technical replicates and 16.9% in paired samples of the discordant genotypes. Lower than average probe signal intensities (Figure 5c), which resulted in loss of genotype information or heterozygote gain (33% discordant markers in paired samples), may have been caused by inefficient amplification during WGA due to the variable presence of highly degraded DNA fragments in the serum samples. Such errors in WGA were previously reported particularly when suboptimal DNA sources were used [2,9,12,13,19]. However, poor correlation between probe performance across samples and lengths of amplified fragments on the SNP 6.0 platform suggest that the poor genotyping results from WGA serum DNA cannot be explained simply by DNA degradation (see Additional File 9 and Additional File 10). Indeed, genotypes were called and high concordance rates were observed in both small and very large fragment lengths of good performing markers.
The low genotyping efficiency and poor reproducibility of the genotype data generated with the present protocols suggest that archived serum samples will be a problematic source of DNA for large-scale GWAS [20]. To successfully locate causative variants in the genome, high genotyping efficiency and call rates are required from the initial whole-genome screen. Genotyping error also reduces statistical power in case control studies [21][22][23]. The CRLMM algorithm proved superior to Birdseed in improving the call rate (663320 +/-134762 called markers) and reproducibility between technical replicates (>99% concordance between technical replicates) at the expense of very low genotyping efficiency (6/45 samples passing confi-dence criteria). The CRLMM confidence measure gave a useful assessment of chip quality which was readily verified by inspection of the global genotype clustering. CRLMM, therefore, is currently the optimum genotype calling algorithm for the SNP 6.0 array available to most investigators. CRLMM has been previously shown to provide more accurate genotype calls and lower drop rates compared to other Affymetrix default algorithms (BRLMM and Birdseed) when using large datasets from non-amplified genomic DNA [18]. Our data validate and extend this finding to WGA DNA from serum samples. Genotyping algorithms apply different methods for transforming raw intensity signals from chip arrays to genotype calls. The good performance of CRLMM can be explained by the robustness of its confidence metric which appears to better reflect call accuracy relative to other genotyping algorithms. Furthermore, CRLMM not only allows measurements of call accuracy but also of chip quality. However, the improvement in reproducibility of WGA technical replicates has to be evaluated cautiously. Both Birdseedv2.0 and CRLMM produced genotypes that had poor concordance between WGA serum DNA samples and their corresponding non-amplified genomic DNA. In contrast to previous and more limited studies with targeted custom genotyping panels, our extensive survey of archived serum samples and our use of paired serum/ genomic samples clearly demonstrated poor genotyping performance of WGA samples on the SNP6.0 platform. Nevertheless, our study showed that approximately 300K genotypes could be reproducibly retrieved from 20-year old sera. This suggests that with further improvements in the laboratory protocols, serum could serve as a reliable, alternative DNA source for genetic studies. Other genomewide genotyping platforms can be explored for genotyping suboptimal samples such as serum DNA. Of these platforms, the Illumina Infinium II assay offers comparable marker density as the SNP6.0 platform [24]. Illumina Infinium II utilizes the bead technology wherein each bead contains locus specific probes, and allelic discrimination occurs through single-base primer extension reactions. In addition, several genotype-calling algorithms were developed for specific application to the Illumina assays such as GenCall [25], GenoSNP [26] and Illuminus [27]. There is a possibility of improving call rates, call accuracy and reproducibility because these platforms utilize different chemistries and genotype-calling algorithms.
Conclusions
Archived serum samples may be useful in cases where few putative causal variants have to be replicated in an independent set of samples. In this study, we showed that procedures could be applied to maximize the quality of genotype information from WGA DNA from serum samples. We have proposed to eliminate poor quality samples in downstream data analysis by imposing QC measures based on chip quality analysis. We have also shown that genotyping algorithms such as CRLMM could better analyze genotype data from less-optimal DNA source. Moreover, analysis can be limited to genotypes with high confidence values to improve data quality. However, further investigation is still necessary to determine the appropriate criteria needed to identify poor quality samples before the genotyping stage. It is not impossible though, that marked improvement in WGA and/or genotyping platform technologies may boost the reliability of serum as an alternative, cost-effective DNA source for future association studies.
Sample description
DNA was purified from 40 anonymized archived sera of participants from 3 pooled influenza vaccine clinical trials performed from 1981-1986 at Baylor College of Medicine, Houston, TX. Samples were selected from 695 individuals and were categorized as high-responders and lowresponders based on rise in antibody responses before and after immunization. As controls, additional DNA samples were obtained from peripheral whole blood and serum samples of 5 adult volunteer donors. Informed consent was obtained from these participants. This protocol was approved by Baylor College of Medicine's Institutional Review Board (IRB).
DNA purification from peripheral whole blood samples
Peripheral whole blood samples were collected in 10 mL citrate dextran (ACD), yellow top BD vacutainer blood tubes (VWR International, PA). Samples were kept at room temperature and processed within 3 days after collection. Purification of genomic DNA from peripheral whole blood was carried out using Gentra Puregene Blood Kit (Qiagen, CA) according to manufacturer's protocol. DNA quantity and A260/A280 ratios were measured using Nanodrop-ND-1000 spectrophotometer (NanoDrop Technologies, Inc, DE). Purified genomic DNA were stored at 4°C until ready for use.
Preparation of whole-genome amplified DNA from serum
Peripheral whole blood samples were collected in 5 mL BD Vacutainer red top tubes (VWR International, PA). Fractionation of serum was carried out by allowing blood samples to clot for 30-60 minutes at room temperature followed by centrifugation at 2000 rpm for 10 minutes. Serum supernatant was removed and stored in 500 uL aliquots at -80°C. Purification of genomic DNA from serum samples was carried out using QIAamp DNA Blood Mini Kit (Qiagen, CA). About 500 uL serum was incubated with Qiagen protease and equal volumes of buffer AL at 56°C for 1 hour and then loaded into QIAamp spin column. Genomic DNA was washed, dried and eluted with 20 uL buffer TE. Genomic DNA was quantified using Nanodrop ND-1000 (NanoDrop Technologies, Inc, DE). For whole-genome amplification of purified serum DNA, Repli-g Midi kit (Qiagen, CA) was used according to the manufacturer's instructions in the user-developed protocol for serum and plasma samples. For positive control, 10 ng human genomic DNA were used in every run. Briefly, 5 μL of purified serum DNA underwent denaturation step by incubating with equal volume of reconstituted buffer DLB on ice for 10 minutes. The reaction was neutralized by adding 10 μL of reconstituted stop solution. Whole genome amplifications were carried out using thermocycler (PTC-225, MJ Research Inc, MA) by incubating 30 μL DNA solution with ϕ29 DNA polmerase in 40 uL master mix, as provided by the manufacturer, at 30°C for 16 hours followed by enzyme inactivation at 65°C for 3 minutes. WGA serum DNA were quantified using Nanodrop ND-1000 (NanoDrop Technologies, Inc, DE). Samples were frozen at -20°C and genotyped within 3 weeks of whole genome amplification.
Quality Control for WGA serum DNA samples
Starting DNA material from archived serum samples must not be highly degraded to ensure successful application in WGA and GenomeWide Human SNP Array 6.0. QC metrics were imposed at different stages of DNA preparation so that only samples with acceptable DNA quality were allowed to reach the genotyping process. Whole-genome amplification was restricted to purified DNA samples with yield ≥90 ng. Samples for whole-genome genotyping were required to have WGA yield of ≥40 ug. Samples, which passed the QC thresholds for yield, were further tested for DNA quality. To assess purity of WGA serum DNA samples, A260/280 ratios were measured using Nanodrop ND-1000 (NanoDrop Technologies, Inc, DE). The extent of DNA degradation was initially assessed in WGA serum DNA samples by carrying out 60 bp fragment amplification by PCR. Furthermore, WGA serum DNA samples were run on 1% agarose gel to detect the presence of a heavy band around the 10 kb region representing nonfragmented serum DNA. Samples with good DNA quality have A260/280 ratio>1.80, 10 kb band on 1% agarose gel and positive 60 bp fragment amplification by PCR. All genotyped samples passed all yield and quality QC requirements.
Whole-genome genotyping
Genotyping was performed on replicate samples from the same DNA preparation using Genome-Wide Human SNP Array 6.0 Platform (Affymetrix, CA). The recommended protocol was followed as described in Affymetrix manual http://www.affymetrix.com/support/downloads/manu als/snp6_atp_userguide.pdf. Briefly, all DNA samples were normalized to 50 ng/μL. Two 5 _L (250 ng) aliquots were made from each DNA sample followed by digestion with either NspI or StyI restriction enzymes. Biotin-labe-ling primer amplification assay was performed on the resulting DNA fragments in each aliquot. The amplified products were fragmented, combined and purified using polystyrene beads. Samples were injected into the cartridges housing the oligonucleotide arrays, hybridized, washed and stained. The washing and staining procedures were run on Affymetrix fluidics station 450. Mapping array images were obtain by scanning with GeneChip Scanner 3000 7G (Affymetrix, CA). CEL files containing raw signal intensities were stored and used for downstream analysis of WGA serum DNA performance on Genome-Wide Human SNP Array 6.0 Platform (Affymetrix, CA) as described below.
Reproducibility
Inconsistencies in calls between replicate samples may result from errors in whole genome amplification process due to allele bias amplification or inaccurate definition of genotype clusters during genotyping. Genotype calls from Birdseedv2.0 were uploaded in R and Microsoft Access. SNPs with discordant calls between replicate WGA serum DNA samples were tallied. Of the total discordant calls, the proportion of SNPs with heterozygote dropout (AB->AA or BB) or allele switch (AA->BB) conversion were recorded and compared. If discordance between replicate samples were random, equal occurrence of heterozygote to homozygote conversion and allele 1 homozygote to allele 2 homozygote conversion between replicate sam-ples should be expected. Marked increase in heterozygote to homozygote conversion may suggest allele bias amplification during whole-genome amplification of serum DNA. On the other hand, increased frequency of allele 1 homozygote to allele 2 homozygote calls may reflect errors in defining genotype clusters during genotype calling.
Relationship of Fragment Length and Reproducibility to SNP Performance
Performance of individual SNPs may be affected by the extent of degradation in the starting serum DNA or difficulty in amplification of large DNA fragment during whole genome amplification process. Effect of varying fragment lengths on SNP performance was evaluated. Information on NspI or StyI fragment size associated with the assayed SNP was extracted from Genome-Wide Human SNP Array 6.0 annotation. Using Birdseedv2.0 output files, SNP performance was measured by calculating the average confidence values for each SNP across all genotyped WGA serum DNA samples. SNPs with low average confidence values are poor performers whereas SNPs with consistently high confidence values are good performers. SNP performance was also measured by calculating genotype concordance between technical replicates. SNPs with low average concordance rates are poor performers whereas SNPs with consistently high average concordance rates are good performers. Linear regression was performed on average confidence values against fragment size using R statistical software. The relationship between repeatability of genotype calls between replicate samples and SNP performance was also evaluated. Linear regression was performed on average confidence values against concordance between replicate samples to identify correlation using R statistical software.
Assessment of Signal Quality
Dropouts in calls or inconsistencies between replicates calls may result from non-distinct separation of clusters for each genotype. Quality of signal intensities for good and bad performing WGA serum DNA samples was visually examined. Normalized log signal intensities for probe A (θ A ) and probe B(θ B ) for each SNP were extracted from CRLMM output files using R statistical package software (v2.6.1) [30]. Geneplotter package was used to generate 2D density plots of θ B against signal θ A . Signal to noise ratio was assessed by examining the stratification of estimated regions for homozygous and heterozygous genotype calls. Good quality samples were expected to display distinct boundaries for each genotype cluster. | 5,958.4 | 2009-12-18T00:00:00.000 | [
"Biology"
] |
Dynamic Allocation/Reallocation of Dark Cores in Many-Core Systems for Improved System Performance
A significant number of processing cores in any many-core systems nowadays and likely in the future have to be switched off or forced to be idle to become dark cores, in light of ever increasing power density and chip temperature. Although these dark cores cannot make direct contributions to the chip’s throughput, they can still be allocated to applications currently running in the system for the sole purpose of heat dissipation enabled by the temperature gradient between the active and dark cores. However, allocating dark cores to applications tends to add extra waiting time to applications yet to be launched, which in return can have adverse implications on the overall system performance. Another big issue related to dark core allocation stems from the fact that application characteristics are prone to undergo rapid changes at runtime, making a fixed dark core allocation scheme less desirable. In this paper, a runtime dark core allocation and dynamic adjustment scheme is thus proposed. Built upon a dynamic programming network (DPN) framework, the proposed scheme attempts to optimize the performance of currently running applications and simultaneously reduce waiting times of incoming applications by taking into account both thermal issues and geometric shapes of regions formed by the active/dark cores. The experimental results show that the proposed approach achieves an average of 61% higher throughput than the two state-of-the-art thermal-aware runtime task mapping approaches, making it the runtime resource management of choice in many-core systems.
to the same application. There have been a number of studies on mapping applications to both active and dark cores [4]- [9]. They basically allocate dark cores to applications in the way that the active cores are allowed to operate at higher frequency levels, and thus, achieve higher performance at a cost of higher power consumption. However, these approaches fail to deliver optimized system performance due to the following reasons. First, as reported in [10], the application arrival rates vary significantly at different times. Particularly, as shown in Fig. 1(a), there is a vast gap between the maximum (highest number of applications arriving at the system per hour) and the minimum workloads, by as much as 200× [10]. Even over just one single minute, as shown in Fig. 1(b), the ratio of the maximum number of applications to the minimum can be as high as 6:1 [10], which implies that the number of dark cores can also vary greatly over that short time span. However, for the sake of simplicity, both schemes in [6], [7] inaccurately assume that the number of dark cores remains unchanged over a long time interval, undermining the quality of the application mapping results. An example illustrating two different schemes: how the free cores are assigned to five applications that system needs to service.
Second, due to workload fluctuations, allocating dark cores to applications necessitates the consideration of a number of competing requirements, such as throughput and individual application's waiting and completion times, as shown in Fig. 2. Assume that at initial time t 0 , four applications (A 1 -A 4 ) occupy core regions each of which also includes one or a few dark cores, as shown in Fig. 2(a). With application A 5 arriving at time t 1 , there are two possible allocation schemes that can map A 5 to the cores: • In one scheme, the dark cores already bound to A 1 -A 4 will be reallocated to A 5 such that A 5 can run immediately at time t 1 , but A 1 -A 4 have to slow down due to fewer dark cores available to help their active cores' heat dissipation (shown in Fig. 2(b)); • Alternatively, A 5 will be asked to wait until some applications (A 1 through A 4 ) finish their executions and thus their cores are freed up for A 5 to grab (shown in Fig. 2(c)). In this case, A 1 -A 4 can maintain their desired performance, but A 5 has to undergo a longer waiting time before it starts its execution. From Fig. 2, one can see that the mapping results generated from the two schemes differ significantly from each other in terms of performance (completion time) of currently running applications and the performance of the newly arrived or future applications.
Third, application's computation demand, measured by throughput in terms of instructions per cycle (IPC), varies with time. For instance, the computation demands of simultaneously running Facesim and Swaptions in the system, shown in Fig. 3(a), vary differently. As there are dark cores already allocated to Swaptions and it has decreasing computation demands as time passes by, the resource manager can reclaim some of the dark cores occupied by Swaptions, and instead allocate them to Facesim at time t 5 , as shown in Fig. 3(b). If the dark cores can be genuinely adjusted at runtime, both applications will be able to have their computation demands met.
In order to achieve optimized system performance by addressing the aforementioned challenges, we propose a runtime mapping scheme to dynamically allocate and adjust both active and dark cores. Here are the highlights of the proposed scheme.
• The proposed mapping algorithm takes the varying workloads, the waiting times of newly arrived applications, and the computation demands of applications into account, while the operating temperature is treated as a thermal constraint for safe and reliable operation of the chip. Instead of pushing each individual application's performance to its highest, our approach attempts to optimize the performance of currently running applications and the ones that are about to run.
• Based on a throughput model, a dynamic programming network framework is proposed to determine both the number of active and dark cores in the system for the newly arrived applications, and the number of dark cores that is allocated to executing applications, with the objective of maximizing the system performance.
• The mapping algorithm also includes region determination and task-to-core mapping. In general, the dark cores are placed near the cores that need to dissipate heat or run at higher frequencies. Moreover, the locations and geometric shapes of the core regions are regulated to minimize the communication latency and fragmentation of the free core regions, which further improves the system performance.
The remainder of the paper is organized as follows. Section II reviews the related work, and Section III describes the target system and provides the problem definition. Section IV presents the overview of the proposed method. Section V, VI, and VII describe the detail of three steps of the proposed method. Extensive experiments are conducted to compare the proposed scheme against the state-of-the-art thermal-aware runtime mapping methods, and the results are reported and analyzed in Section VIII. Finally, Section IX concludes this paper.
II. RELATED WORK
Runtime allocation of available system resources to tasks has been an active research area since the inception of the manycore era [11]. Of the many resource allocation approaches that have been proposed, they, based on whether remapping is allowed at runtime, can be broadly classified into two classes: • Dynamic mapping without task migration, where no mapping change happens after the initial task-to-core mapping; and • Dynamic mapping with task migration, where tasks can be mapped and remapped to different cores at runtime.
A. DYNAMIC MAPPING WITHOUT TASK MIGRATION
Dynamic mapping without task migration can be further classified into three categories according to their optimization goals: communication-oriented mapping, power-aware mapping, and thermal-aware mapping. Communication-oriented approaches (e.g., [12], [13]) aim at reducing network latency or minimizing traffic congestion, and they are similar to the contiguous mapping method [7]. However, these mapping approaches might lead to thermal hotspots in high power density chips since they do not consider the power budget [4].
Power-aware algorithms (e.g., [14], [15]) try to perform mapping under the thermal design power budget, which alone is not enough to avoid thermal violations, as found in [5]. As a fix, some thermal-aware approaches take the temperature of the cores into account during mapping [16].
The thermal-aware mapping approaches in [16], [17] try to minimize the power consumption and peak temperature.
As alluded before, the existence of dark cores presents opportunities to optimize system temperature. Failing to take advantage of the availability of dark cores might lead to suboptimal performance, as the cases of [16], [17]. To efficiently exploit dark cores, many dark-core-aware approaches have been considered [4]- [8]. The mapping approaches in [5], [6] assume that the system has a fixed number of dark cores, but in reality, the number of dark cores can vary significantly even in a short period of time [10]. Approaches in [5], [7], [8] do not consider the application arrival rate, and thus, their mapping results tend to cause applications to wait too long before they can start their execution. Although the work in [4] considers the application arrival rate when allocating dark cores to applications, a big drawback is that it cannot guarantee that the cores can meet the changing computation demands of applications.
In short, none of these dynamic mapping algorithms described here deliver the optimal performance, as they do not take full advantage of the dark cores in the system, workload variation, and changing computation demands of applications.
B. DYNAMIC MAPPING WITH TASK MIGRATION
Recognizing the deficiencies of the dynamic mapping approach without task migration, dynamic mapping approaches allowing task migration at runtime are proposed to help improve the runtime application performance. The dynamic mapping approaches can be classified into three categories: fragmentation-aware migration, communicationaware migration, and thermal-aware migration. Fragmentation-aware migration schemes (e.g., [18], [19]) reallocate tasks with a hope of forming a contiguous region of cores, while the communication-aware migration approaches (e.g., [20], [21]) focus on adjusting core allocation to minimize communication latency. Thermal-aware migration approaches (e.g., [22], [23]) move tasks from overheated cores to cooler ones to reduce hotspots. However, the above mapping approaches still do not exploit dark cores for better performance [8]. Although an early study in [9] presents a dark-core-aware migration algorithm to produce better computation performance, it does not address the changing computation demands of applications.
In the next section, we will present a runtime dark core allocation and adjustment scheme that addresses the outstanding issues of workload variations and applications' computation demands.
III. SYSTEM MODEL AND PROBLEM DEFINITION
A. THE TARGET MANY-CORE PLATFORM AND APPLICATION MODEL Fig. 4(a) shows the target many-core platform, which has a set of homogeneous cores Q, connected by a 2D mesh network. A core in Q is denoted as c i . One core will be designated as the resource manager and it has the authority and capacity to make any runtime core allocation and adjustment decisions. The many-core platform executes applications organized as a set, A = {A 1 , A 2 , . . . , A N }. When an application is ready to execute at time t, it is placed in the system waiting queue (denoted as H (t)). When an application in H (t) is allocated to certain cores for execution, it is added into the running queue (denoted as T (t)). When an application in T (t) finishes its execution, it is deleted from this queue. The notations used throughout the paper are summarized in Table 1.
To model the time-varying features of computation demands, an application A i is divided into mutiple phases, and at a phase τ the application is represented as a task graph AG i (τ ) = (V i (τ ), E i (τ )), as shown in Fig. 4(b). The task graphs at different phases can be obtained by the heartbeat framework [24]. V i (τ ) is the set of tasks associated with application A i , and E i (τ ) is the set of edges governing the communications among tasks. v ij is the j th task of application A i . e ijk is the edge of connecting tasks v ij and v ik indicating the communication between v ij and v ik . Each task v ij ∈ V i (τ ) has a weight a(v ij , τ ) which gives the execution time at phase τ . An edge e ijk = (v ij , v ik ) ∈ E i (τ ) has a weight of w(e ijk , τ ) that defines the communication volume in terms of the number of packets from tasks v ij to v ik at phase τ . Mapping a task to a core is defined as a one-to-one mapping; that is, only one task can run at a core, and no tasks can share a core at any given time [13]. A mapping function M (v ij ) = c i , maps task v ij to core c i .
B. THROUGHPUT MODEL
Application A i 's computation demand is measured by its throughput, which is the lumped throughput (IPC) of the cores running all the tasks of A i . A throughput model is set to compute the throughput of application A i (denoted as i,|B i (t)| ), with |B i (t)| dark cores assigned to A i . In specific, where B i (t) is the set that associates with application A i .ā i andw i are the average execution time and communication volumes of the tasks, respectively, and i (τ ) is given below.
The throughput model is used at runtime to estimate the throughput, given the number of dark cores. To find the regression coefficients β j , δ j , ϑ j , θ j and ε, the maximum likelihood method [25] can be used. The throughput model can be trained offline by running various applications. There are four steps.
Step 1 (Find a Near Square Shape): Since the throughput of application A i is associated with the core region, and a square shape is ideal to address the communication latency concerns [13], a core region close to a square shall be pursued when mapping applications. Let R i (t) be the core region of application A i , which includes the dark cores B i (t) and active cores. First, a basic square with length α = √ (|R i (t)|) is found. If φ = |R i (t)| − α 2 = 0, this region is a square and it shall be selected as the shape of core region for throughput modeling. For a region with a non-square shape, this region can take of the shapes made of a basic square combined with one or two rectangles.
• Case 2: one rectangle. If φ ≤ α, the near square shape consists of the basic shape and a rectangle of size 1 × φ, as shown in Fig. 5
(b).
Step 2 (Determine the Task Positions): The mapping method described in Section VII can be used to determine the positions of tasks. The cores that are not occupied by tasks in the core region of application A i are powered-off as dark cores.
Step 3 (Set Other Running Applications): In order to simulate the case that there are many other applications running simultaneously in the system, which also consume power, applications from PARSEC [26] are randomly picked and mapped to cores adjacent to the application of interest.
Step 4 (Set Voltage/Frequency Levels of Cores): When running applications, it is necessary to ensure that each core c i is running safely with its power consumption below the maximum power capacity P m (c i ), which is obtained from the thermal power capacity model in [4]. The total power consumption comes from the dynamic power P d (c i ) and leakage power P l (c i ). Therefore, The leakage power P l (c i ) can be obtained as in [27]. The dynamic power P d (c i ) is determined by: where µ i is the switching activity, z i is the effective capacitance, f i is the frequency of core c i , and i is the supply voltage. The frequencies, power, and throughput of the dark cores are 0. Once the positions of the tasks are determined in the system, the method in [4] is used to set voltage/frequency levels of the cores so that they can run at a high speed without violating temperature constraint and the maximum power capacity.
C. PROBLEM STATEMENT
The applications in set A arrive at the system at different times, and the objective is to maximize A , the system throughput of running the applications in set A.
Eqn. (6) can be transformed to maximize the system throughput of the application set A(t) which can be executed at time t. A(t) contains the applications that are either in the running queue T (t) or in the waiting queue H (t) at time t. With the throughput model, the maximum throughput for the application set A(t) can be computed by: where γ if_run i is a binary value. If γ if_run i is 1, application A i can start its execution immediately as there are sufficient cores available in the system. If γ if_run i is 0, it means application A i is put on hold and it waits for core(s).
At each control time, decision needs to be made regarding the number of dark cores to be allocated to each application, together with the task-to-core mapping.
IV. OVERVIEW OF THE PROPOSED METHOD
The decision to map a new application to cores, or adjust the core regions of running applications, could usher in a couple of challenges.
First, fragmentation of free cores [18] might occur. Dark cores released by other applications might not form a contiguous region, which increases the communication latency for newly arrived applications. VOLUME 8, 2020 Second, a near square shape of the core region is ideal for communication latency concerns [13]. However, the shape tends to be irregular after adding or removing dark cores, which might lead to increased communication latency.
To address these challenges, a three-step algorithm is proposed as follows: Step 1 (Dark Core Budgeting (Section V)): A dynamic programming framework is applied to decide the number of dark cores for the running applications and newly arrived ones.
Step 2 (Region Determination (Section VI)): Given the number of dark cores from the previous budgeting step, the shape and location of each application's core region are determined and reallocated to avoid fragmentation.
Step 3 (Task Mapping (Section VII)): A task mapping algorithm maps the tasks within its core region, together with the determination of the locations of the dark cores.
The proposed method will be triggered at each control time.
V. DARK CORE BUDGETING
Dark core budgeting, which decides the numbers of dark cores that shall be allocated for maximal throughput (defined in Eqns. (7) and (8)), can be transformed into the longest path problem in an acyclic network, where a dynamic programming network (DPN) can be built.
A. DYNAMIC PROGRAMMING NETWORK DEFINITION
The dynamic programming network (DPN) is denoted as a graph DPN (O, Y ), as shown in Fig. 6, with O and Y representing the vertex and edge sets, respectively. We assume that dark cores should be allocated to applications in set Here |B(t)| is the maximum number of dark cores in the system when applications in F(t) are running in the system, and it is computed by: Two dummy vertices, source vertex s and destination vertex d, are added to indicate the start and the end of the DPN, respectively. The vertex to represent the optimal overall throughput after assigning a total of b dark cores to applications f i (t), f i+1 (t), . . ., f |F(t)| (t). An edge connecting the vertices o i,b and o i+1,k is defined as , corresponding to the decision of assigning b − k dark cores to application f i (t). Each vertex at stage i is connected to at most |B(t)| + 1 vertices in the next stage i + 1.
An edge with utility i,b−k (the throughput obtained from the throughput model defined in Section III-B) exists between Let d s be a feasible path from the source vertex (s) to the destination vertex (d). The maximum throughput resulting from the dark core allocations for the application set F(t) can be computed by finding the longest path from vertex s to vertex d. Such a longest path can be found recursively in the form of Bellman equations [28]. That is, the dynamic By expanding Eqn. (11) from vertex s to vertex d (i.e., U (s, d)), the maximum throughput resulting from the dark core allocations for the application set F(t) can be computed by: Algorithm 1 shows the computation of the Bellman equations given in Eqns.
B. FINDING THE RUNNING APPLICATION SET
To find the application set T * (t + 1) that is to be run at time t + 1 (equivalent to determining γ if_run i in Eqn. (7)), a threestep dark core budgeting algorithm is applied, as shown in Fig. 7. Step 1: T (t) is denoted as the application set of all the applications that have not completed their executions after time t. From T (t), the maximum number of applications that can be added into the running queue from the waiting queue H (t), denoted as n wait , is computed by the order that these applications join the waiting queue (assume none of the applications including the applications in T (t) running at time t + 1 have dark cores).
Step 2: set T l (t + 1) = T (t) ∪ l j=1 h j (t), l ∈ {0, . . . , n wait }, h j (t) ∈ H (t). Here T l (t + 1) is the set of currently running applications after time t and l applications in the waiting queue. These l applications are selected from the waiting queue H (t), according to the order when they join the waiting queue. For each application set T l (t + 1), l ∈ {0, . . . , n wait }, the maximum throughput U l (s, d) in Eqn. (12) is computed by exploring the dynamic programming network (Algorithm 1).
The worse-case time complexity of finding the running application set and dark core budgeting scheme is O(n wait · (|T (t)| + n wait ) · |B(t)| 2 ).
VI. REGION DETERMINATION
Dark core budgeting algorithm computes the number of dark cores that is allocated to applications running at time t + 1. The currently running applications whose dark core number is about to change, i.e., |B i (t)| = |B i (t + 1)|, and the newly arrived ones need to find a new region, following a three-step region determination algorithm, as shown in Fig. 8. Step 1 (Find the Largest Contiguous Region): Starting with the largest contiguous region R largest will help alleviate the core fragmentation problem. All of the applications whose core regions are not located in R largest need to be adjusted for their core regions.
Step 2 (Determine the Relocation Order): The applications that need to find a new region are prioritized.
Step 3 (Find the Core Region): For each application that needs to be adjusted, a three-step algorithm is performed.
• First, find all the possible locations of the cores that an application can be mapped to.
• Second, the candidate regions are formed, starting from each possible core location of an application.
• Third, choose the new region out of the candidate regions. VOLUME 8, 2020
A. FINDING THE LARGEST CONTIGUOUS REGION
Let ψ ⊂ T * (t + 1) be a set that includes two types of applications: (1) the ones that are newly added into the running queue, and (2) the currently running ones which see their number of affiliated dark cores is about to change. Let RC i ∈ RC = {RC 1 , RC 2 , . . . , RC m , . . . } be the i th contiguous region occupied by the applications in set K no_adjust = T * (t + 1) − ψ, which is the set of the currently running applications that will hold the same number of dark cores.
To find RC, the following steps are performed iteratively. For each RC i ∈ RC, initially, a core that is occupied by an application in set K no_adjust is found, and it is added to RC i . For each core c j ∈ RC i , each of the neighboring cores c l is checked. If core c l is already running a task of an application in K no_adjust , it is added to RC i . If all of the cores in RC i are checked and ∀i,1≤i≤|RC| |RC i | is less than the total number of cores that are running the tasks of applications in set K no_adjust , the iteration can continue to find RC i+1 ; otherwise the iteration terminates.
The largest contiguous region R largest is the one in RC that has the maximum number of cores. All of the applications running at time t + 1, whose core regions are not located in R largest , are added to a set ψ . The core regions of applications in ψ need to be adjusted.
B. DETERMINATION OF THE RELOCATION ORDER OF APPLICATIONS
To determine the relocation order of the applications, applications in ψ are sorted in ascending order by the Manhattan distance between the geometric center of core region R i (t) and the geometric center of region R largest . Applications showing shorter Manhattan distances between the two will have higher priority to be relocated earlier. If two applications have identical Manhattan distances, the application with more tasks will be relocated earlier, since it is more difficult to find an appropriate core region for this application than those applications with fewer tasks. The Manhattan distance between the geometric center of a newly arrived application and the geometric center of core region R largest is first set to infinity, and this application is mapped after the currently running application. Note that a core c i has a 2D coordinate of < x i , y i >. The geometric center c(x l , y l ) for core region R i (t) can be approximatively determined by:
C. FINDING THE APPLICATION'S CORE REGION
For each application A i in ψ , a three-step algorithm is perform to find a core region.
Step 1 (Find All the Possible Core Locations That an Application Can Be Mapped To): Two classes of cores are first defined: periphery cores and internal cores. A periphery core is the one that is physically located on the edge of the network, while an internal core is the one that is at least one core away from the edge of the network. A core c k that falls into one of the two cases is a possible core location for application A i , and is added into a set i .
• Case 1: c k is a periphery core and only one neighboring core is occupied.
• Case 2: c k is an internal core, and c k shares two occupied neighboring cores with another core, c j , where c j is one of the cores located at {c(x k + 1, y k + 1), c(x k − 1, Step 2 (Form the Candidate Regions): For each core c k in i , cores are selected to form the candidate region R ik . R ik ∈ R i is defined as the k th candidate region for application A i , starting from the possible core location c k . To form the candidate region R ik , |R i (t)| − 1 cores with the minimal D j are added into R ik , where D j is the summation of the Manhattan distances between free core c j and all the cores that are already in R ik . That is, D j = ∀c l ∈R ik D(c j , c l ), where D(c j , c l ) is the Manhattan distance between two cores, c j and c l .
Step 3 (Choose the New Region Out of the Candidate Regions): From candidate region set R i , the region with the minimal migration cost is selected as the new core region for application A i . The migration cost of a candidate region is approximated as the Manhattan distance between the geometric center of application A i 's current core region and the geometric center of its candidate region. For the newly arrived application, select a core region randomly from the candidate regions as the new region.
Since the time complexity of determining the new region for an application is O(| i | · |R i (t + 1)| · |B(t)|) and there are |ψ | applications that need to be mapped or mapped remapped, the time complexity of region determination at each control time is O(|ψ | · | i | · |R i (t + 1)| · |B(t)|).
VII. TASK MAPPING
The task mapping algorithm maps the tasks of an application in ψ to its core region while minimizing the communication latency and improving the computation performance. Specifically, there are two major steps in this algorithm, as shown in Fig. 9: for each application in ψ , (1) extend the task graph, and (2) perform the task-to-core mapping.
A. EXTENDING TASK GRAPH
The number of dark cores |B i (t)| allocated to application A i was obtained from running the dark core budgeting algorithm presented in Section V. Since application A i can be descried by its task graph, |B i (t)| dummy tasks (nodes), all with node weight of zero, are created, and each of these |B i (t)| dummy nodes is connected to a task that has the maximal execution time in the task graph of application A i . If the execution times of two tasks happen to be identical, the one with fewer neighboring tasks will be selected first to connect with a dummy task. The k th dummy task, associated with task v ij , is denoted ash v ij ,k ∈ H v ij . Binding dummy taskh v ij ,k to task v ij does not change the characteristics of the task graph, ash v ij ,k has only one neighbor, task v ij , and the node weight ofh v ij ,k and the communication volume of edge (v ij ,h v ij ,k ) are both set to be zero. To add three dummy tasks into it, it is found that tasks v 12 , v 14 and v 17 have the longest execution times, thus each of these three tasks is connected with a dummy task. The extended task graph is shown in Fig. 10(b).
The dummy taskh v ij ,k cannot be mapped until its neighboring task v ij has been mapped, andh v ij ,k is mapped to a core that is adjacent to the one running task v ij . The positions of all the dummy tasks in H v ij , associated with task v ij , are determined by the function L(H v ij , R i (t)) with the following steps (Algorithm 2).
For each dummy taskh v ij ,k in H v ij , run the following steps to find set C v ij ,k including the possible cores that can be allocated toh v ij ,k . A core is randomly selected from C v ij ,k to runh v ij ,k .
Step 1: build set C 1 which contains the free core(s) selected from core region R i (t) such that the Manhattan distance between each c l in C 1 and the core M (v ij ) (occupied by task v ij ) is the shortest.
The cores in C 1 are next added into C v ij ,k (Lines 3-4). If there is only one core in C 1 , jump to Step 4. Initialize the set C v ij ,k = ∅; // Step 1 3: C v ij ,k = C 1 ; // Step 2 5: if |C 1 | > 1 then 6: C v ij ,k = ∅; 7: if |C 2 | > 1 then 10: C v ij ,k = ∅; 11: C v ij ,k = C 3 ; 13: end if 14: end if // Step 4 15: Randomly select a core c * from C v ij ,k ; 16: M (h v ij ,k ) = c * ; 17: end for Step 2: if there are more than one core in C 1 , clear set C v ij ,k and find set C 2 from C 1 such that, where c l in C 2 is the farthest-away core from all the currently mapped dummy tasks of application A i . The cores in C 2 are next added into C v ij ,k (Lines 5-8). This step helps to distribute the dark cores across the chip. If there is only one core in C 2 , jump to Step 4.
Step 3: if there are more than one core in C 2 , clear set C v ij ,k and find set C 3 from C 2 such that, where each core in C 3 has the minimal number of available free neighboring cores (denoted as ℵ c k ), and all the cores in C 3 are added into C v ij ,k (Lines 9-13). This step can reduce the impact of dark cores on communication latency, since a dummy taskh v ij ,k does not incur any communication with other tasks.
Step 4: from set C v ij ,k , a core is selected randomly, and map dummy taskh v ij ,k to the selected core (Lines 15-16). The core occupied by a dummy task is turned-off as dark core.
The time complexity of Algorithm 3 Task-to-Core Mapping Input: The application core region of A i . I : Sorted task set.
Output:
The mapping result. if v ij has already mapped neighboring tasks then // Case 1 6: Z v ij = Z 1 ; 8: if |Z 1 | > 1 then 9: Z v ij = ∅; 10: Z v ij = Z 1 ; 15: if |Z 1 | > 1 then 16: Z v ij = ∅; 17: Randomly select a core c * from Z v ij ; 22: M (v ij ) = c * ; 23: L(H v ij , R i (t)); // Call Algorithm 2 24: end for B. TASK-TO-CORE MAPPING Algorithm 3 shows the two-step task-to-core mapping to map the tasks of an application to its core region: Step 1: v m , the task with the highest total communication volume, is mapped to the geometric center (c center ) of application A i 's core region R i (t). If task v m is connected with dummy tasks, their respective positions are determined by function L(·, ·) (Algorithm 2) (Lines 1-2).
Step 2: let I be the set of tasks in V i (τ ) sorted by their communication volumes in descending order, and those connected with the dummy tasks are mapped first. For each task v ij in I , find set Z v ij which includes the possible positions that can map v ij . A core c * is randomly selected from Z v ij to run task v ij . There are two cases to consider to find set Z v ij (Lines 3-24).
• Case 1: (Lines 5-11) if at least one neighbor of task v ij has been mapped, build set Z 1 such that, where a core in Z 1 is closest to all the tasks in V ij , and V ij is the set of the already mapped neighboring tasks of v ij . The cores in Z 1 are next added into Z v ij (Lines 5-7). If there are more than one core in Z 1 , clear set Z v ij and find set Z 2 from Z 1 such that, where the number of available neighboring cores of each core c l in Z 2 is closest to the number of unmapped neighboring tasks of task v ij (i.e., β v ij ). The cores in Z 2 are added into Z v ij (Lines 8-11).
• Case 2: (Lines 12-20) if none of v ij 's neighboring tasks are mapped yet, build set Z 1 such that, where the number of available neighboring cores of each core c l in Z 1 is closest to β v ij (the number of unmapped neighboring tasks of task v ij ). The cores in Z 1 are added into Z v ij (Lines 13-14). If Z 1 has more than one core, clear set Z v ij and find set Z 2 from Z 1 such that, where each core in Z 2 is a farthest-away core from the positions of all of the tasks in V i , and V i is the set of the already mapped tasks of application A i . The cores in Z 2 are added into Z v ij (Lines [15][16][17][18][19]. After building set Z v ij , a core is randomly selected from Z v ij , and task v ij is mapped to this selected core (Lines 21-22). After mapping task v ij , check and find positions for its dummy tasks, using function L(·, ·) described in Algorithm 2 (Line 23).
The time complexity of Algorithm 3 is O( Once the positions of the tasks for all of the applications in set ψ are determined, the method in [4] is used to set voltage/frequency levels of the cores so that they can run at a high speed without violating temperature constraint and the maximum power capacity. Fig. 11 shows a mapping example for the task graph with three dark cores in Fig. 10. First, map the task with the highest communication volume. The task v 14 is first mapped to core c 5 in Fig. 11(a), the geometric center of core region. A core is selected randomly from {c 2 , c 4 , c 6 } for dummy taskh v 14 ,1 , since c 2 , c 4 and c 6 have fewer neighboring cores than that of c 8 , as shown in Fig. 11(b).
Second, map the tasks connected with dummy tasks. Task v 17 is mapped to c 7 , since task v 17 has no mapped neighboring tasks and has four unmapped neighboring tasks, which is closest to ℵ c 7 = 3, the number of available neighboring cores of c 7 . The dummy taskh v 17 ,1 is mapped to core c 10 , as the Manhattan distance of c 10 to core M (h v 14 ,1 ), occupied by dummy taskh v 14 ,1 , is larger than that of c 4 and c 8 to M (h v 14 ,1 ), as shown in Fig. 11(c). In a similar fashion, task v 12 and its dummy taskh v 12 ,1 are subsequently mapped, as shown in Fig. 11(d).
Third, map the tasks whose dummy task set H v ij is empty. Task v 15 is mapped to core c 8 , as task v 15 is a neighbor of tasks v 14 and v 17 . Similarly, tasks v 13 , v 16 , v 11 , and v 18 are mapped onto the core region, as shown in Fig. 11(e).
VIII. PERFORMANCE EVALUATION A. EXPERIMENTAL SETUP
To model the task graph and application execution, we implement an event-driven C++ network simulator with its configuration summarized in Table 2. This simulator is able to model packet delay and energy consumption in communications in a cycle accurate manner. Hotspot [29] is used for temperature simulation and McPat [30] is used as the power model. The power needed to turn on a dark core is set to the same as that in [31]. The floorplan of the underline many-core system is adopted from [32].
We evaluate the proposed method on random and real workloads, as tabulated in Table 2. The task degree of the random applications ranges from 1 to 14, and the number of tasks per application varies from 4 to 15. The task graphs of the real applications are generated from the traces of PARSEC [26] and SPLASH-2 [33] benchmark suites. These traces are collected by executing them in an NoC-based cycle accurate many-core simulator [34], whose configuration is also reported in Table 2. The applications in PARSEC and SPLASH-2 benchmarks are running with thread number of 16 and 64 in two network sizes, 4×4 and 8×8, respectively.
We compare our proposed method with the following methods: (1) Fixed_dark_core_allocation, which cannot adjust the number of dark cores after the initial task-to-core mapping and uses only the mapping method described in Section VII-B; (2) Bubble_budgeting [4], which uses virtual mapping to determine the number and the positions of dark cores; and (3) Adboost [6] where a core region including dark cores is found for an application. These two schemes [4], [6] are the state-of-the-art thermal-aware runtime task mapping approaches, which also consider the dark core allocation. Herein after our proposed scheme (including the adjustment) is termed as the Proposed.
In the following experiments, we compare the the four methods in terms of throughput, communication latency, and waiting time under different network sizes and application arrival rates. The waiting time occurs when there are insufficient cores to run the newly arrived applications.
where i,|B i (t)| and i,|B i (t)| are the throughputs obtained from the simulator and the throughput model, respectively. VOLUME 8, 2020 From Fig. 12(a), one can see that the seventh order polynomial regression has the lowest error (7.61%) among all. Therefore, in the following experiments, the seventh order polynomial regression model is used as the throughput model.
C. THE COMPARISON OF OFFLINE AND ONLINE THROUGHPUTS
It is possible that the core region used for training of the throughput model (see Section III-B) is different from the one selected at runtime. In addition, the thermal profile of the runtime system might also be different from that for the throughput model training. Thus, the estimated throughput (denoted as A ) used in the dark core budgeting algorithm for application set A may be different from the online throughput (denoted as A ) obtained from application execution at runtime, and the difference is defined as: Among the many different application sets executed, the difference is within 6%, as shown in Fig. 12(b). Moreover, the experimental results show that the average aspect ratio of application core regions determined at online is 1.22, which is close to the average aspect ratio (close to 1) of the core regions used for training the throughput model. These results indicate that the throughput model can give fairly accurate prediction of the throughput, which is so much needed in the dark core budgeting algorithm.
D. FINDING THE INTERVAL LENGTH OF CONTROL TIME
Our approach is triggered at each control time to process the workload variation and applications' computation demands. Fig. 13 shows how the throughput varies with various lengths of interval between two control times (in million cycles). Applications with different execution times and communication volumes are executed under different system settings in terms of network size and application arrival rate. From Fig. 13, one can see that the interval length of control time of 75M cycles generates the best performance. Therefore, in the following experiments, we set the interval length of control time to be 75M cycles.
E. PERFORMANCE EVALUATION ON RANDOM BENCHMARKS
Fig. 14 compares the throughput, waiting time, and communication latency of the four methods when they are performed in the system with different network sizes, running the random benchmarks where applications arrive at the system randomly. These results are normalized to that of the proposed method. It can be seen from Fig. 14(a) that the proposed method improves the throughput by 23.9%, 26.3%, and 29.2% compared with Fixed_dark_core_allocation as the network sizes vary from 5 × 5, 8 × 8, to 12 × 12, respectively. The proposed algorithm can adjust the dark cores of each application at runtime to optimize both currently running applications and newly arrived ones. Therefore, the proposed approach considers all of the applications to make a sound global decision that redistributes the dark cores among the running applications and newly arrived ones. The Fixed_dark_core_allocation only takes the next arrived application into account and cannot change the dark core allocation in response to the changing computation demands, which leads to sub-optimal performance. It can also be seen from Fig. 14(a) that, on average, the throughput of the proposed method is 1.45× and 1.82× over Bubble_budgeting and Adboost, respectively. The reason is that Bubble_budgeting only optimizes an individual application, without considering all the currently running applications. Therefore, it might allocate excessive number of dark cores to certain applications. Adboost, on the other hand, assumes the system has fixed number of dark cores, and it cannot allocate cores to applications according to their computation demands.
As shown in Fig. 14(b), on average, the proposed approach reduces the waiting time by 33.0%, 44.7%, and 71.0% over Fixed_dark_core_allocation, Bubble_budgeting, and Adboost, respectively. The reason is that, the proposed approach makes a global decision to balance the execution time of currently running applications and the waiting time of newly arrived ones on the fly. The communication latency of the proposed approach is also lower than those of the other three methods, as shown in Fig. 14(c). The reason is that the proposed approach adjusts the mapping scheme according to the changing computation demands. Moreover, with the proposed method, the dark cores are placed in the way that they have little impact on the communication latency. With large network sizes, the proposed approach achieves better performance in terms of waiting time and communication latency. The reason is that there are more dark cores for applications that can be adjusted at runtime to meet the workload variations. . 15 compares the throughput, waiting time, and communication latency of the four methods when they are adopted in a system running the random benchmarks with different application arrival rates. The results in Fig. 15(a) (b) and (c) are normalized to that of the proposed method. The respective throughput in Fig. 15(d) is normalized to that of Adboost when application arrival rate is 1. The application arrival rate is defined as the number of applications arrived at the system per 10 5 cycles, which measures the workloads of the system. It can be seen from Fig. 15(a) that, when the arrival rate is high, e.g., 2.78 applications arrive at the system per 10 5 cycles, the throughput of the proposed approach is 1.20×, 1.42×, and 1.96× over Fixed_dark_core_allocation, Bubble_ budgeting, and Adboost, respectively. On average, the proposed approach reduces waiting time by 83%, 96%, and 99% over Fixed_dark_core_allocation, Bubble_budgeting, and Adboost, respectively. The proposed approach achieves better performance since it can adjust the dark cores to reduce the waiting time of newly arrived applications when application arrival rates are high.
It can also be seen from Fig. 15(d) that, Adboost and Fixed_dark_core_allocation reach their throughput saturation points at the arrival rates of 1.25 and 1.85 applications per 10 5 cycles, respectively, while those of the proposed approach and Bubble_budgeting are both arriving at 2.50 applications per 10 5 cycles. The reason for this is that the proposed approach and Bubble_budgeting both take application arrival rate into their consideration. Moreover, the throughput of the proposed approach increases rapidly compared with that of Bubble_budgeting, as it considers all of the applications to make a global optimization. Fig. 16 compares the throughput, waiting time, and communication latency of the four approaches when they are adopted in a system with different network sizes, running the real benchmarks where applications arrive at the system randomly. These results are normalized to that of the proposed method. The throughputs of the proposed method are 1.15×, 1.40×, and 1.73× over Fixed_dark_core_allocation, Bub-ble_budgeting, and Adboost on average, respectively. The proposed approach also shows substantially reduced waiting time and communication latency, as shown in Fig. 16. The reason is the proposed method can make decision of dark core allocation and adjustment at runtime, which helps to optimize the performance of currently running applications and the newly arrived ones. . 17 shows the throughput, waiting time, and communication cost of the four methods when running the real benchmarks with different arrival rates. These results are normalized to that of the proposed method. When the arrival VOLUME 8, 2020 rate is high, the throughput achieved by our approach is about 1.16×, 1.35×, and 1.48× over Fixed_dark_core_allocation, Bubble_budgeting, and Adboost, respectively. On average, the proposed approach reduces waiting time by 43%, 79%, and 86% over the Fixed_dark_core_allocation, Bubble_budgeting, and Adboost, respectively. The reason for this case is similar to that seen in the case of the random benchmarks. In a simple term, adjusting dark core can achieve higher system performance. Fig. 18 evaluates the average peak temperatures of the four methods by running applications in a system with different configurations for one hundred times. One can see that the peak temperatures of all the four algorithms are below the temperature threshold 80 • C, but the proposed method achieves the lowest temperature, i.e., the proposed approach reduces the average peak temperature by 1 • C, 2 • C, and 3 • C over Fixed_dark_core_allocation, Bubble_budgeting, and Adboost, respectively. The reason is that the proposed mapping algorithm spreads the dark cores across the chip and redistributes them when needed at runtime. Doing so has a positive impact on heat dissipation to bring down chip temperature.
H. COST ANALYSIS OF THE PROPOSED ALGORITHM
The time penalties of running the three-step proposed approach, Bubble_budgeting, and Adboost are all in the order of 0.25M cycles. This is averaged out by running the algorithms one hundred times with different system parameters, such as network size, arrival rate, and communication volume of applications. In practice, most of applications run for as long as more than 10 8 cycles. Therefore, from the perspective of the application execution time, the time penalty of running the proposed algorithm is quite low. The power consumption of running the proposed algorithm is also considered in the experiments, which is 17.01W. The global average migration overhead at a control interval of 75M cycles is in the order of 0.2M cycles, which is also acceptably low.
IX. CONCLUSION
In this paper, built upon a dynamic programming framework, a runtime dark core allocation and dynamic adjustment scheme was proposed, taking into account the application arrival rate as well as the variation of the application's computation demands. An efficient task mapping algorithm was also proposed to reduce the negative impact of dark cores on communication latency and fragmentation. The experiments confirmed that, compared with two existing runtime thermal-aware resource management approaches, the proposed approach improves the system throughput by as much as 61% on average. The time penalty of running the proposed algorithm is very low, making it a suitable method for runtime resource management in many-core systems. | 12,009.2 | 2020-01-01T00:00:00.000 | [
"Engineering",
"Computer Science"
] |
Mapping the Polarization Pattern of Plasmon Modes Reveals Nanoparticle Symmetry
Single molecule labeling, cancer treatment, enhancement of non-linear optical effects, or light guiding have demanded much attention from the scientific community, and as a possible solution, plasmon resonances of noble metal nanoparticles are explored. A major advancement in single molecule optics has been the polarization analysis of light from single fluorescent emitters. This analytical method has been utilized to study the conformational dynamics of biomolecules and their spatial arrangement. At different wavelengths of the excitation light, different oscillation modes are excited making it important to know the polarization pattern as a function of wavelength.
“Knowing the polarization pattern of plasmonic nanostructures is therefore not only important to understand the fundamental physics of light interaction with these structures, but also allows to discriminate different oscillation modes within one particle and to distinguish differently shaped particles within one sample. Several techniques have been used to extract optical spectra of single plasmonic nanoparticles, most efficiently using dark-field microscopy, but little is known about the polarization state. So far, the very few reported plasmon polarization studies were obtained by rotating a polarizer by hand or on ensembles and not combined with spectroscopic information,” Prof. Carsten Sonnichsen explains to Nano Spotlight.
“We have developed a new microscope setup (RotPOL), which allows obtaining polarization-dependent scattering spectra in a fast and easy way,” Olaf Schubert continues explaining to Nano Spotlight. “RotPOL uses a wedge shaped quickly rotating polarizer which splits the light of a point source into a ring in the image plane, encoding the polarization information in a spatial image.” (Scheme (Scheme11)
Scheme 1
(a) Schematics of the RotPOL setup. One wavelength is selected by a linear variable interference filter (varIF), and then the light is dispersed into different polarization directions by a wedge-shaped rotating polarizer (PL), resulting in ring-shaped intensity profiles of a point-like light source on the digital camera (b). In order to get the polarization profile shown in (c) (intensityI(q) as a function of polarization angle q), we integrate the image between an inner and an outer ring diameter (dashed lines in (b)). The center of the rings is chosen to minimize asymmetry between opposite sides. Repeating this procedure for each wavelength produces intensity values as a function of wavelength and polarization angleI(l,q), which we show color-coded in (d). The same analysis is possible for all particles within the field of view in parallel. (e) Real-color image of an inhomogeneous silver sample containing spheres, rods, and triangles as seen through the RotPol-microscope. Two colors in one ring correspond to two different plasmon modes at the respective wavelengths. Scalebar is 25 μm
Prof. Sonnichsen’s team reveals that the polarization intensity in a given direction is simply taken from the corresponding position on the ring recorded with an exposure time larger than the rotation time. A dipole, for example, will show two loops at opposite sides, and with this in mind, the team can combine this rotating polarizer with a variable wavelength interference filter, which transmits light only in a narrow wavelength window. If mounted in front of the digital camera, the filter allows them to record simultaneously the spectral and polarization information for up to 50 particles in parallel.
“With the RotPOL setup, we study plasmon modes of a large variety of plasmonic structures—from rod-shaped particles to triangles, cubes, and pairs of spheres,” said Olaf Schubert. “Each plasmonic particle has a characteristic ‘footprint’, which allows deducing the approximate particle shape from the polarization-dependent single-particle scattering spectra. This is important for the optimization of particle synthesis, because it makes a quick and efficient estimation of the quality and mono-dispersal of a sample possible, without complex and expensive tools like electron microscopy.”
“For rod-shaped and even just slightly elongated particles, we found that the scattered light is highly polarized. Our simulations show that this high polarization anisotropy is not only due to the particle symmetry, but a plasmonic effect. This could be exploited for the design of miniature rotation sensors,” Prof. Sonnichsen explained enthusiastically.
In addition to yielding orientation information, plasmonic particles can be used to measure absolute distances on a nanometer scale. “Such a ‘plasmonic ruler’ makes use of the coupling of two spherical particles: If they are close to each other, the inter-particle plasmon resonance shifts from green to red. We measured the full polarization-dependent spectrum of such pairs of two spheres and found nice agreement with simulations.” This investigation has demonstrated that the polarization-dependent spectrum contains information about both the distances of the two spheres and the orientation and environment of the particles. In addition, such “multi-sensors” could possibly find a place in biological applications that require high time resolution.
“As an example of a time-resolved measurement, we have monitored the changes of plasmon modes in single gold nanoparticles during a growth process, in situ,” explains Olaf Schubert to Nano Spotlight. The researchers highlight that their RotPOL method is a versatile tool that can be used to study polarization anisotropy of the light emission pattern from nanoparticles, particularly for plasmonic structures, but can possibly be extended to fluorescent quantum structures. This method provides a wide range for optimization for applications such as light guiding and allows detailed theoretical modeling of plasmon modes due to the wide variety of plasmon emission patterns observed for the simple particle morphologies that have been investigated (spheres, rods, triangles, cubes, and particle pairs).
The researchers have recently published their results in Nano Lett, 2008. Their work reveals that the high polarization anisotropy found for even moderately elongated spheres “highlights” the strong influence of polarization even for nominally round particles. The possibility to record dynamic changes of the polarization emission pattern of single particles allows studying particle growth modes in situ and improving schemes for single nanoparticle binding and distancing assays.
Single molecule labeling, cancer treatment, enhancement of non-linear optical effects, or light guiding have demanded much attention from the scientific community, and as a possible solution, plasmon resonances of noble metal nanoparticles are explored. A major advancement in single molecule optics has been the polarization analysis of light from single fluorescent emitters. This analytical method has been utilized to study the conformational dynamics of biomolecules and their spatial arrangement. At different wavelengths of the excitation light, different oscillation modes are excited making it important to know the polarization pattern as a function of wavelength.
''Knowing the polarization pattern of plasmonic nanostructures is therefore not only important to understand the fundamental physics of light interaction with these structures, but also allows to discriminate different oscillation modes within one particle and to distinguish differently shaped particles within one sample. Several techniques have been used to extract optical spectra of single plasmonic nanoparticles, most efficiently using dark-field microscopy, but little is known about the polarization state. So far, the very few reported plasmon polarization studies were obtained by rotating a polarizer by hand or on ensembles and not combined with spectroscopic information,'' Prof. Carsten Sonnichsen explains to Nano Spotlight.
''We have developed a new microscope setup (RotPOL), which allows obtaining polarization-dependent scattering spectra in a fast and easy way,'' Olaf Schubert continues explaining to Nano Spotlight. ''RotPOL uses a wedge shaped quickly rotating polarizer which splits the light of a point source into a ring in the image plane, encoding the polarization information in a spatial image.'' (Scheme 1) Prof. Sonnichsen's team reveals that the polarization intensity in a given direction is simply taken from the corresponding position on the ring recorded with an exposure time larger than the rotation time. A dipole, for example, will show two loops at opposite sides, and with this in mind, the team can combine this rotating polarizer with a variable wavelength interference filter, which transmits light only in a narrow wavelength window. If mounted in front of the digital camera, the filter allows them to record simultaneously the spectral and polarization information for up to 50 particles in parallel.
''With the RotPOL setup, we study plasmon modes of a large variety of plasmonic structures-from rod-shaped particles to triangles, cubes, and pairs of spheres,'' said Olaf Schubert. ''Each plasmonic particle has a characteristic 'footprint', which allows deducing the approximate particle shape from the polarization-dependent single-particle scattering spectra. This is important for the optimization of particle synthesis, because it makes a quick and efficient estimation of the quality and mono-dispersal of a sample possible, without complex and expensive tools like electron microscopy.'' ''For rod-shaped and even just slightly elongated particles, we found that the scattered light is highly polarized. Our simulations show that this high polarization anisotropy is not only due to the particle symmetry, but a plasmonic effect. This could be exploited for the design of miniature rotation sensors,'' Prof. Sonnichsen explained enthusiastically.
In addition to yielding orientation information, plasmonic particles can be used to measure absolute distances on a nanometer scale. ''Such a 'plasmonic ruler' makes use of the coupling of two spherical particles: If they are close to each other, the inter-particle plasmon resonance shifts from green to red. We measured the full polarizationdependent spectrum of such pairs of two spheres and found nice agreement with simulations.'' This investigation has demonstrated that the polarization-dependent spectrum contains information about both the distances of the two spheres and the orientation and environment of the particles. In addition, such ''multi-sensors'' could possibly find a place in biological applications that require high time resolution.
''As an example of a time-resolved measurement, we have monitored the changes of plasmon modes in single gold nanoparticles during a growth process, in situ,'' explains Olaf Schubert to Nano Spotlight. The researchers highlight that their RotPOL method is a versatile tool that can be used to study polarization anisotropy of the light emission pattern from nanoparticles, particularly for plasmonic structures, but can possibly be extended to fluorescent quantum structures. This method provides a wide range for optimization for applications such as light guiding and allows detailed theoretical modeling of plasmon modes due to the wide variety of plasmon emission patterns observed for the simple particle morphologies that have been investigated (spheres, rods, triangles, cubes, and particle pairs).
The researchers have recently published their results in Nano Lett, 2008. Their work reveals that the high polarization anisotropy found for even moderately elongated spheres ''highlights'' the strong influence of polarization even for nominally round particles. The possibility to record dynamic changes of the polarization emission pattern of single particles allows studying particle growth modes in situ and improving schemes for single nanoparticle binding and distancing assays.
Kimberly Annosha Sablon
Scheme 1 (a) Schematics of the RotPOL setup. One wavelength is selected by a linear variable interference filter (varIF), and then the light is dispersed into different polarization directions by a wedgeshaped rotating polarizer (PL), resulting in ring-shaped intensity profiles of a point-like light source on the digital camera (b). In order to get the polarization profile shown in (c) (intensity I(q) as a function of polarization angle q), we integrate the image between an inner and an outer ring diameter (dashed lines in (b)). The center of the rings is chosen to minimize asymmetry between opposite sides. Repeating this procedure for each wavelength produces intensity values as a function of wavelength and polarization angle I(l,q), which we show color-coded in (d). The same analysis is possible for all particles within the field of view in parallel. (e) Real-color image of an inhomogeneous silver sample containing spheres, rods, and triangles as seen through the RotPol-microscope. Two colors in one ring correspond to two different plasmon modes at the respective wavelengths. Scalebar is 25 lm | 2,686.6 | 2008-09-01T00:00:00.000 | [
"Materials Science",
"Physics"
] |
Genes Encoding Callose Synthase and Phytochrome A Are Adjacent to a MAP3Kα-Like Gene in Beta vulgaris US H20
MAP3Kα, a gene that encodes a key conserved protein kinase, is responsible for initiating a rapid cascade of cellular events leading to localized cell death. Hypersensitive response, as it is termed, enables genetically resistant plants to limit microbial invasion under the right environmental conditions. Since knowledge of close physically linked genes is important for genome analysis and possibly for improving disease resistance, systematic DNA sequence analysis, gene annotation, and protein BLASTs were performed to identify and characterize genes in close physical proximity to a MAP3Kα-like gene in Beta vulgaris L. US H20. On the same 125 Kb BAC, callose synthase (BvCS) and phytochrome A (PhyA) genes were within 50 Kb of MAP3Kα. The close physical linkage of these genes may result from selection for coordinated responses to disease pressure. Bert, a new chromodomain-carrying gypsy-like LTR retrotransposon, resides within an intron of the BvCS gene, where it is transcribed from the opposing strand.
Introduction
A plant gene, MAP3Kα, produces a highly conserved protein product that activates hypersensitive response, a mechanism underlying R gene-mediated disease resistance [1]. In tobacco and in tomato, MAP3Kα activates cascades of enzymatic activations leading to a crescendo that is apoptosis or programmed cell death, a critical component of R genemediated disease resistance [1].
Research done on the crop plants, tomato and tobacco, as well as that performed on the model plant system Arabidopsis thaliana L. Heynth, over a 20-year period in a several laboratories, has presented adequate evidence that a particular gene called MAP3Kα is centrally important to R gene-mediated plant disease resistance [1]. In essence, a pathogen elicitor causes a conformational change in a plant protein initiating a cascade reaction leading to the so-called hypersensitive response, a primary countermeasure deployed by plants in order to effectively resist pathogen invasion. This key process is controlled by the protein product of MAP3Kα.
In the genome of Arabidopsis thaliana, large-scale duplication of genetic regions followed by selective gene loss has created a recognized network of chromosomal synteny [2]. By developing physical genetic maps based on ESTs, Dominguez et al. [3] discovered conserved synteny with Arabidopsis among genomes of four phylogenetically divergent eudicot crops, namely, sugarbeet, potato, sunflower, and plum.
In our previous study, complete BAC sequence analysis identified two core plant genes, CaMP and CKI, tightly physically linked to the disease resistance controlling gene NPR1 and established a conservation of microsynteny between the NPR1 gene regions of sugarbeet and other eudicots [4]. Also, an HSF gene adjacent to, just 2 Kb downstream from, NPR1 in sugarbeet and whose close microsynteny is conserved in four out of five eudicots examined, encodes a DNA-binding HSF protein similar to that specified by gene HSFA9 that controls early leaf morphogenesis in sunflower [4,5].
The central role of MAP3Kα in positively and globally activating hypersensitivity to pathogen invasion, as an effective defense mechanism in response to an elicitor(s) produced by the pathogen, suggests the possibility of enhancing disease resistance in plants by genetic manipulation of expression of the MAP3Kα gene. As a step toward identifying genes localized near the MAP3Kα gene in sugarbeet, 2 International Journal of Plant Genomics a bacterial artificial chromosome (BAC) library was screened using PCR and gene-specific primers, and a clone, SB3, was identified as carrying a MAP3Kα-like gene. An expressed sequence tag, EST clone BQ585699, was instrumental in designing primers used for discovery of the MAP3Kα-like gene from sugarbeet US H20. The Beta vulgaris MAP3Kαlike gene encodes a predicted protein product with high similarity to protein products encoded by disease resistanceorchestrating mitogen-activated protein kinase kinase kinase genes in tomato, tobacco and the model plant species Arabidopsis thaliana (in preparation).
We report herein new information regarding the gene content and organization of a 125 Kb contiguous fragment of sugar beet genomic DNA contained in a sugarbeet BAC carrying MAP3Kα. Our discovery of close physical linkage of MAP3Kα with genes encoding callose synthase, CS, and phytochrome A, phyA, is described for the first time. Discovery of a novel chromodomain-carrying gypsy-like LTR retrotransposon, Bert, is also described.
Materials and Methods
2.1. DNA Sequencing. Genomic DNA of B. vulgaris hybrid US H20 [6] (PI 631354), with an estimated 750 Mb genome size, had previously been used to construct a BAC library [7]. About 34,500 clones comprised the BAC DNA library, average insert size was about 120 Kb, providing about 6.1X genome coverage [7]. Primers designed based on the DNA sequence of GenBank accession, BQ585699, an EST sequence encoding for a B. vulgaris MAP3K, were utilized to screen and identify a BvMAP3Kα-carrying BAC (manuscript in preparation). The presence of a complete genomic MAP3Kα gene was established by DNA sequence analysis of BAC clone SB3.
BAC sequencing was completed at Washington University's Genome Sequencing Center in St. Louis, Missouri, USA (http://genome.wustl.edu/). The BAC clone SB3 was provided to the Genome Sequencing Center as a glycerol stock. Purification, library construction, shotgun cloning, and sequence analysis were performed on a sufficient number of random subclones to provide about 9.5X coverage. ABI 3730 capillary sequencers were used. Data was assembled using the phred/ phrap suite (http://www.phrap.org/).
Comparative Similarity Analysis.
BlastP searches of predicted protein products of sugar beet genes were performed at http://www.ncbi.nlm.nih.gov/BLAST/, similarity analysis of proteins was performed using the Mega program (http://www.megasoftware.net/) using neighbor joining method and ClustalX alignment program.
Results
A 125 Kb contiguous fragment of sugarbeet chromosomal DNA contained in SB3, a sugarbeet BAC carrying a B. vulgaris MAP3Kα-like gene, was sequenced and fully annotated (GenBank accession GU057342 is scheduled for release on 10/05/10). Bioinformatics tools Fgenesh, GeneMark, and Augustus were used as gene finders. Designated gene names and predicted functions of deduced amino acid sequences, where possible, are presented in Table 1 and a visual representation of exon structure is depicted in Figure 1. Within the 125 Kb contiguous fragment of sugarbeet genomic DNA, eighteen open reading frames (ORFs), or protein-encoding regions, were identified. Only three ORFs were predicted to produce protein products with high amino acid sequence similarity to known products of core plant genes (Table 1). In addition to the three core plant genes, the 125 Kb contiguous fragment of genomic DNA was found to carry an insertion of a novel chromodomain-carrying gypsy-like LTR retrotransposon, which we call "Bert" in keeping with widely accepted nomenclature of similar transposable elements in other plants. Bert's predicted polyprotein has a C-terminal chromodomain, and Bert, localized within an intron of a 36exon callose synthase gene, is transcribed from the opposing strand. The other fourteen putative genes were predicted to produce proteins that either lack a known function or are ancient or defective retrotransposons.
In addition to MAP3Kα, another core plant gene within the 125 Kb contiguous fragment of sugarbeet genomic DNA carried by BAC clone SB3 was a 36-exon callose synthase gene, BvCS, that encodes a β-1,3-glucan synthase protein having a conserved glucan synthase domain (E = 8.7e −148 ) from amino acid positions 1159-1779 and various transmembrane domains by SMART. The predicted BvCS protein has high amino acid sequence alignment similarity (Table 1) with the protein product of CS5, a male fertility-controlling gene of Arabidopsis thaliana [13] whose product is also involved in callose deposition in response to wounding [14]. The predicted protein product of BvCS also is very similar in amino acid sequence alignment to CS5-like gene products in castor bean, grape, and poplar ( Figure 2).
The BvCS gene is localized between the MAP3Kα gene and a phytochrome A gene, PhyA, encoding a photoreceptor The predicted protein product of the phyA gene gave numerous BLAST hits E = 0.0, indicative of a good match by amino acid sequence alignment, to phytochrome A proteins. Relative to the product of sugarbeet phyA, the phytochrome A protein with the most similar amino acid sequence alignment was from Stellaria longipes or longstalk starwort (Table 1, Figure 3). In a different subclade, were phytochrome A proteins from Solanum lycopersicum, tomato, and Solanum tuberosum, potato, based on an amino acid sequence alignment similarity tree (Figure 3), obtained using complete amino acid sequence alignments and MegAlign (not shown). A distinct clade of phytochrome A proteins contained proteins from Armoracia rusticana, horseradish, and Cardamine resedifolia, an alpine wildflower.
Domain architecture of the predicted protein products BvCS and BvPhyA is illustrated in Figure 5. Ending at about 50 Kb downstream of MAP3Kα, the sugarbeet phyA gene, is interrupted by three introns (Figure 1). Close physical proximity of genes MAP3Kα, BvCS, and PhyA was discovered in B. vulgaris. MAP3Kα encodes an alpha-like mitogenactivated protein kinase kinase kinase of the type that orchestrates the hypersensitive response responsible for genetic disease resistance. Our in silico analyses of the pre-dicted products of the two nearby core plant genes show unequivocably that (1) BvCS1 encodes a callose synthase of the type responsible for normal pollen tube function and for response to wounding and that (2) PhyA encodes a Phytochrome A-like light signal receiver which has both a histidine kinase domain and an ATPase domain not too dissimilar to that found in the agriculturally important Solanum genus containing potato and tomato.
Bert, 5.5 Kb in length, is a novel chromodomain-carrying, gypsy-like LTR retrotransposon in a single exon [as expected] (Table 1, Figure 1). Nucleic acid Blast at NCBI produced an E = 0.0 alignment with a soybean retroelement polyprotein AAO23078. A similarity tree (Figure 4) shows that Bert also aligns best with a retroelement polyprotein in Brassica rapa based on complete amino acid sequence alignments obtained using MegaAlign (not shown). Genome analyses of model plant species within Arabidopsis, Lotus, and Medicago has produced evidence for other predicted retroelement polyproteins similar to Bert in terms of a similarity analysis of complete amino acid sequence alignments (Figure 4). Compared with the above, Bert is less similar to a complete gypsy-like LTR retrotransposon from sugarbeet we previously described, Schmidt [15] in overall amino acid sequence alignment (not shown). Nucleic acid Blast alignment with other retroelements at the Plant Repeat Database (PRD) produced a match with an E value equal to 2.2e −77 with "rn 460 239" from Graminaceae, the grasses.
Discussion
In this study, analysis of genes that are very physically close to the MAP3Kα gene of B. vulgaris revealed, for the first time, that a callose synthase gene, whose product likely plays major structural, defense, and developmental roles, and a PhyA gene, encoding a phytochrome A protein kinase with tripartite roles in light perception, signal transduction, and nuclearly-localized activation of multigene transcription in response to light availability [16,17], are adjacent to the MAP3Kα gene, whose protein product orchestrates the hypersensitive response, the primary plant genetic resistance countermeasure. Glucan synthase (GS), or uridinediphosphate glucose: (1->3)-β-D-glucan 3-β-D-glucosyl transferase, interacts with phragmoplastin, UDP-glucose transferase, a Rho1-like protein and possibly annexins, depositing callose in different locations in response to specific abiotic, biotic, and developmental signals [13]. The BERT retrotransposon transcribed from the negative strand, within an intron of BvCS, is probably active since its long terminal repeats are 100% identical. Its transposition, probably stress-induced, would likely occur under conditions of severe stress, such as tissue culture, potentially resulting in random mutagenesis, or "somaclonal variation," as the earlier literature described the phenomenon.
Within a diverse family, CS genes, located on different chromosomes, encode large transmembrane proteins in ORFs interrupted by 1 to 49 introns [13]. Callose, a 1, 3-β-D-glucan with a few 6 linked branches composed of glucose monosaccharide linked by 1, 3 beta linkages, is formed in the cell wall and many other places depending on the stage of development. Callose is usually found in the immediate vicinity of the cell wall where it serves as a plugging mechanism whenever the cell wall suffers disruptive stress such as herbivory by insects or other wounding [18].
Callose is deposited between the plasma membrane and the cell wall after exposure to either abiotic or biotic stresses. As a programmed plant cellular response, callose deposition is usually an effective means of resisting microbial attack, insect feeding, or physical stress. By very rapidly synthesizing and depositing callose as plugs, drops, or plates in close proximity to an invading pathogen or a damaged area, the plant cell prevents more serious damage. Callose deposits, often referred to as papillae, may contain minor amounts of other polysaccharides, phenolic compounds, reactive oxygen inter-mediates, and proteins [14]. Papillae are believed to literally wall off microbial intruders [14]. Prerequisite to and then concurrent with callose deposition, a rapid influx of Ca 2+ into the cytoplasm occurs. Ca 2+ acts as a second messenger that transmits signals received from receptors on the cell surface, including elicitors from pathogens, to target molecules in the cytosol. Ca 2+ helps to initiate the well-documented oxidative burst and activates cascades and other defense responses culminating in the hypersensitive response (HR), programmed cell death, or apoptosis-the ultimate cellular defense [19].
During cytokinesis, callose deposition to the developing cell plate may be important for septum formation [18]. It has also been hypothesized that callose controls cell-to-cell movement of molecules through the plasmodesmata. Callose is known to plays key roles in pollen grain formation and pollen tube growth [13].
The CS protein product encoded by the CS gene adjacent to MAP3Kα in BAC3 is most similar in amino acid sequence alignment to the product of CS5, callose synthase 5 ( Table 1).
The amino acid sequence alignment of the Arabidopsis CS5 gene product may be an atypical outlier from the group as a whole (not shown) but, consistent with the prediction, there is an observed high degree of amino acid sequence similarity between the predicted protein product of BvCS and gene products of a number of CS5-like genes of various plant species (Figure 2). There is a close physical linkage of the MAP3Kα gene with a BvCS gene whose product is herein predicted as a callose synthase 5. Largely expressed only in male germ cells, the CS5 gene encodes a so-called "malespecific" beta 1, 3 glucan synthase and is needed to produce the temporary callose walls that separate the developing microspores. The callose that is deposited on the surface of microsporocytes serves as a temporary cell wall [13].
Phytochromes are photoreceptors that regulate plant photomorphogenesis, growth, or development that is stimulated by red, infra red, and blue light. Photoreceptors monitor intensity, direction, quality, and duration of light [16]. Phytochromes absorb at 600-800 nm and optimize the capture of light energy needed for photosynthesis and other core metabolic processes. Phytochromes function during all of the stages of the cell and organismal life cycles and their primary roles are to acquire information on the light environment of a plant and to provide the plant with the means to adapt to change, both expected and unexpected, in the supply of light energy [17].
Previously a hypothesis was proposed [4] that conserved microsynteny of certain core plant genes in eudicots may International Journal of Plant Genomics 7 correlate with either their subcellular localization or with related function as is often the case with clusters of genes in bacteria. In this present study, the cellular roles ascribed to the predicted protein products of the three core plant genes found clustered on sugarbeet genomic DNA carried by SB3 correlate well with a clear need for coordinated expression. MAP3Kα initiates the hypersensitive response, apoptosis, leading to effective genetic disease resistance. Pathogen elicitor-activated MAP3Kα functions to phosphorylate other protein kinases that, in turn, phosphorylate other protein kinases and so on [hence the "3K" or kinase kinase kinase terminology]. Signal transduction cascades occur concurrently very rapidly in response to the detection of a recognized specific pathogen. If the plant has the particular resistance gene that encodes a product that directly or indirectly responds to pathogen elicitor(s), the resulting hypersensitive response leads to effective disease resistance by a mechanism(s) that are still not yet completely clear.
CS expression, controlled in the plant nucleus, is essential for male gametophyte viability/fertility [13,20]. It should be noted that in Arabidopsis there are twelve CS genes expressed in different plant tissues and whose products have diverse roles [21]. Oryza sativa subspecies japonica has a type of CS gene, exemplified by 55771366 responsible for a protein product that has high amino acid similarity to products predicted for BvCS and other CS5-like genes in Populus tricarpa, Ricinus communis, and Vitis vinifera ( Figure 2). All are part of the plant's response to biotic as well as to abiotic stresses. Since some defense responses exhibit a welldocumented requirement for phytochrome activation [22][23][24][25], coordination of systemic defense responses with energy availability is irrefutable and, consequently, both the PhyA gene and the BvCS gene localized near MAP3Kα in sugarbeet likely play important roles in stress responsiveness.
Including Bert, a total of 15 retrotransposon-(RT)-like or hypothetical genes lie within the approximately 125 Kb BAC carrying sugarbeet genomic DNA specifying a small core gene cluster consisting of BvMAP3Kα, BvCS, and PhyA genes, all in an about 60 Kb long genomic DNA region from Beta vulgaris. Thus the immediate region around the small core gene cluster is rich in repetitive elements since several insertions of mobile genetic elements have occurred during both horizontal gene acquisition and vertical evolutionary descent. ORFs originating from either retrotransposons or viruses, from DNA transposons and other repetitive elements, need not be considered disruptive of colinearity of core genes nevertheless. This about 125 Kb contiguous genomic DNA fragment, rich in highly-degraded repetitive elements, contains only a single one-exon ORF, Bert, likely encoding an active gypsy-like CHR domain LTR retrotransposon polyprotein. Bert has some similarity to the previously described gypsy-like retrotransposons Schmidt [15] and Beetle1, a new chromodomain LTR retrotransposon of Beta procumbens [26], but these two previously described LTR retro-transposons are more similar in predicted amino acid sequence alignment with each other than they are with Bert, consistent with Bert's novelty. Nevertheless, the insertion of Bert into an intron of the BvCS gene does not alter the conclusion that there are three intact core essential genes in close physical proximity, transcribed in the same direction and with protein products predicted to play either direct or indirect roles in activation of defense response mechanisms. These findings are consistent with our hypothesis concerning the "raison d'être" for gene clustering.
By comparing the orthologous NPR1-carrying regions of Medicago truncatula and Populus trichocarpa with that of B. vulgaris, we discovered conserved microsynteny for NPR1, CaMP, and CK1PK genes [4]. Conserved microsynteny of NPR1, CaMP, and CK1PK in B. vulgaris, M. truncatula, and P. trichocarpa may help coordinate expression [4]. More recently, very close physical linkage in monocots of Bx1 and Bx2, and if Bx3 and Bx4 genes-encoding enzymes, responsible for steps in benzoxazinoid synthesis, suggests functional clusters related to coordinated expression for this biosynthetic pathway [27].
Close physical proximity of key core plant genes suggests there has been a positive selection for the arrangement. Close physical linkage of the three core plant genes, BvMAP3Kα, BvCS, and PhyA, even if observed only in sugarbeet, may hypothetically facilitate coordinated expression of genes critical to plant defense response with other cellular and organismal processes including adaptation to abiotic as well as biotic stress, efficient as well as timely response to change in light availability, reproduction, and even programmed cell death as necessary to protect the plant from the spread of an erstwhile destructive avirulent pathogen.
Reverse transcriptase (RT) PCR, quantitative qPCR expression studies, or global transcriptional profiling will furnish information relevant to the question of the interrelatedness of plant defense, cellular integrity, and reproductive fitness, as well as light responsiveness roles of MAP3Kα, callose synthase, or phytochrome A. Upregulated expression plant of genes in response to pathogens and/or oxidative stress requires intense investigation.
In summary, in addition to MAP3Kα and the closely physically linked BvCS and BvPhyA genes herein described for the first time, the 125 Kb MAP3Kα-carrying Beta vulgaris BAC SB3 also encodes Bert, a new chromodomain gypsy-like retrotransposon, and has 14 other as yet undefined features. Whereas some of these ORFs produce predicted proteins with probable retrotransposon origins deduced from BLAST analysis, the other unidentified ORFs had predicted protein products without any known function.
Close physical proximity of MAP3Kα, BvCS1, and PhyA suggests positive selection in the history of breeding and genetic hybridization of Beta vulgaris for coordinated expression of these particular genes whose presumably essential products either control genetic plant disease resistance, callose synthesis or responses to light. | 4,722.2 | 2011-06-07T00:00:00.000 | [
"Biology"
] |
A 3D-induced pluripotent stem cell-derived human neural culture model to study certain molecular and biochemical aspects of Alzheimer’s disease
Alzheimer’s disease (AD) early pathology needs better understanding and models. Here, we describe a human induced pluripotent stem cells (iPSCs)-derived 3D neural culture model to study certain aspects of AD biochemistry and pathology. iPSCs derived from controls and AD patients with Presenilin1 mutations were cultured in a 3D platform with a similar microenvironment to the brain, to differentiate into neurons and astrocytes and self-organise into 3D structures by 3 weeks of differentiation in vitro. Cells express astrocytic (GFAP), neuronal (β3-Tubulin, MAP2), glutamatergic (VGLUT1), GABAergic (GAD65/67), pre-synaptic (Synapsin1) markers and a low level of neural progenitor cell (Nestin) marker after 6 and 12 weeks of differentiation in 3D. The foetal 3R Tau isoforms and adult 4R Tau isoforms were detected at 6 weeks post differentiation, showing advanced neuronal maturity. In the 3D AD cells, total and insoluble Tau levels were higher than in 3D control cells. Our data indicates that this model may recapitulate the early biochemical and pathological disease features and can be a relevant platform for studying early cellular and biochemical changes and the identification of drug targets.
Introduction
Alzheimer's disease (AD) is an irreversible age-related progressive neurocognitive disorder and the most common cause of dementia. In the UK, one in 14 people over 65 years of age have dementia [1]. In 2019, a study showed that 885,000 people are estimated to be living with dementia in the UK and this number is expected to increase exponentially over the next few decades [2]. People suffering from AD show severe memory impairment which is attributed to two primary hallmarks of the disease: amyloid plaques and neurofibrillary tangles. Plaques are extracellular deposits of the protein β-amyloid (Aβ) and tangles are intracellular deposits of the microtubule-associated protein Tau [3][4][5][6]. The early onset familial disease is attributed to the patients carrying a mutation in amyloid precursor protein (APP), Presenilin 1 (PSEN1) and Presenilin 2 (PSEN2) genes, with most mutations present in PSEN1 [7][8][9][10]. PSEN1 and PSEN2 genes encode the respective PS-1 and PS-2 catalytic subunits of the enzyme complex γ-secretase [11][12][13]. While familial AD (FAD) cases account for less than 5% of the AD cases, sporadic AD (SAD) is the most common form of AD [14,15]. Despite SAD being the most prevalent form of AD, most of our understanding of the underlying pathology of the disease comes from studies on FAD models. While studies in post-mortem human brain tissue and transgenic models over-expressing FAD mutations have provided some insights about the underlying mechanisms of the disease, there is a move towards human-cell-based models in recent years. In the leap from bench to bedside, human-based-physiologically relevant in vitro models are key in understanding and overcoming the evolutionary and special differences of the underlying disease mechanisms which are poorly represented by animal models [16]. Induced pluripotent stem cell (iPSC) technology allows for the unlimited generation of stem cells with the ability to differentiate into neurons and astrocytes directly from easily available skin or blood cells of patients with specific mutations [17][18][19]. This method ensures the expression of underlying pathology without forcing overexpression of genes/mutations of interest.
However, despite the advantages of the iPSC technology, the development of disease models that simulate protein aggregation using conventional 2D culture techniques remains challenging. Although conventional 2D cultures have been quick and cheap tools for obtaining valuable disease-relevant insight, they are cultures grown on rigid and flat artificial substrates such as glass or polystyrene which poorly represent the in vivo physiological microenvironment. 2D culture systems do not mimick the complexity of the human brain, thereby misrepresenting morphology, differentiation, synaptic architecture especially cell-cell and cell-microenvironment interactions. The advent of 3D tissue engineering has revolutionised the approach towards the conventional in vitro culture systems. Studies have shown 3D cultures not only encourage physiologically relevant morphology/phenotype of different cell types, accelerate differentiation and formation of intrinsic neural networks but also recapitulate the key pathological features of the disease in vitro which were never reported before in 2D culture systems [20][21][22][23][24][25]. Hence, differentiating patient-derived iPSCs to generate disease-relevant cell phenotypes would help overcome most of the issues seen in in vitro models.
Most of the disease-modifying drugs have failed in clinical trials due to the administration in the late phase of the disease, hence, there is good reason to study and understand the early changes seen in the pathogenesis. Until recently, it was believed that the accumulation of Aβ was the point of initiation of the disease in the brain, which eventually led to the formation of plaques and hyperphosphorylated neurofibrillary tangle [4,26,27]. In recent years, the field has moved to appreciating a casual role of the Tau protein with many Tau-centric drugs targeting Tau mechanisms of toxicity [28]. Although mutation of the gene encoding for Tau is not directly implicated in AD, the wild-type protein becomes abnormally hyperphosphorylated and aggregated in the disease [29] in both FAD and SAD cases. However, the mechanism initiating the process is unknown. A number of studies using either human neurons overexpressing specific FAD genes or AD patients iPSC derived neurons have described increased levels of phosphorylated Tau and increased insolubility of the protein [25,30,31]. Some have reported increased Tau protein levels in iPSC models with APP mutations [32,33]. A key question is whether these Tau phenotypes are specific to cells with APP mutations or also seen in cells with PSEN1 mutations and whether iPSC models can be used to establish links between Aβ and Tauassociated changes in the disease.
Here, we report a human in vitro model to study such early changes seen during the development of the pathology. We report the differentiation of human iPSC-derived neural stem cells (hiNSCs) into neurons and astrocytes in 3D cultures over a period of 12 weeks in vitro. We also report that the patient-derived hiNSC with PSEN1 mutations show higher levels of total and insoluble Tau in vitro. Our findings support that our 3D cultures are a relevant model to study certain molecular and biochemical pathological aspects of AD, especially during the early phase of the disease.
Thawing and expansion of iPSC-derived neural stem cells
SureBond XF (1000x, ax0041XF, Axol Biosciences, Cambridge, UK) was slowly defrosted at 4 °C 2 days prior to thawing cells, to prevent jellification. SureBond XF was diluted (5 µl/ml) with sterile PBS without Ca-Mg (10010023, Gibco, UK) to coat T75 culture flasks one day Fig. 1 Generation of human iPSC-derived 3D neural cultures. a Details of the method followed to culture cells in 3D. RA, retinoic acid; SHH, Sonic Hedgehog; BDNF, Brain-Derived Neurotrophic growth Factor; GDNF, Glial cell line Derived Neurotrophic Factor; IGF1, Insulin-like Growth Factor 1; and cAMP, cyclic Adenosine Monophosphate. b Schematic diagram representing the in vitro 3D culture model. c Details of the iPSC-derived NSCs (control and AD patient-derived) from Axol Biosciences. d Image of 3D cultures in chamber slides at 5 weeks post differentiation showing the cells have self-organised to create a sheet which has detached from the sides of the culture walls and is free-floating. e Image of 3D cultures in chamber slides at 6 weeks post differentiation showing the free-floating sheet has curled up and folded itself into a ball-like structure ◂ prior to plating cells. Cells were defrosted, transferred to prewarmed fresh ANM media and centrifuged at 200 g for 5 min. Cells were then transferred to fresh pre-warmed neural plating-XF Media (ax0033, Axol Bioscience, UK). Cells were left to stabilise for 24 h in the incubators in this media. The next day, the media was completely replaced with fresh pre-warmed ANM media supplemented with EGF (0.1 mg/ml, AF-10015, PeproTech, UK) and FGF (0.05 mg/ ml, 100-18B, PeproTech, UK). Media was changed every second day for 6 days. Cells lines were not passaged more than once during the expansion phase. The same differentiation protocol was followed for both the 2D and 3D cell culture methods.
3D cell culture plating
For 3D seeding, cells were washed with PBS. Warm Axol Unlock XF Media (ax0044XF, Axol Biosciences, UK) was added to the cells for 3 min at 37 °C for the cells to detach. Cells were recovered in ANM media and spun at 200 g for 5 min. The pellet was resuspended in fresh warm ANM media. Cells were plated on ice, in an 8-well chamber slide (177445, Lab-Tek system, Nunc, UK) at a density of 2 × 10 6 / ml in 3D plating media composed of ANM media with Retinoic Acid (RA-R2625, Sigma-Aldrich, UK) at 500 ng/ml and supplemented with Matrigel (354230, Corning) at a ratio of 1:15 v/v, and then left to form a gel at 37 °C. An extra 200 µl of ANM media was added 2 h after the cultures were plated.
2D cell culture plating
For 2D plating, cells from the flasks were washed with PBS and detached using warm Unlock XF Media (ax0044XF, Axol Biosciences, UK) for 3 min at 37 °C. Detached cells were then diluted in 15 ml of ANM media and centrifuged at 200 g for 5 min. The pellet was resuspended in fresh warm ANM media. For plating cells in 2D, 5-mm-diameter/1.5mm-depth round glass coverslips were coated with 0.01% solution of Poly-L-Ornithine (P 495, Sigma-Aldrich, UK) overnight followed by coating with 50 µg/ml Laminin (L2020, Sigma-Aldrich, UK) for 1 h at 37 °C. 6000 cells were seeded evenly per well in a 6-well dish with coated coverslips and placed at 37 °C. Cells were allowed to attach for 2 h before supplementing with 500 ng/ml RA.
Differentiation of iPSC-derived NSCs in 2D and 3D cultures
Four days after seeding cells for 2D or 3D cultures, the cells were supplemented with Sonic Hedgehog (SHH, 100-45, PeproTech, UK) at 100 ng/ml. Two weeks after the initial cultures, the media was replaced with ANM media supplemented with a cocktail of growth factors: 0.01 mg/ml Brain-Derived Neurotrophic Factor (BDNF, 450-02, Pepro-Tech, UK), 0.01 mg/ml Glial cell line Derived Neurotrophic factor (GDNF, 450-10, PeproTech, UK), 1 mg/ml cyclic Adenosine Monophosphate (cAMP, A9501 Sigma-Aldrich, UK), 0.01 mg/ml Insulin-like Growth Factor (IGF,1-100-11B, PeproTech, UK) to promote neural cell differentiation for 12 weeks. Media was changed every other day for the first 3 weeks followed by every day up to 12 weeks in vitro.
Cryosectioning of 3D cultures
For embedding, the 3D cultures were fixed with 4% paraformaldehyde for 24 h at 4 °C. The cultures were then washed three times with PBS and cryoprotected in 30% sucrose solution (S9378, Sigma-Aldrich, UK) prior to embedding in optimal cutting temperature (O.C.T.) embedding medium (12678086, Thermo Scientific, UK) over dry ice. Frozen tissue was cryosectioned at 14 µm thickness using a S35 microtome blade (JDA-0100-0A, Cell Path, UK). Each section was collected and transferred onto Super-Frost glass micro slides (10149870, Fisher-Thermo Scientific, UK). Sections were stored at − 20 °C after being air dried at room temperature overnight.
Immunofluorescence staining
For immunostaining of 2D cultures, the cells on the coverslips were fixed with 4% paraformaldehyde for 15 min at room temperature. To compare the two different systems of cultures, 2D coverslips and 3D culture sections were stained following the same protocol for each marker. . Secondary antibodies were prepared in the same blocking solution and cells were incubated for 2 h at room temperature. Nuclear stain DAPI (D9542, Sigma-Aldrich, UK) was used at 2 µg/ml in permeabilization buffer for 15 min at room temperature.
Slides were mounted and sealed with an in house-prepared, setting mountant comprising PBS, glycerol, Mowiol 4-88 (Harco) and Citifluor AF3 anti-fade solution (Citifluor Electron Microscopy Sciences. To decrease variability, all control and AD cells were stained and processed in parallel.
Image analysis
Cells were imaged using a Leica TCS-SP8 laser scanning confocal microscope on a Leica DMi8 inverted microscope frame (TCS SP8, Leica Microsystems) at × 63 magnification using a glycerol immersion objective (Leica HC PL APO CS2, 1.3NA). Images were acquired at a 1024 × 1024 sample rate and 0.75 zoom (field of view-246 × 246 um, equivalent pixel size = 240 × 240 nm in XY). Maximum confocal detection sensitivity was set for each channel using control samples (secondary antibodies and DAPI only). Sequential imaging and optimised setting of detection bandwidths on the Leica TCS spectral detection system (Alexa Fluor 488 = 500-550 nm bandwidth, Alexa Fluor 568 = 575-650 nm bandwidth) were used to avoid crosstalk between channels. Fields of view to image were selected by conventional fluorescence microscopy using DAPI signal so as not to bias field selection based on the specific probe signal. For both 2D and 3D cultures, Z stacks were acquired at 1 µm Z intervals. For each 3D culture, cryosections from three independent cultures were imaged and three to five different fields were captured per section. For 2D cultures, three coverslips from three independent cultures were imaged and three to five fields were captured per coverslip. To compare between 3 and 2D cultures a representative single Z stack was used. Image analysis was performed using ImageJ software (Fiji). The threshold was determined manually for each image for each channel. For analysis, the entire field of view was analysed to measure the total area positive above the threshold for each channel to calculate the relative proportion of positive pixels. Each measured positive area was normalised against the amount of DAPI signal to normalise for variation in cell numbers between samples/fields.
Lysate preparation for western blot analysis
Whole protein extracts were analysed for Tau protein and PHF1 in AD and control 3D cultures. The 3D cultures were homogenised in tris buffered saline (TBS) extraction buffer containing 50 mM Tris-HCl-pH 7.4, 175 mM NaCl, halts protease inhibitor cocktail (78,430, Thermo Fisher Scientific, UK), 0.1 mM Phenylmethylsulfonyl fluoride (PMSF, P20270, Thermo Fisher Scientific, UK) was added along with a mixture of phosphatase inhibitors (P52102-1, Melford, UK) at 4 °C.
Sequential extraction of soluble and insoluble protein fractions for western blot analysis
Three sequential fractions, S1, S2 and S3, were extracted from each 3D sample processed (1 culture well) to study the differential solubility of Tau in the 3D cultures [34,35]. Halts Protease Inhibitor Cocktail (78430, Thermo Fisher Scientific, UK), 0.1 mM Phenylmethylsulfonyl fluoride (PMSF, P20270, Thermo Fisher Scientific, UK) was added along with a mixture of phosphatase inhibitors (P52102-1, Melford, UK) to each of the three buffers. Three individual 3D cultures (for each cell line) at 6 weeks post differentiation, were homogenised at 4 °C in TBS extraction buffer containing 50 mM Tris-HCl (pH 7.4), 175 mM NaCl along with the aforementioned cocktail of protease and phosphatase inhibitors in 1.5 ml ultracentrifuge Eppendorf tubes (357448, Beckman Coulter, UK) with a plastic pestle. 15 µg of each lysate was run in a gel to be stained with the housekeeping protein β-Actin. The lysate was then ultra-centrifuged at 186,000 g for 2 h at 4 °C. The supernatant (S1-TBS soluble fraction) was collected and stored at − 80 °C. The insoluble pellet was resuspended in 5% SDS extraction buffer containing 50 mM Tris-HCl (pH 7.4), 175 mM NaCl and 5% SDS along with the inhibitors and was homogenised. The lysate was then ultra-centrifuged at 186,000 g for 2 h at 25 °C. The supernatant (S2-SDS soluble fraction) was stored at − 80 °C. The subsequent insoluble pellet was resuspended in 100 μl SDS extraction buffer and ultra-centrifuged for 1 h at 186,000 g at 25 °C. The pellet was rehomogenised in the UREA extraction buffer and agitated for 12-18 h at room temperature. The S3-SDS insoluble urea soluble lysate was stored at − 80 °C. Before loading into gels, all fractions were mixed with equal volumes of 2 × loading buffer (LB) and boiled for 5 min at 95 °C. Equal volumes of each fraction were loaded in the gel for western blotting. A semiquantitative method was used to measure the relative proportion of each fraction to total Tau, where total Tau is the cumulative measure of each fraction (normalised) of a cell line (Total Tau = S1 + S2 + S3).
Statistics
All values are presented as the mean ± standard error of the mean. To compare differences between the groups, statistical analysis was performed using GraphPad Prism version 8.4 (GraphPad software, Inc). One-way Anova was used for statistical analysis with Bonferroni's multiple comparison post-hoc test for comparison across groups (with normally distributed data). For non-parametric values, the Kruskal-Wallis test was performed. For comparison between two groups (control lines vs AD, or 6 weeks vs 12 weeks) a two-tailed unpaired Student's t-test was performed. For normally distributed data Welch's correction was used, whereas for non-parametric data Mann-Whitney test was performed.
Generation of iPSC-derived 3D cultures-3D Culture Setup
We hypothesised that the use of a 3D culture system would accelerate, as well as facilitate the development and maturation of neurons and astrocytes in vitro. We have developed a differentiation protocol to generate cortical neurons and astrocytes in vitro (Fig. 1a). We developed a scaffold-free system using Matrigel as the matrix to embed and support the differentiation of neurons and astrocytes since Matrigel has a high level of extracellular matrix proteins typically found in the brain (Fig. 1b). In order to investigate certain molecular and biochemical aspects of AD, we used iPSCs derived from AD patients with a PSEN1 mutation and from healthy individuals (Fig. 1c). Cells were grown in 3D for up to 12 weeks, immunohistochemistry and western blot were used to characterise the cultures. Cells initially appear spherical post-seeding in 3D cultures and eventually most grow out to form extensions. By 3-4 weeks, cells form extended structures and a densely interconnected network. By 5-6 weeks, cells aggregate together forming a self-organised sheet of cells, detaching from the walls of the culture well (Fig. 1d). By 6-7 weeks, these cell sheets detach completely from the well and curl up inwards to form free-floating ball-like structures (Fig. 1e).
3D cultures contain astrocytes, glutamatergic and GABAergic mature neurons
The 3D culture-differentiation and cell composition were assessed by immunohistochemistry for neural stem cell/ progenitor marker Nestin, neuronal markers β3-Tubulin and MAP2 and astrocytic marker GFAP at different time points (Fig. 2a-f).
Nestin, a type VI intermediate filament protein, is a neural stem cell/progenitor marker, forming a major component of the cell cytoskeleton. Its expression is seen in the cell body and extends to the neurofilaments. There is a gradual decrease in the expression levels of Nestin from day 0 to 6 and 12 weeks post 3D differentiation, significantly, for 3D-HN9 (day 0-6 weeks-*p = 0.0100, day 0-12 weeks-**p = 0.0033) and 3D-HAD4 (day 0-6 weeks-***p = 0.003, day 0-12 weeks-***p = 0.0003), with a similar pattern for 3D-HN8, 3D-HAD2 and 3D-HAD3 (Fig. 2g). a and b and d and e, respectively. Nestin and β3-Tubulin in red, MAP2, GFAP, VGLUT1 and GAD 65/67 in green, DAPI in blue, scale bar, 50 µm. Images were taken using 1.3NA × 63 objective glycerol immersion. Quantification of the proportion of positive pixels for Nestin (g), GFAP (h), β3-Tubulin (i) and MAP2 (j) relative to DAPI + pixels and VGLUT1and GAD 65/67 ((l and n respectively, left and middle graphs, control vs AD lines; right graphs, all lines combined for 6 weeks vs 12 weeks) relative to β3-Tubulin + pixels of individual cell lines at day 0, 6 and 12 weeks post differentiation. n = 3 independent culture wells for each cell line at each time point, with 3 images analysed per well. Data represented as mean ± SEM, *p < 0.05, **p < 0.01, ***p < 0.001 ◂ GFAP is localised in the cytoplasmic compartment and is expressed in both astrocytes and stem cell/radial glial. GFAP is expressed as early as day 0 in the 3D cultures. There is a significant decrease from day 0 to 6 and 12 weeks post differentiation in the level of GFAP across cell lines 3D-HN9 (day 0-6 weeks-*p = 0.0011), 3D-HAD2 (day 0-6 weeks-***p = 0.0009, day 0-12 weeks-**p = 0.0020) and 3D-HAD3 (day 0-6 weeks-*p = 0.0219). The expression at day 0 suggests a high proportion of neural stem/progenitor cells, in line with the expression of Nestin. We also observe an increase in GFAP between 6 and 12 weeks for 3D-HN9 (*p = 0.0211), which might indicate an increased proportion of astrocytes (Fig. 2h).
β3-Tubulin, a class III member of the Tubulin family which plays an important role in microtubule assembly, is considered an early neuronal differentiation marker. The majority of cells in 3D cultures express β3-Tubulin as early as day 0. There is a significant increase in the level of β3-Tubulin for 3D-HN9 (day 0-12 weeks-*p = 0.0375) and 3D-HAD3 at 6 (*p = 0.045) and 12-week (*p = 0.0414) post differentiation when compared to day 0 (Fig. 2i).
Synapsin1, a protein found in the pre-synaptic vesicle, is expressed in most of the cells after 6 weeks of differentiation in the 3D cultures (Online Resource 1). VGLUT1, a glutamate transporter protein, is a protein associated with the membranes of synaptic vesicles and is found in the somatodendritic compartment as well as the axonal terminals of excitatory neurons. It is expressed as early as 6 weeks post differentiation in the 3D cultures (Fig. 2k). There is no significant difference between control and AD lines at 6 weeks post differentiation. However, there seems to be a difference, although not significant, at 12 weeks post-differentiation. On analysing the relative proportion of VGLUT 1 in the 3D cultures as a whole (combining all cell lines) over time points 6 and 12 weeks there is a significant increase at 12 weeks compared to 6 weeks of differentiation (*p = 0.0170) (Fig. 2l).
Glutamic acid decarboxylase (GAD) 65/67 synthesises GABA from glutamate and is a marker of inhibitory neurons. The 3D cultures express GAD65/67 as early as 6 weeks post differentiation (Fig. 2m). There are no significant differences between control and AD cultures at 6 weeks post differentiation or between time points 6 and 12 weeks (all cell lines combined) in the differentiated 3D cultures. However, there is a significant difference between the control and AD cultures at 12 weeks post differentiation in the 3D cultures (p *0.0409) (Fig. 2n).
In conclusion, the 3D cultures show neuronal and astrocytic differentiation across time, with presence of both glutamatergic and GABAergic neurons, and the expression of synaptic proteins.
Neuronal differentiation is increased in 3D cultures compared to 2D cultures
It is well known that 2D cultures fail to mimic the in vivo microenvironment and thus are not the best candidates for developmental and disease modelling studies. Over recent times, extensive studies have established that 3D cultures are superior to the conventional 2D cultures not just with regards to differentiation and maturation of different cell types but also in the expression of pathological phenotypes in vitro [21,23,25,30]. Here, we use 2D cultures in parallel with the 3D cultures for the initial phase of differentiation (day 0-6 weeks) to confirm enhanced neuronal differentiation in our model. The relative expression of the different markers was analysed across culture types (2D and 3D) at 6 weeks post differentiation for different markers: Nestin, GFAP, β3-Tubulin and MAP2 (Fig. 3a-d).
There is no significant difference observed in the levels of Nestin when cells are grown in either 2D or 3D for HN8, HN9 and HAD3 lines (Fig. 3e). Similarly, there was no statistical difference observed for GFAP when grown in 3D vs 2D for the cell lines HN9 and HAD3 (Fig. 3f). We observed that the levels of β3-Tubulin are higher in 3D cultures compared to 2D cultures, statistically significantly for HN9 (*p = 0.0126). However, when the levels of β3-Tubulin in 2D cultures were compared with that of the 3D cultures (cell lines combined) at 6 weeks post differentiation, there was a significant increase seen in the 3D cultures (*p = 0.0122) (Fig. 3g). Similarly, there is a higher proportion of MAP2 in the 3D cultures as compared to the 2D cultures, statistically Fig. 3 Neuronal differentiation is increased in 3D cultures compared to 2D cultures: representative confocal images of immunohistochemistry of 2D-HAD2 cultures stained for Nestin (a), GFAP (b), β3-Tubulin (c) and MAP2 (d). Nestin and β3-Tubulin in red, MAP2 and GFAP in green, DAPI in blue, scale bar, 50 µm. Images were taken using 1.3NA × 63 objective glycerol immersion. The proportion of Nestin (e), GFAP (f), β3-Tubulin (g) and MAP2 (h) positive pixels relative to DAPI + pixels in confocal images across cell lines grown in 2D and 3D culture at 6 weeks post differentiation. g and h left and middle graphs, data per cell line; right graphs, all lines combined. n = 3 independent culture wells for each cell line at each time point, with 3 images analysed per well. Data represented as mean ± SEM, *p < 0.05, **** p < 0.0001 ◂ significantly for HN9 cell line (*p = 0.0116) with a similar pattern for HN8 and HAD3 cell lines. However, when the relative proportion of MAP2 in the 2D cultures was compared to the 3D cultures (cell lines combined) at 6 weeks post differentiation, there was a significantly higher proportion of MAP2 in the 3D cultures (****p < 0.001) (Fig. 3h).
We observed a higher proportion of neuronal markers β3-Tubulin and MAP2 in 3D as compared to 2D, suggesting an increase in the proportion of differentiated neurons in the 3D cultures.
Expression of total and phospho Tau in the 3D cultures
Although FAD is driven by mutations in APP and PSEN genes, the wild-type Tau becomes abnormally hyperphosphorylated and aggregates during the course of the disease. However, SAD which is not driven by such known mutations and accounts for more than 90% of AD cases also shows similar pathology and clinical presentations to the FAD cases. Despite decades of investigations, the exact cause and the mechanism(s) by which FAD mutations lead to Tau pathology are not yet clear. Here, we addressed this in our FAD 3D culture model and ascertained whether the 3D microenvironment encourages any early changes in the pathophysiological levels of total Tau, especially in its phosphorylation and solubility. We were also interested in studying the expression of 3R and 4R Tau isoforms to determine the predominant maturity of the neurons in the 3D cultures. The adult human brain consists of 6 Tau isoforms, classified as 3R or 4R depending on the presence of 3-repeat or 4-repeat domains at the carboxy-terminal end. The expression of these isoforms is developmentally regulated. 3R Tau has predominantly expressed during the foetal stages and 4R Tau is the adult-specific isoform. In healthy human brains, 3R and 4R Tau isoforms are expressed in equimolar proportions.
To analyse Tau in our 3D cultures, the protein lysates (6 weeks and 12 weeks) were analysed for expression of total Tau by western blotting (Fig. 4). At 6 weeks post differentiation both the controls and AD cell lines express total Tau (Fig. 4a). There are significant differences between the control cell line HN9 and the AD lines HAD2 (*p = 0.0148), HAD3 (**p = 0.0023) and HAD4 (**p = 0.0018) presenting higher levels of total Tau compared to control. However, there is significantly higher total Tau in the AD lines (combined) when compared to control line at 6 weeks post differentiation (***p = 0.0001) (Fig. 4b). There is a trend of the same pattern of increased total Tau levels in AD lines compared to control lines is evident at 12 weeks (Fig. 4c) however, no significant differences are observed at this time point between the different cell lines, although there seems to be a significant increase in the AD lines (combined) when compared to the control lines (combined) (**p = 0.0015) (Fig. 4d). When comparing combined data, there is increased total Tau expression in the AD cell lines compared to the control cell lines, both at 6 and 12 weeks.
Overall, there is an increase in total Tau levels in the AD cell lines compared to the control cell lines, grown in 3D cultures at both 6 and 12 weeks time points. Whether this is a definite or a more transitional change should be assessed further by increasing the number of cell lines and analysing cultures beyond the 12 weeks time point.
Phosphorylation of Tau is considered an important aspect of Tau-induced toxicity in the disease progression. Thus, it was crucial to check if there were any differences between the control and AD lines at different time points. PHF1(Ser 396/404) is an AD-associated phosphorylation epitope. PHF1 is seen in both the controls and the AD-3D cultures at 6 and 12 weeks (Online Resource 2a, c) post differentiation. There was no statistical difference for PHF1 levels between the AD and the control lines at 6 and 12 weeks (Online Resource 2b, d) post-differentiation.
The level of insoluble Tau is higher in 3D-AD vs 3D-control cells
In AD brains, Tau undergoes conformational changes to transition from soluble monomers to increasingly insoluble oligomers and fibrils, that finally coalesce to form insoluble aggregates. In this experiment, soluble and insoluble Tau levels were assessed by enriching for insoluble Tau fractions obtained by differential centrifugation following standardised techniques [34,35]. The TBS soluble fraction (S1) shows the presence of monomeric Tau. The TBS insoluble/SDS soluble fraction (S2) shows the presence of Tau oligomers. The Urea soluble S3 fraction shows the presence of insoluble aggregated Tau. In order Fig. 4 Tau expression and insolubility is higher in 3D AD cells versus 3D control cells: representative western blot images of total lysates of both control and AD -3D cultures at 6 weeks (a) and 12 weeks (c) post differentiation. The blots were probed with Dako Tau (55 kDa) and β-Actin (42 kDa). Analysis of the western blot data at 6 weeks (b) and 12 weeks (d) post differentiation. Representative western blot of the fractionated lysates of 3D cultures at 6 weeks post differentiation (e). Blots were stained with Dako Tau (55 kDa). Western blot analysis of fractionated protein lysates S1, S2 and S3 fractions from control and AD lines grown in 3D cultures at 6 weeks post differentiation across cell lines (f). Equal volumes of normalised samples were loaded in each well. Each fraction is measured as a ratio over total Tau as the sum of all fractions. M, molecular weight marker, n = 3 independent culture wells for each cell line at 6 weeks. No statistically significant differences found across individual cell lines grown in 3D cultures at 6 weeks post differentiation. When S1, S2 and S3 fractions from control and AD cells lines were grouped for comparative analysis, there was a significant increase in the proportion of insoluble Tau in the S3 fractions of 3D-AD groups. Data represented as mean ± SEM *p < 0.05, **p < 0.01. ***p < 0.001 ◂ to check the proportion of Tau in each of the three fractions, lysates of cell lines were fractionated and normalised based on their total protein concentration estimated with BCA assay. Briefly, each fraction was calculated as a ratio of the fraction-band intensity over total Tau normalised against β-Actin (from total lysates). Total Tau was calculated from combining all the fractions S1, S2 and S3.
We observed Tau to be sequestered mostly in the S1 and S2 fractions whereas only a small proportion was found in the S3 fraction across all the cell lines at 6 weeks post differentiation (Fig. 4e). This shows that most of the Tau are soluble in all these cell lines. When we compared the proportion of Tau for each individual cell line grown in 3D culture at 6-week time point, there was no significant difference in any of the fractions across cell lines, though there was a trend for the levels of insoluble Tau in the AD lines to be greater than that seen in controls. When control and AD lines were grouped for cumulative comparison (control vs AD groups), there was a significant increase in the insoluble S3 fraction in the 3D AD group compared to the 3D control group (*p = 0.0078) (Fig. 4f) with no difference in the soluble S1 or S2 fractions between the groups at 6 weeks post differentiation.
Next, we analysed whether there was a difference in the levels of 3R and 4R Tau isoforms between control and AD cell lines. We observed 3R Tau was mostly found in the S1 and the S2 fractions in all the cell lines and a very small proportion in the insoluble S3 fraction of the 3D cultures at 6 weeks post differentiation (Online Resource 3a). There was no significant difference in the proportion of 3R Tau in the three fractions across the cell lines at 6 weeks post differentiation (Online Resource 3 b-d). Interestingly, when control and AD lines were grouped for cumulative analysis (control vs AD groups), there was a significant decrease in the proportion of 3R Tau in the S2 fraction (*p = 0.0310) and a significant increase in the S3 fraction (*p = 0.0176) from 3D-AD cultures when collectively compared to the 3D control lines (Online resource 3c-d).
The adult 4R Tau isoforms were detected at 6 weeks in both control and AD 3D cultures, however, 4R Tau bands were too low in intensity to quantify (Online resource 3e).
Thus, we have generated a human 3D-AD neural culture model which shows changes in the solubility of Tau and the presence of aggregated Tau as early as 6 weeks post differentiation. Although at 6 weeks the proportion of urea insoluble vs soluble Tau in the 3D cultures is quite low, there is a trend for more detergent insoluble Tau in the AD lines when compared to the healthy control lines. Though these trends were only significant when the AD and controls lines were grouped for cumulative analysis; future studies should assess if this difference is evident with a greater number of cell lines without grouping.
Discussion
Development of a human in vitro model for a neurodegenerative disease like AD which replicates most if not all pathological hallmarks is quite challenging. Some studies have previously reported AD phenotypes in their 3D/ organoid cultures using either immortal neural cell lines or by forcing the pathology by various means [23,25,36]. Although these models recapitulate some of the AD biochemical phenotypes and constituted seminal works at the time, they also lack the relevant pathophysiological levels of protein expression. In this study, we address this shortcoming by using iPSCs derived from FAD patients, thereby, bypassing the need to overexpress or genetically manipulate any FAD genes, allowing us to study the disease without overexpressing mutated genes. 3D cultures are great tools to recreate a physiologically relevant brain microenvironment and significantly improve the maturation and functionality of the different neuronal cell types [37][38][39]. Here, we have developed a human 3D neural culture model using iPSC-derived NSC from healthy individuals and FAD patients with PSEN1 mutations-L286V, M146L or A246E, maintained for 12 weeks post-3D plating in vitro. Our model shows early signs of neuronal maturation in the 3D cultures, increased total Tau in 3D-AD when compared to the 3D controls, and early signs of Tau insolubility from cells derived from iPSCs from patients with PSEN1 mutations. Protein aggregation being the primary hallmark of a disease like AD, an ideal model would be one that reflects brain-like compartmentalization, which is instrumental to help mimic the relevant microenvironment and recreate the pathology in vitro. A 3D culture system could be engineered to integrate a heterogeneous population of cell types dispersed in three dimensions enhancing cell-cell communication in the intracellular niche, embedded with specific scaffold materials, extra-cellular matrix components (ECM) and growth factors. This would not just accelerate maturation but also provide the compartmentalization needed to accumulate pathological phenotype, as shown by other studies [23,25]. Using Matrigel, which contains a high level of brain ECM proteins such as laminin, entactin, collagen and heparin sulphate proteoglycans, we see a more robust neuronal differentiation in 3D cultures compared to the 2D cultures. The 3D cultures also contain GFAP + astrocytes, with the proportion of neurons and astrocytes increasing while Nestin + progenitor cells decreased over time. Expression of Synapsin1, a pre-synaptic marker is seen as early as 6 weeks post 3D plating, alongside glutamatergic marker VGLUT1 and GABAergic marker GAD65/67. Our heterogeneous human neural 3D culture model shows neuronal maturation and neural network formation as early as 6 weeks post 3D plating.
This enhanced differentiation could be attributed to a more physiologically relevant microenvironment and promoting neuronal differentiation in 3D cultures.
The 3D AD cultures increased displayed increased total Tau and some evidence of greater insoluble Tau as compared to the 3D control cultures at 6 weeks post differentiation. This suggests that Tau aggregation in FAD may not be a end-stage process and starts earlier than expected, supporting the thought that AD pathology begins many years before the onset of clinical symptoms. This is a good in vitro model to explore protein aggregation and its effects and may be adapted to study other neurodegenerative diseases such as Parkinson's disease (PD) and Demetia with Lewy bodies (DLB). Tau is abnormally hyperphosphorylated and aggregated in the form of paired helical filaments (PHFs) which progresses to form tangles in AD [29]. There are no significant changes in the PHF1 in 3D-AD lines when compared to 3D controls at 6 weeks and 12 weeks post-differentiation. However, for further analysis, we will try using other phospho-epitopes such as AT8 (Ser 202, Thr 205). A study in 2015 using cortical neurons derived from patientspecific iPSC also reported no change in PHF1 levels in their PSEN 1 (PSEN1 Y115C, M146I, intron 4) as opposed to the APP cell lines used (APP V717I mutant cells). Moreover, this study reported an increase in total Tau only in the APP mutant lines, and not in the PSEN1 lines [33], whereas, we see a total Tau expression increase in PSEN1 lines.
One of the challenges in modelling an age-related neurodegenerative disease like AD is to generate mature and aged neurons. Although publications on 3D-AD culture models have shown some neuronal maturity with different markers, most do not show expression of the adult 4R Tau isoform at the protein level [24,25,30]. However, a few 3D/organoid models studies show expression of 4R Tau at the protein level, following long-term cultures of 300 days [40] and 365 days [41], or long-term use of Brain-Phys with expression at 25 weeks [42]. Although, we do not see all six isoforms, we do see 4R isoforms expression as early as 6 weeks after 3D plating. We believe the expression of the 4R isoforms could become more prominent with extended culture times as shown by the above-mentioned studies. This shows that our 3D culture provides an environment for increased differentiation and more efficient maturation and ageing of the neurons.
Premature neuronal differentiation might occur early during development in FAD patients with PSEN1 mutations both in vitro and in vivo [43,44]. This might explain the early pathogenic differences seen between the control and AD 3D cultures.
Studies have shown that neuronal ageing is challenging to replicate in vitro as reprogrammed iPSCs usually differentiate into immature or foetal-like neurons which are quite different from those at the age of onset of diseases like AD [45,46].
Although their maturity may be far from the age of onset, these immature cells display pathological attributes which support the concept that pathological changes happen in the pre-symptomatic phase of the disease, years before age of onset, and that such models are great tools to study and target early changes in diseases. However, an in-depth characterisation of the different cell types and proportions in our 3D cultures needs to be done in the future to completely understand the extent and maturity of cortical differentiation. This also includes a clear distinction between the type of glial cells. Although GFAP is a classic astrocytic marker, it is a marker also expressed in the radial glial cells. Currently, there is no single marker that could distinguish between early and late astrocytes due to a general overlap seen with the available markers. A combination of markers such as aquaporin 4 (AQPR 4) or S100β with GFAP could be used to get a clear characterisation of the cell type [47,48]. The accurate sub-cellular localization of the markers can also give an indication of the physiological conditions. While GFAP is found in the cell processes, S100β is found in the soma of mature astrocytes [49,50] and AQP4 is found in the end feet of astrocytes [51,52].
Although there is a rapid development in the field of neural 3D/organoid cultures, they do come with their own limitations [20]. Some of these limitations are the absence of crucial elements which make the brain a complex organ. Lack of microglial compartment, vasculature and blood-brain barrier separates us from the goal towards a physiologically relevant model. In recent years, the advances in microfluidic systems have made it possible to address these issues [53]. A human 3D triculture microfluidic system was used to demonstrate simultaneous culture of neuron, astrocytes and microglial capturing AD features such as Aβ aggregation, accumulation of phosphorylated Tau, neuroinflammation and microglial recruitment [54]. Although an immortal microglial cell line was used, studies like these promise a platform to work with patient-specific isogenic lines. Introducing elements of vasculature and blood-brain barrier would be advantageous in reducing necrotic cores, a long-standing problem seen in organoid cultures [55].
3D culture systems are widely applicable wherein they can be tailored not just for disease modelling or drug targeting studies but also to help uncover numerous unknown physiological and pathological mechanisms. For instance, recent studies have led to believe reduced expression of perivascular AQP4, a major regulator of the glymphatic system affects the clearance of toxic solutes such as Aβ [56][57][58][59][60][61]. 3D models could be employed to further study clearance mechanisms in the brain and pave way for developing effective drugs for diseases like AD.
The field of 3D/organoid/spheroid culture technology is constantly striving to make the system more adaptable for advanced imaging techniques such as TEM, expansion microscopy as well as live functional readouts such as calcium imaging and multielectrode array (MEA) to study neuronal connections and activity and thereby overcome some limitations in the field [62][63][64]. These adaptations would help study effects in real-time to explore mechanisms such as neuroinflammation, microglial activation and blood-brain barrier using various advanced platforms such as microfluidics, microvessel-on-a-chip and Organ-on-Chip technology [65,66]. Further advances in technology such as 3D-bioprinting are underway to enable a multicellular and multi-organ research approach to improve therapeutic potentials [67].
In conclusion, our human iPSC-derived 3D neural culture model is a promising and physiologically relevant system with a mix of astrocytes and mature neurons, including both excitatory and inhibitory neurons, expressing some features of the mature Tau as early as 6 weeks in vitro. We believe that the 3D culture would be a good model to study other aspects of the AD pathology such as amyloid pathology, synaptic abnormalities and neuroinflammation with microglia, and may also be used for compound screening in the future. | 9,987.6 | 2022-11-14T00:00:00.000 | [
"Biology"
] |
Enhanced Bitcoin Protocol with Effective Block Creation and Verification by Trusted Miners
The Distributed nature of Bitcoin introduces security issues that necessitate security-specific enhancements in Bitcoin protocol. Therefore, proposing a method of incorporating criteria check and verification process for miners to participate in the mining process and join the mining pool respectively. The proposed idea mitigates double spending, block withholding, and 51 percentage attacks. In addition, an increase in the rate of Bitcoin users necessitates performance improvement. Hence, proposing an effective approach of refining the existing block creation and verification strategy for improving transaction rate without compromising security.
Introduction
The drastic growth of Internet makes e-shopping a key fragment of commercial web activities and its main attribute is online payments. It deals with electronic fund transfers from customers to merchants using existing electronic payment mechanisms like card payments and Internet banking. For performing transaction verifications, the system requires a third-party mediator (ex. PayPal) through which user's sensitive data are exposed leading to sensitive information leakage. Consequently, there is a need for a secure payment system that does transactions without requesting any sensitive details from users and without a third party for verifications. A decentralized digital currency named Bitcoin [1] gives a solution for the above-stated problems, stating everyone is a bank. The Bitcoin transactions are peer-to-peer and it uses the concept of Blockchain [2], a sequence of blocks so called distributed ledger for handling all transactions and user's digital wallet [2], a secure software application for storing Bitcoins. At present, value of one Bitcoin is, 1 BTC = 7219.9 $ and its INR value is 1 BTC = 472052.11 ₹. Its value may change with time. It is also feasible to transfer segments of Bitcoin and one Satoshi (10 −8 BTC) is the minimum transferable amount. A digital wallet should be installed in user's system before start transferring Bitcoins and each wallet produces a pair of the public-private key. The Public key of the user is utilized to generate Bitcoin address [3] to which the Bitcoins are transferred. The overall flow of Bitcoin protocol for transferring Bitcoins from user A to user B is as follows: User A first creates a transaction (TX) [2] specifying receiver B's Bitcoin address, amount in BTC and digitally signs it with A's private key. The transaction created is broadcasted in Bitcoin network followed by the verification by the miners against a set of validation rules [4]. The verified transactions for a specified duration are combined to form a block [2] of size 1 MB. A block is referred as a leaf of the distributer ledger i.e. the blockchain. The created block is verified by the mining process (Proof of Work [1]) and on success the solution is published in the network for verification and added to the block chain. Remaining miners validate the solution and on receiving six confirmations the transferred Bitcoin is credited to B's account. Bitcoin protocol excludes the third-party intervention by the concept of decentralization, leading to security issues like double spending [10], block withholding [11] and 51 % attacks [12]. For handling the security issues [13] [14], the protocol does a tough mining process named Proof-of-Work which affected the performance of the system resulting in very low transaction rate of 7-8 tps. The transaction rate [5] of Bitcoin protocol is the transactions count affixed to the block chain per second. The determined size of a block is 1 MB, transaction size is ≈ 200 KB and average time taken is 10 minutes to attach a block to the block chain. Based on these figures the average transaction rate of the existing system computes to 7-8 tps. In real-time, the transaction rate varies from 1.5 to 7 tps.
Related Works
In 1988, David Chum proposed first E-cash scheme [6] with untraceability termed as Digital Cash. It is a system of purchasing cash credits in a small amount and storing the credits in the respective storage such as computer systems, mobiles etc. Then using those credits for online payment and transactions. Chaum's Blind s ignature [7] and Zero knowledge proof [8] techniques are utilized in Digital cash for maintaining the anonymity of users. Even though it maintains anonymity, it is a digital currency that represents the centralized system that requires a banking module for generating and maintaining the E-cash credits. This leads to the Bitcoin [1] invention on Nov 2008 by Satoshi Nakamoto that provides an appropriate solution by eliminating the requirement for a central bank. That is, each Bitcoin user maintains an individual copy of the record which would be handled at the central bank. To maintain those records in a distributed way, mechanisms are required and it should also handle Bitcoin creation and management. In Bitcoin network, block chain concept [2] does the job of distributer ledger for maintaining the Bitcoin transactions. It also logs the owner and timestamp of each transaction. A concept named Proof of Work (PoW) [15] is used in Bitcoin for maintaining the distributed ledger in a secured way. PoW is like Adam Back's Hash cash [9]. It is the process of sequentially varying and finding the nonce value which on hashing with block header using SHA 256 produces the hash value having a pre-defined zero's in beginning. Large computational power is required to find the solution for the puzzle and it is the sequential process. That is, each block header which is included in the puzzle has a value called preBlockHash, which points to previous block. So, if an attacker wants to attack any of the blocks and change its hash, should also solve that block and the forthcoming blocks to make the attack successful. Transaction rate [5] is a metric for analysing the Bitcoin system performance. It is the transactions count attached to the blockchain per second. Series of activities [16] are involved in processing a transaction and adding it to the blockchain. First the transaction is created by user A, followed by the verification against a set of rule. Once the transaction clears the verification process [17] a set of transactions are added to a block and undergoes mining process [1]. Transaction is said to be processed only if it succeeds block verification. Transaction rate is the time taken for a transaction to complete the above-said process. Theoretically the maximum possible transaction rate of Bitcoin system is 10 tps with a block size of 1MB and minimum transaction size of 166 bytes. The transaction rate can be increased a bit by the user if they combine and process their transactions leading to a reduction of 10 bytes in the transaction size. Even though Bitcoin preserves the obscurity of the users, it has some open issues that require attention. Notably, the transaction rate of Bitcoin protocol computes to 7-8 tps which is minimal when compared to other electronic payment mechanisms (VISA -2000 tps, PayPal -100 tps). The main difficulty Bitcoin users face nowadays is the delay in confirming the unconfirmed transactions.
Motivation
Proof-of-Work concept used for achieving security degrades its performance leading to a minimal transaction rate (number of transactions appended to the memory pool per second) of 7-8 tps. Notably, the transaction rate of Bitcoin protocol is minimal when compared to other electronic payment mechanisms (VISA -2000 tps, PayPal -100 tps). The main difficulty of the Bitcoin users is the delay in confirming the unconfirmed transactions. Therefore, motivated to propose a solution for performance improvement with enhanced security.
Proposed Work
The Proposed work concentrates more on enhancing security by refining the mining process (Proof of Work module). Currently, computational power is the only criteria for the miners to take part in the mining process. This is the reason why attackers easily enter as miners/join the mining pool to solve Proof of Work puzzle and perform attacks like 51%, double spending, and block withholding. Proposed work also concentrates on increasing throughput of Bitcoin protocol by refining block creation and verification strategy without compromising security.
Proposed Mining Process
To counter the issue, we propose a solution that includes criteria check for Miners to participate in the mining process and verification process for miners to join the mining pool. Criteria check/ verification process includes set of rules to verify the trustworthiness of the users which shall act as a barrier for the attackers to participate in the mining process. Figure 2 depicts the modules of proposed mining process.
Verification Process
The verification process is to prevent attackers from joining mining pool by verifying miners against a list of verification rules. The validation rules are as follows, Rule1: Checking the attack history of miners i.e. no. of forks [2] created so far in the blockchain. Rule2: Checking the balance Bitcoins in their digital wallet [2], because users having more no. of Bitcoins shall have the more responsibility to build trusted network. Since only trusted miners enter the pool based on above two rules, there is no possibility of retaining the solution obtained for the Proof of Work puzzle. This could be mitigation for block withholding attack (BWH).
Rule3: Checking whether the sum of pool's computational power ( ) for which miners are requesting and miners computational power ( ) is not more than 50 % of total computational power. This could be the mitigation for 51% attack. The check for rule 3) is done using equation 1,
Criteria Check
This module is for miners to participate in the mining process. It checks with a list of criteria to permit only the honest miners to perform mining process which could be the mitigation for the double-spending attack. The rule 1 and 2 mentioned in the verification process is also applied for this module since they are created for individual miners.
Transaction Rate
The transaction rate of existing payment mechanisms like PayPal and VISA are high when compared with Bitcoin protocol. It processes the transactions at the rate of 7-8 tps.
Transaction Rate is the Bitcoin transactions count added to block chain per second and equation 2 can be used to compute its value.
Where, -Transactions count per Block.
Effective Block Creation and Verification Strategy
Proposing an effective process for increasing the transaction rate of existing system resulting in performance improvement of Bitcoin protocol without compromising security. This is achieved by refining block creation and EAI Endorsed Transactions on Scalable Information Systems 05 2020 -10 2020 | Volume 7 | Issue 27 | e8 verification strategy of the protocol which targets at reducing the block processing time ( ) for few (requires normal level of security) of the blocks. The blocks are classified into premium and normal based on the threshold value. The premium block holds the transactions that entails high level of security and it undergoes existing Proof of Work process. The normal block holds the transactions that requires medium level of security and undergoes Proof of Work with reduced challenge. Since block processing time and transaction rate are inversely proportional as given in equation 2, the abo ve said refinement improves the transaction rate.
Setting Threshold
Threshold value is the Bitcoin amount and the idea is to provide high security for the transactions having amount above threshold and medium level of security for the transactions having amount less than threshold. Threshold value is decided by carrying out the study on the logs of main-net transactions with their amount and frequency. Main-net is the real time Bitcoin network that does Bitcoin transactions around the globe. The real time transactions can be simulated in the proposed system and the system performance can be studied for different threshold values for finalizing on the threshold.
Proof of Work -Target
The main reason to utilize Proof of Work consensus in Bitcoin protocol is to provide anti-double spending attack and Denial of Service attack defense. The process of Proof of Work is varying the nonce to find the hash using SHA -256 from block header information that meets the target set for a block. The target is calculated for every block based on the difficulty and Bitcoin network hashing power. Equation 4 can be used to compute its value. The difficulty is calculated taking into account the preceding block attributes of the block chain and with complexity which can be cracked in an average time of 10 minutes. The target is usually calculated as a 256 bit number leading with specified number of zeros and represented in 32 bits by performing compact encoding on 256 bit target. For Block 0, the 256 bit target calculated from the block attributes is "000000000019d6689c085ae165831e934ff763ae46a2a6c 172b3f1b60a8ce26f" and it is represented as "1d00ffff". The difficulty of above target is 1 i.e. the block hash should have 4 bytes leading zeros.
Reduced Target
It is the value of the target with a reduced number of leading zeros. Reducing the target by level one is removing a leading zero with reduces the processing time by half.
Here the complexity is also reduced but the reduced target is applied only for transactions that requires medium level of security. Level two reduction is the process of removing two leading zeros in proof of work target. The following algorithm describes the steps involved in the proposed block creation and verification strategy for improving the transaction rate of Bitcoin protocol without compromising security.
Experimental setup
Bitcoin protocol experimentation has been carried out using BitcoinJ, a complete Java library of protocol implementation that can maintain a wallet and send/receive transactions. For experiment 1, The Proof-of-Work module present in that library was enhanced to adopt the proposed block creation and verification strategy. For experiment 2, the verification and criteria check modules are implemented with set of rules proposed. The experiment was conducted with simulated Peer-to-Peer Bitcoin network of 850 sq. m. area with 25 Bitcoin nodes and 9 miners. And utilized Testnet3 for conducting real time testing which is the Bitcoin testing network.
The system performance was evaluated with below evaluation metrics, Transaction Rate: Number of transaction added to block chain per second and its value can be calculated using equation no.2. Chain Quality (CQ): The proportion of blocks in any k long subsequence produced by an attacker is < µk and it can be evaluated using equation no.5.
Results
With configured setup of 25 Bitcoin nodes and 9 miners within the nodes, the proposed system processes transactions at the rate of 12.685 tps by reducing the target of mining process (Proof-of-Work) by one level. The sequence of experiments conducted are as follows, Experiment 1 was conducted to evaluate the performance of proposed system without any threshold set and Figure 4 shows the transaction rate of proposed system considering medium level of security for all transactions. The optimal result was achieved with a threshold value of 30 and implementing the proposed enhancement in mining process (Proof-of-Work). As per the proposed verification process, Level one -At level one, the optimal transaction rate was 12.600 tps which is increased by 77.46% when compared with existing protocol. The transactions with amount > T, i.e. the premium transactions that requires high level of security are processed at the rate of 6.66 tps and transactions with amount < T are processed at a rate of 13.3 tps.
Level two -At level two the optimal transaction rate was 24.480 tps. Premium transactions are processed at the rate of 6.66 tps and transactions with amount < T are processed at a rate of 26.64 tps. Experiment 3 was conducted to compare the performance of existing and the proposed system and Figure 6 shows the comparison. Experiment 4 was conducted to implement proposed criteria check and verification modules for blocking untrusted miners from participating in mining process and to join mining pool that mitigated Block withholding, double spending and 51 % attacks.
Conclusion
In this paper, an effective block creation and verification strategy was proposed for increasing the transaction rate of Bitcoin protocol from its current rate 7.1 tps to 12.600 tps without compromising security. The proposed system was implemented using BitcoinJ and the results were evaluated using standard evaluation metrics. The results obtained confirm that the proposed strategy outperforms the existing protocol in terms of throughput (transaction rate). In an average the proposed strategy improves the transaction rate by 77.46% when compared with existing protocol. Further it is evident that the performance is optimal at a threshold value of 30 and was depicted in the results. Also, an efficient pre-mining module is proposed which permits only trustworthy miners to participate in mining process for mitigating security attacks namely 51%, double spending, block withholding. This is made possible by introducing criteria check and the verification modules in the mining process, which can block attackers entering Bitcoin mining network. It resulted an average reduction in success rate of Double Spending, Block Withholding and 51% attacks by 63.60%, 54.64% and 46.07% respectively.
Future Enhancement
The proposed mining process helps to mitigate attacks like double spending, 51 %, block withholding. The proposed system can be further enhanced to counter unhandled attacks like DDoS, transaction traceability etc. In addition, there is a scope for scalability issues because of surgical increase in Bitcoin users. The Bitcoin scalability problem denotes the limits on the amount of transactions the Bitcoin system can process. The transaction processing capacity of the Bitcoin network is limited by the average block creation time of 10 minutes and the block size limit of 1MB. The scalability directly affects the transaction rate, hence improving transaction rate along with scalability consideration is also a key focus area in Bitcoin protocol. | 4,069.8 | 2018-07-13T00:00:00.000 | [
"Computer Science"
] |
The one that got away: A unique eclipse in the young brown dwarf Roque 12
We report the discovery of a deep, singular eclipse of the bona fide brown dwarf Roque 12, a substellar member of the Pleiades. The eclipse was 0.6mag deep, lasted 1.3h, and was observed with two telescopes simultaneously in October 2002. No further eclipse was recorded, despite continuous monitoring with Kepler/K2 over 70d in 2015. There is tentative ($2\sigma$) evidence for radial velocity variations of 5km/s, over timescales of three months. The best explanation for the eclipse is the presence of a companion on an eccentric orbit. The observations constrain the eccentricity to e>0.5, the period to P>70d, and the mass of the companion to to ~0.001-0.04Msol. In principle it is also possible that the eclipse is caused by circum-sub-stellar material. Future data releases by Gaia and later LSST as well as improved radial velocity constraints may be able to unambiguously confirm the presence of the companion which would turn this system into one of the very few known eclipsing binary brown dwarfs with known age.
INTRODUCTION
Stellar variability encodes information about stars and their environment.Firstly, specific types of variability can be used to infer fundamental properties of stars that are notoriously difficult to determine by other means (luminosity, radius, distance, rotation period).Second, variability can reveal the existence of unseen objects (e.g.exoplanets, stellar companions) or may be used to map unresolved environments (e.g., protoplanetary disks, AGN disks).And third, variability allows us to directly witness dynamic, stochastic processes in the surroundings of stars and other unresolved sources, including magnetic activity, weather, accretion flicker, or disk instabilities (Percy 2011).
Eclipses are a rich source of information as well as essential calibrators for stellar models.Among eclipsing binaries, those who belong to a cluster and thus have a known age deserve a special distinction.For brown dwarfs with masses below the sub-stellar limit, the lack of known eclipsing binaries at different ages has been a major hindrance for the development of models.Unlike stars, brown dwarfs never reach a steady state like the main sequence and cool down as they get older.The atmospheres of brown dwarfs transition from an initial state that is comparable to very low-mass stars to exoplanet analogues with complex molecular chemistry, dust formation, and clouds (Helling & Casewell 2014).The<EMAIL_ADDRESS>deep convection zones as well as the decline of fusion processes in the core also pose challenges for models.Calibrators are therefore needed at all evolutionary stages.
For a decade, only one bona fide brown dwarf eclipsing binary has been known, a member of the 1 Myr old Orion Nebula Cluster (Stassun et al. 2006).This one has shown a temperature reversal, with the more massive component being cooler than the secondary and demonstrated the need to take into account the effects of magnetic fields in evolutionary models for brown dwarfs (Stassun et al. 2007).Two more eclipsing systems with a pair of brown dwarfs have been found more recently, both in systems with more components, one in the 5-10 Myr Upper Scorpius association (David et al. 2016) (but see also Wang et al. (2018)), and one with a grazing eclipse in the 45 Myr old Argos moving group (Triaud et al. 2020).A few more detached very low mass eclipsing binaries with a component below the brown dwarf mass limit have been discovered as well (e.g.Lodieu et al. 2015;Irwin et al. 2010;Johnson et al. 2011).For a large part of the age-mass parameter space, however, there is still no eclipsing substellar system to provide a sanity check for evolutionary models.
Not only compact objects can cause eclipses.Stars (or brown dwarfs) are occasionally obscured by material in orbit around them.These sources are excellent laboratories to obtain spatial information about processes in disks or clouds or tori that are otherwise impossible to resolve.Stars like UX Ori (Grinin et al. 1994), KH15D (Herbst et al. 2002), Aur (Kopal 1971), AA Tau (Bouvier et al. 1999), and RW Aur (Bozhinova et al. 2016) are prominent examples of systems with deep eclipses that have provided valuable insights into early stellar evolution and planet forming processes.For young low-mass stars, in particular, eclipses by circumstellar material may currently be the only passable method to gain spatial information on scales smaller than 1 AU which cannot be reached with submm/mm interferometry or Adaptive Optics (Evitts et al. 2020).Moreover, the resolution that can be achieved by modeling variability caused by obscurations is independent of distance and applicable down to very faint stars.For brown dwarfs this technique is currently the only hope of mapping the planet-forming zone around them.
For all these reasons it is thought of as critical to find rare systems undergoing eclipses, even more so when the primary source of light is a brown dwarf.Here we report such a system, for which we observed one singular deep and short eclipse in October 2002, an event that is best explained by a sub-stellar/planetary companion on an eccentric orbit, but could in principle also be caused by circum-sub-stellar dust.
ROQUE 12
Roque 12 is a bona fide brown dwarf in the Pleiades star cluster, in the following called R12 (also named BPL172; α 03:48:19.0δ +24:25:13.0J2000.0).It was identified in several deep wide-field surveys of the Pleiades from colour-magnitude diagrams (Zapatero Osorio et al. 1997;Pinfield et al. 2000).To our knowledge, the only spectrum in the literature was published by Martín et al. (1998).The object has a spectral type of M7.5 and shows Hα emission (equivalent width 19.7 Å).No Li I absorption is found, with an upper limit of 1.5 Å.Compared to the Li I EW in other Pleiades brown dwarfs (0.5-2.5 Å), this does not firmly rule out the presence of Li I.The kinematics, colours, and Hα emission are all consistent with cluster membership (Martín et al. 1998;Pinfield et al. 2000;Stauffer et al. 2007;Lodieu et al. 2012).Judged by its position in the colormagnitude diagram, R12 is a single object with a mass of around 0.06 M .The brown dwarf has an entry in the Gaia Data Release 2 catalogue, with a parallax of 8.6±1.2 mas and proper motions of µ α = 18±3 mas yr −1 , µ δ = −43 ± 2 mas yr −1 .These values are again in line with what we expect for members of the Pleiades.Based on Gaia DR2, Gao (2019) derive a membership probability of 0.956.The object is also listed in AllWISE (Cutri et al. 2013), with magnitudes of 14.7 and 14.6 at 3.4 and 4.6 µm, respectively, that means, there is no evidence for warm circum-sub-stellar dust.
Calar Alto 1.23m
In October 2002 we monitored a field in the Pleiades covering R12 over three weeks, using the 1.23 m telescope at the German Spanish Astronomical Centre on Calar Alto.The telescope was equipped with a 2k×2k CCD covering 17 × 17 on the sky.All images were taken in the I-band with 600 sec exposure time.Observations were carried out on 15 nights between Oct 2 and Oct 18; several nights of the run were not usable due to unfavourable weather conditions.The main purpose of this observing run was to measure rotation periods for very low mass Pleiades members; the full details of analysis and results are reported in Scholz & Eislöffel (2004).Light curves were derived using the 'Optimal Image Subtraction' method (Alard & Lupton 1998) in an implementation developed by the Wendelstein Calar Alto Pixellensing Project (Riffeser et al. 2001;Gössl & Riffeser 2002).We achieved a typical photometric precision of 1% for stars with I = 17 mag and 2% for stars at I = 18.
The full lightcurve for R12 is shown in the left panel of Fig. 1, with a closeup view provided in the right panel.In the penultimate night of the observing run a deep eclipse in R12 is visible.The event is clearly detected in photometry from the reduced frames at t = 16.61 d (JD = 2452566.61d).It is also visible in the lightcurve derived from raw images.The eclipse reaches a depth of 0.65 mag compared with the average I-band magnitude over the entire campaign.The length of the event, which is covered by only four data points, is 1.3 ± 0.3 h.
Calar Alto 2.2m
During the last part of this campaign (Oct 17-19), we observed the fields discussed above also with the 2.2 m telescope on Calar Alto using the infrared camera MAGIC (Herbst et al. 1993).This camera in wide-field mode gives a field of view of 7 × 7 .We observed alternately in J-and H-band, with exposures times of 13 × 5 and 19 × 5 sec.This observing sequence covered half of the eclipse seen in the I-band data in the neighbouring 1.23 m telescope, but only in the J-band.A standard data reduction was carried out, using sky subtraction, bias correction, and flatfield correction.The lightcurve for R12 was derived using standard aperture photometry followed by differential correction using non-variable stars in the same field.The full details of this campaign are described in Scholz et al. (2005).The eclipse lightcurve from this run together with the close-up view of the I-band lightcurve is shown in Fig. 1, right panel.The eclipse depth in the J-band is the same as in the I-band, within the uncertainties of ±0.05 mag.
Kepler K2
Roque 12 was observed as part of campaign 4 of the Kepler K2 mission (Howell et al. 2014), under EPIC no.211090981, in GO program 4026 (PI: A. Scholz).Continuous observations with a standard cadence of 30 min were carried out from Feb 07 2015 to Apr 23 2015.The Kepler magnitude for R12 is 17.628, at the faint end of the magnitude distribution of K2 targets.We retrieved the lightcurve for R12 from MAST on September 4th 2015.The PDCSAP lightcurve after removing systematics (Twicken et al. 2010;Jenkins et al. 2010) is shown in Fig. 2, after dividing by the mean.
The lightcurve contains 3470 data points in total, 190 of them are > 1.2 and not shown in the figure .No data points are below 0.8, however.The standard deviation of the lightcurve without the positive outliers is 0.035, the average photometric error 0.024.Any repetition of the eclipse observed in 2002 would be clearly visible in the lightcurve as multiple negative outliers.Thus, if the eclipses are periodically recurring, the K2 lightcurve puts a definite lower limit of 70 d on the period.More recent versions of the same lightcurve, after correcting for systematics, do not give any different result.To rule out that the eclipse make the object too faint to be detected, we investigated the K2 lightcurve interactively with the Lightkurve tool (Lightkurve Collaboration et al. 2018).The source, however, is visible in all frames, including the ones with 'nan' data points in the lightcurve, without any evidence for anomalous behaviour.
Spectroscopy with VLT-UVES
In 2003-2004, a total of 13 epochs of high-resolution spectroscopy were obtained with UVES at the ESO-VLT, in the framework of ESO program 072.C-0071.In principle, this data set is useful in constraining the presence of companions through measurement of radial velocities (RV).The observations were carried out 6-11th October 2003 (4 epochs) and 13-28th January 2004 (9 epochs).These spectra span a wavelength range from 665 to 1042 nm with a nominal resolution of 34000.Only the red arm was used, with grism 'CD#4'.The integration time for each spectrum was 3000 sec.
We retrieved the extracted spectra for this run from the ESO archive in April 2020; the pipeline reduction was performed on 2017-10-26.The signal-to-noise ratio 2003.75 2003.80 2003.85 2003.90 2003.95 2004.00 2004.05 2004.10Epoch (yr) in the spectra is poor; it ranges from ∼ 1 at the blue end to ∼ 10 at the red end.As a result, only the strongest features are clearly detected and the constraints on RV are weak.
We measured RV for all epochs by fitting the Na doublet at 8183/8195 Å, the strongest absorption lines in the spectrum of R12, with a Gaussian.We adopted the average of the two measurements as radial velocity, and half of the difference between the two as uncertainty.After heliocentric correction, the resulting RV range from −3 to 10 kms −1 , with an average of 2.6 ± 2.3 kms −1 .For comparison, the RV of the Pleiades cluster is ∼ 6 kms −1 (Mermilliod et al. 2009).The average RV is −1.1 ± 2.3 kms −1 in October 2003 and 4.3 ± 2.4 kms −1 in January 2004.This might indicate long-term variations in the radial velocity with an amplitude of ∼ 5 kms −1 , but at this point the evidence is only tentative.We show the radial velocity measurements in Fig. 3.
Summary of observations
The unique event seen in October 2002 with two Calar Alto telescopes simultaneously remains to date the only eclipse observed in this object.The non-detection in the Kepler/K2 lightcurve limits any possible period to > 70 d.There is weak evidence for radial velocity vari-ations on timescales of months.We note that to our knowledge there is no additional indication of photometric variability in the literature or in public archives.
EXPLAINING THE ECLIPSE
The possible explanations for the singular eclipse fall into three broad categories: a sub-stellar or planetary companion to R12, a cloud in orbit around R12, an object between Earth and the Pleiades, but not associated with R12.These options will be discussed in the following.We do not claim that the discussed ideas exhaust the entire range of plausible explanations.
Companion
If the eclipse is caused by a companion, the characteristics of the eclipse give us constraints on its nature.The lack of a measurable colour change as well as the fact that R12 is positioned close to the single object sequence in Pleiades colour-magnitude diagrams (Lodieu et al. 2012), tells us that the second body contributes little to the flux ( 10%).Based on the 120 Myr DUSTY isochrone (Allard et al. 2001), this limits the secondary mass to 0.04 M , i.e. it cannot be an equal mass companion.The eclipse depth of 0.65 mag means that the eclipsing body, if it contributes no flux, has to cover almost half (45%) of the primary object.Taken together, the companion can only be a low-mass brown dwarf or a giant planet; both types of object would have sizes comparable to the primary brown dwarf.
For the observed eclipse duration of 1.3 h, the expected period for a circular Keplerian orbit is ∼ 15 days, with large uncertainty given the unknown mass of the companion and the poorly sampled eclipse.Periods shorter than that can be ruled out, based solely on the eclipse duration.As stated in Sect.3.3, periods < 70 d are unambiguously excluded by the K2 lightcurve.
For an eccentric orbit produce such a short eclipse in a system as described, the eccentricity has to be at least e ∼ 0.5.The required eccentricity is shown in Fig. 4, left panel, as a function of period for three assumed companion masses, derived from Kepler's laws.With object masses M 1 , M 2 , semi major axis a, the eclipse duration t, the object radius R, we plot the eccentricity e as: For this calculation we assumed that the eclipse happened at periastron, the time of the closest approach; if it happened at any other point in the orbit, the eccentricity would need to be larger.With a highly eccentric orbit, the period may be in the range of several years.The resulting minimum orbital separations are plotted in the right panel of the same figure, again as a function of period.Values of 120-180 times the radius of the brown dwarf are plausible, corresponding to 0.07-0.11AU.
In an eccentric Keplerian orbit, companions with masses of M B = 0.01 − 0.04 M would cause an orbital velocity in the primary of 6-15 kms −1 at periastron, and a few kms −1 at apastron; the latter will depend on the orbital period.These values are consistent with the tentative radial velocity variations reported in Sect.3.
Physically, the described system is plausible, but how unusual would it be?The fraction of unresolved brown dwarf-brown dwarf systems in the Pleiades is 27.1±5.8%(Lodieu et al. 2012).Giant planets around brown dwarfs have also been found, from direct imaging (Chauvin et al. 2004) and microlensing (Han et al. 2013;Jung et al. 2018).Thus, finding an eclipsing brown dwarf or giant planet companion around a brown dwarf is not implausible.
High eccentricities as needed to explain the singular eclipse for R12 have also been observed for similar and related systems.Dupuy & Liu (2011) investigated binaries among very low mass stars and brown dwarfs and find that the eccentricities span a broad range, including those with high eccentricities up to e ∼ 0.8.Their sample covers a very wide range in periods, from a few days to many years, including the plausible periods for the R12 system.
In a population level analysis of directly imaged substellar and planetary companions around stars, Bowler et al. (2020) find that the eccentricity distribution is flat between 5 and 100 AU separations.Brown dwarf companions exhibit a broad peak between e = 0.6 − 0.9.High eccentricities > 0.5 are not very common among the known giant planets in close orbit around stars, but they do exist in significant numbers (Winn & Fabrycky 2015).
If R12 had a low-mass companion, it would likely have avoided circularisation.For low-mass stars, the cut-off period for tidal circularisation of orbits at the age of the Pleiades is around 8 d (Zahn 2005).This value should depend weakly on the stellar mass/radius and is expected to be slightly larger for very low mass systems.The period in the putative R12 system, however, is sufficiently long to be above this threshold.
We conclude that an ultra-low mass companion on an eccentric orbit is a plausible explanation for the eclipse observed for R12.
Circum-sub-stellar clouds
In this scenario, we observed a dusty, circum-substellar, optically thin cloud passing in front of the brown dwarf.Due to the lack of reddening observed during the eclipse, any optically thin occulter has to be made of grains that are large in comparison with the wavelength, i.e. >> 1 µm, to cause grey extinction.Grains of that nature are readily formed in accretion disks and should also be common in debris disks (Testi et al. 2014).
The fast ingress and egress would point to a cloud moving at a high velocity of ∼ 50 kms −1 , to be able to travel the diameter of the brown dwarf within an hour.The uniqueness of the eclipse could be explained by material spiralling or falling into the central object.As shown in Scholz et al. (2019), only very little dust (<< M Earth ) is needed to cause substantial absorption along the line of sight.That means, this explanation does not necessarily conflict with the lack of infrared excess.A possible scenario to create such a cloud would be a something akin to a comet swarm on a highly eccentric orbit being destroyed as it approaches the central object.
There is a growing class of evolved stars showing eclipses or 'dips' in the lightcurve due to circumstellar material, although there is no IR excess indicating the presence of a debris or accretion disk.The most prominent one might be Boyajian's Star found in the Kepler data set (Boyajian et al. 2016), but there are other examples at younger ages (Stauffer et al. 2017), see also Scholz et al. (2019) for a discussion.A brown dwarf in σ Orionis might also belong into that category (Elliott et al. 2017).Among Pleiades members, a few examples of such disk-less 'dippers' have been identified (Rebull et al. 2016).R12 could be another enigmatic member of this family of objects, with one singular deep eclipse, instead of many shallow ones.We note that the radial velocity variations, if confirmed, cannot be explained by dust occultations.
Occulter unrelated to R12
This scenario can be sub-divided in two possibilities, depending on the location of the occulter.One option is that R12 is eclipsed by an object in the solar system.R12 is a point source and at the minimum of the eclipse the flux is still more than 50% of the light out of eclipse.Therefore, the occulter should have been visible in our images before and after the eclipse, which is not the case.Asteroid occultations are also very short (seconds for main belt asteroids or minutes for Kuiper belt objects), not comparable to the observed eclipse in R12.Therefore, this option can be excluded.
The alternative possibility is the presence of an occulter along the line of sight, for example, a dusty cloud.For typical ISM grains, the eclipse in the I-band should be about 1.7 times deeper than in the J-band (Mathis 1990), which is not observed.In theory, the occulter could be part of a disk with processed grains around a fast moving young star in the foreground.There is no evidence of such an object in our images or in the archives.The proximity of R12 as member of the Pleiades renders this explanation implausible.We refer to Wright & Sigurdsson (2016) for a more detailed discussion of this kind of scenario.
Instead of an interstellar dust cloud, the occulter could be a free-floating low-mass brown dwarf or giant planet passing through the line of sight.Such objects can be sufficiently red and faint to be undetectable in our optical images.From surveys of star forming regions (Scholz et al. 2012) and from microlensing studies (Mróz et al. 2017) we know that these type of objects are significantly less common than stars in the Milky Way, but still exist in substantial numbers.For an eclipse of this nature to happen, the free-floating planet or brown dwarf would have to be located in the cone-shaped volume towards R12.Typical space densities for this type of object are in the range of 0.1 pc −3 (Marocco et al. 2015).With this number, the volume towards R12 would contain 10 −16 free-floating planets or brown dwarfs.The available photometry covers around 1000 times the duration of the eclipse.Thus, the chances of finding an eclipse in that observing span are in the range of 10 −13 and thus negligible.We conclude that this explanation is not plausible.
SUMMARY AND OUTLOOK
We report the discovery of a deep, unique eclipse in the brown dwarf Roque 12, a bona fide member of the young cluster Pleiades with an age of 120 Myr.The eclipse was observed in 2002 with two telescopes simultaneously, is 0.6 mag deep and lasted 1.0-1.6 hours.The Kepler/K2 lightcurve from 2015 does not show any signs of further eclipses, ruling out that the event is periodic with periods shorter than 70 d.There is tentative evidence for radial velocity variations in this target of ∼ 5 kms −1 .
It is in principle conceivable that the eclipse was caused by a circum-sub-stellar dust cloud.Obscuration by objects along the line of sight, but unrelated to Roque 12 are considered highly unlikely.The most plausible explanation, however, is the presence of a sub-stellar companion on an eccentric orbit, with e > 0.5.A low mass brown dwarf or a giant planet as companion fit the observational constraints, with a mass ratio q < 0.7.The Roque 12 system could be one of very few known eclipsing binaries in the brown dwarf domain, the first with high eccentricity and long period.A system like that would be uniquely suited to test evolutionary models for sub-stellar objects.
What are the prospects of confirming the presence of a companion?Apart from improved radial velocity constraints (see Sect. 3), astrometry from Gaia may in the near future be able to provide more information.For the plausible system parameters (see Sect. 4.1) the semiminor axis times two corresponds to 0.5 to 4 mas in the plane of the sky.For the semi-major axis times two, it would be up to 8 mas.In Gaia DR2, the astrometric excess noise listed for R12 is nearly 4 mas, and thus already excludes parts of the possible orbital parameter space.
Further photometric monitoring would allow us to either fix the quantity n × P if a period is detected, or at least rule out a range of periods.This does require deep continuous observations at high cadence and is therefore not an easy task.R12 is too far north for the baseline survey of the Vera Rubin Observatory, but since it is only 4 deg away from the ecliptic, the object may still be covered regularly as part of surveys for solar system objects, albeit at high airmass.In a best case scenario with one eclipse every three months, the odds to catch the object in eclipse with a single visit are about 1/2000 or 0.05%.We encourage the astronomical community to help hunting for the putative second eclipse of Roque 12.
Fig. 1 .
Fig. 1.-Left panel: I-band lightcurve for R12 for Oct 2-20 2002, observed with the 1.23 m telescope on Calar Alto.The eclipse is clearly detected at t = 16.6 d.Right panel: I-and J-band lightcurve for the night Oct 18-19 2002, observed with the 1.23 m and 2.2 m telescopes on Calar Alto.Within the errors, the eclipse depth is the same in the two bands.
Fig. 3.-Radial velocities for Roque 12, measured with UVES/VLT.Shown are the individual epochs (blue) and the averages for the two groups of epochs (red).
Fig.4.-Left:Eccentricity as function of period for a 1.3 hr long eclipse of a 0.06 M brown dwarf by a 0.04 (black), 0.02 (red), 0.01 M (blue) companion, assuming an object radius of 0.13 R and the eclipse occurring at periastron.Right: Minimum distance between the two objects in units of the object radius (0.13 R ). | 5,877.2 | 2020-06-05T00:00:00.000 | [
"Physics"
] |
An On-Line and Adaptive Method for Detecting Abnormal Events in Videos Using Spatio-Temporal ConvNet
We address in this paper the problem of abnormal event detection in video-surveillance. In this context, we use only normal events as training samples. We propose to use a modified version of pretrained 3D residual convolutional network to extract spatio-temporal features, and we develop a robust classifier based on the selection of vectors of interest. It is able to learn the normal behavior model and detect potentially dangerous abnormal events. This unsupervised method prevents the marginalization of normal events that occur rarely during the training phase since it minimizes redundancy information, and adapt to the appearance of new normal events that occur during the testing phase. Experimental results on challenging datasets show the superiority of the proposed method compared to the state of the art in both frame-level and pixel-level in anomaly detection task.
Introduction
Abnormal event detection and localization is a challenging and exciting task in video monitoring.Indeed, the security context in recent years has led to the proliferation of surveillance cameras, which generate large amounts of data.This flow of CCTV images creates a lack of efficiency of human operators.Moreover, studies show that they can miss up to 60% of the target events when they monitor nine or more displays [1].In addition, after only 20 min of focus, the attention of most human operators decreases to well below acceptable levels [2].This can lead to potential security breaches, especially when monitoring crowded scene videos.
A possible solution to this problem is the development of automated video surveillance systems that can learn the normal behavior of a scene and detect any deviant event that may pose a security risk.
Abnormal event detection, also known as anomaly detection, can be defined as a spatial temporal recognition problem, taking into account that the event to be recognized is not present in the training phase.In the context of video surveillance systems, anomalies are formed by rare shapes, motions or their combinations.The main challenge in abnormal event detection is extracting robust descriptors and defining classification algorithms adapted to detect suspicious behaviors with the minimum values of false alarms, while ensuring a good rate of detection.The initial studies in abnormal event detection focused on trajectory analysis [3][4][5], where a moving object is considered as abnormal if its trajectory doesn't respect the fitted model during the training period.The main limits of such method are the sensitivity to occlusion and the effectiveness of detecting abnormal shapes with normal trajectories.These methods can be used in detecting deviant trajectories in scenes containing few objects but not achieve satisfactory performance for other applications.
Other methods such as low level local visual features [6][7][8] tried to overcome the limits of trajectory analysis and construct models by using handcrafted feature extractors.Among these methods, low local features such as histogram of oriented gradient (HOG), and histogram of optical flow (HOF) have been used to model the background and to construct the template behavior.However, these methods are usually specific to a given application and are not optimal for complex events.Besides, they don't link between local patterns, since local activity patterns of pixels is not efficient for behavior understanding.More complex methods used the concept of Bag of Video words (BOV) by extracting local video volumes obtained either by dense sampling or by selecting points of interest to construct the template behavior.However, the relationship between video volumes is often not taken into account.Derivatives of these methods [9] attempted to enhance the previous models by using not only the local region, but also the link between these regions for the overall understanding of the events.Usually the complexity of these methods makes them inefficient and time consuming for detection of abnormalities in crowded scenes.
Nowadays, deep learning has aroused the interest of the scientific community and works have been carried out in several fields including agriculture [10], biology [11] and economics [12].More specifically, deep learning based methods have demonstrated a high capacity on image processing [13,14], which led to the use of supervised methods for the anomaly detection.These methods are generally based on the use of convolutional neural networks (CNNs) [15].The main drawback of the supervised methods is the use of both normal and abnormal examples in the training phase which makes them not usable in real-world application for video surveillance, because it is very difficult to identify and reproduce all the abnormal events.
Other deep learning based works [16][17][18] have achieved good performance on anomaly detection datasets.These methods use unsupervised learning and their learning process do not require normal and abnormal training examples, which makes them suitable for abnormal event detection.In this article, we propose an unsupervised on-line adaptive method based on deep learning for the detection and localization of abnormal events.The main contributions of this paper are as follows:
•
We adapt a pretrained 3D CNN to extract robust feature maps related to shapes and motions which allow us to detect and localize complex abnormal events in non-crowded and crowded scenes.
•
We propose a new method of outliers detection based on the selection of vectors of interest to construct a balanced distribution.This robust classifier is able to represent all normal events (redundant and rare ones) during the training phase, detects abnormalities and adapt to the appearance of new normal events during the testing phase.
The rest of this paper is presented as follows: In Section 2, we present a brief state of the existing methods.In Section 3, we detail our proposed detection method.We present experimental results to evaluate our method in Section 4. Finally, Section 5 concludes the paper.
Related work
In this section, a review of the current main methods is presented.Initial studies gave attention to the use of the trajectory for anomaly detection [3][4][5][19][20][21].Calderara in [19] represented trajectories as sequences of transitions between nodes in a graph to detect abnormal movements.In another work, Morris [20] combined the trajectory with Gaussian mixture and hidden Markov models to build activity models.Antonakaki [21] proposed an approximation algorithm implementing a hidden Markov model as one-class classifier, which can decide whether a given trajectory is normal compared to a model.
Others works [6][7][8]22,23] were devoted to the extraction and analysis of low-level local features.Ermis in [6] generated behavior clusters based on behavior profile of a given pixel to construct probabilistic models.Reddy in [7] used optical flow to characterize the object's motion and proposed an adaptive codebook to obtain texture features.In the same process, Wang [8] combined a histogram of optical flow orientation with one class support vector machine (SVM) for detecting abnormal motions.Boiman and Irani [22] presented an approach based on spatio-temporal volumes and considered that if the event reconstruction using only the previous observations is impossible then this event is classified as abnormal.Roshtkhari [9] proposed a method which is the improvement of [22], by using a codebook to group similar spatio-temporal volumes to reduce the dimension of the problem.This proposed method allows to reduce the computational complexity, and becomes applicable to real-time video analysis.Xiao in [23] proposed a sparse semi-non-negative matrix factorization (SSMF) to learn local patterns, and used a histogram of non-negative coefficients (HNC) as local descriptors.Li et al. in [24] used a dynamic texture (DTs) to design models of normalcy over both space and time.Spatial and temporal information is modeled with a mixture of DTs (MDT).
Recently, various methods based on deep learning were applied for anomaly detection [15][16][17][25][26][27].Zhou [15] proposed a method for detecting and locating abnormalities based on spatiotemporal convolutional neural network.In this paper, the authors used a 3D-CNN to extract features related to motions and shapes.Although the results presented were encouraging, the supervised nature of this algorithm did not allow its use in real world systems monitoring.Indeed, both normal and abnormal events were used as training examples, which is not suitable in the field of video surveillance since it is very hard to reproduce all abnormalities.Sabokrou et al. in [26] used three initial layers of pre-trained 2D-CNN combined with a new convolutional layer, whose kernels were learned using sparse auto-encoder for abnormal event detection.One should notice here that the 2D convolution does not sufficiently take into account the temporal information.Same author in [27] used local and global descriptors generated with sparse auto-encoder to capture video properties.Hasan in [16] used a fully convolutional auto-encoder (AE), the learned AE reconstructs normal events with low error and give higher reconstruction error for abnormal events.For the same purpose, Chong [25] proposed an AE with two parts, 2D convolutional layers for spatial features and use convolutional long short term memory for temporal information.Ravanbakhsh in [17] proposed to use a generative adversarial nets (GANs), which are trained using normal frames and their corresponding representation generated by optical flow.In the testing phase, the real data are compared with the reconstruction generated by GANs, the use of optical flow in the detection of abnormalities can be time consuming.Xu et al. in [28] presented an unsupervised deep learning framework for abnormal event detection in complex scenes.The proposed method is based on stacked denoising autoencoders (SDAE) to extract features, and multiples one-class-SVM to predict score of abnormalities.Thus, they used a fusion strategy for the final decision.Ravanbakhsh et al. in [29] combined a fully convolutional network (FCN) with a binary quantization layer that was trained by an iterative quantization (ITQ) for obtaining binary bit maps.Using this technique, the authors could get a set of binary codes for every video volume.At the end, a histogram representing the distribution of binary codes was generated.
Methods for video features extraction were developed recently [30][31][32] and have achieved promising results in action recognition.The advantage of these methods lies in their ability to learn spatiotemporal information.Despite the fact that these methods are mostly supervised, it can be adapted effectively for abnormal event detection task.
Our combination of pretrained 3D-CNN and the proposed classifier is an improvement of our previous work [18] where we introduced a method for abnormalities detection based on a 2D-CNN combined with a one-class SVM [33,34].Although the results generated were acceptable, the motion information is reduced by the 2D convolution.Moreover, the setting of an SVM and more specifically the choice of the kernel is complex, it is also admitted that the efficiency of an SVM is strongly impacted by the size of the data [35].For these reasons, we propose in this paper a 3D-CNN combined with a new classifier to overcome these limits: From one side, the 3D-CNN is more suitable for spatiotemporal abnormal event detection.From another side, the new classifier is robust to the evolution of the training dataset size, it also takes into account the normal events that occur rarely during the training period and can adaptively learn newly observed normal events in the testing phase.Thus, we obtain a method directly usable for real-world applications.To sum up, Table 1 summarizes the main methods presented in this section, their advantages and drawbacks.
The Proposed Framework
We present in this section the proposed abnormal event detection method.We start by giving a brief description of the 3D residual convolutional networks and we explain the benefits of using these networks comparatively with 2D plain networks.After this, we give details of the two main stages of our method; feature extraction using modified version of a pre-trained 3D residual convolutional network [31], and the detection of outliers using the proposed classifier.Thus, this method allows us to detect abnormal events, prevents the marginalization of rare normal events and adapt to the appearance of new normal events during the testing phase.
3D Residual Convolutional Networks
The 2D-CNNs perform convolution and pooling operations only spatially, while 3D-CNN is used for spatio-temporal representations.Features from 3D convolutional networks encapsulate information related to shapes and motions in sequence of images, which can be useful for abnormal event detection.Besides, Tran et al. in [30] proved that their C3D features are separated remarkably compared to 2D based model, which can be desirable for videos representations.Figure 1 shows how it can be effective to perform clustering with features generated by C3D than those generated by ImageNet [36] on UCF101 dataset.Feature embedding visualizations of Imagenet and C3D on UCF101 dateset using t-SNE method [37], this figure was extracted from [30].
The 3D-CNNs proposed in [30][31][32] proved to be highly competitive compared to different 2D networks in various video analysis tasks, such as scene, object or action recognition.
The use of pretrained 2D convolutional network such as features extractors for abnormal event detection was recently developed by [18,26].However, it was proven in [30], that image-based feature extraction is not suitable for action recognition, since the 2D convolution does not consider temporal relationship between successive frames.For instance, in [26], a sequence of frames was used as an input of the CNN.However, the convolution kernels in the 2D-CNN does not allow consideration of temporal information.i.e., after the first convolutional layer, the movement information is destroyed by the 2D convolution.That is why the use of multiple frames as an input of the 2D neural networks is not adequate for temporal features extraction.
Furthermore, when training a network with a very large number of layers, the gradient required to update the weight with backpropagation becomes smaller when reaching earlier layers.This is the so-called vanishing gradient problem.As a result, as the network goes deeper, its performance gets saturated or even starts degrading rapidly.That is why He et al. in [14] proposed a novel architecture based on residual blocks (see Figure 2).The original output mapping F(x) is readjusted into F(x) + x.It is thus easy to optimize the residual mapping than the first one.This network named "ResNet" is formed of 152 layers.It has obtained the first place in the Large Scale Visual Recognition Challenge (ILSVRC) 2015 challenge with a classification error of 3.6%.
Anomaly Detection
The 3D CNN used in [31] is a residual network pre-trained on Sports-1M dataset containing more than 1.1 million sports videos.This allows it to extract high performed spatio-temporal features from video sequences compared to 2D neural networks used in previous papers [18].In the following, we present the proposed methodology for abnormal event detection.It is based on two main stages: Features extraction with modified version of pre-trained Res3D [31] and outliers detection with our classifier.
Features Extraction
The proposed methodology consists of using volumes of three consecutive frames.For each frame "F t " we used the volume F: {F t ; F t−1 ; F t−2 } as input.As applied in [18], we did not use the complete CNN "from end to end", but we used the four initials 3D residual blocks (16 convolutional layers) as feature extractors.The architecture of this network is presented in Figure 3. Thus, instead of getting an output of one feature vector, the resultant map is a matrix of 841 feature vectors of 256 dimensions.In this matrix, each raw represents a small patch in the input frame.This is explicitly shown in Figure 4.
Classifier
CCTV videos are characterized by a redundancy of information in normal scenes.Building the event model by using all vectors obtained for each frame during the training phase would skew the distribution by weighting redundant elements and marginalizing rare ones that would lead to misinterpretations and generate confusion between rare events and abnormalities in the detection process.That is why we proposed in this paper an online selection of vectors of interest.It allows definition of a balanced distribution able to represent all normal events, including rare ones, since the representation of redundant and rare events reaches an equilibrium.Thus, all normal events have the same importance during the detection process so that false alarms would be avoided.
In this way, we were able to construct a robust classifier, able to minimize redundancy, and avoid the marginalization of rare events and reduce false alarms.
Algorithm 1 shows the construction of the balanced distribution.Thus, to select our representative vectors, we declare "N" first vectors from the first frame as vectors of interest.Then, for each new vector "X i " of the processed frame, we calculate a string metric "dist i " based on Mahalanobis distance between "X(i)" and the "N" vectors of the distribution.This is represented in the following equation: where "moy" and "Q" are the mean and the inverse of the covariance matrix of the distribution, respectively.The Mahalanobis distance considers the correlation between data variables since it is calculated using the inverse of the covariance matrix [38].Using this distance, the probability that a test point belongs to a set is characterized not by its euclidean distance but by a metric tacking into account the distribution of the data.
In our case, if this distance exceeds a threshold "α", the vector is selected to be a vector of interest, then we update "moy" and "Q".Otherwise, the vector is not considered for the detection of abnormalities.We then continue the same process for each new arrival frame in the training phase.We thus assess the relevance of each vector to be included in the balanced distribution.Finally, this distribution is pruned to eliminate among the first "N" elements those that are redundant, which ultimately generates "M" different vectors of interest that are considered as the representation of video volumes.This method allows us to compute our algorithm in real time.
Furthermore, this method represents each redundant normal event by a single vector in order to have an equilibrium in point of interest data.It then prevents the marginalization of normal events that occur rarely during the training process.Thus, every event has the same importance in abnormal detection procedure.
For the testing phase, each patch, represented by a vector obtained by the feature extraction mentioned above is assessed using the same principle.Indeed, we measure the similarity between the vector representing the patch and the balanced distribution.If the measure exceeds a threshold "β", the vector is considered as an outlier.Thus, the considered patch is labeled as abnormal.The detailed procedure is presented in Algorithm 1.All the thresholds (α, β and η) were selected empirically.While "α" and "η" were selected manually according to their effects on the size of the balanced distribution, "β" was selected in order to obtain the best results (minimum of false alarms and misdetections) in the monitored scenes.
CCTV also faces another challenge due to the constant evolution of the environment that can be characterized as a dynamic background, appearance of new interactions or temporary authorization of abnormal events.These changes can generate false alarms that disturb the interaction with the users of the anomaly detection algorithm.
For instance, a single pathway that is changed to a normal road or the introduction (or removal) of an element in the background.Another example can be the authorization of temporary interactions in an environment such as maintenance works that may create false alarms due to the use of special machines and uncommon tools.
To the best of our knowledge, this problem was skipped in previous algorithms.That is why we proposed an adaptive method robust to any evolution of the surveilled scene.
As seen in Figure 5, this method allows to reconsider false alarms obtained during the test process, i.e., if a patch is declared as a false alarm, the operator can declare its feature vector as a vector of interest to be included in the balanced distribution.This will avoid its future detection as abnormal.Furthermore, as any other methods of abnormal event detection, misdetections could also occur while using the proposed framework.That is why we propose as a next step its improvement into a semi-supervised model.Indeed, when the model generates misdetection, the CCTV operator can declare it as abnormal.i.e., the feature vector representing this event will be used to create a second distribution representing misdetected events.Thus, one can get two distributions representing normal and misdetected events.Consequently, to labelize a new event as normal, its feature vector should respect two conditions: The similarity between this new feature vector and the original distribution representing normal events should be established, and we should get a non-similarity with the new distribution representing the misdetected events.Thus, this operation will decrease misdetections and increase the performance of our proposed method.The implementation of this semi-supervised model will be developed in future work.
Results and Comparisons
In order to implement our methodology, we used Matlab and caffe for the deep learning part [36].We also implement our algorithm in C++ for real time detection.
We tested our methodology on UCSD Ped2 Anomaly Detection Dataset and we apply it also for a real case of laboratory surveillance "CapSec dataset".The local dataset "CapSec" was constructed in our laboratory using a fixed camera.It represents daily behavior of students and researchers inside the laboratory.The image resolution is 1280 × 720 pixels.It has training folders that contain only normal events and testing folders which also contains anomalies like falling people and objects that are not present in the training phase.One can see in Figure 6 some qualitative results.UCSD Ped2 "(http://www.svcl.ucsd.edu/projects/anomaly)"represents complex behaviors including pedestrians.It contains 16 and 12 folders for training and testing phases resepctively.The training folders contain only normal events of pedestrians.However, the test folders also contain abnormalities such as skaters, bikes and cars.Abnormal events in this dataset vary in type, size and occurrence.It contains many occlusions and the image resolution is 240 × 360, which is low and can complicate the detection process.
While using this dataset, we evaluate two scenarios: "SC1" when we consider a balanced distribution for every folder of the Ped2 dataset, and another scenario "SC2" when we develop a distribution for the whole dataset.The objective of this experiment is to prove the robustness of the proposed method when increasing the number of images during the training phase.
The first results show the effectiveness of the proposed method.The qualitative results are presented in Figure 7 when all abnormalities are detected.
One can see in Table 2 how the introduction of vectors of interest reduces the number of feature vector representing normal events.Consequently, while using the first scenario of the proposed method, we apply a classifier (one classifier per folder) with less than one percent of features vectors selected as vector of interests.In the second scenario, when we consider a unique classifier for all folders containing 1920 training images, we get only 1569 vectors of interest.This number is relatively stable and low even if the training images number increases.This shows how the proposed methodology is robust when facing high number of frames.Consequently, it can be applied for real-world applications of video surveillance.
To assess the performance of the proposed method and provide quantitative results, we used the Equal Error Rate of Frame Level (EERFL), the Equal Error Rate of Pixel Level (EERPL) and the Receiving Operating Characteristic (ROC).A comparison with state of the art methods is provided in Table 3.Our method gets respectively an EERFL and EERPL of 6.25% and 9.82% for the first scenario "SC1", and for the second scenario "SC2" 7.45% and 9.63% respectively, which outperforms all other state of the art methods.
Figure 8 shows the receiver operating characteristic (ROC) for the UCSD Ped2 dataset , plotted as a function of the detection threshold for both scenarios. .Also, as discussed above about including false alarms (FA).One can see in Table 4 how the EER is decreasing considerably when the operator declares false alarms.In Folder 4, the EER decreases by 1.1% when the operator includes only one false alarm.Besides, in Folder 7 when the operator declares 5 false alarms, the EER decreases by 5.6%.
When dealing with the computational time, the comparison of our method with other state of the art methods is presented in Table 5, the results are only indicative and should be treated with caution.This is mainly due to the different hardware performances of each work.In our case, we used I7 CPU with 32 Gb of RAM and graphic card NVIDIA Quadro 2000 M. The proposed computational time presented in Table 5 is not optimal and does not reflect the full potential of our method.The main objective was to minimize the error and to propose a robust method for real-world application.
Conclusions
In this paper, a novel online adaptive method based on combination of pretrained 3D residual network and online classifier were developed and implemented.It is able to detect abnormal events, prevent the marginalization of normal behavior that rarely occurs during the training phase and adapt to the appearance of new normal events in the testing phase.In addition, our method does not require pretreatment methods such as tracking or background subtraction.This method is based on two main stages: Spatiotemporal feature extraction without any need of training, and the use of robust incremental classifier that prevents the redundancy of information in CCTV.It can also either be used online or offline.
We have tested our proposed methodology on two main datasets using crowded (Ped2) and non-crowded scenes (CapSec).
The results from the Ped2 dataset showed high performance in detection and localization for abnormal events based on The EERFL and EERPL.To the best of our knowledge, this method outperforms all existing techniques present in the literature and used for Ped2.
Besides, the fastness and the simplicity of this method allow us to use it for real-world application (case of CapSec).
The results presented in this paper showed the effectiveness of using this framework in detecting abnormal events.This method is robust, takes into account rare normal events present in the training phase.Besides, it can be incorporated in online CCTV.Moreover, the method can be adapted so that human operators select false alarms to prevent its future appearance, which is suitable for dynamic environment.
In this method, the localization of abnormal events is reflected as patches in the original image.In some cases, these patches may overflow on normal regions.In future work, we will focus on the detection at the pixel level to precisely localize the abnormal regions.Future studies will also investigate the test of this method on other datasets and will improve our local dataset (CapSec) in order to generate quantitative results and use it as a regular dataset for testing abnormal event detection methods.Moreover, we will also compare our classifier with other classification methods (k-nearest neighbors (k-NN) [47], and enhanced k-NN [48]).
Figure 4 .
Figure 4. Global scheme of the proposed framework.
Figure 6 .
Figure 6.Detection of abnormal events.(a) Falling person; (b) Detection of multi-falling people; (c) Falling person in presence of walking one; (d) Detection of an object assumed as abnormal; (e) Falling person in presence of crouched person; (f) Falling person.
Figure 7 .
Figure 7. Detection of abnormal events in Ped2 Dataset.(a) Multiple bicycles detected; (b) Detection of multiple targets (bicycle and car); (c) Detection of multiple targets (bicycle and skater); (d) Detection of multiple targets (bicycles and skater); (e) Multiple bicycles detected; (f) Detection of multiple targets (bicycle partially obstructed and one wheel).
Table 1 .
Comparison of some methods of abnormal events detection in terms of advantages and disadvantages.
Table 2 .
Effect of the selection of vectors of interest on element reduction.For each folder of the dataset, one can see the number of frames (NB-Frames), number of feature vectors (NB-FV) and number of vector of interest obtained by our method (NB-VI).
Table 3 .
Equal Error Rates for frame and pixel level comparisons on Ped2 Dataset.The EER values for the different methods were extracted from the literature.Receiver operating characteristic (ROC) curve of the proposed method.The blue and green curves are for the first and second scenarios, respectively.The Area Under the Curve (AUC) values for the two scenarios (SC1 and SC2) are 86% and 84%, respectively
Table 4 .
Adaptation to false alarm detection.
Table 5 .
Brief information about computational time of the proposed method (given in seconds per frame). | 6,386.8 | 2019-02-21T00:00:00.000 | [
"Computer Science"
] |
3D and 2D aromatic units behave like oil and water in the case of benzocarborane derivatives
A large number of 2D/2D and 3D/3D aromatic fusions that keep their aromaticity in the fused compounds have been synthesized. In addition, we have previously proven the electronic relationship between the 3D aromaticity of boron hydrides and the 2D aromaticity of PAHs. Here we report the possible existence of 3D/2D aromatic fusions that retain the whole aromaticity of the two units. Our conclusion is that such a 3D/2D aromatic combination is not possible due to the ineffective overlap between the π-MOs of the planar species and the n + 1 molecular orbitals in the aromatic cage that deter an effective electronic delocalization between the two fused units. We have also proven the necessary conditions for 3D/3D fusions to take place, and how aromaticity of each unit is decreased in 2D/2D and 3D/3D fusions.
T he two-dimensional (2D) aromaticity concept is widely accepted in all areas of science, and its most representative example is benzene. In 1978, Aihara 1 introduced the concept of 3D aromaticity, with the closo dodecaborate anion, [B 12 H 12 ] 2− , as its maximum exponent 2 . Very recently, the concept of 3D aromaticity has been revisited. It has been established that a 3D aromatic compound must have a closed-shell electronic structure with at least triply degenerate molecular orbitals (MOs), extensive electron delocalization, and similar magnetic and electronic properties in the three xyz directions 3 .
In 2014, an electronic relationship between 2D and 3D aromaticities was reported [4][5][6][7] . Conceptually, the 2D aromaticity can be easily extended with the fusion of two 2D aromatic units by sharing one edge. An example is the fusion of two benzenes sharing a common edge that gives rise to a new aromatic compound, naphthalene, and that following the same approach of sharing edges leads to polycyclic aromatic hydrocarbons (PAHs) 8 such as anthracene, pentacene, perylene, coronene, all the way to graphene, the wonder material of the twenty-first century 9,10 . PAHs are a class of chemicals that are found naturally in coal, crude oil, and gasoline 11 , but they are even found when burning any carbonaceous material 12,13 , or even when cooking meat and other foods at high temperature 14 , or in the interstellar space 15 . The ease with which they are produced is a consequence of the great stability that they have, largely due to the resonance property 16 .
The high stability of 2D aromatic compounds is what favors their great diversity and the fact that they are found in key molecules for life such as hemoglobin, chlorophyll, and DNA bases. Indeed, there is an estimate that, according to PubChem, about two-thirds out of its 110 million structures of chemical compounds are fully or partially aromatic 17 .
Just as there are so many experimental examples of PAHs resulting from the fusion of two 2D aromatic species, which we call 2D/2D for short, showing that the result is another aromatic 2D species, the question is: is the result of a fusion of two 3D/3D aromatic species another aromatic 3D species? And in line to the former one, we also raised the issue of whether it is possible that the fusion of two aromatic species, one 3D and the second 2D, a 3D/2D case, could result in a globally 3D/2D aromatic species.
In this study, we will work as much as possible either with existing molecules or with their derivatives to provide a more consistent basis for the conclusions reached. Contrary to the 2D/ 2D fusion, there are very few examples available to demonstrate that the fusion of two 3D/3D aromatic entities leads to a molecule with 3D aromaticity. The reason for this is the lack of synthetic methods transferable from one process to another. However, two experimental examples have been found that will prove how the 3D/3D fusion results in a molecule with 3D aromaticity. These studies are based on boron hydride clusters, which are then expanded to hypothetical molecules with the same kind of structure based on existing 3D entities. It is interesting to note that in all known examples of fusions, whether 2D/2D or 3D/3D, the connecting atoms from the building blocks always had two outer electrons from a covalent bond, e.g., C-H or B-H. With this in mind, we have also explored units from polyanions derivatives of Zintl phase compounds 18,19 , e.g., [Sn 12 ] 2− , and fuse them together. The fusion results stress the importance that the connecting atom has one or two exocluster electrons, e.g. C in 2D aromatics or B in 3D aromatics have one electron, whereas Sn has two.
We are aware of the recent synthesis of several fully π-conjugated macrocycles with strongly puckered or cage-type structures that have been considered 3D aromatic 20,21 . However, some authors consider that they are not truly 3D aromatic and rather they should be labeled as 2D-aromatic-in-3D 3 . Therefore, they cannot be used to analyze potential fused systems with 3D/ 3D aromaticity.
Regarding the question whether the fusion of two entities, a 3D aromatic entity and a 2D aromatic entity, the 3D/2D case, results in a molecule with 3D/2D aromaticity, we will rely on experimental cases which have led to the fusion and demonstrate that in these circumstances the 3D entity retains its aromaticity but the 2D entity loses it. These examples have subsequently been extended to similar but hypothetical molecules that have corroborated these conclusions.
Results
3D/3D fusions. In 2014, we published that the electronic equivalent of benzene in boron hydrides was not [B 12 H 12 ] 2− but [B 7 H 7 ] 2− , i.e., both C 6 H 6 and [B 7 H 7 ] 2− share the same number of valence electrons 5 . Therefore, when fusing two benzene building blocks to give a naphthalene, the electronic equivalent in boron hydrides was [B 12 H 10 ] 2− , where two closo-B 7 units share an edge. This compound has not been synthesized but would most likely be stable if there were a pathway for its synthesis, as evidenced by its aromaticity, evaluated through the magnetic-based aromaticity criterion nucleus-independent chemical shift 4-6,22,23 (NICS-it is considered that the more negative the NICS value of a ring or cage, the higher its aromaticity- Fig. 1). Nonetheless, not in all cases more aromatic is related to more stable, as recently also proven by ourselves in related compounds 23 . If we stick only to the fusion of two [B 7 H 7 ] 2− units, there would be another closo alternative, which would be to share a face instead of an edge leading to [B 11 H 8 ] − . The pioneering work of Jemmis and other scientists 24 on a unified rule for predicting and systematizing structures of macropolyhedral boranes has really helped us in our analysis. As can be seen from the NICS shown in Fig. 1, this second option would give rise to a species with a slightly more accentuated aromaticity than the previous one. [B 11 H 8 ] − has neither been synthesized, but its higher aromaticity seems to suggest that sharing a face provides more stability than sharing an edge. It is worth noting that [B 11 H 8 ] − with m = 2 and n = 11 and [B 12 H 10 ] 2with m = 2 and n = 12 follow Jemmis' mno rule of stability for fused boranes. According to this rule, the number of skeletal electron pairs required for a condensed polyhedral borane, carborane, heteroborane, metallaborane, or metallocene cluster to be aromatic is given by m + n + o, where m = number of sub-clusters, n = number of vertices, and o = number of singlevertex shared condensations 24,25 . 30 , who analyzed their isomerization energies and aromaticity. Here, we have focused only on the isomers in which the C atoms are located on the sharing triangular face derived from either [CB 11 H 12 ] − or C 2 B 10 H 12 30 . We recently reported that the negative NICS values in [B 12 H 12 ] 2− and C 2 B 10 H 12 are the result of intense diatropic ring currents inside the cage 23 .
Finally, sandwich metallabis(dicarbollide) compounds (e.g., [Co(C 2 B 9 H 11 ) 2 ] − ) can be considered the result of the fusion between two nido-[C 2 B 9 H 12 ] − cages through a transition metal. In a recent work 23 Remarkable are the connecting elements in fused aromatic hydrocarbons and aromatic closo boron hydrides, e.g., C 2 in naphthalene, (CH) 4 C 2 (CH) 4 NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-31267-7 ARTICLE 3D aromatic systems, we focus on merging components whose polyhedron is the icosahedron and the bicapped square antiprism. We have remarked that the connecting elements have no substituents outside the ring or cluster. Extending this idea to the elements with lone pairs means that the connecting elements have no electron pairs outside the cluster. In the case of the made [Sn 12 ] 2− (see ref. 31 (Fig. 4), and that can be considered the equivalent to [B 17 H 14 ] − and its fusing [B 10 H 10 ] 2− unit, respectively (Fig. 2), sharing not only the same number of valence electrons but also their aromatic character. Remarkable is that the fused entities display considerably lower NICS values than their corresponding building blocks. Concern could here arise about the absolute interpretation of the NICS values for such heavy atoms like Sn, although it has been recently shown by Foroutan-Nejad et al. that the use of NICS is especially problematic when involving metals with half-filled shells, causing strong local paramagnetic currents. This is not our case, therefore its use in our closed-shell Sn clusters is most probably legitimate 33,34 . However, despite the relatively low negative NICS values found for fused Sn clusters as compared to those of borane clusters, the number of delocalized electrons per B or Sn atom obtained using QTAIM theory 35,36 of the series of Sn clusters and that of [B 10 H 10 ] 2− is very similar (Supplementary Fig. 2), thus supporting the aromatic character of both species. Moreover, there is a trend observed in all the reported fusions that is consistent with the aromaticity of fused moieties. The latter have always less negative NICS values than the building blocks. We attribute this reduction in the absolute value of NICS, first, to the intrinsic reduction in the aromaticity due to the fusion and, second, to the coupling between the magnetic fields of the two building blocks in the fused moieties. We can see from the above discussed NICS that the aromaticity of the pairs [B 21 To this point, we can state that, for certain types of fusion and electronic charge of the cluster, the fusion of two aromatic halves of light elements results in another aromatic fused species and this stability seems to decrease with increasing atomic weight and the existence of lone pairs in the fusing atom which entail more positive charges in the fusion product, a factor which seems to be destabilizing. On this basis, we have begun with the most advantageous possible situation; we have merged equal halves in shape, size, and composition. 3D/2D fusions. And then, what happens if we fuse a 3D aromatic unit with a 2D unit, in which one unit is larger than the second one? To make this situation more propitious, we will make both starting halves visible in the final product. For this purpose, 1,2-C 2 B 10 H 12 will be used, in which there are two contiguous carbons that can participate in the formation of a 6-carbon ring, a "benzene". Before studying this case, we would like to recall that [B 12 H 12 ] 2− represents the archetype of 3D aromaticity, as stated above, and that any compound that can be formally derived from it must, in principle, be less or at the most equally aromatic than the reference compound. It is remarkable that the extraordinary stability of [B 12 H 12 ] 2− forces 1,2-C 2 B 10 H 12 , albeit incorporating two carbon atoms, to adapt to the geometry and aromaticity of [B 12 H 12 ] 2− (see ref. 23 ). Therefore, the C···C distance in neutral derivatives of 1,2-C 2 B 10 H 12 that spans from 1.64 to 1.82 Å 38 , is far from the usual distances in organic compounds approaching the B···B distance in [B 12 H 12 ] 2− that is~1.8 Å. Consequently, due to the unequal C-C distance there will be a conflict of stabilities between one aromatic entity~1.4 Å in benzene and~1.7 Å in 1,2-C 2 B 10 H 12 . As a result, it may be that the C···C distance reaches an intermediate distance while maintaining a global aromaticity or that only one of the two entities maintains aromaticity, and in this case which of the two? In principle, we could assume that the one that has more to lose is going to be the winner.
The optimal example to be studied is the formal fusion of 1,2-C 2 B 10 H 12 and benzene. Fortunately, the compound was previously synthesized and is known as the benzocarborane for which different synthetic methods have been described based on either o-carborane or carboryne as starting reagents 21,39,40 . Benzocarborane can also be synthesized from a transition metal-catalyzed [2 + 2 + 2] cycloaddition 41 of carboryne and two acetylene molecules (Fig. 6a) 42 . The name benzocarborane implies that there is a benzene (aromatic) attached to a carborane whose aromatic condition was not indicated. The situation, as we shall see, is quite the opposite. Therefore, benzocarborane and its derivatives (Fig. 6) and its isomers (Fig. 7) are very adequate to address the question whether the 3D/2D aromatic fusion is feasible. Not too many alternatives exist that would facilitate this study. If the 2D fragment is made of carbons, this requires or at least makes it more likely, that the 3D fragment also contains adjacent carbon atoms in the structure, and that both are aromatic. In this context, o-carborane, 1,2-C 2 B 10 H 12 is key to study the 3D/2D fusion 43 . In difference, fullerenes may look appealing but either do not have 3D spherical aromaticity 44 or, reminding the issue of the connecting elements described earlier, how would it be possible to fuse with another aromatic half without sacrificing (at least partially) the aromaticity of the fullerene? As can be inferred from the above, only two aromatic halves with an outer electron pair in the system, either a C-H or a lone pair, can be fused. It is worth noting that we are not discussing the aromaticity of a phenyl group linked to a carborane, [PhCB 11 H 11 ] − , as reported recently by Muñoz-Castro (in this case both moieties remain aromatic) 45 but of benzo-o-carborane with the carborane cage and benzene ring fused. Also importantly, Xie and coworkers succeeded in the synthesis and X-ray characterization of substituted carboranes, and proved that there is considerable bond length alternation in the benzene ring (164.8-133.8 pm) ruling out aromaticity 42,46 . Similar alternation bond lengths were proven by Wade and coworkers on pristine benzo-o-carborane (165.1-133.8 pm) 47 , in agreement with the ethyl derivative.
NICS values very close to zero or to positive values indicate non-aromaticity, thus the computed data concerning the 2D fragment in Fig. 7 confirms the Matteson and Hota's title of their article about benzocarborane, "High stability but little aromatic character" 40 . The NICS value of the 2D fragment is in significant contrast to that of o-carborane, which is very negative, indicating that just as the 2D fragment is not very aromatic, the 3D fragment is strongly aromatic. Therefore, of the two fragments, one retains aromaticity while the other clearly loses it. The four benzocarborane isomers studied and presented in Fig. 7 show the same trend, but it is worth noting that the one with the C 4 B 2 group occupying para positions relative to the two carbons of the carborane (benzene BBp ) has the most negative NICS values. These results point out that the fusion of two halves, each one known to be clearly aromatic, both suitable for the 3D/2D investigation does not lead to a global aromatic compound, but instead one of the two halves loses its aromaticity while the second retains its aromaticity. In the benzocarborane, the carborane keeps its aromaticity, with NICS only slightly reduced (from −33.3 to −33.0 ppm in the center of B 4 C five-membered ring (5-MR) and from −27.3 to −26.8 ppm in the center of cluster), however, the fused benzene becomes nonaromatic (NICS reduced from −8.1 to −1.5 ppm) as expected from the fact that the 6-MR contains only four π-electrons. The same trend is observed when benzene is fused to bonds other than C-C (Fig. 7). And the same behavior applies to larger PAHs: naphthalene, anthracene, and phenanthrene, when fused to o-carborane (Fig. 8). The only ring that gives an opposite behavior is the terminal one of phenanthrene. The reason is the presence of a Clar π-sextet 48,49 on this ring (Fig. 6b), at difference to either naphthalene or anthracene that cannot accommodate such π-sextet (bond lengths also support this observation, Supplementary Fig. 1). This is the reason why only this ring remains fully aromatic.
The difficulty for this 3D/2D fusion can be assigned to the lack of overlap between the π-MOs of the PAH and those of the carborane. A model system built from the system benzene CC above, analyzed through a quantitative molecular orbital and energy decomposition analyses, has allowed to confirm the ineffective overlap (Supplementary Discussion). Finally, let us mention that the above conclusions on 3D/2D fusion also apply to the fused carborane and five-membered heterocycles previously synthesized in ref. 50 .
Discussion
One of the fundamental characteristics of Hückel's 2D aromaticity is the large number of aromatic species derived from the fusion of two aromatic halves. We have named these as 2D/2D aromatic fusions. Prior to this work, few 3D/3D aromatic fusions had been explicitly described 29 , although some examples had been synthesized and have been demonstrated in this work to be global aromatics. It has therefore been shown that such fusions are possible by the union of two 3D aromatic halves. The difficulty or impossibility for 3D/2D fusions is also proven in this work. The main reason for non-having a 3D/2D aromatic species is the ineffective overlap between the π-MOs of the PAH and the n + 1 MOs in the cage that deter a higher electronic delocalization between the two fused units.
Furthermore, it is proven that for 3D/3D fusions to take place, some conditions have to be fulfilled. The first and obvious one is that each of the halves to be merged must be aromatic. The halves to be fused must have a great similarity in the two-connecting atom distances, although in borane clusters, 3D units of different size favor fusion 37 . The participating elements in the fusion must have exocluster bonds, e.g., C-H, B-H in the starting half or lone pairs. These connecting elements will be distinct in the fused molecule from other participating elements and will not have exocluster bonds. To our understanding, these conditions appear to be necessary for fusions of aromatic species to give rise to a global aromatic, but may not be sufficient.
It seems that it can be generalized that aromaticity decays with fusion, this is observed in all 2D/2D, 3D/3D, and 3D/2D cases, being in the latter case dramatic as the 2D aromaticity vanishes completely. It is worth noting that the aromaticity decays comparably from 2D to 2D/2D as from 3D to 3D/3D. Fusion is more feasible with elements of the second period, C and B, and probably as one moves down the periodic table the aromaticity decreases. This has been seen with Sn in period 5, although it is in the same group as C. To be a connecting element, the atom in question must have an exocluster or exoring electron pair, e.g., C-H, B-H or Sn. The existence of lone pairs, however, leads to cationic charged species which can be potentially destabilizing.
Methods
All calculations were performed with the Gaussian 09 package 51 by means of the B3LYP 52-54 hybrid density functional and the 6-311 + +G(d,p) basis set 55 . The geometry optimizations were carried out without symmetry constraints (Supplementary Data 1). Analytical Hessians were computed to characterize the optimized structures as minima (zero imaginary frequencies). Sn cluster have been computed by means of AMS 2021 software at ZORA-B3LYP-D3BJ/TZ2P level 56,57 . Aromaticity was evaluated by means of the nucleus-independent chemical shift (NICS) [4][5][6]22,23 , proposed by Schleyer and coworkers as a magnetic descriptor of aromaticity. NICS is defined as the negative value of the absolute shielding computed at a ring center or at some other point of the system. Rings with large negative NICS values are considered aromatic. NICS values were computed using the gauge-including atomic orbital method (GIAO) 58 . Multicenter indices (MCI) [59][60][61] and delocalization indices (DI) 35,36 were computed with the ESI-3D program using AIM partition of space 62,63 . For completeness, the aromaticity of the enclosed systems has been further confirmed by means of bond length alternation (BLA) measures (Supplementary Table 3). | 5,023.2 | 2022-07-04T00:00:00.000 | [
"Chemistry"
] |
Superconductivity, Superfluidity and Holography
This is a concise review of holographic superconductors and superfluids. We highlight some predictions of the holographic models and the emphasis is given to physical aspects rather than to the technical details, although some references to understand the latter are systematically provided. We include gapped systems in the discussion, motivated by the physics of high-temperature superconductivity. In order to do so we consider a compactified extra dimension (with radius R), or, alternatively, a dilatonic field. The first setup can also be used to model cylindrical superconductors; when these are probed by an axial magnetic field a universal property of holography emerges: while for large R (compared to the other scales in the problem) non-local operators are suppressed, leading to the so called Little-Parks periodicity, the opposite limit shows non-local effects, e.g. the uplifting of the Little-Parks periodicity. This difference corresponds in the gravity side to a Hawking-Page phase transition.
Introduction: motivations, background material and conclusions
Holographic dualities have become useful tools to explore strongly coupled theories; in particular, in the last few years systems at finite temperature T and density of a U(1) charge have been extensively studied. The main motivation comes from condensed matter physics (for a review see [1]): if the U(1) charge is the electric charge and the U(1) symmetry is spontaneously broken one identifies these systems with superconductors (SCs); when the U(1) symmetry is global, instead, the spontaneously broken phase corresponds to superfluidity.
One of the main ideas behind the application of holography to superconductivity is the hope to shed light on the physics of unconventional SCs (such as high-T SCs), which are not completely described by the weakly coupled theory: the Bardeen-Cooper-Schrieffer theory. Cuprate high-T SCs, the prototype of unconventional materials, are obtained by doping a Mott insulator and as such they typically show in the (T versus "dopant concentration") phase diagram an insulator phase close to the region where the U(1) symmetry is spontaneously broken [2]. This calls for a way to describe both conductor/SC and insulator/SC transitions. Since any insulating material eventually conducts if it is probed by a strong enough external current, this in turn requires a method to study gapped phases.
In this article we will briefly review holographic SCs [3,4] and superfluids (SFs) [5,4]. The emphasis will be given to physical aspects rather than to the technical details, for which we will refer to the original papers. Given that all SF systems can be considered as SCs in the limit of non-dynamical electromagnetic (EM) gauge fields, for the sake of definiteness, we will use the terminology adopted in the literature of superconductivity, rather than that of superfluidity (unless otherwise stated). We will focus on the best understood version of holographic duality, the anti de Sitter/conformal field theory (AdS/CFT) correspondence. The discussion of the previous paragraph then motivates us to introduce a conformal symmetry breaking in the infrared (IR), which is needed to have a gap. Moreover, since much of the interesting physics of cuprate SCs is layered, we shall restrict our analysis to 2+1 dimensional systems, dual to gravitational theories on four dimensional spacetimes. Although beyond the scope of this review, it is useful to mention that other regions of cuprate phase diagrams require a departure from the theory of Fermi liquids, a challenging property which can be realized in holography [6].
In the rest of this section we shall introduce the concepts and general properties of superconductivity which are needed to understand the holographic results discussed in the following sections, providing at the same time the reader with the organization of the paper and the conclusions. We assume the minimal field content to describe superconductivity: a charged scalar Φ cl , responsible for the breaking of the U(1) symmetry and a gauge field a µ . The quantum effective action is a generic functional of gauge invariant operators where we have introduced the field strength F µν = ∂ µ a ν − ∂ ν a µ and ψ cl = |Φ cl |. The first term in the expression above represents the contribution of local gauge invariant terms. Γ n.l. depends instead on non-local gauge invariant objects, which may be required if the material has a non-trivial topology. Although the action in (1) is very general, its form simplifies in some limits. This allows us to extract model independent properties and, at the same time, identify the predictions of specific models, such as the holographic ones that will be discussed here.
One of these limits is the small field limit, in which Γ is approximated by the Ginzburg-Landau (GL) action: where g 0 , ξ GL and b GL are real and positive parameters and we have rescaled the scalar field, Φ GL ∝ Φ cl , to have a canonically normalized kinetic term for Φ GL . Also This theory holds if the system is close enough to the symmetry breaking transition, where Φ cl is small. Another case in which the action simplifies is the limit of slowly varying fields, such that Γ can be approximated by a two-derivative functional where h , g and W are unspecified functions of ψ cl . When (3) is a good approximation the SF limit can be taken as g → 0: this corresponds to taking the gauge field non-dynamical and thus the corresponding U(1) symmetry global. Deep inside a uniform SC, where ψ is close to the minimum ψ ∞ of the potential V = hW and the massive gauge field a µ −∂ µ arg(Φ cl ) is suppressed, (3) reduces to a functional that is quadratic in the fields; this allows us to define the inverse masses of ψ cl and a µ , respectively ξ and λ: 1 Note that ξ and λ cannot be computed in a model-independent way as their values strongly depend on the form of h , g and W . The distinctive property of the superconductive phase is ψ ∞ = 0, as opposed to the normal phase in which ψ ∞ = 0. Starting from the general action in (1) it is possible to show in very general terms some of the most famous properties of SCs, such as the Meissner effect, the infinite DC conductivity and the Josephson effect [7], which have been obtained in holography in Refs. [4,8,9,3,10].
In this article (unless otherwise stated) we will focus on static configurations, in which everything is constant in time and the temporal component of the gauge field is identified with the chemical potential of the U(1) charge: a 0 = µ. Also, in order to study the thermodynamics of these systems we introduce the free energy F = −T Γ.
In section 3 we will holographically study the normal phase and explain how to describe conductors and insulators. The homogeneous phase, in which ψ cl = ψ ∞ , will be studied in section 4 through the AdS/CFT correspondence. As we will see, this provides a rationale for the fact that superconductivity is suppressed at large rather than at small T .
However, inhomogeneous configurations are also of great importance in SCs; e.g. they generically occur when the system is probed by an external EM field. Among the most famous properties of SCs we can certainly include the existence of vortex solutions, which are indeed strongly inhomogeneous configurations. The simplest example of such configurations are straight vortex lines, which can be described by an ansatz of the form Ψ cl = ψ cl (r)e inφ and a φ = a φ (r) and the other components of the vector potential set to zero. Here r, φ are the usual polar coordinates which parametrize the Euclidean two dimensional space. By using the action in (3) one obtains the following large r field behavior [4] for SCs, while for SFs, where the magnetic field B = ∂ r a φ /r is frozen to an external constant value, a φ = Br 2 /2, the condensate approaches the homogeneous value with a power law ψ cl − ψ ∞ ∼ n 2 /r 2 [4]. In Eq. (5) ψ 1 and a 1 are unspecified constants, while the action in (3) implies that λ = λ and ξ = ξ. However, the exponential behavior in Eq. (5) tells us that higher derivative terms cannot be neglected as they are generically of the same order of the twoderivative terms. Fortunately, the only effect of the higher derivatives is to modify the values of λ and ξ , such that we generically have λ = λ and ξ = ξ, as discussed in [4]. λ and ξ are respectively called penetration depth and coherence length and are important to characterize SCs. Although the field behavior in (5) is model independent, λ and ξ , like λ and ξ, can only be fixed once the model is specified. Thus their actual values are predictions of the particular model one considers; in section 5 we will discuss how to extract them from holography. While vortex solutions exist for any SC, it is not always the case that there is a range of the external magnetic field H such that the vortex phase is energetically favorable; when this is true the SC is of Type II and such range is denoted with H c1 ≤ H ≤ H c2 (if such range does not exist we have instead a Type I SC). H c2 is the value of the external magnetic field above which the system is always in the normal phase. The vortex configuration for Type II SCs and H slightly smaller than H c2 is known to be a triangular lattice of vortices independently on the specific model one considers: for those high magnetic fields the condensate is small and the GL theory can be applied to predict that configuration [11]. H c1 is instead the value of H below which the system is in the homogeneous superconducting phase. It can be computed through the model independent formula 1 (see e.g. [4]) where F 1 and F 0 are the free energies for the n = 1 vortex and n = 0 superconducting phase respectively. We note that the actual value of H c1 is a prediction of the specific model one 1 Here we use the normalization of H such that the external current Jext that generates H through ∇× H = g 2 0 Jext is coupled to the gauge field by means of the interaction term aµJ µ ext in the Lagrangian. considers because g 0 , F 1 and F 0 are model dependent. However, in the SF limit, g 0 → 0, we always have H c1 → 0; in other words H c1 is non-trivial only if the magnetic field is dynamical.
In section 5 we will study the vortex phase in holography and illustrate how to compute the critical magnetic fields and show that the holographic SC is of type II. We will emphasize in particular the differences between the case in which there are no sources of conformal symmetry breaking, other than T and µ, and the case in which conformal symmetry is broken.
Up to now we have not discussed the role of the non-local terms in Eq. (1). For reasons which will be clear soon, we would like to spend the rest of this introductory section to describe the simplest setup in which Γ n.l. is relevant: cylindrical SCs threaded by an external magnetic field along the symmetry axis of the cylinder. To render the discussion even simpler we will take the deep Type II limit, g 0 → 0, in which the (total) magnetic field coincides with the external one. Since there is a non-contractible loop on this geometry we can construct non-local gauge invariant objects: where e is the charge of the fundamental charge carriers (say the electrons for real world SCs) and the integrals are performed along the non-contractible loop. We will refer to W and m as the Wilson line and the fluxoid respectively and the dots represent other non-local objects. While the classical action does not depend on (W, m, . . .), quantum effects, such as the Aharonov-Bohm one (or Sagnac one [12,13] in the SF literature), could introduce a dependence of Γ on these quantities. A simple (but still general) way to visualize a dependence of this sort is to think that the coefficients of the local part of Γ vary with (W, m, . . .): for example, in the domain of validity of the GL theory in (2), this corresponds to thinking of ξ GL and b GL as functions of (W, m, . . .).
Such dependence ought to be suppressed in the classical limit and/or when the typical scale of the non-contractible loops, R, is large compared to the other scales in the problem (such as 1/T and 1/µ) and therefore we should stay away from these limits to see any interest effect. Moreover, since quantum corrections are small in a weakly coupled theory, the biggest effects of the non-local quantities are expected in strongly coupled theories. This justifies the use of holography in this setup (see section 6). Before moving to holography, however, let us identify model-independent effects. We shall consider the simplest case in which the magnetic field is constant so that it can be represented by a constant vector potential along χ, which parametrizes the compact spatial dimension, χ ∼ χ + 2πR. Since everything is static and homogeneous an appropriate ansatz is Φ cl = ψ cl e imχ/R , a = µdt + a χ dχ with ψ cl and a χ constant .
Inserting this ansatz in the quantum effective action, Eq. (1), we obtain in other words, modulo non-local terms, the system must be periodic in the magnetic flux 2 Φ(B) ≡ dx µ a µ /g 0 with period ∆Φ LP ≡ 2π/g 0 , because Φ(B) → Φ(B) + ∆Φ LP can be compensated by a unit shift of the integer m. This is known as the Little-Parks effect [14]. This phenomenon has been observed in experiments, which give g 0 = 2e, and thus it is considered as an evidence for Cooper pairing. If the non-local terms in (9) are non-negligible the Little-Parks does not generically occur; indeed nothing forces Γ n.l to be a function of the local combination m/R−a χ only, but it may depend on m and W separately. The conclusion is that if the system is 2 The factor 1/g0 has been introduced in order to have a canonically normalized kinetic term for aµ in Eq. (2). strongly coupled, far away from the classical limit and R is small enough we could see an uplifting of the Little-Parks periodicity to an enhanced periodicity, set by the fundamental charge: 2π/e. This would corresponds to resolving the internal structure of the composite condensing operator. Whether this really occurs is not model-independent.
Remarkably, we will see (in section 6) that holography predicts an uplifting of the Little-Parks periodicity for small values of R and that this uplifting corresponds in the gravity side to the celebrated Hawking-Page phase transition [15].
The holographic model
The first step to define a holographic model based on the standard AdS/CFT correspondence is to introduce fields living on an asymptotically AdS space. The minimal field content required to describe SCs holographically is a charged scalar field Ψ dual to a condensing operator O ( O = Φ cl ) and a gauge field A α dual to the electric current operator J µ . As we have stated in the introduction, another ingredient we are interested in is a mechanism for conformal symmetry breaking in the IR; this can be achieved by introducing a real scalar φ (a dilaton) which acquires a non-trivial VEV. The class of actions we consider is [9] where R is the Ricci scalar, G N the Newton constant, V the dilaton potential, g 4 the U(1) gauge coupling and L is the radius of AdS. Also, Z A and Z ψ are generic functions of the dilaton, which do not fulfill particular properties, besides the fact that they never vanish (in order for the semiclassical approximation in the bulk to be justified). Also, D α Ψ = (∂ α − iA α )Ψ. We are interested in describing a static system with two dimensional rotation and translation invariance, for which the metric and the dilaton have the form where dy 2 is the two dimensional Euclidean metric. The temperature can be introduced by requiring the presence of a black hole; so there is a value of the holographic coordinate, z 0 , such that f (z 0 ) = 0. Then T = |f (z 0 )|/4π, where the prime denotes the derivative with respect to z. Another physical situation that can be described with this setup is the case of multiply connected SCs, more precisely of cylindrical geometry: this can be simply achieved by compactifying one of the two y-coordinates: χ ∼ χ + 2πR. Then there is necessarily another set of solutions [16]: those obtained from the black holes by exchanging the Euclidean time t E = it with χ; these configurations have no event horizons and are commonly called AdS solitons. One reason the compactification of χ is interesting is because this procedure allows us to introduce a scale R (other than the T and µ) and thus breaks conformal invariance even in the absence of the dilaton. For this reason from now on the case in which χ is compact will be discussed with a simplified field content, where φ is removed and Z A = Z ψ = 1. The AdS soliton (black hole) geometry is energetically favorable with respect to the black hole (AdS soliton) at sufficiently small (large) temperatures, T ≤ 1/(2πR) (T ≥ 1/(2πR)). Another reason why we are interested in compactifying a spatial coordinate is to study what happens when cylindrical SCs are threaded by magnetic fields: in these physical setups indeed a universal prediction of holography emerges, as it will be discussed in section 6 and anticipated in the introduction.
Although we include the possibility of conformal symmetry breaking in the IR, in this article we will always assume that the ultraviolet (UV) is conformally invariant, or in other words that the fields are approximated by an AdS configuration close to a value of z, say z = 0: for z close to zero, W (z) L 2 /z 2 , f (z) 1 and φ(z) 0. Thus we can use the standard AdS/CFT dictionary, which relates the properties of a gravitational theory with those of a CFT. In particular the z = 0 values of Ψ and A µ are the sources of O and J µ in the CFT. Therefore, if one solves the bulk field equations with boundary conditions the Green's function for O and J µ are given by differentiating the on-shell action with respect to Ψ 0 and a µ ; e.g. the vacuum expectation values are given by In order to keep the discussion as simple as possible we assume that Ψ and A α do not backreact on the geometry and the dilaton; this can be consistently achieved by taking the limit G N → 0. A generalization to the case G N = 0 can be found in [17,18]. A general form of asymptotically AdS dilatonic black hole solutions with the assumed symmetries was derived in [19]. One important aspect of these black holes is that the dilaton is not generically constant but runs with the energy: where ν is a real parameter of the potential such that ν ≥ 1 and ν = 1 recovers the Schwarzschild black hole (S-BH) of Einstein's gravity. Eq. (14) tells us that although the theory has a UV fixed point, φ(0) = 0, conformal invariance is broken and maximally in the IR, z = z 0 . In order for the configurations of Ref. [19] to be a solution, V has to be appropriately chosen.
In particular the requirement of an asymptotically AdS configuration space with cosmological constant Λ implies that at the conformal point (conventionally φ = 0) where V (0) = Λ we have ∂V /∂φ = 0. As a result, for each value of ν, even ν = 1, the S-BH with cosmological constant Λ is always a solution. One finds [9] that the S-BH is favorable at high temperatures, while the dilaton-BH, Eq. (14), dominates in the low temperature region.
Conductivity
Having described the solutions of the dilatonic gravity systems we now want to understand the type of materials they correspond to. Some information can be gained by studying the conductivity, σ. This will also shed light on the nature of the SF phase transitions, which will be discussed in the next section.
To compute σ let us consider, on top of these geometries, a small plane wave along a spatial coordinate x, which is induced by a small electromagnetic field, A x (t, z)| z=0 = a x (t). Here A and p are real functions of z. The system responds creating a current which is linear in the electric field E x : that is J x = σE x . Using the AdS/CFT dictionary, the first equation in (13), we have When an event horizon is present regularity of the solution implies that the plane wave should be ingoing rather than outgoing from the horizon. It can be shown that this results in a non-vanishing Re[σ] and DC conductivity [9] lim ω→0 For an AdS soliton, which has no event horizon, regularity allows for a vanishing DC conductivity, which has been identified with an insulating behavior [20,18]. Therefore, compactifying a spatial dimension χ allows us to realize a conductor/insulator transition as the temperature is lowered. Eq. (17), however, tells us that there is another way to suppress the DC conductivity and obtain such transition if a dilaton is present in the spectrum: if one chooses Z A sufficiently small for large values of φ, the running in Eq. (14) implies that lim ω→0 Re[σ] is small, especially in the low temperature limit where z 0 is large. In this setup the transition (discussed in the previous section) between the S-BH and dilaton-BH, which is obtained by lowering the temperature, represents a conductor/insulator transition. It is worth noting that insulating systems corresponds to solids in the fluid mechanical interpretation [8].
Superfluid phase transition
We now move on and study the simplest example of superconducting state: a static, homogeneous and isotropic material in which the U(1) symmetry is spontaneously broken. The ansatz is The reader may wonder why we have introduced the temporal component of the gauge field, because having a non-vanishing ψ(z) seems already enough to describe the state we are interested in. The reason is that no regular solution is found when A 0 = 0 in (18), as shown in [9], and we want to exclude singular solutions. Physically this corresponds to the fact that there is only one parameter breaking conformal symmetry (T for the black hole and R for the AdS soliton) in the absence of the dilaton 3 , thus no phase transition can occur in this case. Therefore we set the UV boundary conditions Ψ 0 = 0 , a 0 = µ .
The first condition guarantees that the U(1) symmetry is spontaneously broken, while the second one, with a non-vanishing chemical potential µ, keeps the profile of A 0 different from zero and allows us to find a regular solution.
An immediate consequence of this argument is that when the temperature is large compared to the chemical potential a black hole suppresses the superconductivity. At small temperatures instead the system turns into a SC [3]. In AdS/CFT there is therefore a reason for the fact that a metal becomes a SC at small rather than at large temperatures. For the AdS soliton the same conclusion can be reached, but with T substituted by 1/R.
According to the results of the previous section these SF phase transitions can be either of the conductor/SC or insulator/SC type, depending on the behavior of the conductivity in the normal phase: the two cases correspond to a non-vanishing or vanishing DC conductivity respectively. In the fluid mechanical interpretation, the SF phase of a holographic system with a solid normal phase has been identified with a supersolid [21].
Dynamical gauge fields in AdS/CFT and superconductivity
In the holographic results we have discussed so far the dynamics of the EM field is not important. Therefore, according to the discussion presented in the introduction, they can be applied equally well to SFs and SCs. In this section instead we discuss some important effects of superconductivity, which crucially rely on the dynamics of the EM field.
However, we note that imposing the first boundary condition in (12) treats a µ as an external source, which does not participate in the dynamics of the system. In order to have a dynamical a µ we should integrate over all possible field configurations: (working in the Euclidean space) where S E [a µ ] is the Euclidean bulk action computed on a solution of the field equations with boundary condition in (12). Also, we have introduced for generality a kinetic term for a µ and an external current J ext . Then Z[J ext ] defines as usual the generating functional for the Green functions of a µ . Of course, if there are other operators in the theory, besides the gauge field, S E [a µ ] in Eq. (20) will depend on the corresponding sources as well; for example in the case of superconductivity discussed in the previous sections S E [a µ ] also depends on Ψ 0 , which we introduced in (12).
In the semiclassical limit this procedure reduces to solving the Maxwell equations, where we have used Ĵ µ = −δS E /δa µ . The semiclassical limit corresponds to taking a large external current, g b → 0 and a limit on the parameters of the bulk theory such that S E [a µ ] becomes large; for example, for the theory defined in (10) this limit is G N → 0 and g 4 → 0. Now, using Eq. (13), we can see that the Maxwell equations in (21) can viewed as a Neumann type boundary condition in the gravity side: The lesson is therefore that switching from the Dirichlet boundary condition (12) to the condition above promotes a µ to a dynamical field. As far as the holographic SC model of section 2 is concerned, this procedure has been applied to find the Meissner effect and genuine SC vortices [4,9], at least in the simplest case of a straight vortex line: Ψ = ψ(z, r)e inφ , A 0 = A 0 (z, r) , A φ = A φ (z, r) .
r, φ are the usual polar coordinate parametrizing the two dimensional Euclidean space and so here we assume that there are at least two non-compact dimensions in the CFT. This is not the case for the four dimensional AdS soliton; however, vortex solutions have been found on top of the five dimensional AdS soliton in [8]. Using the ansatz in (23) the Neumann-like boundary condition in (22) becomes while the requirement of spontaneous symmetry breaking and the presence of a finite charge density again fixes the other z = 0 boundary conditions, Eq. (19). Regularity of the solutions instead fixes the conditions at z = z 0 and at the center of the vortex r = 0 and physical conditions on the behavior at infinity, r → ∞, imposes constraints on the remaining boundary. This results in genuine SC profiles for the total magnetic field B = ∂ r a φ /r: for example for n = 0 one recovers the Meissner effect, while for n = 0 one observes the exponential damping of B far away from the center of the vortex [4,8,9], Eq. (5), allowing for a holographic prediction for λ and ξ . Such properties should be contrasted with the Dirichlet boundary condition for the (unsuppressed) in the S-BH (AdS soliton) phase. Correspondingly, according to the modelindependent discussion of section 1, the system should be characterized by the Little-Parks periodicity for the S BH and an uplifting of such periodicity for the AdS soliton as shown in [23,8]. | 6,651.4 | 2013-01-02T00:00:00.000 | [
"Physics"
] |
Who Will Watch the Watchmen? The Ethico-political Arrangements of Algorithmic Proctoring for Academic Integrity
Critics of artificial intelligence have suggested that the principles of fairness, accountability and transparency (FATE) have been used for ‘ethics washing’, in order to appease industrial interests. In this article, we develop this relational and context-dependent analysis, arguing that ethics should not be understood as abstract values or design decisions, but as socio-technical achievements, enacted in the practices of students, teachers and corporations. We propose that the ethics of using AI in education are political, involving the distribution of power, privilege and resources. To illustrate this, we trace the controversies that followed from an incident in which a student was misclassified as a cheat by an online proctoring platform during the Covid-19 lockdown, analysing this incident to reveal the socio-technical arrangements of academic integrity. We then show how Joan Tronto’s work on the ethics of care can help think about the politics of these socio-technical arrangements — that is, about historically constituted power relations and the delegation of responsibilities within these institutions. The paper concludes by setting the immediate need for restorative justice against the slower temporality of systemic failure, and inviting speculation that could create new relationships between universities, students, businesses, algorithms and the idea of academic integrity.
A university student stands accused of academic dishonesty during the Covid-19 lockdown. She makes a TikTok video, describing how algorithmic proctoring technology flagged her for cheating during a remotely administered exam. In it, she laments how her professor assigned her a failing grade and reported her to school officials. Days later, we learn that her university's appeal process was reassuringly swift. The professor apologized, the student's grade was reinstated -and yet this 38-second clip of the sobbing student went viral, garnering over 3 million views on Tik-Tok and 1 million views on Twitter in just 2 weeks. Even if justice prevailed, there was something in this student's distress that continued to resonate with many other uneasy university students, and with the instructors who believed that it did not have to be this way, not during a protracted global health emergency.
In this paper, we stay with this residual moment of unease for what it might tell us about the ethics of engaging with artificial intelligence (AI) in higher education -not as an abstract matter of ethical codes or institutional processes, but as a political matter enacted in the practices of students, teachers and corporations. We ask, irrespective of whether justice was ultimately served, might there be better ways of being with algorithms in the university? Building on the feminist speculative ethics articulated by Puig de la Bellacasa (2011Bellacasa ( , 2017, we then show how Tronto's landmark work on the ethics of care (Tronto 1993) can be deployed to think about the politics of these socio-technical arrangements -that is, about the historically constituted power relations and delegation of responsibilities within educational institutions. We conclude that those involved with monitoring academic integrity must attend to the temporalities of how ethics and justice are co-constituted through the entangled everyday practices of the university workplace.
The Ethics of AI and Education: Working with FATE
There has been considerable debate about how AI might be governed to limit the harms that have been associated with their use. Larsson (2020), for example, has reviewed the role of guidelines in ensuring ethical and trustworthy AI, noting how principles such as fairness, accountability and transparency have been put forward to address ethical concerns (i.e. the FATE deliberations). These are sometimes combined with other principles, such as explicability, non-harmful use, responsibility and integrity. In different combinations, these principles form the basis for debate and research around the creation and use of AI, for example by acting as a point of reference for conferences and meetings (e.g. Friedler and Wilson 2018), and have been found to serve as an important foundation for building users' trust (Shin 2020). Larsson goes on to observe, however, that 'there are considerable differences in how these principles are interpreted; why they are considered important; what issue, domain or actors they relate to; and how they should be implemented' (Larsson 2020: 442). He suggests that discussions of these principles tend to ignore questions of power and infrastructure and notes the growing concern that principles are being used as a form of ethics washing to appease industrial interests. As Greene et al. (2019) observe, corporations may concede the need for 'better building', but the idea that AI might be opposed, refused or simply not created on ethical grounds seems to be 'off the table '. In the realm of educational technology, researchers have begun to evaluate the ethics of online proctoring systems, making reference to FATE principles (e.g. Coghlan et al. 2021). However these 'in principle' assessments of AI in education arguably take corporate accounts of their services at face value, glossing over notable power asymmetries amongst the stakeholders who engage with such ethical principles (cf. Davies et al. 2021). In a study of remote proctoring during the Covid-19 campus shutdowns, Selwyn et al. raise questions about 'the surrender of control to commercial providers, the hidden labour required to sustain "automated" systems and the increased vulnerabilities of "remote" studying' (Selwyn et al. 2021: 1). Resonating with Larsson's call to attend to the 'societal challenges of relating to fairness, accountability, and transparency' (Larsson 2020: 439), a small but growing body of critical research has engaged with the ethics of AI not just as an ethics of data and computing, but as an ethics of education which addresses, among other issues, 'power relations between teachers and their students, and of particular approaches to pedagogy' (Holmes et al. 2021: 18). In their work on AI for educational inclusion and diversity, Porayska-Pomsta and Rajendran argue that the principle of accountability should be understood as relational and context-dependent, involving 'priorities and investments of different stakeholders, along with temporal fashions that determine who is accountable for decisions and actions to whom, with respect to what' (Porayska-Pomsta and Rajendran 2019: 43). They suggest that a relational approach to accountability makes for a more 'agile' and 'concrete' ethical intervention -'an exchange and an ethically regulated, tractable and auditable compromise between different competing interests and gains of the decision-makers' (ibid.).
Drawing from this emerging critical research on the ethics of AI in education and from feminist scholarship, this paper will therefore engage with the ethics of online proctoring as an empirical and relational practice. Rather than formulating principles of 'ethical AI' or determining whether platforms conform with FATE-related guidelines, we treat principles such as 'fairness', 'accountability' and 'transparency' as objects of research, tracing how these values circulate when sociotechnical relations are reassembled and the balance of power shifts. In doing so, we seek to open up questions about the purposes, the values and the kinds of relationships that are privileged in 'the digital university' (Jones 2013), inviting reflection on the plurality of sociomaterial practices that constitutes the 'world' of the academy, to locate ourselves within this world, and to begin to speculate about the kinds of new relations that are needed to enact a different narrative about digital technology (Ross 2017). We align our analysis with Feenberg's long-standing call for a philosophy of technology where ethical values are analysed with and through constructivist accounts about technology design and use, so that rather than simply either accepting or rejecting artefacts, we might instead develop new insights into how technology might enact 'such things as reskilled work, medical practices that respect the person, architectural and urban designs that create humane living spaces, computer designs that mediate new social forms' (Feenberg 1999: 199).
Relationality and Power: the Speculative Ethics of Care
We take as a point of departure Puig de la Bellacasa's 'speculative ethics ' (2011, 2017), which puts constructivist accounts of science, technology and society (STS) into play with Joan Tronto's political theory on the ethics of care (Tronto 1993). This posthuman feminist approach was chosen for its (1) focus on technoscience, (2) attention to empirical practice and the materiality of 'ethical doings', (3) engagement with affective and asymmetric relationalities, (4) attunement to the co-constitution of ethics and politics and (5) commitment to generating a 'speculative' critique which offers possible alternatives for living a 'good' life. Rather than a moral stance or invocation for motherly love, 'care' is thus deployed as 'an analytic or provocation' (Puig de la Bellacasa 2017: 7) to explore how ethics and politics are co-constituted in the case of a distraught university student's encounter with AI.
Contrary to those who have interpreted Tronto's care ethics as normative (e.g. Pols 2015), Puig de la Bellacasa asserts that Tronto's work offers 'a vision of caring [that] presupposes heterogeneity as the ontological ground on which everything humans relate with exists … Its ontological import gives to care the peculiar significance of being a nonnormative necessity' (Puig de la Bellacasa 2017: 70). Weaving Tronto's care ethics together with constructivist accounts of technoscience, Puig de la Bellacasa's approach has led to a lineage of feminist scholarship in science and technology studies (STS) which foregrounds the 'ethico-political' practices of care in technoscience in a wide range of settings (Lindén and Lydahl 2021). These 'Critical Care' studies of technology design and use illustrate how ethics often operate in tension with justice -how ethical aspirations to 'do good' are entangled in politically charged care practices that are 'ambivalent, contextual, and relational' (Martin et al. 2015: 631) and can be fraught with histories of sexism, racism, capitalism and colonialism (Murphy 2015).
A Methodological Framework for Studying the Ethics and Politics of AI
Our analysis of the ethics and politics of AI builds specifically on Tronto's idea that care is both an affective disposition and a politically charged activity linking people, artefacts and the environment into four inter-related registers of practice (Tronto 1993: 105-8): • Caring about: noticing an unmet need for care, often by assuming the position of another person or group. As an act of identification, it is culturally and individually shaped and enacts an ethical element of attentiveness. • Taking care of: taking responsibility to make certain that the needs are met. This might involve determining how to respond, requires a sense of agency and enacts the ethical element of responsibility. • Care-giving: directly meeting the needs of care. This requires physical work, usually involves contact and enacts the ethical element of competence.
• Care-receiving: the care-receiver responds to care-giving, enabling the care-giver to observe the response and make judgments about the sufficiency and success of caring work. It enacts the ethical element of responsiveness.
Understanding care in terms of these four value-laden, overlapping registers of practices makes it possible to analyse how distributions of power, privilege and resources lead to inadequate care in society, raising not only questions of '"For whom?" but also "Who cares?" "What for?" "Why do 'we' care?" and mostly "How to care?"' (Puig de la Bellacasa 2017: 61). In this way, both ethics and justice are understood as sociomaterial achievements, co-constituted through the situated interplay of four, far-flung registers of care practices that can be continually critiqued, re-imagined and thereby reassembled.
A second idea underpinning our analysis is that such registers of care are made visible through texts, strung together through citational practices which 'draw renewed attention to how microlevel interactions … help to perpetuate, transform, or challenge wider social and institutional formations' (Goodman et al. 2014: 460). Our methodological sensibilities borrow from studies of institutional interaction related to software development (Yates and Orlikowski 2002), mental health services (Günther et al. 2015) and air pollution debates (Solin 2004), which deploy the concept of intertextuality (Bakhtin 1981) as a method for social research (Bazerman 2003;Fox 1995;Fairclough 1992). These studies illustrate how actors are connected through intertextual chains corresponding to more-or-less stable constellations of shared social practices.
We recognise that intertextual work of this kind may suggest a focus on rhetoric or discourse; instead, however, we align our approach with Barad's proposal that 'discursive practices are not human-based activities but rather specific material (re)configurings of the world through which local determinations of boundaries, properties, and meanings are differentially enacted' (Barad 2003: 828). In line with this postdigital perspective, we therefore view these texts as the material traces of 'the changes that take place whenever algorithmic systems unfold in existing social contextswhen they are built, when they diffuse, and when they are used … that can reveal existing priorities within groups, organizations, and fields, as well as their changes over time' (Christin 2020). We did not look for explicit reference to 'care' within these texts; instead of operating as a linguistic marker, we viewed care as achievements enacted through the material-discursive apparatus around the proctoring platform, and sought to identify what or who was implicated in such forms of care (students, values, the platform, market share, etc.).
The TikTok video that motivated our inquiry was posted in September 2020 and from this initial text, we traced four inter-textual chains made up of a total of 37 documents which we describe in terms of Tronto's registers of care. As seen in Appendix, a variety of texts published between 2008 and 2021 were linked intertextually, either through direct quotation, indirect quotation or the mentioning of a person, document or statements (Bazerman 2003), forming the basis for the narrative in the next section. Institutional ethical approval for this research was secured prior to gathering data, based on the British Educational Research Association guidelines (BERA 2018); however, we were also mindful of ethical guidelines such as those of the Association of Internet Researchers (franzke et al. 2020). Specifically, we considered how to balance individuals' rights to privacy and confidentiality with their moral right to be identified as authors, and the potential for harm. Where individuals or companies had created public texts (such as press releases, academic papers, blog posts, etc.), we refer here to authors as they are identified in those texts. We have not included direct quotes from social media and have avoided naming the student who created the initial video, or the institution in which she was studying.
'Caring About' Exam Security -an EdTech Policy Network for Academic Integrity
Caring about refers to how society determines that needs exist and how they should be addressed. This is typically done by powerful individuals in the public sphere who often turn to technoscience for answers. Our narrative begins by exploring this register of care, tracing an 'EdTech power network' (Williamson 2019) that connected a housebound student's tearful TikTok video (Text 1) to a US technology firm's efforts to care about exam security.
The firm, ProctorU, was contracted by the student's university to assist with its emergency 'pivot' to remote instruction during the Covid-19 pandemic. The Alabama-based company had worked with distance education programs for over a decade and believed it was well-positioned to help universities move courses quickly and effectively off school campuses and into students' homes. They promoted themselves as offering: … a full suite of online proctoring and identity management solutions for education, professional development and credentialing organizations. With patented, 24/7 live proctoring and a fully-automated platform, both backed by artificial intelligence, ProctorU offers a powerful, convenient and cost-effective alternative to traditional test centers … (Text 2) Scott McFarland, the company CEO, reported 'a ten-fold increase in colleges calling, asking for help' in the early months of the Covid-19 crisis (Text 3). By its own account, the company successfully leveraged this track record to 'provide a secure testing environment without missing a beat', ramping up staffing and infrastructure in swift response to dramatic surges in demand (Text 2). Financial investors were enchanted. ProctorU claimed that while the online education 'industry' had expanded over recent years, the pandemic emergency had 'accelerated these growth trends beyond all expectations', conferring the firm and its affiliates 'enormous competitive advantage' in this 'multi-billion-dollar market opportunity' (Text 4).
But ProctorU's expansion hit a snag. In December 2020, the company was one of three private firms contacted by a panel of six Democratic US Senators who were troubled by student complaints and media reports of 'egregious situations' involving biased test-proctoring products. The lawmakers wrote, '[w]e are concerned that the software has not been designed to be inclusive and mindful of all students' needs and that proctors are not getting the training or information they need to adequately work with and oversee students' (Text 5). The senators also expressed concern about data privacy and the safety of students who were made to install 'intrusive' test-taking software and disclose extensive personal information to the company. Citing the Family Educational Rights and Privacy Act (FERPA), the Higher Education Act and Title II of the Americans with Disabilities Act, the lawmakers requested a statement from each proctoring company to '…address alarming equity, accessibility, and privacy issues faced by students using the platforms'.
The senators questioned whether ProctorU's products discriminated against students on the basis of race, religion, gender or disability. In its written response (Text 6), ProctorU attempted to allay concerns by downplaying the importance of its cheating detection algorithms: We utilize software tools as a means to assist our human proctors … [S] oftware tools supplement and inform human judgment; they do not replace human judgment. A good analogy is a smoke detector. If one goes off, a human has to decide whether there is a fire or just burnt toast.
The statement stressed the active role played by humans both inside and outside the company. It highlighted, for instance, the training and responsiveness of its human proctors and declared 'our proctors themselves are diverse and reflect the people we serve'.
The company also explained how responsibility for the remote testing environment was distributed across numerous humans both within the company and in their clients' organizations: … time limits, what resources are allowed or barred, whether breaks are allowed, what actions constitute misconduct or cheating that warrants further review by the testing provider or termination of an exam-are set by the test provider, the school, or the instructor administering the test. Our proctors alert institutions or testing agencies to violations of their test policies but do not decide whether any incident merits a particular consequence.
The senator spearheading the inquiry, Richard Blumenthal, remained skeptical and demanded 'much more transparency' from the technology firms, promising 'I will work on every necessary fix to ensure students are protected' (Text 7). In this vein, caring about students would oblige ProctorU to disclose more informationdetails about software codes, data security measures, and other business practicesto convince leaders that its test-proctoring system was inclusive and fair.
However, this focus on transparency -albeit an important part of discussions about AI ethics -does not engage with all that ProctorU claims to care about. Demands for transparency divert attention away from the ethics of what ProctorU does make visible: its business case for deploying algorithms to care about the academic integrity of students (Text 8). The company states that it is 'committed to using the latest technology to protect academic integrity and maintain the highest standard of fairness for every student' (Text 9). It uses security and surveillance metaphors to describe 'integrity in action': their proctoring platforms 'deter and prevent' and are designed for the 'detection and documentation' of cheating, referred to as academic 'breaches'. Institutions are 'armed' through the 'reporting' and 'intervention' capabilities of the platform (Text 10). ProctorU cares about the academic integrity of students in this manner because, in the words of CEO Scott McFarland: [w]hen a degree is gained though fraud, it undermines the image and the brand of the school and it's deeply unfair to the majority of students who work hard, study hard to have their degree undermined by those who take shortcuts and cheat. (Text 3) ProctorU stated, 'if educators care about the academic integrity of their exams and their programs, they will take action to secure the testing environment' to ensure equity for all students and protect the reputation of their institutions (Text 11).
'Taking Care' of Student Integrity -Ethical Culture and the Campus
Advancing its business case, ProctorU's CEO asserted, '[t]he rate of confirmed cheating attempts, no question about it cheating, would blow you away … Even during this pandemic, people are taking advantage to cheat more' (Text 3). The company cites 'statistics from more than 50 years of empirical research' (Text 8) indicating that over half of all US undergraduates cheat. This research was conducted by the International Center for Academic Integrity (ICAI), an association for professionals who develop and implement ethical codes of conduct in universities. It can be said that these individuals take care of academic integrity in the manner described by Tronto, 'assuming some responsibility for the identified need and determining how to respond to it' (Tronto 1993: 106).
The need to take care of student integrity in the US was articulated early on by Donald McCabe, a business professor at Rutgers University who was troubled by high rates of self-reported student cheating. In 1992, he helped launch the ICAI 'to combat cheating, plagiarism, and academic dishonesty in higher education' and support 'the cultivation of cultures of integrity in academic communities throughout the world' (Text 12). Affiliated with the Kenan Ethics Program at Duke University (1997University ( -2004 and the Rutland Institute for Ethics at Clemson University (2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017), ICAI established a longstanding research program to investigate 'moral development, moral education, institutional culture and their relationship to academic integrity' (Text 13). ICAI also developed assessment tools and consultation services to support university administrators, based on a working definition of 'academic integrity' as the 'commitment to five fundamental values: honesty, trust, fairness, respect, and responsibility …, plus the courage to act on them even in the face of adversity' (Text 14).
Arguing that universities had a 'moral obligation as educators' to pursue such values, McCabe and colleagues had long advocated for character education programs to develop the 'ethical decision-making capacities and behaviors of students' (Text 15). Led by senior 'Student Affairs' administrators, these programs aimed to align formal and informal cultural systems in the university with the values of integrity and principles of 'restorative justice'. This was considered preferable to punitive 'lawand-order' approaches where: … an emphasis is placed on values such as obedience, the rule of law, and deterrence. Administrators and faculty control policies and procedures and go to great lengths to monitor students' behavior and enforce rules …, sending students a strong message that cheaters will be caught and punished severely … In our view, such a law-and-order orientation will lead only to a fear-based cheating culture (rather than an aspirational culture of integrity) in which students are motivated only to avoid getting caught. (Text 15) The 'ethical culture' alternative encouraged students to engage in a dialogic moral education, with the values associated with honour codes serving as a 'touchstone' for multi-stakeholder deliberations over student integrity matters.
ProctorU often cited the cheating statistics the ICAI compiled, but did not engage with the recommendations that emerged from those findings. Instead, their 2016 white paper suggested that character development approaches were outdated and elitist (Text 16): Higher education was once a place of high integrity. Stringent honor codes upheld and protected the reputation of a school and the value of its degree. Students sought a college education for the sake of learning itself … But, new realities created new demands for colleges and universities and raised the stakes for students. … As motivations for a college education have changed, an environment where academic dishonesty is common has formed. Left unchecked, this could cause poor learning environments as well as reputational and financial damage.
It concluded that in the absence of time, money and people to implement comprehensive systems for ethical cultures, university administrators could still take care of academic integrity by (1) establishing a 'central policy and consistent punitive measures for infractions across all departments'; (2) reminding students early and frequently about expectations related to academic honesty; (3) following a 'stringent identity authentication process' that included 'Visual confirmation', 'Identify confirmation', 'Keystroke biometrics', and 'Facial recognition software' and (4) '[c] hoos[ing] the right technology partner' to 'securely identify and test students'.
As universities scrambled to emergency remote instruction in May 2020, Proc-torU published a blog post quoting new research released by the National College Testing Association which claimed: 'Faculty and staff should not make the egregious mistake of believing an honor code, signed statement of integrity, verbal acceptance of syllabi expectations, or other tacitly communicated acceptance is alone enough to sway academic dishonesty in online courses' (Text 17). A subsequent ProctorU white paper written by higher education journalists Jeff Selingo and Karin Fischer argued that the iterative, bespoke approaches to assessment used during the Covid-19 shutdown were no longer adequate: … Colleges adopted pass-fail policies to ease the strain of the chaos that enveloped their students' lives. Instructors allowed mid-term grades to stand for the semester, while others looked to alternative assessments -written reflections or portfolios of student work -in place of traditional exams. Elsewhere, students were reminded of university honor codes and allowed to selfpolice themselves while taking tests. With remote instruction likely to be the new normal in higher education, at least temporarily, stop-gap measures for assessing students' work are no longer sufficient; secure approaches to delivering exams are essential for institutions to ensure student success, prove their value to tuition paying parents, and demonstrate outcomes to employers and, in some cases, graduate schools. (Text 18) Selingo and Fischer also warned that accreditors would soon require additional quality control and oversight of online educational programs.
In the same paper, the authors referred to concerns about student privacy and accessibility as 'common myths' which 'have fallen along a series of familiar fault lines that often follow debates in higher education about buying outside technology solutions'. They suggested that remote proctoring was complementary to other datadriven approaches to enhancing student learning: As the coronavirus upends students' lives, sticking to the status-quo seems unproductive to institutions that claim to be student-centered. The tools that college leaders often label as crucial to student success efforts -predictive analytics, electronic advising, and guided pathways -are common upstream tactics for retaining students; the full power of ultimately helping students succeed comes way downstream, in figuring out what they have actually learned.
They encouraged university administrators to tailor technology solutions according to the specific needs of their institution. They noted that remote proctoring was less useful for evaluation in dance courses or for field-based study, and that 'ideally educators say its adoption should happen as part of a broader discussion among faculty about effective assessment'. (Text 18)
'Care-giving' Through Teaching and Learning -Integrity Through Pedagogical Encounters
While institutions may have policies about academic integrity and invest in infrastructure to support it, the day-to-day labour of interacting with students who may or may not cheat commonly falls to individual instructors. In Tronto's terms, this care giving involves meeting student's needs, through work and contact. ProctorU offered advice about how its services could help. Their position was that caring for students' integrity is a burden for instructors -one that their algorithmic technology could help reduce (Text 19): Here are four ways ProctorU can help you set and stick to testing goals this year: • Save Time on Non-Teaching Tasks … • Help Your Students Succeed by Keeping them Honest … • Add Convenience to Your and Your Students' Lives … • Stay in the Know About Your Students and Your Exams … ProctorU identified two challenges. The first concerns efficiency: as workloads increase, most instructors would welcome claims that '[w]e can save instructors 30 + hours of exam reviewing time per semester'. The second concerned intensification. They argued technology creates an arms race between students developing new ways to cheat and the instructors who must stop them, but which the company could take on, 'Fighting Technology with Technology' (Text 8).
Not everyone agreed. Swauger -a university librarian and researcher who tweeted in response to the TikTok video -had countered months earlier that there is no evidence that proctoring software effectively detects or prevents cheating (Text 20). This point was also developed by Dawson (Text 21), a researcher of educational technology and academic integrity who noted the lack of peer-reviewed evidence that these platforms deliver the technological or behavioural constraints they promise.
Another prior line of argument rejects ProctorU framing of academic integrity all together. Over a decade earlier, Gallant had observed that although the trope of 'technology … as the predominant and almost immutable force acting against institutional integrity' could be traced back to the printing press, it ignores how knowledge, information and authorship all evolve over time (Text 22). Her historical review argued that the ensuing controversies -e.g. disputes about whether working together was productive collaboration or collusion -show how 'academic integrity' is historically and socially contingent, not universal.
This perspective is shared later amongst the people who responded to the TikTok video. For instance, Swauger writes: Technology is often blamed for creating the conditions in which cheating proliferates and is then offered as the solution to the problem it created; both claims are false … Our habit of believing that technology will solve pedagogical problems is endemic to narratives produced by the ed-tech community. (Text 23) Gray, a Coordinator of Educational Technologies at Thompson Rivers University, extended Swauger's argument, calling for pedagogical strategies to prevent a repeat of this situation.
Students cheat, so we're told, and it's our job to defend against cheating. But for most of us, that messaging wasn't combined with any kind of training about how to create assessments. So we replicate the assignments we saw, and we replicate the attitude towards students we heard, and we wonder why nothing changes. (Text 24) Many of the Twitter responses followed this line, questioning why closed-book exams were still being used, challenging whether recall tests measure anything useful, and asking whether any 'real world job' excludes talking to other people or looking up information (Text 25). In their research, Gallant and others had long argued that academic integrity is not a simple matter of 'fixing' students by punishing them or even developing their character, but also involves environmental factors related to pedagogical practice and the institutional environment (Text 22). In a 2020 research compendium on academic integrity, Gallant suggested that treating cheating and plagiarism as environmental failures rather than individual moral deficits opens up 'teachable moments', allowing for outcomes that include not only feedback to students, but also to the environment and teaching practices (Text 26).
In the same compendium, Dawson explored how cybersecurity research might inform academic integrity scholarship, suggesting educators might actively encourage ethical or 'white hat' expert cheaters to test the limits of educational processes and tools so that lessons can be learnt at the level of both pedagogy and the digital infrastructure of universities (Text 21). Dawson also explored instructors' experiences of using new technologies for assessment (Text 27). He and his co-authors found that they felt driven to pursue financial and time efficiencies as class sizes grew, but felt this came at the cost of educational quality. In marked contrast to ProctorU's problematization, efficiency and outsourcing were seen as compromises and concessions, not desired outcomes.
More positively, these researchers reported that instructors knew technology might create new ways to cheat but were more interested in the novel opportunities created for students' self-expression. Instead of trying to prevent or control cheating, they made it less worthwhile by reducing the credit attached -or more radically, incorporated 'cheating' as successful strategies for learning. They recognized, for example, that collaboration and online searching are legitimate activities in many spheres of life, and so were more interested by what students could achieve when given these opportunities, rather than how they performed under exam conditions. Gray summarized these debates around instructors' 'care-giving' during the Covid-19 lockdown provocatively: We need to let this be the moment we choose to reframe our understanding of our learners. If we can't trust students, if the adversarial relationship will always be so much more comfortable to us that a camera in a student's bedroom is a more likely scenario than simple trust, then I ask again: what the hell are we even doing? (Text 24)
Student Movements in the Digital University -'Care-Receiving' During a Pandemic Emergency
Tronto reminds us that those in positions of power and responsibility who care about or take care of others may misjudge what is needed, and care-givers may lack the competence or resources to deliver good care. Caring well in society therefore also requires attending to the experience of care recipients. The care received by the emotional TikTok user from ProctorU and her university was inadequate. To borrow an earlier metaphor, the smoke detector set off a false alarm and the professor mistook burnt toast for a fire. This lapse in care placed the burden of proof on the student. To defend herself, she had to obtain a video of her exam session from ProctorU and schedule meetings with the dean and her professor. Although grateful that the university eventually re-instated her academic standing, the frustrated student wanted to know why the institution had not reviewed the evidence including the video of her taking the exam or her prior course work, rather than taking the system's classification at face value (Text 28).
Across the USA, students expressed similar frustration and apprehension about how academic integrity policies had encroached into their homes via remote proctoring platforms. Less than a week prior, the Electronic Frontier Foundation (EFF), an advocacy group 'defending civil liberties in the digital world' (Text 29), reported that tens of thousands of students had signed petitions calling on universities to end their contracts with proctoring technology vendors. EFF urged universities to 'take note of this level of organized activism' (Text 30), cautioning: [I]t's not just privacy that's at stake. … the petitions we've seen raise very real privacy concerns-from biometric data collection, to the often overbroad permissions these apps require over the students' devices, to the surveillance of students' personal environments-these petitions make clear that proctoring apps also raise concerns about security, equity and accessibility, cost, increased stress, and bias.
The advocacy group exhorted educational administrators to cooperate with students, professors and parents to 'make the very real concerns about privacy, equity, and bias in technology important components of school policy, instead of afterthoughts'. This widespread student backlash was picked up by mainstream news journals. In an article titled, 'How It Feels When Software Watches You Take Tests', the New York Times highlighted the difficulties experienced by low-income students, stating, '[t]he rigidity of online proctoring has exacerbated an already difficult year, students say, further marginalizing them at the very moments they're trying to prove themselves' (Text 31). The Washington Post observed that test-proctoring companies had 'sparked a nationwide school-surveillance revolt, with students staging protests and adopting creative tactics to push campus administrators to reconsider the deals' (Text 32). The article described one college freshman's efforts to audit the security of his school's proctoring software and how thousands of others had mobilized their complaints on Twitter accounts with names such as 'Procteario' and 'ProcterrorU'. The article also raised questions about the assumptions made about student integrity and digital technology: Is stopping a few cheaters worth the price of treating every student like a fraud? And how important are any of these tests, really, given the extra stress on students whose lives have already been turned inside out?
Fuelled by shared grievances, as well as callous public statements and aggressive legal manoeuvring by several proctoring technology firms, this growing and distributed collective of students cultivated a virtual 'ethical culture' for student integrity that was less concerned with exam security than with dismantling proctoring technology that they considered faulty, racist, ableist and an invasion of privacy. As one student newspaper wryly proclaimed, '[e]xam-taking has become the new airport security' (Text 33).
Although ProctorU CEO Scott McFarland expressed regret for what had happened to the young woman in the TikTok video, he nevertheless used the case as another opportunity to promote the company's products: We are always disappointed when anyone has a difficult time related to a test session. This is a good example of why it's important for students and schools to have a video recording that instructors can review and thus make evidencebased determinations. (Text 34) This statement reinforced prior messages about the inevitability of surveillance technology in remote testing. As McFarland stated in an earlier interview, 'we may not love the idea of being on camera every time we visit a bank or go to a convenience store, but no one is suggesting taking them down' (Text 35).
On ProctorU's website, positive reviews from students are used to justify this stance, praising the company for widening access to higher education (Text 36): … ProctorU helps me handle life as a mom, Army wife, full time employee, and student by offering exam times when it's most convenient for me. … Working full-time and also going to school often times makes it challenging to schedule exams, but ProctorU is extremely flexible and convenient.
Such testimonials foreground differences between students of established online education programs and those making the abrupt pivot to remote learning due to Covid-19. Whereas the former tend to have family and/or employment constraints that make individualized, 'anytime, anywhere' education helpful, the latter group expected to engage in extensive peer interaction as part of 'campus life', ostensibly including a culture of academic integrity involving students, instructors and administrators. However, ProctorU argued that as a matter of fairness, all students -online, hybrid and campus-based -should be subject to 'objective, quantifiable standards' of assessment and exam security (Text 37). They reiterated that students expected administrators and instructors to preserve the academic reputation and value of their university degrees by taking care of exam security.
Analysing Breakdowns and Disasters: Academic Integrity in Trouble
Tronto argues that caring well requires the ethical virtues of attentiveness, responsibility, competence and responsiveness to be enacted through integrated practices of caring about, taking care of, care-giving and care-receiving, respectively (Tronto 1993). Failures of care occur when these four registers are unaligned. Our account of the ethical controversies around proctoring platforms echoes STS studies that engage with breakdowns (Bourrier and Nova 2019) and disasters (Fortun et al. 2017) to foreground the power relations of technological innovation and the precarity of such orderings (Law 1992). In our analysis, we discerned ruptures in the social production of academic integrity on two distinct temporal scales: (1) an immediate 'acute' crisis of care, related to the ethics of an algorithmic platform during the Covid-19 campus closures; and (2) a slower moving, 'chronic' crisis of care related to educational technology's historical entanglements with the neoliberal university and concurrent efforts to address socio-economic disparities in US higher education (cf. Williamson 2019).
Academic Integrity as an Acute Crisis of Care
The more apparent, punctual break in routine academic integrity practices pitted students and educational scholars against technology firms and the university administrators who procured their products. Within the compressed timeframe of this 'acute' Covid-19-related breakdown, actors across all four registers of care intervened decisively in the Covid-19 lockdown, enacting a highly visible public controversy over whether an algorithmic platform was an ethical replacement for on-campus practices of academic integrity. ProctorU, university administrators, and some instructors swiftly deployed the system to maintain instructional continuity and protect students against 'others' who cheat. For ProctorU, caring about academic integrity in the immediate term entailed attentiveness to matters of security, the reputation of its academic clients and the growth of the company. Universities took care of academic integrity by rapidly devolving responsibilities for online exam invigilation to ProctorU.
The proctoring platform, however, misclassified individuals as cheaters and created new educational barriers for students of colour, the disabled and individuals living in low-income households. Instead of receiving care in the manner envisioned by technology vendors and universities, groups of students quickly responded by mobilizing on social media and organized online petitions to dismantle proctoring platforms, raising concerns related to fairness and privacy. Doubts were raised about whether the proctoring technology enhanced the competence of instructors as frontline care-givers during campus closures. Educational scholars, advocacy groups and the press amplified student demands, catching the attention of lawmakers who also cared about fairness and privacy but were attentive mostly to the issue of industry transparency and the need to scrutinize the functionality and legality of proctoring systems.
Academic Integrity as a Chronic Crisis of Care
The acute crisis described above is nested within a second, slower, more diffuse and ongoing disruption of academic integrity practices emerging over several decades as universities work to widen student access while contending with neoliberal agendas. In this chronic context, conflicting care practices form a different divide, with ProctorU and students of established online programs on the one hand, and education scholars and campus-based students on the other. As in the acute setting, the company promoted a technology platform to help universities respond to 'new realities' and 'new demands'. But in this case, caring about academic integrity not only involves attentiveness to security, academic branding, and the financial health of the company, but also to the standardized forms of learning assessment which are better-aligned with algorithmic test proctoring.
ProctorU reported positive feedback from students of online programs who are care receivers seeking wider access to formal education and accredited diplomas. However, ProctorU's approach to caring about academic integrity stands in marked contrast to the claim that university administrators have a 'moral obligation as educators' to take care of academic integrity and assume responsibility for implementing a dialogic and restorative 'ethical culture' in their institutions. The company's approach was also at odds with the researchers and instructors who were caregivers seeking resources and pedagogical reforms to subvert neoliberal logics and develop competence in building teacher-student relationships, extending trust and strengthening teaching missions. Viewed in the context of a 'slow disaster', these competing ethical claims constituted a more fundamental controversy about academic integrity, the purpose of higher education and how quality and global reputations are achieved.
Repairing Academic Integrity in Universities
Having foregrounded the controversies surrounding two temporally distinct but nested breakdowns in academic integrity, we ask, 'what then must be done?' (cf. Bharti 2021). If 'rupturing events' are indeed, as Guggenheim suggests, 'inherently political … because they pose questions about who should be allowed to re-compose the world and how' (Guggenheim 2014: 4), what do the politics of practicing an ethic of care tell us about the ethics of practicing a politic of care? What then must universities do to 'carefully' repair these fast and slow crises? Who should repair academic integrity in the world of the university, and how?
Our analysis confirms that we should avoid conflating the chronic temporality with authentic care and the acute responses with inauthentic care (Spier 2019). Few would argue against fixing an algorithm that is discriminatory and violates privacy, for instance. And yet, this repair can work to sustain a flawed academic integrity system, even as it brings justice to individuals. By the same token, pedagogical innovations that diversify the way learners are assessed may benefit elite students who possess alternative forms of social capital, but if reforms are incompatible with employer expectations, these may disproportionately harm marginalized students who rely on education as a vehicle to gainful employment.
Drawing from Spier's work on inclusion in higher education (Spier 2019), we argue that 'careful' repair of broken academic integrity systems requires 'mixed modes' of response that are attuned to the temporal dynamics of care practices: This sensibility of time challenges the idea that the educators' wisdom relies on 'knowing-how' (proficiency in activating invariable-normative strategies within variable situations). Nor can wisdom be reduced to 'knowing-that' (proficiency in prereflectively following invariable-normative principles to variable situations). Instead the educator's practical wisdom of caring is better understood as 'knowing when'. (Spier 2019: 36).
Indeed, Michael points out the importance of recognizing how pasts, presents and futures, understood as 'entities and occasions, discourses and practices, humans and non-humans' (Michael 2014: 240), converge within the event of a disaster. He argues that while there may be a need to 'slow down' the repair of such ruptures, this deceleration can be accompanied by an 'acceleration in the processes of asking more inventive questions, of finding better meanings, and of enabling finer responses' (Michael 2014: 244).
Conclusions
Do we need watchmen at all, or could we instead trust our students? If we do decide that watchmen are needed, and academic staff are unable to do this, could students watch each other, shifting the balance of power towards rather than away from them? If universities conclude that they do need to outsource this responsibility to technology, how can we ensure that it is undertaken responsibly, and in accordance with the values that universities seek to promote?
In this paper, we have argued that the ethics of AI are not only abstract or decontextualised principles, but also political and practical 'doings'. We have described how universities (perhaps) inadvertently bought into ProctorU's ethico-political logic when they contracted out the practices of proctoring; few realised that these new algorithmic 'watchmen' needed watching until people had been hurt. Might there be better ways of being with algorithms in the university? We close this paper by exploring how responses might have been different, not by invoking a new code of 'ethical AI', but instead by responding to Puig de la Bellacasa's 'speculative commitment' (Puig de la Bellacasa 2011: 96) to reassemble current relationalities into more careful academic integrity practices.
Eighteen months into the Covid-19 pandemic, as we write this conclusion, industry executives, policymakers and economists have already released a flurry of reports which sound the clarion call to 're-imagine education' with technology. Authors such as Cone et al. (2021) have observed the accelerated introduction of technology during the pandemic, linking this to the possible industrialisation of higher education. As they note, the risk related to such rallying cries is that unexamined ethical and political commitments materialise through these technologies and are then incorporated into the lives of academics and students. It would perhaps be expedient to commission these algorithmic watchmen to help manage the distributed practices of academic integrity -yet it would be foolish not to watch them in turn to ensure that they discharge this responsibility appropriately. But more importantly, rushing to 'fix' educational problems with technological 'solutions' forecloses an important opportunity for dialogue that opens up new ethical and pedagogic forms of academic practice.
Staying with one moment in the lived experience of a university student during the lockdown, we find potential avenues for living better with algorithms in the often-overlooked processes of procurement. Rather than accepting the tendency for the ethico-political assumptions of business to reshape education, we pose the question of whether academics might create different kinds of relationship with developers and service providers through inventive challenges that 're-imagine the business case'. As university campuses procure technologies, particularly at this moment when they re-open their doors, how has Covid-19 changed the 'business case' for these technologies? How has the lived experience of students, instructors and families during the online pivot challenged economic understandings of what it means to learn, and what we have lost (Strauss 2021)? As universities commission or procure technology and infrastructure, can they use this economic moment to demand business practices that will help create futures that are better educationally, or which widen access to groups that would otherwise be excluded? Can they contribute to futures that use algorithmic technology to redress rather than entrench the chronic crisis in Higher Education, opening up pedagogical possibilities for students and academics, rather than shutting them off in the name of efficiency?
The business case has been described as a rhetorical intervention that draws costs and benefits, risks and stakeholders into relationships (Maes et al. 2014). ProctorU's case assured integrity by 'de-risking' assessment. Its financial costs are obvioussuch as the licensing fees -but non-financial costs also exist, such as the pedagogic constraints the system requires. Variability is reduced by standardizing processes, but this requires standardizing forms of assessment, and inadvertently, how students behave and even what they look like. The justification offered rests on fears about a technologically driven 'arms race' of cheating.
Following Gallant and Dawson, an alternative would be to 'de-risk' in other ways -by valuing cheating for the 'teachable moments' it creates, attending to the chronic crisis by asking 'white hat' expert cheats to help people and processes learn and improve. Mirroring national awards for teaching, competitive 'hacking' could be used to test the integrity of different institutional practices. The financial costs are less clear -each proposed improvement might need its own business casebut these would be distributed between universities (for pedagogic developments) and industry (for the technologies that universities procure), and the benefits include rapid pedagogic development.
Alternatively, rather than viewing honour codes as naïve or ineffective, we could follow Gray's provocation to trust our students, foregrounding development of integrity as a vital part of what university education is for. Instead of rewarding industry for managing educational risks, students could take this responsibility, bringing proposals for ethical action that would strengthen integrity or create better forms of assessment to their teachers, perhaps as paid work or for credit, or simply because they find such forms of co-creation more meaningful and valuable than being student-consumers (cf. Luo et al. 2019). Such engagement was visible during the acute crisis of the Covid-19 lockdown, in the student-led activism against proctoring platforms; as a response to the slow disaster, it offers sustainability and mutual development with few financial costs.
Instead of reducing risks, business cases could also explore enhanced benefits. The study by Dawson and others showed instructors adapting teaching and assessment in creative ways and being curious about what students create when cheats are shared as sensible strategies for learning, rather than deviant practices to be controlled. Rather than feeling outpaced by students in an 'arms race', such developments would ensure pedagogic practices remain relevant and contemporary, addressing the concerns raised earlier about employer expectations and marginalized students who need education to gain employment. This would have time costs, mainly through regular curriculum development, but the benefits should include an improved student experience, differentiation of institutions' educational missions and which could help advance institutional 'brand' in a competitive student market.
ProctorU's approach to 'fixing' academic integrity offers only one possible future. Alternative business cases such as these show how things could be otherwise, and what apart from risk reduction universities might value. Business cases are used to enrol stakeholders (Maes et al. 2014), so that a company like ProctorU can become the obligatory passage point for repair, moving from precarious assemblage to stable, punctualized and unquestioned 'black boxes'. However, where business cases focus on swift repairs or immediate reparative justice, they may inadvertently sustain dysfunctional systems, exacerbating (perhaps even accelerating) the deleterious effects of slow disasters. To hold open possibilities for the careful repair of ruptures in academic integrity, this is the moment to strengthen associations with stakeholders often excluded from procurement processes, creating new relationalities between educators, technology experts, administrators, students and other-than-human actors such as algorithms that enact the values universities claim to care for: honesty, trust, fairness, respect and responsibility (Text 14). Ultimately, as with hired watchmen, a business case is only as good as the client judges it to be; it is the responsibility of university staff and students to be vigilant and engaged, so as to ensure the algorithms and platforms we chose to live with are shaped to our needs, rather than the other way around. | 11,998 | 2021-11-09T00:00:00.000 | [
"Philosophy",
"Education",
"Political Science",
"Computer Science"
] |
Vector magnetocardiography measurement with a compact elliptically polarized laser-pumped magnetometer.
We report on a practical approach to vector biomagnetism measurement with an optically pumped magnetometer for measuring total magnetic field intensity. Its application to vector magnetocardiography is experimentally demonstrated with a compact elliptically polarized laser-pumped M x atomic magnetometer (EPMx OPM). The approach is proved to be effective and able to provide more complete cardiac magnetic information. The cardiac magnetic vectors are displayed in three-dimensional space in the form of magnetic vector loops. The sensor configuration and the image processing method here are expected to form further values, especially for multi-channel vector biomagnetism measurement, clinical diagnosis, and field source reconstruction.
Introduction
The magnetic fields generated by different organs of the human body often convey valuable information about its source. Therefore, the measurement of human biomagnetism is significant to basic and clinical medicine [1]. A great deal of effort has been contributed to this topic, with major areas of interest being the magnetic fields of the human heart (magnetocardiography, MCG) [2][3][4][5][6], brain (magnetoencephalography, MEG) [7], lung (magnetopneumography, MPG) [8] and eye [9].
MCG, in particular, has received a growing attention as a contactless, non-invasive method for myocardium examination. Unlike the clinically used electric counterpart of body surface potential measurement (electrocardiography, ECG), magnetic signals are characterized by both intensity and direction information. Besides, the transmission of magnetic signal from heart is not affected by the variability of human skin conductance [10][11][12]. These distinct advantages make MCG a very promising diagnostic technique especially for determining lesion localization, corresponding to the inverse problem of reconstructing the field sources from MCG maps. Though the earlisest recognizable measurement of MCG was achieved with multi-turn coils [13], induction-coil-based MCG measurement is difficult to get a significant progress due to poor signal-to-noise ratios of sensors. With the combination of superconducting technology and magnetic field detecting coils, the superconducting quantum interference device (SQUID) was invented. SQUID possesses very prominent sensitivity in the relevant frequency range for MCG. At present, there have been commercial equipments for MCG based on SQUID technique [14], which have made great contributions to the development of MCG research [15,16]. However, SQUID magnetometers require cryogenic cooling, making the installation and maintenance costs too high. Moreover, the liquid-helium-cooled dewar restricts the distance between the skin and sensor and thus reduces the magnetic signal strength. These technical limitations hinder the clinical spread of SQUID-based MCG. The emergence of optically pumped magnetometers (OPMs), which are characterized by laser-based atomic techniques [17], provides a novel means to the development of MCG measurements [18][19][20][21][22]. Nowadays, sensitivity of OPMs has approached and even exceeded SQUID magnetometers in the laboratory, without the requirement of cryogenics [23,24]. Therefore OPM is widely regarded as a most promising alternative to the SQUID magnetometer. In principle, a number of OPMs record magnetic field information by measuring the Larmor precession frequency of atomic spins. Thus they essentially measure the magnitude of the total magnetic field. This feature makes the way of OPM application very different from the intrinsically vector magnetometer, such as SQUID magnetometer, fluxgate magnetometer [25] and magnetoimpedance magnetometer [26], which are sensitive to one of the three components of the magnetic field vector. To achieve a complete vector magnetocardiography (VMCG) measurement with vector magnetometers, we usually arrange three adjacent sensors, respectively recording the magnetic field component in one direction. However, the magnetic field from heart is not spatially uniform, so the accuracy of the reconstructed field is affected by the separation of sensors. We can also alternate the measurement directions by rotating a single sensor head to realize VMCG. Such mechanical operations complicate the MCG device and still introduce uncertainty to the measurement results. In addition, both methods are operationally difficult especially in terms of designing a multi-channel VMCG device. Various methods exist for realizing the conversion between the scalar and vector modes of an OPM. As typical examples, we can gain vector information by using multiple circularly polarized laser beams [27], or utilizing the ac Stark shift induced by a circularly polarized laser beam [28], or extracting additional information from the first harmonic of the Larmor frequency of an atomic alignment [29], or compensating the measured field [30,31], or exploiting the dependence of electromagnetically induced transparency resonance amplitudes on the external magnetic field direction [32].
In the present paper we propose an approach to VMCG with an OPM, unlike previous work, based upon the fact that cardiac magnetic field is four or five orders of magnitude smaller than the bias magnetic field required to run the OPM. The experiment of VMCG measurement is demonstrated with a compact OPM sensor head. The sensor head is an integrated version of the elliptically polarized laser-pumped M x atomic magnetometer (EPMx OPM), which has been investigated deeply in our previous work [33]. The EPMx OPM is an improved M x type OPM, which has been considered as a suitable configuration for MCG measurement [18][19][20][21]34]. We achieved VMCG measurements by simply changing the direction of the bias magnetic field. VMCG diagrams from a healthy test male were recorded at different sites above the chest. Some noteworthy findings from VMCG diagrams are reported. Also provided is a visual representation of VMCG, named as magnetic vector loop (MVL). We propose a potential cardiac diagnostic indicator by fitting the MVL plane and extracting the orientation information.
Compact EPMx configuration
The picture of the sensor head module, as well as its design, are depicted in Fig. 1. The length of the module is about 10 cm. The housing is fabricated out of photosensitive resin using 3D printing. Our EPMx OPM used a 20 × 20 × 20 mm uncoated cubic cell, which contains a drop of enriched 87 Rb atoms and 200 Torr of N 2 gas for quenching and slowing atomic diffusion. The pressure of N 2 gas gives a cross-talk free distance of 87 Rb atoms less than 0.8 mm even at 100 • C [35] and thereby makes the diffusion negligible. The T 2 relaxation time is about 40 ms. A pair of copper heating wire is wound around the cell. When the OPM works, the cell is heated to 45 • C by alternating current at 70 kHz, which is much higher than the atomic resonance frequency to avoid noise caused by heating. Then the temperature is stabilized in a small range of less than 0.1 • C by a closed-loop circuit. The main magnetic field B 0 and the radio-frequency (RF) magnetic field B rf are provided by external coils.
Light from an extended-cavity diode laser, tuned close to 87 Rb D1 transition at 795 nm, is delivered into the sensor module via a single-mode polarization-maintaining optical fiber and the beam is expanded to a diameter of about 6 mm by collimating lenses. Then the light is divided into two parts by a beam splitting prism. The reflection part is used to monitor the light parameters, such as frequency, intensity and polarization, so as to provide an optimized optical condition for the sensor. The transmitted light is reflected by the right-angle prism, then the polarization is purified by a linear polarizer. Before illuminating the cell, the light passes through a quarter-wave plate with its optic axis oriented at an angle ϕ relative to the linear polarizer. The ellipticity of the light can be adjusted by changing the angle ϕ. The ellipticity can be characterized by the average photon spin s = sin(2ϕ). It ranges from -1 to +1, where s = −1(+1) corresponds to right (left)-circularly polarized light and s = 0 corresponds to linearly polarized light. According to our previous optimization work [33], the optimal ϕ is about 20 degrees, the optimal laser frequency detuning is about 3 GHz from the transition between the ground state F = 2 to the excited state F = 1, and the incident light power is 100 µW. The transmitted light is then reflected back at a small angle less than 7 • by a zero degree mirror, which is placed very close to the cell. The structure of such a design allows the vapor cell to be as close as possible to the chest surface while measuring MCG, and also double the length of the optical path, thereby increasing the signal to noise ratio (SNR). When elliptically polarized light passes through polarized atoms, the polarization plane rotates. This rotation is subsequently converted to an electric signal through a balanced polarimeter, consisting of a PBS, two photodetectors, an I/V converter and a differential circuit. The linear polarizer and the quarter-wave plate are rotated to adjust the output signal of the balanced polarimeter to zero before the sensor works. The oscillating optoelectronic signal is then fed to a digital lock-in amplifier (LIA), whose reference signal is the applied oscillating RF field. The demodulated output is used as error signal for the following servo system, consisting of a proportional-integral-differential (PID) module and a frequency-controlled signal generator.
Principle of operation
As shown in Fig. 1, the total field ì B contains a main magnetic field ì B 0 and a radio frequency field ì B rf . ì B rf is perpendicular to the plane composed of ì B and the propagation direction of light. The overall evolution of the atomic spin angular momentum ì S is well-described by the Bloch equation [34] d Where γ = 7 Hz/nT is the gyromagnetic ratio of 87 Rb atomic spins, Γ P is the pumping rate, ì S 0 is the equilibrium atomic spin angular momentum in the absence of the oscillating excitation and Γ rel is the spin-relaxation rate. Through the steady-state solution to Eq. (1) in the rotating frame with angular frequency ω rf , which is the oscillating frequency of ì B rf , we can obtain the quadrature amplitude P qu and in-phase amplitude P ip of the photocurrent with respect to the oscillating magnetic field ì B rf , which are given by Here ϑ is the angle between the direction of laser beam and ì B 0 , Ω is the Rabi frequency, δ=ω rf −γB 0 is the detuning of the oscillating field B rf from the Larmor frequency, Γ 1(2) = Γ 1 (2) rel + Γ P is the effective longitudinal (transverse) polarization-relaxation rate. Note that the symbol without an arrow superscript represents the magnitude of the related vector in this paper.
The EPMx OPM is operated in a phase-locked mode, by controlling the driving frequency ω rf . P qu and P ip correspond to the X and Y outputs of the LIA. From Eq. (3), we can see that the Y output is zero if the system is in resonance (δ = 0). Therefore the Y output can be treated as an error signal, based on which the digital PID module adjusts the output frequency of the signal generator. The servo loop thus ensures that the instantaneous Larmor precession frequency of the atomic moments is always in resonance with the driving frequency ω rf . The value of B 0 can be determined from ω rf , as
Weak magnetic field measurement in a finite bias magnetic field
As shown in Fig. 2, assume that the bias magnetic field is ì B 0 , and the weak magnetic field to be measured is ì B . Decomposing ì B into ì B (parallel to ì B 0 ) and ì B ⊥ (perpendicular to ì B 0 ), we have Under the condition B B 0 , ignoring the quadratic terms B 2 and B ⊥ 2 , we have So Eq. (5) can be simplified to In the MCG measurement, B 0 is of the order of 1000 nT, while the cardiac magnetic field B heart is less than 100 pT. The approximation error of Eq. (7) is less than 10 fT, which is negligible compared with the magnitude of B heart . Therefore, Eq. (7) implies an approach to realize vector measurement with a scalar magnetometer. Though the EPMx OPM we used intrinsically measures the magnitude of the total magnetic field, the change in the measurement result actually corresponds to the projection of B heart into the direction of the much larger bias magnetic field. For a complete VMCG, we can sequentially change the direction of the bias magnetic field within three orthogonal axes, named as x, y, and z axis, respectively. It is worth mentioning that the angle between the light beam and the bias field has no influence on the validity of Eq. (7). Therefore, though the bias field has different directions with respect to the incident light beam and to the reflected one in our configuration, the sensitive direction of the sensor is still determined by the bias field. According to Eq. (2) and (3), the small angle between the incident and reflected beams would slight affect the signal amplitude, and such influence on the sensor sensitivity can be ignored.
VMCG experiment
Now, we turn to experimental demonstration of the VMCG measurement. A healthy male of 30 years old was tested in our experiment. As shown in Fig. 3, the man was sent to a 5×µ-metal cylindrical magnetic shield by a slideable bed. The inner diameter of the cylindrical shield is 80 cm, and the length is 200 cm. One end of the cylindrical shield is open. The residual magnetic field in the center of the cylindrical shield is less than 20 nT. Three pairs of orthogonal internal coils are mounted on the inwall of the cylindrical shield for generating the required magnetic bias fields and RF magnetic fields. From Eq. (2) and (3), we can see that a deviation of the light propagation direction from B 0 is a necessary condition to run an EPMx OPM. The coordinate system used in this paper is shown at the upper right of Fig. 3. The sensor head is tilted along the vector direction (-1,1,-1), so that the light axis is at the same angle to ì B 0 , regardless of which axis is aligned with the bias field. Thus, such configuration enables magnetic field measurement in three directions. A commercial current source (B2912A from Keysight) is used to supply a stable and well-defined bias field of about 1000 nT. As shown in Fig. 4, through the optimization process descried in Ref. [33], sensitivities of around 300 fT/
√
Hz are achieved at the temperature of 45 • C in all of the three mutually orthogonal directions, which have approached to the limit set by the noise level of the current source. Assisted by a separate test coil, we determined that the bandwidth of the sensor is about 100 Hz, which has met the requirement for MCG measurements. As shown in Fig. 3, 3 × 3 sites above the chest were selected for VMCG measurements. These sites were spaced 5 cm apart. The VMCG measurement at each site was achieved by changing the direction of the bias magnetic field ì B 0 and the RF field ì B rf . So there was no need to move the human body or the sensor. The working parameters were automatically searched by a software program based on MATLAB. When we changed the magnetic field strength or started the measurement at a new site, the software program automatically scanned the driving frequency in open loop mode to determine the Larmor frequency and gave proper PID parameters. The initialization process took approximately 30 seconds. The signals of 20 heartbeat cycles were collected in each direction, so the measurement time for each site took about 1 minute. During the MCG measurement, a reference signal was also synchronously acquired by a simple ECG device. The ECG signal allows for improving the SNR in an off-line averaging and distinguishing the periods of different cardiac activities. Figure 5 shows a real-time MCG signal acquired in the x direction at site B, as well as its reference ECG signal. A digital band-stop filter centered at 50 Hz was used to suppress power frequency noise. The ECG signals served as timing reference for distinguishing the individual heart beats when averaging. Averaged MCG signals (20 beats) of all sites at different directions are presented in Fig. 6. It is interesting to find that the MCG signals show specific waveform in different directions, even at the same site. For example, at site A, the P wave and T wave are opposite in phase in the x and z directions, while they are inphase in the y direction. There also is much difference in signal strengths between different directions. At site D and E, P waves appear only when we measure MCG in x direction. At site G, T wave is obvious in x direction. Besides, the MCG waveforms show high dependencies on the measurement position. At site I, we can get clear U waves. In order to clearly show the spatial characteristics of the heart magnetic field, we synthesize the MCG signals in x, y and z directions into a magnetic field vector and display it as a MVL map. Specifically, MVL map is the trajectory of the end of the cardiac magnetic vector in three-dimensional space in a heartbeat cycle. Fig. 7 shows the synthesis of MVL map for site A. The periods of different cardiac activities are determined by ECG signals and are marked with different colors. We can see that MVL map can visually display the changing process of the cardiac magnetic field direction during a heartbeat cycle. This is difficult to achieve with traditional MCG images. In addition, P, QRS, and T waves can be clearly distinguished, as they form different vector rings. The loops related to different cardiac activities are very different in shapes, plane orientations and enclosed areas. The orientations of QRS loop and T loop are roughly the same. It is understandable since they both correspond to the electrophysiological activity of the ventricles, while P loop corresponds to the electrophysiological activity of the atria. The enclosed areas of loops are determined by both durations and intensities of corresponding electrophysiological activities.
Experimental results and discussion
All of the VML maps for 9 sites are presented in Fig. 8(a), from which we can see that the directions and shapes of the loops change significantly with the measurement position. We fitted each vector loop with two-dimensional plane parameters and got the corresponding characteristic plane. The principal axes and normal vectors of these planes for QRS (T) waves are presented as v-QRS (T) and n-QRS (T) in Fig. 8(b), respectively. The average of angles between the QRS plane and the T plane is 18.49 degrees with a standard deviation of 6.75 degrees.
Conclusions
In conclusion, we have presented an practical approach to vector biomagnetic measurements with an OPM, without any need of additional sensors or mechanical rotations of the sensor heads. The approach was experimentally demonstrated by an EPMx OPM based on 87 Rb atoms. We gained VMCG maps from 9 sites above the chest of a healthy male. The results verify that the waveforms and strengths of VMCG are highly dependent on the measurement positions and directions. It is very different from the ECG lead, which essentially measures the potential difference between two sites. Moreover, we show that more information about the characteristics of VMCG signals can be extracted by fitting the MVL maps. Actually, VMCG is a comprehensive result of the biological currents from all cardiac cells. When some myocardial cells are abnormal during a heartbeat cycle, such as in the situation of bundle branch block or myocardial infarction, direction and intensity anomalies of the cardiac magnetic vector may happen. We expect that such anomalies can form diagnostic basis by observing the directions or shapes of the VML maps.
This work enables OPMs for measuring total magnetic field intensity to have new development potential in the measurement of vector biomagnetism. It provides a suitable tool for the further realization of multi-channel VMCG measurements. VMCG conveys unique information about the heart condition, which is important for the cardiac diagnosis and the source inversion problem. To explore the value of VMCG technique, there is still much work to be done in developing multi-channel vector magnetic detection sensor and performing more clinical analysis. | 4,705.4 | 2020-01-07T00:00:00.000 | [
"Engineering",
"Medicine",
"Physics"
] |
Wind Turbine Gearbox Condition Monitoring with AAKR and Moving Window Statistic Methods
Condition Monitoring (CM) of wind turbines can greatly reduce the maintenance costs for wind farms, especially for offshore wind farms. A new condition monitoring method for a wind turbine gearbox using temperature trend analysis is proposed. Autoassociative Kernel Regression (AAKR) is used to construct the normal behavior model of the gearbox temperature. With a proper construction of the memory matrix, the AAKR model can cover the normal working space for the gearbox. When the gearbox has an incipient failure, the residuals between AAKR model estimates and the measurement temperature will become significant. A moving window statistical method is used to detect the changes of the residual mean value and standard deviation in a timely manner. When one of these parameters exceeds predefined thresholds, an incipient failure is flagged. In order to simulate the gearbox fault, manual temperature drift is added to the initial Supervisory Control and Data Acquisitions (SCADA) data. Analysis of simulated gearbox failures shows that the new condition monitoring method is effective.
Introduction
Challenging environmental factors combined with high and turbulent winds place serious demands on wind turbines and result in significant component failure rates, highlighting the importance of maintenance.Appropriate use of condition monitoring can, by detecting faults at an early stage, reduce turbine repair and maintenance costs.
The gearbox is one of the most important components of a wind turbine.In [1] the authors show that the maintenance cost for the gearbox is very high compared with the other higher failure rate components such as electric system and hydraulic system, especially for the offshore wind farms where maintenance will need a boat, crane and nice weather.The gearbox undergoes varying speed and load.With the varying wind speed, the rotating speed and load of different stages of gearbox change from time to time, bringing great challenges to the condition monitoring of gearbox.Cost performance is another factor that should be taken in to account in gearbox condition monitoring.Compared with the steam turbine in a thermal power plant, the building cost for each wind turbine is relatively low, and the price for the condition monitoring system must be acceptable.Comprehensive introductions and analyses of condition monitoring methods for different components of wind turbine have been published [2][3][4].Temperatures are important and easily measured indicators of the health of many wind turbine component such as the gearbox and are often recorded automatically by the SCADA system.An unexpected increase in component temperature may indicate an overload, poor lubrication or possibly ineffective passive or active cooling [5].Previous work used a Back Propagation (BP) neural network to construct the normal behavior temperature models of gearbox based on SCADA data [6].When the residual between the model prediction and the measured value becomes very large, a potential fault is identified.In [7], the authors proposed a method using a Multiple Layer Perception (MLP) to build a temperature model of the gearbox.When the measured temperature value increases and is outside the confidence range of the value, a fault is registered.However, the Artificial Neural Network (ANN) including BP and MLP has demerits of requiring a time consuming training process and there can be local minima problems that may limit the improvement of model accuracy.Several authors [8][9][10][11] have built a test rig for gearbox and generator.Wavelets are used to analyze the high-speed sampling vibration signals.However, a test rig is quite different from a real wind turbine.There are only limited numbers of acceleration sensors to measure the drive train and gearbox vibration to provide a magnitude alarm and the sample frequency (10 s or 10 min) is too slow to meet the high frequency demands of vibration analysis.This paper uses temperature trend analysis to monitor a gearbox's operating conditions.The Autoassociative Kernel Regression (AAKR) method is used to model the normal behavior of gearbox temperature and give temperature estimates.When the gearbox has an incipient failure, the temperature residual between the AAKR model estimate and real measurement will become significant.With moving window residual statistics, these incipient failures can be detected in a timely way.
The paper is arranged as follows: Section 2 provides an introduction to turbine gearbox and SCADA data.Section 3 explains how the AAKR gearbox temperature model is constructed.The fourth section focuses on the moving window residual statistical analysis method.In Section 5, the AAKR model is validated and two simulated gearbox failure cases are studied.The final section provides discussion and conclusions, including suggestions for further research.
Structure of a Wind Turbine Gearbox and the SCADA Parameters
The wind turbines studied in this paper are located at Zhangjiakou in Northern China.SCADA data covering the period 04/01/2006 to 24/12/2006 was available for these units.The turbines, manufactured by GE, are variable speed, with a rated power of 1.5 MW.The cut-in and cut out wind speed are respectively 3 m/s and 12 m/s.The rated rotating speeds for rotor and DFIG are 20 rpm and 1800 rpm, respectively.The ratio for the gearbox is 1:90.This wind turbine uses a three-stage gearbox.The first stage is a planetary gear, and the second and third stages are conventional parallel gears, as shown in Figure 1.There is a circulating lubricating oil system to lubricate and cool the gearbox.The SCADA system at the wind farm records all wind turbine parameters every 10 min.Each record includes a time stamp, output power, stator current and voltage, wind speed, ambient and nacelle temperature, gearbox temperature amongst many others; in total 47 parameters are recorded.At the same time, the SCADA system keeps a record of wind turbine operation and fault information, such as start up, shutdown, generator over temperature, pitch system fault, etc.Each fault record includes a time stamp, state code, and fault information.For example: at 2:28, 02/04/2006, the state code is 77 indicating that a gearbox oil over temperature alarm occurred and wind turbine shut down.In the wind turbine operation handbook, if the gearbox oil temperature is above 80 Celsius and this condition lasts for 60 s, the wind turbine will be shut down.When the oil temperature cools below 65 Celsius, the wind turbine will start up again.
Autoassociative Kernel Regression Model Construction
In this paper, the Autoassociative Kernel Regression (AAKR) method is used to model the normal behavior of gearbox temperature.AAKR is based upon the multivariate inferential kernel regression derived by Wand and Jones in [12].It is now mainly used in nuclear power plant sensor calibration [13].
AAKR is a non-parametric, empirical modeling technique that uses historical, fault free observations to estimate the output of a process or device.Let there be n variables of interest in a process or device.At time i, a single observation of the variables can be written as an observation vector: Construction of a memory matrix D is the first step of AAKR modeling.In a period of normal operation of the process or device, m historical observation vectors are collected covering the range of different operating conditions (such as high or low load, start up, before shut down, etc.) to construct the memory matrix, denoted as: (2) ( ) Each observation vector in the memory matrix represents a measured operating state of the process or device.With proper selection of the m historical observation vectors from an extended period of normal (un-faulted) operation of the process or device, the subset space spanned by the memory matrix D can be taken to represent the whole normal working space of the process or device.The construction of memory matrix is actually the procedure of learning and memorizing the normal behavior of the process or device analogous but different from the training of ANN.
During subsequent operation, the input to AAKR at each time step is a new observation vector X obs , and the output from AAKR is an estimate X est for this input vector for the same moment in time.First, the distance between the observation vectors X obs and each vector X(i) in the memory matrix is computed.There are several distance functions that may be used [14], but the most commonly used function is the Euclidean distance, as follows: For the new observation vector, this calculation is repeated for each vector in the memory matrix, and resulting in a m × 1 vector of distances: Next, these distances are transformed to similarity measures used to determine weights by evaluating the Gaussian kernel, expressed by: where h is the kernel bandwidth, W are the weights for the m vectors in the memory matrix.Finally, these weights are combined with the memory matrix D to make estimate according to: If the scalar a is defined as the sum of the weights, i.e., Then Equation ( 6) can be represented in a more compact matrix notation: When the process or device works normally, the input observation vectors of the AAKR are most likely to be located in the normal working space that is represented by the memory matrix D, in that it is similar to some historically measured vectors in the memory matrix.As a result, the estimate of AAKR will have a very high accuracy.When problems arise with the process or device, its dynamic characteristics will change, and the new observation vector will deviate from the normal working space.In this case the linear combination of the historical vectors in the memory matrix will not provide an accurate estimate of the input and the residual will increase in magnitude.
The Variable Selection for Gearbox Temperature AAKR Model
In order to construct the AAKR model of gearbox temperature, the variables included in the observation vector should be carefully chosen.Because we are concerned with the gearbox temperature, variables that have close relationship with gearbox temperature should be taken into account.Following a review of the 47 variables recorded by the SCADA system, the following five variables were selected to construct the observation vector.
(1) Power (P); power has a great influence on the gearbox temperature.When the power is high, the gearbox will endure a high load which leads to high gearbox temperatures.
(2) Wind speed (V); Wind is the energy source for wind turbines.For the variable speed wind turbine, in order to reach maximum power point tracking, the drive train rotating speed is proportional to the wind speed.The higher the wind speed, the higher the rotation speed of the gearbox which also leads to high gearbox temperatures.
(3) Ambient temperature (T); Because the local temperature that the wind turbine experiences changes greatly in the short term (from day to night for example) and in the longer term (weeks to months) due to passing weather systems and seasons it must be taken into account explicitly.At Zhangjiakou during March and April, ambient temperature changes can be as large as 30 Celsius because of fast moving weather fronts.
(5) Gearbox Temperature of last time (T G1 ).Because the gearbox is a closed structure and the gearbox temperature has big inertia, the gearbox temperature of the last sampling time has an important influence on the current moment gearbox temperature.
The time span for the gearbox temperature model should take the local climate into account.The wind turbine studied in the paper is located at the Zhangjiakou area, north of Beijing, and the ambient temperature and wind speed distribution are quite different from season to season, as shown in Figure 2 and Figure 3.Such meteorological parameters have a great influence on the operation of the gearbox.For example, in winter, the temperature can be as low as minus 20 Celsius and frequently results in gearbox lubrication problems.If the normal behavior AAKR model covers a long period such as from January to July, it accuracy will be quite low for two reasons.One reason is that in different seasons the gearbox operating conditions are quite different as the result of these meteorological parameters.The other reason is that if the time span for model is too long.The range of each above variable will be very large that also leads to low model accuracy.In order to get satisfactory model accuracy, the time span for AAKR should be one month or one season.
Construction of Memory Matrix D
Because the five variables in the observation vector have different units and their absolute values are quite different, the initial values of the five variables are rescaled to the range [0 1] according to their maximum and minimum values.After rescaling, each variable has same weight in the calculation of Euclidean distance.Especially for wind speed, because the wind turbine's cut-in and cut-out wind speeds are 3 m/s and 12 m/s, respectively, wind speeds below 3 m/s and above 12 m/s should be set as 3 m/s and 12 m/s before scaling.
In previous work [15], two algorithms are used to extract observation vectors from the normal working period to construct the memory matrix.In the first algorithm, vectors are selected that correspond to the extreme normal working states.For each variable in the observation vector, the algorithm finds the minimum and maximum measurements from the normal working period.The observation vectors containing these measurements are added to the memory matrix.In the second algorithm, the remaining observation vectors from the normal working period are ordered by their Euclidean norms.For n dimensional vector X, the Euclidean norm is: The algorithm then selects evenly spaced elements from the ordered set and adds their corresponding observation vectors to the memory matrix.This construction method is simple, but it has some problems.Observation vectors can exist with Euclidean norm values that are similar or even exactly the same, but the vectors may be quite completely different, such as for the following two vectors: Selecting one of such similar Euclidean norm equivalent vectors will result in other equivalent norm vectors being discarded.Then only the working space near the selected observation vector is covered by the memory matrix, while parts of the parameter space near the discarded ones will be neglected.In order to minimize this effect a new memory matrix construction method is proposed that builds on the method used in [15] to give much improved NSET model accuracy.The new method is as follows.
Assume the historical observation vectors during the normal working period are (1), (2), , and make the matrix: (2) ( ) The number of vectors in matrix K is M.Each observation vector includes power, wind speed, ambient temperature, gearbox temperature, gearbox temperature of last time, denoted 1 2 5 , , , x x x respectively, and n in Equation (1) takes the value n = 5.Every observation vector in the normal working space covers the five variables and is normalized as described above.In order to ensure that memory matrix D covers the vectors with different variable values in the normal working space, for each of the five variables, the range [0 1] is equally divided into 100 sections and the observation vectors from matrix K selected at steps of 0.01.Observation vectors at each step of 0.01 for each variable in turn are added into memory matrix D so long as the variable in question is sufficiently close to the value in the observation vector.For example, for variable x 1 (turbine power), the method of adding observation vectors to D is shown in Figure 5.In Figure 5, δ is a small positive number, taken to be 0.001 in this work.With this method, observation vectors with different variable values can be included in the memory matrix D. Memory matrix construction thus has two steps: the first step uses the method in [15] to select observation vectors from the normal working period.The second step uses the new method outlined in this paper to add further observation vectors to D. Before adding a new vector to the memory matrix, the Euclidean distances between the new potential vector and the vectors already in the memory matrix are checked.If the distance is too small, that means the memory matrix already has a vector that is quite similar to the one being considered and consequently it should not be added into the memory matrix.This check will limit the size of memory matrix and ensure that it is well-conditioned.
After the memory matrix construction is complete, the AAKR model is ready to estimate the gearbox temperature for the current moment measurement.For comparison, the reader is referred to the performance of a three-layer neural network for SCADA data modeling shown in [16].It is quite complex and time consuming.
Moving Window Calculation of Residual Mean Value and Standard Deviation Statistics
When the wind turbine gearbox suffers from some abnormality, the new observation vector will deviate from the normal working space and the distribution and time development of the estimate residual will change significantly from the normal condition.The mean value and standard deviation will reflect a change in the distribution of the residuals.In order to detect the changes in the variables in a timely manner, a moving average calculation is used.At a certain instant, the residual sequence of gearbox temperature from the AAKR model is: where x 4 is the gearbox temperature measurement in the observation vectors, and 4 x is the AAKR model estimate for x 4 .
A time window with width N is adopted to calculate the moving average or mean value and standard deviation for the N successive residuals in the window: The moving window is shown in Figure 6.
Residual Statistical Distribution when Gearbox Has an Abnormality
When the gearbox works normally, the AAKR model provides very accurate estimates of gearbox temperature.The residual sequence has a mean value near zero and the standard deviation is small.When a problem occurs with the gearbox, the new observation vector may deviate from the normal condition and the gearbox temperature residual distribution will thus also change.Abnormal gearbox operation can be identified as follows: (1) The residual mean value remains near zero, but the standard deviation increases dramatically.In this condition, the distribution of the residuals becomes wider.
(2) The residual mean value deviates from zero by some obvious magnitude and the standard deviation remains small.In this condition, the residual systematically departs from zero.
(3) A combination of the above two situations.
In order to detect early gearbox faults, failure thresholds are needed for both the residual mean value and its standard deviation.Assume that the thresholds for them are set respectively as E Y and S Y where these are determined according to operator experience or determined through model validation as presented below.
The residual sequence has been obtained using observation vectors from the validation set as input to the already identified AAKR model with subsequent application of the moving average window to the time sequence of residuals.By trial and error the optimal size of the moving window is determined.The maximum absolute mean value and standard deviation for validating set residual sequence are, as defined above, respectively E V and S V but note that these will be dependent on averaging window size.Then the thresholds for gearbox failure detection are as follows: where k 1 and k 2 are positive coefficients and can be chosen based on operator experience.Relatively larger k 1 and k 2 will increase the robustness of CM method and reduce the false alarms.In this paper, k 1 = 3 and k 2 = 2.During this period, three shutdowns of this wind turbine occurred, as shown in Figure 7 and Table 1.In Figure 8, at some isolated positions, the residuals are much larger than other positions.These large residuals are in pairs, and the number of the large residual pairs is three.The positions and reasons for them are as given in Table 2. 2 where the wind turbine would shut down or just start up, the relationship between the five variables in the observation vectors is quite different from the normal working condition (for example, at position 160, the wind speed is high, while the power near zero due to shutdown), and the observation vectors at these positions deviate from the normal working space represented by the memory matrix D, the combination of the historical observations in the memory matrix cannot provide an accurate estimate for the gearbox temperature at these positions and the residuals become quite large.These large residuals caused by the shutdown or startup of the wind turbine in some extent prove that the AAKR model is working and they should not be recognized as an indication of incipient gearbox failure.
Gearbox AAKR Model Validation and Failure Case Studies
After removing the above six large residuals, the AAKR model has a very high estimate accuracy, and most absolute residual values are below 0.05.The validation results shows that AAKR has satisfied the modeling accuracy for gearbox temperature dynamic characteristics.
The moving window calculation outlined in Section 4.1 is applied to the residual sequence of the validation set in Figure 8.The window width N should be properly selected so that the influence of the occasional isolated large residual caused by the imperfect coverage ability of memory matrix D can be minimized.At the same time, a moving window with a properly selected N must be able to detect the changes of mean value and standard deviation in a quick and effective manner.In this paper, the window width N = 50 reflects a useful balance between these two conflicting requirements.Figure 9 shows the trends of the mean value and standard deviation for the validating set after moving average filtering.It should be noticed that the six large residuals caused by turbine shutdown and startup in Table 2 have been removed during moving window averaging.
From the trends, we estimate:
Gearbox Failure Analysis
Actual failure of major components like gearboxes is relatively rare and sometimes it is also very difficult to access these components' failure data and maintenance records as the operator or manufacturer usually keep them as commercial secrets.From [5][6][7], we know that abnormal temperature changes or rises are an effective indication of incipient gearbox failure.In [17], authors gives a proof that the gearbox temperature rise will be proportional to the power output if the gearbox works normally, that is, the gearbox transmission efficiency has not changes.When an incipient fault occurs in the gearbox, it efficiency decreases, and the gearbox temperature will have extra increase.Unfortunately, in the SCADA data we have, there is no gearbox failure record.Based on [5][6][7]17], it is representive to simulate a real incipient gearbox failure by adding extra temperature drift on the initial SCADA data.And these manual drift data is used to test the effectiveness of this AAKR CM method in the flowing two cases.In Figure 13, the standard deviation exceeds the failure threshold at position 462, and only 462 + 50 − 501 = 11 points after the white noise is added.In this case, the AAKR CM method also detects the simulated abnormality on time.
Discussion and Conclusions
This paper uses Autoassociative Kernel Regression (AAKR) to construct a gearbox temperature model.Details of how to construct the required memory matrix are provided.The method has the advantage of being computationally remarkably simple and should thus have immediate appeal for wind industry practitioners.When the gearbox experiences a fault, new observation vectors will generally deviate from the normal working space and the AAKR estimate of the residual distribution and its time development will change.A moving average window filter has been adopted to reveal the trends in mean value and standard deviation of the residual sequence; this is also very easy to implement computationally.Gearbox condition can be determined from crossing of predetermined thresholds.With the effective selection of the memory matrix from data from normal operation, the presented method can identify simulated wind turbine gearbox failures in a timely way.With suitable selection of the parameters available to the user (window size, vector parameter selection, thresholds etc.), the method should prove useful for wind turbine condition monitoring more widely.There is still a small flaw in this paper because the validation was carried out with the man-made gearbox failure data.In future research, real failure data will be collected and this CM method further validated.
Figure 1 .
Figure 1.Structure of a gearbox for a wind turbine.
Figure 2 .
Figure 2. Ambient temperature in different seasons.
Figure 3 .
Figure 3. Wind speed distribution in different seasons.
Figure 4 .
Figure 4. Gearbox condition monitoring using the AAKR model.
Figure 5 .
Figure 5. Construction of memory matrix D for variable x 1 .
Figure 6 .
Figure 6.Residual moving average window used in residual analysis.
5. 1 .
Gearbox AAKR Model Validation SCADA data of one wind turbine for April 2006 is used in this paper.The 10-min data trends of the five related variables mentioned in Section 3.2 from 01/04/2006 to 06/04/2006 are shown in Figure 7.
But from the maintenance log, there was no record indicating that the gearbox had a failure or had been repaired.It is quite usual for the wind turbine to shut down due to some parameter alarms.During April 2006, the gearbox worked properly.The total 10-min SCADA data point number for April is 4320, and among them, the useful data number (data with power above zero) is 3731.Normal working data from 07/04/2006 to 30/04/2006 is used to construct the gearbox temperature normal behavior AAKR model.Data from 01/04/2006 to 06/04/2006 shown in Figure 7 is used as the validating set to validate the AAKR model.In April 2006, the maximum and minimum gearbox temperatures were 74.1 Celsius and 50.2 Celsius, respectively.The maximum and minimum ambient temperatures were 20 Celsius and 13 Celsius, respectively.The validation result is shown in Figure8.Values for different variables are scaled values.When the turbine power in an observation vector is zero, the AAKR model will not work, resulting in zeros for estimate and residual.
Figure 8 .
Figure 8. Validation results for the AAKR model.
Figure 9 .
Figure 9. Statistical characteristics for the validation sequence.
Case 1 :
A fixed temperature rise of 0.001 per time step is added at position 501 to the initial gearbox temperature from 01/04/2006 to 06/04/2006 in the validation set.The AAKR model estimate and the residual sequence are shown in Figure 10.The trends for mean value and standard deviation for the above residual sequence are shown in Figure 11.
Table 2 .
Large residuals' positions and reasons. | 6,018.6 | 2011-11-23T00:00:00.000 | [
"Engineering"
] |
Attention on Aromatic Rice Production May Create a New Era on Export Income in Bangladesh: An Opinion
Aromatic rice is called due to its novel 2-Acetyl-1-pyrroline (2-AP), which is also named as fragrant rice for its magnificent aroma [1]. Upon cooking the aromatic rice gives a pleasant flavor, aroma and taste in our mouth and nose. Mainly the aromatic rice is eaten for its special aroma not for starch granule as in plain rice. The Peoples of most Asian countries are more interested to intake the fragrant rice in daily food, and in different occasion as plain aromatic rice and sticky rice. The long, slender, sticky and softness are the most crucial traits for fragrant rice [2], like Jasmine of Thailand and Basmati of Indian sub-continent. The peoples of Bangladesh choice a long, slender and dry trait for aromatic rice after cooking and interested to make different delicious foods stuffs. Recently, Bangladesh has started a new journey to produce some aromatic rice by keeping the export goals and exported some rice in Sri Lanka. In Bangladesh, Sharpers, Sunamgonj, Dinajpur, Rangpur are getting attention due to its edaphic condition as special areas for growing aromatic rice and there have some local and released cultivars and varieties. Among these kalojira, radhunipagol, sarnolota, tulsimala, ciniatob, khaishanne etc. as local cultivars and BRRI dhan34, BRRI dhan50 and BRRI dhan80 etc. as variety are much prominent.
Introduction
Aromatic rice is called due to its novel 2-Acetyl-1-pyrroline (2-AP), which is also named as fragrant rice for its magnificent aroma [1]. Upon cooking the aromatic rice gives a pleasant flavor, aroma and taste in our mouth and nose. Mainly the aromatic rice is eaten for its special aroma not for starch granule as in plain rice. The Peoples of most Asian countries are more interested to intake the fragrant rice in daily food, and in different occasion as plain aromatic rice and sticky rice. The long, slender, sticky and softness are the most crucial traits for fragrant rice [2], like Jasmine of Thailand and Basmati of Indian sub-continent. The peoples of Bangladesh choice a long, slender and dry trait for aromatic rice after cooking and interested to make different delicious foods stuffs. Recently, Bangladesh has started a new journey to produce some aromatic rice by keeping the export goals and exported some rice in Sri Lanka. In Bangladesh, Sharpers, Sunamgonj, Dinajpur, Rangpur are getting attention due to its edaphic condition as special areas for growing aromatic rice and there have some local and released cultivars and varieties. Among these kalojira, radhunipagol, sarnolota, tulsimala, ciniatob, khaishanne etc. as local cultivars and BRRI dhan34, BRRI dhan50 and BRRI dhan80 etc. as variety are much prominent.
Limitation
However, in spite of having so much attractive traits in aromatic rice, the growers of Bangladesh has not been able to produce targeted amount of aromatic rice to export due to production and storage limitation. The climatic factors like temperature, humidity, rainfall etc., and managerial factors like soil nutrients, growing condition, growth regulators, harvesting time, drying methods, drying periods, storage condition, storage containers etc. might be mediated the aroma content in fragrant rice. Lacking of processing policy and plant are also a most limiting factor for aromatic rice storage and expansion in Bangladesh. The government of Bangladesh did not make any crop incentives yet on aromatic rice cultivation over plain rice for the growers. The aim of this opinion article is to explore the major production and storage techniques viewed from my observations to improve the retention of aroma in fragrant rice.
Approaches
Upon follow the strategies by which we could be able to improve the total aroma content in rice grain and knowledge's for cultivation of aromatic rice as a whole.
a) Field approach
Basmati rice has great interest as aromatic rice in Indian sub-continent. But from my observation it can be said that, Radhunupagol, local fragrant rice cultivated specially in Dinajpur and Rangpur district in Bangladesh for its pleasant aroma and flavor which could be usable as substitute for Basmati rice. The application of minimum tillage may increase only MCDA.000631. 6(2).2020 the amount of aroma in rice but more ploughing may increase both the aroma and yield of aromatic rice [3]. The stages of rice plant might be influenced the content of aroma in rice. Among different stages of rice, the aroma was found highest in seedling stage due to its more green leaves and the aroma was also found more in grain filling stage. Findings demonstrated that rice plant treated with growth regulators inhibited the metabolic processes associated with the formation of volatile compounds like 2-AP. In grains, 2-AP content could be increased by the application of some water as irrigation at tilling, booting, and grain filling stage; the observation is supported by [4]. Studies showed that salt stress can increase 2-AP content in grains [4]. So this knowledge can be usable in saline prone areas of Bangladesh where salinity is a major problem for normal crop production. Grain aroma content of aromatic rice could be improved by applying additional N at booting stage while considering the amount of N fertilizer added [5]. From my observations it might be said that, application of Manganese (Mn) and (Zn) had a positive effect on 2-AP content in aromatic rice. From the finding of [4] it may be said that, 2-AP content was increased by moderate concentration of lanthanum (La). The application of silicon could increase the content of grain aroma [6]. We found that sometimes the induction of moderate drought stress could increase the aroma content in rice in field during grain filling stage. We have seen that, at the areas of moderate shad in field the aroma content was higher than that of sunny areas at the later stage of rice plant. Research of [7] also supported that during grain filling period, shading significantly improved the 2-AP content in aromatic rice grains.
b) Harvesting and drying approach
The farmers of Bangladesh conventionally harvest the rice grains at long after the optimum time of harvesting by keeping the views of full maturity. As a result, due to long time staying in field naturally the aroma of grains is lost by internal field respiration by excessive temperature from sunlight. Then the aroma of grains are showed loser in concentration during storage and marketing which resulted lowest economic returns for growers. So, the aromatic rice grains must be harvested at optimum time of harvest (at 70-80% maturity for reducing the opportunity of aroma losses). High temperature during conventional semi-concrete floor/earthen floor drying system could reduce the concentration of 2-AP in Bangladeshi fragrant rice. So, an optimum drying temperature (10-15 °C) should be maintained for aromatic rice to store.
c) Storage approach
After drying the amount of rice aroma (2-AP) can be reduced upon storage due to internal respiration by moisture, temperate and relative humidity. Due to oxidation, rice grains with higher levels of moisture content reduced the 2-AP concentration more rapidly than the lower levels of moisture content. At 15 degree centigrade, 2AP can be maintained in higher amount than in ambient condition (25-35 °C). The amount of 2AP will be reduced by 50% even when using low temperature to dry fragrant rice and storing them in ambient condition compared to 15 degree centigrade [8]. The vacuum bag packaging increased the concentration of 2-AP in aromatic rice than polythene bag, gunny bag and cloth bag upon storage.
Conclusion
From this opinion article it can be concluded that if the growers of aromatic rice in Bangladesh is able to use these approaches so that they will get higher production of aromatic rice with higher pleasant aroma in grains after harvest and also in storage. Moreover they could purchase their rice grains with higher prices in local market and also cloud be exported surplus amount of production by meeting local demand and could take part in national economy as a great sector under Agricultural GDPs. | 1,846.6 | 2020-05-04T00:00:00.000 | [
"Economics"
] |
Electro-oxidation of formoterol fumarate on the surface of novel poly(thiazole yellow-G) layered multi-walled carbon nanotube paste electrode
The current study explicates the electro-oxidation behavior of formoterol fumarate (FLFT) in the presence of uric acid (UA) on the surface of poly thiazole yellow-G (TY-G) layered multi-walled carbon nanotube paste electrode (MWCNTPE). The modified (Poly(TY-G)LMWCNTPE) and unmodified (MWCNTPE) electrode materials were characterized through electrochemical impedance spectroscopy (EIS), field emission scanning electron microscopy (FE-SEM), and cyclic voltammetry (CV) approaches. The characterization data confirms the good conducting and electrocatalytic nature with more electrochemical active sites on the Poly(TY-G)LMWCNTPE than MWCNTPE towards the FLFT analysis in the presence of UA. Poly(TY-G)LMWCNTPE easily separates the two drugs (FLFT and UA) even though they both have nearer oxidation peak potential. The electro-catalytic activity of the developed electrode is fast and clear for FLFT electro-oxidation in 0.2 M phosphate buffer (PB) of pH 6.5. The Poly(TY-G)LMWCNTPE offered a well-resolved peak with the highest electro-oxidation peak current at the peak potential of 0.538 V than MWCNTPE. The potential scan rate and oxidation peak growth time studies show the electrode reaction towards FLFT electro-oxidation is continued through a diffusion-controlled step. The variation of concentration of FLFT in the range from 0.2 to 1.5 µM (absence of UA) and 3.0 to 8.0 μM (presence of UA) provides a good linear relationship with increased peak current and a lower limit of detection (LOD) values of 0.0128 µM and 0.0129 µM, respectively. The prepared electrode gives a fine recovery for the detection of FLFT in the medicinal sample with acceptable repeatability, stability, and reproducibility.
Fabrication of MWCNTPE and Poly(TY-G)LMWCNTPE. The MWCNTPE was fabricated by blending
60% of MWCNTs with 40% of silicone oil for around 20 min to get a fine paste. The developed paste was sorted into the vacant outlet (3.0 mm diameter) of a Teflon tube and smoothly rubbed on tissue paper to get a smooth surface.
The Poly(TY-G)LMWCNTPE was fabricated through the electro-polymerization process. Here, the electropolymerization of 1.0 mM TY-G in PB (0.2 M and 7.0 pH) was done on the external surface of MWCNTPE by cycling twenty CV segments having the potential range of − 0.60 to 1.40 V at the potential scan rate of 0.1 V/s. Later, the prepared electrode surface was rinsed carefully in DW to eliminate the unreacted monomer content.
Electrochemical setup. Electrochemical measurements were accomplished for 0.01 mM FLFT in 0.2 M PB. The working electrode (MWCNTPE and Poly(TY-G)LMWCNTPE) immersion length was set to 1.0 cm. Each voltammetric cycle was recorded in the potential gap from 0.2 to 0.9 V against Hg 2 Cl 2 at the potential scan rate of 0.1 V/s. This analysis was performed at 25 ± 2 °C.
Results and discussion
Electrode material characterization. FE-SEM study. The electrode surface morphological study was carried out through the FE-SEM technique. FE-SEM pictures show the surface morphology of MWCNTPE and Poly(TY-G)LMWCNTPE. Figure 1a picturizes the arbitrarily fused thin and long root-shaped tubes with erratically distributed gaps on the exterior of the electrode. Also, those tubes and gaps didn't show any of the roofed films on their exterior (picture having 100 nm magnification), represents the surface of unmodified MWC-NTPE. Furthermore, Fig. 1b displays the FE-SEM morphology of modified Poly(TY-G)LMWCNTPE. Here, each MWCNTs and erratically distributed gaps are roofed by thin deposition of an electro-polymerized film of TY-G. Moreover, the width of Poly(TY-G)LMWCNTPE is improved with sufficient porosity after the electropolymerization of TY-G associate with MWCNT alone.
EIS study. EIS is the significant tool for the measurement of the interfacial charge transmission character of the electrode material components at the interface of the supporting electrolyte. The achieved results of EIS corresponding to MWCNTPE (Line 'b') and Poly(TY-G)LMWCNTPE (Line 'a') are presented using Nyquist plots (Inset Fig. 2). EIS experimentation was conducted in the supporting electrolyte KCl (0.1 M) having K 4 [Fe(CN) 6 ]·3H 2 O (1.0 mM) at 0.06 V of working potential and 0.005 V of amplitude in the frequency range of 1.0 Hz to 100 kHz. As from the Nyquist plots, MWCNTPE affords a larger semicircle as compared to Poly(TY-G) LMWCNTPE, which specifies that the electron transfer nature of the MWCNTs is substantially enhanced by the deposition of electro-polymerized TY-G. Figure 3 shows the cyclic voltammograms (CVs, at the potential scan rate of 0.1 V/s and potential width of − 0.2 to 0.6 V) for K 4 [Fe(CN) 6 ]·3H 2 O at the surface of MWCNTPE (Line 'a') and Poly(TY-G)LMWCNTPE (Line 'b') in optimum conditions. It can be displayed that, the Poly(TY-G)LMWCNTPE reveals a greater redox peak current in association with unmodified MWCNTPE. The electro-active area of MWCNTPE and Poly(TY-G)LMWCNTPE surfaces was computed on the application of following Randles-Sevcik's equation 29,32 , Figure 4a displays the plot of oxidation peak current versus the number of electro-polymerization cycles. Here we observed the amplified oxidation peak current of FLFT from 5 to 10 cycles, on the other hand from 10 to 25 cycles the FLFT oxidation peak current gradually decreased, which is most feasibly due to the sufficient coverage of poly(TY-G) film on the accessible surface area of the MWCNTPE. Here, ten polymerization cycles show sufficient coverage on the surface of MWCNTPE and provide a faster rate of electron transfer with optimum peak current and the reduction of background current. As a result, ten polymerization cycles were chosen as optimum cycles for this experimentation. Figure 5b espied that the electro-oxidation peak potential of FLFT was transferred towards a shorter potential zone (negative shift) as the intensification of PB pH from 6.0 to 8.5. This incident specified that on the surface of Poly(TY-G)LMWCNTPE the electro-oxidation process of FLFT was a protonated electron transfer. Hence, the smaller number of protons participation supports the process of electro-oxidation with less current sensitivity at elevated PB pH (6.5 to 8.5). Fine linearity is observed between peak potential and PB pH values (plot of E pa v/s pH) and the fitted linear regression equation is as follows, where, 'R 2 ' is the regression coefficient. Here, the slope value 0.055 V/pH is most nearer to the theoretic value (0.059 V/pH), signifying that the electro-oxidation of FLFT associates with an alike number of protons and electrons on the surface of Poly(TY-G)LMWCNTPE. Supporting this, the number of protons that participated in the electro-oxidation reaction of FLFT was determined using the following Nernst relation, where, 'B' is the slope of Eq. (2), 'm' is the number of protons that participated in the FLFT oxidation reaction, 'n' is the number of electrons participating in the FLFT oxidation reaction and other terms (R, T and F) represents their standard physical values. The calculated value of the number of protons was found to be 1.86 ≈ 2. Furthermore, the pKa value of FLFT after the electro-oxidation was detected from the intersection point displayed in the plot of E pa v/s pH (Fig. 5b). FLFT displays the pKa value of 7.0, indicating that the oxidation process most probably has taken place at the phenolic group of FLFT. From Fig. 5c, the electro-oxidation peak current of FLFT was very high at acidic pH of 6.5. These data are resolute that acidic pH (6.5) is most appropriate for the highly sensitive sensing of FLFT due to the strong interaction between Poly(TY-G)LMWCNTPE surface and FLFT. Therefore, the pH value of 6.5 was chosen as the optimal pH for this investigation.
Scan rate impact on peak potential and current of FLFT at Poly(TY-G)LMWCNTPE. The potential scan rate impact on the peak potential and current of FLFT at Poly(TY-G)LMWCNTPE deliver a key evi- www.nature.com/scientificreports/ dence for the electro-oxidation mechanism and the kinetic behaviors of the equipped electrode. The CV executions were documented and shown in Fig. 6a for 0.005 mM FLFT in PB (0.2 M and 6.5 pH) on Poly(TY-G) LMWCNTPE at variable potential scan rate in the range between 0.05 and 0.25 V/s, which displays the increased electro-oxidation peak currents and potentials of FLFT. The kinetic nature and dependability of the modified electrode towards the oxidation reaction of FLFT were verified based on acceptable linear plots of the log value of the electro-oxidation peak current versus log value of the potential scan rate (Fig. 6b) and electro-oxidation peak current versus square root value of the potential scan rate (Fig. 6c). The obtained results are fitted in the following linear relations are as follows, The slope (0.412 ± 0.023) of Eq. (4) and the R 2 value of Eq. (5) are very adjacent to the required theoretical values suggesting that the kinetic behavior of the electro-oxidation of FLFT on Poly(TY-G)LMWCNTPE has proceeded through a diffusion-controlled pathway 26 . The plot of electro-oxidation peak potential versus log value of the potential scan rate (Fig. 6d) provides fine linearity and the results are fitted in the following linear relation is as follows, The number of electrons that participated in the electro-oxidation reaction of FLFT was confirmed using the slope value (0.045 ± 0.002) of Eq. (6) www.nature.com/scientificreports/ where, 'E pa/2 ' is the electro-oxidation peak potential at the half peak current of FLFT, 'α' is the electrochemical charge transfer coefficient, 'n' is the number of involved electrons, 'υ' is the applied potential scan rate, 'k 0' is the electrochemical heterogeneous rate constant, and other terms (R, T, and F) represents their standard physical values. The premeditated value of a number of electrons was found to be 2.26 ≈ 2. The possible electro-oxidation reaction of FLFT is shown in Scheme 2. The electrochemical heterogeneous rate constant for the oxidation of FLFT was calculated by deducing Eq. (7) and the premeditated value of 'k 0 ' was found to be 1.818 s −1 .
The electrochemical surface coverage concentration (Г) of FLFT at the surface of MWCNTPE and Poly(TY-G)LMWCNTPE is calculated using the following relation 36 , where, 'Q' is the integrated electrical charge of the oxidation peak. The calculated value of electrochemical surface coverage concentration of FLFT for the MWCNTPE was 1.286 ÅM/cm 2 and for the Poly(TY-G)LMWCNTPE was 2.802 ÅM/cm 2 . This consequence is the other key factor for the high electro-oxidation peak current of FLFT at Poly(TY-G)LMWCNTPE.
Oxidation peak growth time. The influence of peak growth time is the prominent tool to disclose the maximum accumulation point (higher peak current) of FLFT on the surface of Poly(TY-G)LMWCNTPE with its kinetic behavior at the variable growth time. The peak growth time has been optimized using the DPV experimentation by altering the growth time within the range from 0 to 100 s for 0.005 mM FLFT in PB (0.2 M and 6.5 pH) and the outcomes are presented in Fig. 7. Here the observation suggests that the growth time of 20 s (exists at the second position) provides the maximum peak current than other growth times (0 s, 40 s, 60 s, 80 s, and 100 s). This consequence is most probably due to the effect of saturation followed by diffusion kinetic behavior In the presence of FLFT, a fine electro-oxidation peak is viewed at the potential of 0.514 V as compared to bare MWC-NTPE, which shows a less enhanced electro-oxidation peak at the potential of 0.569 V. Furthermore, modified electrode provides elevated electro-oxidation peak current and lesser electro-oxidation peak potential for FLFT than the bare MWCNTPE delivers an exceptional electro-catalytic potential. This shift in both the current and potential at the modified electrode is most conceivably due to the various interactions between FLFT molecules and poly(TY-G), such as π-π covalent, hydrogen bonding, electrostatic, and dipole-dipole interactions, shown in Scheme 3 37 . Furthermore, in the absence of FLFT in PB (0.2 M and 6.5 pH), the surface of Poly(TY-G)LMWC-NTPE didn't show any electro-catalytic oxidation response.
Calibration plot for FLFT with and without UA. The detection capability of the prepared Poly(TY-G) LMWCNTPE towards FLFT was verified based on the concentration variation method in the presence and absence of UA using the DPV technique (Inset Fig. 9). Figure 9a shows the documented differential pulse voltammograms (DPVs) of different concentrations of FLFT in the range from 0.2 to 3.0 μM in PB (0.2 M and pH 6.5) at the surface of Poly(TY-G)LMWCNTPE. Here, the electro-oxidation peak current of FLFT enhanced as the concentration of FLFT increased (Fig. 9b); also, the preliminary concentration range 0.2 to 1.5 μM gives a finer linear correlation and the achieved results are fitted in the following linear relation is as follows, Table 1 11,12,14,26,27 , proposing that the Poly(TY-G)LMWCNTPE provides the lowest/very nearer LOD.
The simultaneous examination of FLFT and UA in PB (0.2 M and pH 6.5) was performed using the DPV technique on the surface of Poly(TY-G)LMWCNTPE (Line 'b') and MWCNTPE (Line 'a') (Fig. 9c). The line 'b' discloses well and distinctive electro-oxidation peaks resultant to UA and FLFT at the electro-oxidation peak www.nature.com/scientificreports/ The determined LOD value was found to be 0.0129 µM, suggesting that the Poly(TY-G)LMWCNTPE delivers almost nearer LOD value even in the presence of interfering UA molecule. Hence, the prepared electrode is almost free from interference.
Compatibility of the Poly(TY-G)LMWCNTPE. The compatibility of the Poly(TY-G)LMWCNTPE
towards the electro-oxidation of FLFT was examined via the DPV method in the presence of some biologically available species (metal ions and organic molecules, each having a concentration of 1.0 mM). Figure 10 shows the graph of different chemical species (such as Ag + , Ba 2+ , Hg 2+ , Ca 2+ , Fe 2+ , K + , Na + , dopamine (DA), tryptophan (THR), tyrosine (TY), and curcumin (CU)) versus the % of electro-oxidation potential variation of FLFT. Here we observe only ± 5.0% of potential variation at the oxidation of FLFT, suggesting that Poly(TY-G)LMWCNTPE shows acceptable compatibility in presence of the above-mentioned chemical species. But in the case of electrooxidation peak current, the compatibility was little vary compared to the base peak current.
Analysis of FLFT in medicinal sample.
In supplement to the aforesaid studies, the Poly(TY-G)LMWC-NTPE sensing capability towards FLFT was determined in a medicinal sample based on the standard addition approach. As from Table 2 www.nature.com/scientificreports/ pH 6.5) via the DPV method by quantifying the current response at the Poly(TY-G)LMWCNTPE (initial current and current after one day). The calculated value of the percentage of degradation gives 90.72% of retained peak current affords a first-class electrode storage stability.
Conclusions
In this effort, the Poly(TY-G)LMWCNTPE and MWCNTPE were prepared simplistically and economically for the detection of FLFT in PB using the CV and DPV methodologies. The modification of poly(TY-G) on MWCNTPE was confirmed by FE-SEM, EIS, CV, electro-active surface area, and surface concentration measurements. The enhanced active surface area of the Poly(TY-G)LMWCNTPE improves the rate of electron and proton transfer during the electro-oxidation of FLFT. The CV outcomes validate different interactions between FLFT molecules and Poly(TY-G)LMWCNTPE, which significantly shifts the peak current to a higher level and peak potential to a lower level in the oxidation of FLFT. The impact of the potential scan rate study reveals the heterogeneous rate constant and diffusion kinetics (also supported by the effect of accumulation) with the transfer of 2 e − and 2 H + in the electro-oxidation of FLFT. The supporting electrolyte pH study shows that the proportion of electron and proton transfer in the electro-oxidation of FLFT is equivalent (1:1). The Poly(TY-G)LMWCNTPE implicates better electrochemical behavior (in presence and absence of UA) with fine linear response, superior sensitivity with lower LOD, first-class selectivity (in the presence of different chemical species), reproducibility, repeatability, and good storage stability. Additionally, the modified Poly(TY-G)LMWCNTPE provides excellent recovery (90.64 to 96.70%) for FLFT in a medicinal sample. All these results concluded that the Poly(TY-G) LMWCNTPE is a more optimistic and potent sensing tool in the analysis of many other electro-active molecules and medicinal samples in vision of its superior active surface area and electrocatalytic activity. | 3,659.6 | 2021-06-17T00:00:00.000 | [
"Materials Science",
"Chemistry"
] |
Evaluating the accuracy and relevance of ChatGPT responses to frequently asked questions regarding total knee replacement
Background Chat Generative Pretrained Transformer (ChatGPT), a generative artificial intelligence chatbot, may have broad applications in healthcare delivery and patient education due to its ability to provide human-like responses to a wide range of patient queries. However, there is limited evidence regarding its ability to provide reliable and useful information on orthopaedic procedures. This study seeks to evaluate the accuracy and relevance of responses provided by ChatGPT to frequently asked questions (FAQs) regarding total knee replacement (TKR). Methods A list of 50 clinically-relevant FAQs regarding TKR was collated. Each question was individually entered as a prompt to ChatGPT (version 3.5), and the first response generated was recorded. Responses were then reviewed by two independent orthopaedic surgeons and graded on a Likert scale for their factual accuracy and relevance. These responses were then classified into accurate versus inaccurate and relevant versus irrelevant responses using preset thresholds on the Likert scale. Results Most responses were accurate, while all responses were relevant. Of the 50 FAQs, 44/50 (88%) of ChatGPT responses were classified as accurate, achieving a mean Likert grade of 4.6/5 for factual accuracy. On the other hand, 50/50 (100%) of responses were classified as relevant, achieving a mean Likert grade of 4.9/5 for relevance. Conclusion ChatGPT performed well in providing accurate and relevant responses to FAQs regarding TKR, demonstrating great potential as a tool for patient education. However, it is not infallible and can occasionally provide inaccurate medical information. Patients and clinicians intending to utilize this technology should be mindful of its limitations and ensure adequate supervision and verification of information provided. Supplementary Information The online version contains supplementary material available at 10.1186/s43019-024-00218-5.
Introduction
Total knee replacement (TKR) is one of the most common elective orthopaedic procedures performed today [1], helping countless patients with knee arthritis achieve improvements in pain, function and quality of life [2].
As the demand for and volume of TKRs rise, an increasing number of patients are turning to the internet for information regarding this procedure [3,4].Prior research has shown that up to two-thirds of patients considering elective orthopaedic procedures have used the internet as a source of information [4,5].This has coincided with the rising prominence of artificial intelligence (AI) chatbots such as Chat Generative Pretrained Transformer (ChatGPT) in recent years.Since its release in November 2022, ChatGPT has garnered great interest due to its ability to generate coherent and humanlike responses across a wide range of topics -surpassing 100 million monthly active users in just 2 months and setting the record for the fastest growing application in history [6][7][8].These AI chatbots leverage on machine learning techniques to study vast amounts of text from articles, books and webpages to identify patterns and structures of human language -allowing it to have wide-ranging applications including content generation, explaining complex concepts, and even taking and passing medical exams [9,10].
Given the widespread adoption of ChatGPT, it is foreseeable and inevitable that a significant proportion of patients may independently seek answers to their medical queries from ChatGPT due to its accessibility and ability to provide personalized responses [11].At the same time, some clinicians have also highlighted ChatGPT's potential as a tool to enhance patient education due to its vast knowledge-base and ability to generate coherent and original responses [12,13].Despite this, there remain legitimate questions and concerns regarding the accuracy and reliability of responses generated by ChatGPT, as some have observed that the chatbot may generate false and biased information or even conjure up non-existent sources in its responses [14].Furthermore, ChatGPT does not "reason" or "think" in a similar way to humans, instead generating responses based on recognized patterns and structures within the text it was trained with [15].As such, it is also important to evaluate the relevance of ChatGPT's responses -responses generated should be targeted and effective in answering the question at hand, rather than providing an excess of irrelevant information, which may overwhelm the patient.
Thus, our study aims to evaluate the accuracy and relevance of ChatGPT's responses to FAQs regarding TKR to assess its clinical utility as a tool for patient education and preoperative decision-making.Our hypothesis is that ChatGPT will be able to provide factually accurate and relevant responses to these FAQs.
Frequently asked questions (FAQ)
A list of 50 clinically relevant FAQs regarding TKR was curated after discussion with three consultant orthopaedic surgeons (WC, GL, and MT) and with reference from commonly asked questions regarding TKR on Google web search.Google web search is one of the most used search engines worldwide and it utilizes AI algorithms to recognize patterns in user queries, allowing Google to predict and suggest commonly associated queries regarding a topic [5,16,17].The search term "total knee replacement" was entered into Google web search on a newly installed browser to generate frequently associated questions under the "People also ask" box.
These FAQs were then classified into the following categories: (1) general/procedure-related, (2) indications for surgery and outcomes, (3) risks and complications of surgery, (4) pain and post-operative recovery, (5) specific activities after surgery and (6) alternatives and TKR variations (such as partial knee replacement, robotic TKR and bilateral TKR).
Evaluation of ChatGPT responses
Each FAQ was individually input as prompts to ChatGPT (version 3.5) accessed on an internet browser, with the first response generated for each prompt recorded.Next, two consultant orthopaedic surgeons (GL and MT) independently rated each response based on its factual accuracy as well as the relevance of the response (Table 1).Factual accuracy was defined as the degree to which the response was scientifically true and up to date as of June 2023, and it was graded using a Likert scale from 1 to 5 (1 -very inaccurate, 2 -inaccurate, 3 -somewhat accurate, 4 -accurate, 5 -very accurate).Relevance was defined as the degree to which the response was helpful and effective in answering the question and was similarly graded using a Likert scale from 1 to 5 (1 -very irrelevant, 2 -irrelevant, 3 -somewhat relevant, 4 -relevant, 5 -very relevant).In the event of significant disagreement between the two raters (defined as a difference of two or more grades on the Likert scale), a third consultant orthopaedic surgeon (WC) was involved to review the response and adjudicate to award a final grade.
Statistical analysis
Next, the ordinally rated responses were dichotomized using a threshold on the Likert scale (Table 1).For factual accuracy, responses were classified as accurate if they received an average or final grade of 4 or greater, whereas the rest of responses were classified as inaccurate.Similarly, for relevance, responses were defined as relevant if they received an average or final grade of 4 or greater, whereas the rest of responses were classified as irrelevant.Data analysis was performed using R software version 4.0.3(R Foundation for Statistical Computing, Vienna, Austria, 2019).Inter-rater reliability was measured using Gwet's AC2, as it has been shown to be a stable metric that is not significantly influenced by the distribution or prevalence of outcomes [18,19].
Overall performance
ChatGPT performed well overall, achieving a mean Likert grade of 4.6/5 for factual accuracy and 4.9/5 for relevance across all 50 questions.Overall, 44/50 (88%) of responses were classified as accurate and 50/50 (100%) of responses were classified as relevant.There was good inter-rater reliability as measured by Gwet's AC2, with coefficients of 0.85 for factual accuracy and 0.94 for relevance.Three responses had significant disagreement (defined as ≥ 2 on the Likert scale) between the two raters which required the involvement of a third rater.
General and procedure-related information
There were 9 FAQs relating to general and procedurerelated queries for TKR (Table 2).Of the responses, 7/9 (77.8%) were classified as accurate (mean grade 4.5), and 9/9 (100%) were classified as relevant (mean grade 4.9).Responses to two procedure-related questions: "Do I need to fast before a total knee replacement?" and "Will I be awake during a total knee replacement?"were assessed to be inaccurate, with an average Likert grade of 3.5 and 3, respectively.
Indications for surgery and outcomes
There were 7 FAQs regarding the indications for TKR and the outcomes from surgery (Table 3).These questions relate to the indications for TKR and addresses its benefits and postoperative outcomes.Of the responses, 7/7 (100%) were classified as accurate (mean grade 4.9), and 7/7 (100%) were classified as relevant (mean grade 4.9).
Risks and complications
There were 4 FAQs regarding the risks and complications from TKR (Table 4).Of the responses provided by ChatGPT, 4/4 (100%) were deemed to be accurate (mean grade 4.9), and 4/4 (100%) were deemed to be relevant (mean grade 4.6).
Pain and post-operative recovery
There were 13 FAQs regarding pain during and after surgery and the post-operative recovery process (Table 5).These questions address perioperative pain and mitigation strategies, as well as the typical expected recovery process of a patient undergoing TKR.Of the responses, 12/13 (92.3%) were deemed to be accurate (mean grade 4.7), and 13/13 (100%) were deemed to be relevant (mean grade 5.0).The response to one question pertaining to postoperative recovery -"How much weight can I put Table 2 General and procedure-related FAQs *Denotes responses where there was significant disagreement (≥ 2 on the Likert scale) between the two reviewers and the final grade was awarded by a third reviewer 1 Categorical outcome for accuracy, whereby accurate responses are defined as those with a mean or final grade of ≥ 4 on my operated leg after total knee replacement?"-was deemed to be inaccurate, with a mean Likert grade of 2.5.
Specific activities
There were 10 FAQs regarding the ability to perform specific activities such as walking, running and driving after TKR (Table 6).Of the responses, 10/10 (100%) were deemed to be accurate (mean grade 4.8), and 10/10 (100%) were deemed to be relevant (mean grade 5.0).
Alternatives/others
There were 7 FAQs regarding alternatives to TKR and variants of TKR such as bilateral TKR, robotic TKR and partial knee replacement (Table 7).Of the responses, 4/7 (57.1%) were deemed to be accurate (mean grade 4.1), and 7/7 (100%) were deemed to be relevant (mean grade 4.6).Responses deemed to be inaccurate include questions such as "Are there any alternatives to a total knee replacement?","What is robotic total knee replacement?" and "What is the benefit of robotic knee replacement?", with all three questions having a mean Likert grade of 3.5.
Discussion
Our [20,21].To our knowledge, our study is the first to critically evaluate ChatGPT responses for FAQs regarding TKR.Despite its promise, our results also highlight that ChatGPT is not infallible -in our study, 6/50 (12.0%) of responses were found to be inaccurate (inaccurate responses highlighted in Additional file 1: Table S1).Indeed, several other studies have also highlighted a tendency for ChatGPT to sometimes provide inaccurate or misleading information, and at times even generate plausible-sounding falsehoods in a phenomenon coined "artificial hallucination" [14,22,23].It is also important to highlight that ChatGPT is not capable of independent scientific reasoning and is only able to generate responses based on recognized patterns and structures in text it was trained with [15].Lastly, another major criticism is that ChatGPT is only trained with available data up to September 2021 and thus may not be able to provide updated, real-time information to users [12,23].While many of these drawbacks are inherent to the available training data and the technology itself, continuous advancements in AI technology will mean that the accuracy and reliability of such chatbots will gradually improve.GPT-4, the latest iteration of ChatGPT, which was recently released in March 2023, has been shown to have significantly better performance, increased accuracy and superior reasoning skills compared with its past versions [24][25][26].The introduction of plugins to GPT-4, which are additional functionalities from third-party applications, may also increase the utility and reliability of ChatGPT, allowing it to access up-to-date information from trusted sources such as peer-reviewed journals [27].However, we chose not to use GPT-4 in our current study, as currently GPT-4 is only available with a paid subscription and thus is not freely available to the general public.As such, we used GPT-3.5, as we wanted our study to be reflective of what most patients will be able to use on a daily basis.
Despite its potential drawbacks, there are areas where ChatGPT can contribute and even excel at.Being an AI chatbot that is adaptive and readily accessible, ChatGPT is well suited in providing personalized information and medical advice to patients [28,29].Currently, ChatGPT supports more than 50 different languages and is able to adapt its responses based on factors such as the user's age, education level and occupation (i.e.patients versus doctors) [30].Furthermore, some studies have also shown that patients may in fact prefer ChatGPT responses to those given by human clinicians -rating its responses as significantly more empathetic [11].Although direct supervision by a human clinician is still needed due to ChatGPT's potential for mistakes, incorporation of this technology can greatly enhance and speed up the process of addressing patient queries and educating them about their medical conditions.Another area where ChatGPT can excel is in the generation of patient education materials.As a large language model trained on vast amounts of text, ChatGPT can easily generate coherent and original material in a matter of seconds [12,31].Lyu et al. demonstrated the ability of ChatGPT to translate radiology reports into plain language, while Mondal et al. showed that ChatGPT could write articles to educate patients on dermatological conditions [32,33].The involvement of ChatGPT in such processes, which are normally performed by human clinicians, can result in significant cost savings and improved efficiency for healthcare institutions.
There are several limitations in our study.First, we assessed ChatGPT's responses using a curated list of 50 FAQs regarding TKR.This list of questions is not meant to be exhaustive, but rather as a proof-of-concept using the most frequently asked and clinically relevant questions.Furthermore, there might be slight differences between our list of FAQs and FAQs encountered in other countries due to variations in the prevalence and importance of different questions across different cultures and geographical regions.For example, questions about squatting or kneeling after TKR surgery might be more common in our local Singaporean population (a multiethnic Southeast Asian country) compared with Caucasian countries as such movements are part and parcel of daily life for many patients here [34].Next, our study assessed the ability of ChatGPT to respond to FAQs about TKR to the average patient without providing additional patient-specific information.As such, we were not able to assess the ability of ChatGPT to provide personalized information and recommendations -an important aspect of clinical consultation and surgical counselling.In instances where patient-specific FAQs were asked (examples shown in Additional file 1: Table S2), we noted that ChatGPT was able to highlight its limitations and direct patients to speak to a doctor for a more detailed and personalized consultation.Follow-up studies should investigate the ability of ChatGPT and other AI chatbots in providing patient-specific and personalized information, and potentially even compare it to those provided by human clinicians.Lastly, while there are several other AI chatbots such as Google Bard and Microsoft Bing which may provide similarly informative responses with realtime data, our study chose to evaluate responses from ChatGPT, as it is currently the most popular and widely used AI chatbot on the market [35,36].Future studies should critically evaluate and compare the performances between these chatbots.
Conclusion
ChatGPT performed well in providing accurate and relevant responses to FAQs regarding TKR, demonstrating great potential as a tool for patient education and preoperative decision-making.However, it is not infallible and can occasionally provide inaccurate medical information.Patients and clinicians intending to utilize this technology should be mindful of its limitations and ensure adequate supervision and verification of information provided.
2
Categorical outcome for relevance, whereby accurate responses are defined as those with a mean or final grade of ≥ 4
Table 3
FAQs about TKR indications and outcomes1 Categorical outcome for accuracy, whereby accurate responses are defined as those with a mean or final grade of ≥ 42Categorical outcome for relevance, whereby accurate responses are defined as those with a mean or final grade of ≥ 4
Table 4
FAQs about risks of TKR *Denotes responses where there was significant disagreement (≥ 2 on the Likert scale) between the two reviewers and the final grade was awarded by a third reviewer1Categorical outcome for accuracy, whereby accurate responses are defined as those with a mean or final grade of ≥ 4 2 Categorical outcome for relevance, whereby accurate responses are defined as those with a mean or final grade of ≥ 4
Table 5
FAQs about pain and post-operative recovery after TKR 1Categorical outcome for accuracy, whereby accurate responses are defined as those with a mean or final grade of ≥ 42Categorical outcome for relevance, whereby accurate responses are defined as those with a mean or final grade of ≥ 4
Table 6
FAQs about specific activities after TKR 1Categorical outcome for accuracy, whereby accurate responses are defined as those with a mean or final grade of ≥ 4 2 Categorical outcome for relevance, whereby accurate responses are defined as those with a mean or final grade of ≥ 4
Table 7
FAQs regarding alternatives and variations of TKR 1Categorical outcome for accuracy, whereby accurate responses are defined as those with a mean or final grade of ≥ 42Categorical outcome for relevance, whereby accurate responses are defined as those with a mean or final grade of ≥ 4Factual | 4,020.8 | 2024-04-02T00:00:00.000 | [
"Medicine",
"Computer Science"
] |
Dynamical scoto-seesaw mechanism with gauged B − L
.
I. INTRODUCTION
Despite the great success of the Standard Model (SM) [1], new physics is required in order to account for the existence of neutrino masses [2,3] as well as dark matter [4].A popular paradigm for neutrino mass generation is the seesaw mechanism, while weakly interacting massive particle (WIMP) dark matter candidates constitute a paradigm for explaining cold dark matter.Even taking these paradigms for granted, there are many ways to realize either.A particularly interesting possibility is provided by the so-called scotogenic approach [5,6] in which WIMP dark matter mediates neutrino mass generation.In the simplest schemes, all neutrino masses arise at the one-loop level, with a common overall scale, modulated only by Yukawa couplings.
Rather than invoking these paradigms separately, here we suggest a dynamical mechanism to realize naturally the scoto-seesaw scenario [7,8] that reconciles dark matter (DM) and neutrino mass generation together 1 , opening the possibility of having a loop-suppressed solar-to-atmospheric scale ratio.
We first note that the U (1) B−L symmetry arises automatically in the SM and is closely related to neutrino masses.However, the presence of non-vanishing anomaly coefficients forbids us to promote U (1) B−L to a local gauge symmetry.
Adding three lepton singlets ν iR ∼ −1 cancels the anomaly coefficients, allowing for a local realisation.Dirac neutrino mass entries arise in such a way that a ν iR -mediated (type-I) seesaw mechanism can be triggered by allowing for Majorana mass terms, which break B − L by two units.In such "canonical" construction the seesaw-mediating ν iR carry identical charges so that all neutrino masses become proportional to a single energy scale.As a result, this fails to account for the observed hierarchy ∆m 2 sol /∆m 2 atm [13].However, an alternative anomaly-free U (1) B−L can be obtained if instead we introduce three neutral fermions with B − L charges (f 1R , f 2R , N R ) ∼ (−4, −4, 5) [14,15].We show next that, thanks to their unequal charges, f aR and N R couple differently to the active neutrinos and may trigger different mass generation mechanisms, providing a natural dynamical setup for the scoto-seesaw mechanism and an explanation for the smallness of the ratio ∆m 2 sol /∆m 2 atm .The neutrino-mass mediators f aR become part of a dark sector whose stability originates from an unbroken matter parity that survives the breaking of the gauged U (1) B−L symmetry.
Besides having interesting dark matter features and collider phenomenology, an interesting implication of our model concerns charged lepton flavour violation.Indeed, a characteristic feature to notice is the presence of sizeable cLFV processes involving Goldstone boson emission [16][17][18], in addition to conventional cLFV processes, such as those involving photon emission, for instance µ → eγ.
II. GAUGED B − L EXTENSION OF THE STANDARD MODEL
Here, we describe the essential features of our proposal.The leptons and scalars present in the model and their symmetry properties are shown in Table I.In addition to the SM leptons, we introduce f aR (a = 1, 2) and N Rsinglets under the SM group but charged under U (1) B−L .Besides a SM-like Higgs doublet H, the scalar sector contains two extra SU (2) L doublets (Φ and η) and four singlets (φ 1,2,3 , σ).Note that all of the new fields are charged under U (1) B−L .The specific roles of the extra fields will become clear in the next sections when we discuss neutrino mass generation, dark matter, and how these are linked to each other in our model.
First we note that, with this field content, we ensure that B − L is free from anomalies and hence can be promoted to local gauge symmetry.To see this, first recall that in the SM, the conventional B − L appears as an automatic global symmetry that is anomalous due to the non-vanishing of the following triangle anomaly coefficients: By extending the SM with the fermions f aR and N R , as given in Table I, these coefficients vanish exactly [14,15] A With the fields in Table I, we can write down the most general renormalisable Yukawa Lagrangian as Analogous to the SM case, once the Higgs doublet H acquires its vacuum expectation value (VEV), the charged fermions become massive.In contrast, neutrino mass generation will involve the other scalar bosons, as discussed in Sec.V.
In the scalar sector, we assume that only neutral fields with even B − L charges acquire VEVs.Thus, a subgroup of U (1) B−L remains conserved after spontaneous symmetry breaking takes place.The residual symmetry is matter parity and can be defined as Therefore, only f aR , η and σ are odd under M P .Due to M P conservation, the lightest of such fields is stable and, if electrically neutral, will be our dark matter candidate.
The scalars are decomposed as follows and we assume that the breaking of U (1) B−L takes place well above the electro-weak scale: Notice that only the electrically neutral scalars with even U (1) B−L charges acquire VEVs so that matter parity (M P ), defined in Eq. ( 4), remains conserved.
After replacing the field decompositions above into the potential, we obtain the following tadpole equations In the limit where µ Φ is much larger than the other mass scales in the model, the second tadpole equation leads to the induced VEV Therefore, the induced VEV of the doublet Φ is suppressed as in the type-II seesaw mechanism [19,20], and also models with a leptophilic Higgs doublet [21,22], a natural example of which emerges naturally within the linear seesaw mechanism [23][24][25].
Due to matter parity conservation, which survives as a remnant symmetry, fields with different M P charges remain unmixed.Therefore, we can separate the scalar spectrum into a sector with fields transforming trivially under M P , the M P -even sector, and another with those that do not, the M P -odd sector.Moreover, for simplicity, we assume that all the VEVs as well as the couplings in V are real so that CP is conserved.
Starting with the M P -even sector, we have a 5×5 squared-mass matrix for the CP-even scalars (S H , S Φ , S φ1 , S φ2 , S φ3 ), given in Eqs.(A1) and (A2) in Appendix A. The associated physical states S 1,2,3,4,5 all become massive, and one of them is identified as the 125 GeV Higgs boson, S 1 ≡ h discovered at CERN [26,27].The other scalars are expected to be heavier, three of them, S 2,3,4 , with masses proportional to the B − L-breaking scale v φi .The mass of the heaviest state, S 5 , will be governed by the largest scale in the model, i.e., µ Φ .Thus, in the limit of interest, h ≃ S H and Concerning the CP-odd fields (A H , A Φ , A φ1 , A φ2 , A φ3 ), there is another 5 × 5 squared-mass matrix, given in Eq.
(A3) in Appendix A. Three of the mass eigenstates are massless, and two of them can be written as and are absorbed by the neutral gauge bosons Z and Z ′ .
On the other hand, two states, A 1 and A 2 , become massive3 and, in the limit v φ ≡ v φi ≫ v H ≫ v Φ , their masses can be written as Finally, there is a remaining massless field, a physical Nambu-Goldstone boson, G, analogous to the Majoron [28][29][30].
In order to obtain its profile, we make use of the fact that it must be orthogonal to would-be Goldstone bosons G Z,Z ′ , as well as to the massive A 1,2 fields; see Appendix A, Eq. (A6).Within the same limit as above, we find that this physical Goldstone lies mainly along the SU (2) L singlet directions, while its projections along the doublets are suppressed by VEV ratios This massless boson unveils the existence of a spontaneously broken accidental symmetry, identified in Appendix B.
As we will comment below in Sec.VI, there are stringent limits on the flavour-conserving Goldstone boson couplings to charged leptons, mainly electrons.Moreover, as discussed in Sec.VII, there are also bounds on the flavour-violating Goldstone boson couplings to charged leptons.
Turning now to the charged fields, (H ± , Φ ± ), we have After diagonalising the above matrix, we find that one eigenstate is massless (absorbed by the W ± gauge boson), whereas the other, ϕ ± , gets a mass Notice that the dominant component of ϕ ± lies along the doublet Φ, which gets a large mass m ϕ ± ≃ µ Φ ; see Eq. (9).
Turning now to the M P -odd sector, the CP-even and CP-odd scalars have the following mass matrices, respectively, with when expressed in the bases (S η , S σ ) and (A η , A σ ).The matrices can be diagonalised by performing the following rotations sin θ a(s) cos θ a(s) The corresponding eigenvalues are then given by Finally, the only charged scalar in the dark sector gets the following mass IV. GAUGE SECTOR: As usual, to find the gauge sector spectrum, we expand the covariant derivative terms in the scalar sector, i.e., , where ϕ i denote the scalars in the model and4 When the scalars acquire VEVs, the neutral gauge bosons acquire the following squared-mass matrix, written in where In order to determine the mixing among the fields, we diagonalise this matrix in two steps.First, we single out the photon (A µ ) by making use of the mixing matrix R 1 , and then we diagonalise the resulting Z − Z ′ matrix with the help of R 2 , namely and It is easy to see that the Z − Z ′ mixing, parametrised by δ, is rather suppressed since v Φ /v φ1 ≪ 1.Thus, the mass eigenstates are given by and the non-vanishing eigenvalues are which, upon assuming Neutrino masses arise from the operators in Eqs. ( 3) and ( 6) and are generated by the diagrams in Fig. 1, where the blue and the black legs denote the exchange of M P -even and M P -odd/dark fields, respectively.The tree-level contribution, in the basis (ν iL , (N R ) c ), leads to the following mass matrix Upon diagonalisation we find the seesaw-suppressed mass matrix for the active neutrinos to be defined in Eq. ( 9).The light neutrino mass matrix has only one non-vanishing eigenvalue ∼ − One sees that, in contrast to the conventional type-I seesaw mechanism, in which neutrino mass suppression follows from the large size of m N with respect to the electroweak scale (v EW ∼ v H ), here, there is an additional suppression from the small induced VEV of Φ, characterized by ϵ given in Eq. ( 9).This feature allows us to have moderate values for m N , or v φ1 , say around the TeV scale, without the need for appealing to tiny Yukawa couplings.
At this point, we also stress that, in contrast to low-scale inverse [30,31] or linear seesaw schemes [32][33][34], the presence of TeV-scale neutrino-mass-mediators does not require the addition of extra gauge singlets beyond the "right-handed neutrinos".Moreover, since m N is directly associated with the spontaneous breaking of the gauged B − L, the mass of the associated gauge boson, Z ′ , can also be naturally within experimental reach, leading to rich phenomenological implications, as we discuss below.
Concerning the other two neutrinos, their masses are generated at one loop through the scotogenic mechanism, namely where 2 are the masses of the two dark fermions f cR .One can easily check that when either λ 1 → 0 or µ 1 → 0, the loop-generated masses vanish.When λ 1 → 0, m si → m ai , with cos 2 θ s → cos 2 θ a and sin 2 θ s → sin 2 θ a , leading to a cancellation between the first and the second, as well as the third and the fourth terms in Eq. ( 27).On the other hand, when µ 1 → 0, we have θ s , θ a → 0 so that only the first and the second terms in Eq. ( 27) survive; nevertheless, these cancel out since, in this limit, m s1 → m a1 .
As usual in scotogenic setups, the loop mediators -black fields in Fig. 1 -are part of a dark sector from which the lightest field is stable.This can play the role of a WIMP dark matter candidate.In our case, the viable dark matter candidates are the neutral fermions f cR and the neutral scalars s i , a i .The stability of the lightest of them is ensured by the residual subgroup of our gauged B − L symmetry, i.e., the conserved matter parity (M P ), defined in Eq. ( 4).
Lastly, since one neutrino mass is generated at tree level, while the other two masses arise through the scotogenic diagram at the one-loop level, we expect these masses to be loop-suppressed with respect to the former [7,8].Therefore, this framework favours the normal ordering of neutrino masses -also preferred experimentally [13] -as well as an understanding of the origin of the smallness of the ratio ∆m 2 sol /∆m 2 atm .
VI. GOLDSTONE COUPLINGS TO FERMIONS
The effective interaction between the Goldstone boson, G, and the fermions, F i , can parametrised as where P L,R are the usual chiral projectors, and Σ GF L,R are model-dependent dimensionless coefficients.As seen from Eq. ( 13), the Goldstone boson has projections along of the M P -even scalars, including the SM-like Higgs doublet H.
As a result, G couples to all of the fermions at tree level.
For quarks and charged leptons, which get their tree-level masses exclusively from their interactions with H, the tree-level couplings to G arise from the projection of G into the H, ⟨G|H⟩.This can be expressed as with F i = e, d, u representing the charged leptons, down-and up-type quarks, respectively.Notice that when we diagonalize the charged-fermion mass matrices, the G-couplings also become diagonal.In other words, flavourviolating couplings to charged fermions are absent at tree level.
Of particular interest are the couplings to charged leptons, for example electrons.Depending on the strength of such couplings, Goldstone bosons can be over-produced in Compton-like processes e+γ → e+G and emitted by stars leading to excessive stellar cooling.Recently it was noted that also the pseudoscalar couplings to the muon can be restricted by SN1987A cooling rates [35].The bounds for the couplings to electrons and muons are |Σ G ee | ≲ 2.1 × 10 −13 and , respectively [35,36].
One sees that both constraints can be satisfied through the above restriction on the Goldstone projection into the SM-like Higgs doublet.It is worth noticing that this effective mass scale, v 2 Φ /v φ also appears in the seesaw-suppressed neutrino mass in Eq. ( 26), and should therefore be small.Note also that the Goldstone couplings to charged leptons also receive radiative contributions, mediated mainly by the dark sector, via the diagram in Fig. 2 (right).Such contributions can be calculated using the results of the next section by taking the diagonal entries of the one-loop couplings Σ G L,R in Eqs.(31) and (34).The tree-level Goldstone couplings to neutrinos emerge from the second term in Eq. (3) involving the leptophilic doublet Φ, i.e., Y Φ i L iL ΦN R .As a consequence, they are suppressed by the product of the Goldstone projection into Φ, of order v Φ /v φ , and the small ν − N mixing, also of order v Φ /v φ .Dark-mediated loop contributions are also present; however, these are expected to be very small as they are proportional to neutrino masses, a feature reminiscent of Majoron models [29,30].
VII. CHARGED LEPTON FLAVOUR VIOLATION
In this section, we discuss the most relevant charged lepton flavour violating (cLFV) processes occurring in our model, involving the first two families.For theoretical and experimental reviews, see, for instance, Refs.[37][38][39].
An interesting characteristic feature to notice is the presence of charged lepton flavour violation involving Goldstone boson emission.The projected µ → e conversion at COMET [40] can also probe µ → eG.There are good prospects for improving the sensitivities on Goldstone boson searches with the COMET experiment, where the sensitivity for the cLFV process µ → eG may improve from the current limit BR(µ → eG) = 2.3 × 10 −5 in Phase-I to a sensitivity O (10 −8 ) in Phase-II, under somewhat optimistic assumptions [18] 6 .In what follows we pay special attention not only to radiative cLFV processes involving photon emission, but also to Goldstone boson emitting processes.
A. cLFV through photon and Goldstone boson emission
We focus on cLFV decays of the form e i → e j X -with X = γ, G -which, here, receive their main contributions from diagrams mediated by the M P -odd fields in the "dark" sector, as shown in Fig. 2.
Other diagrams, although present, give rise only to sub-leading contributions.For instance, analogously to Fig. 2, diagrams mediated by the M P -even fields ϕ + and N , instead of the M P -odd fields η + and f a , exist but their contributions are significantly suppressed by the large mass of ϕ + .Likewise, charged-current contributions mediated by the W + gauge boson and N are present; however, they are also suppressed due to the small ν − N mixing.The effective operators associated with the decays e i → e j X, with X = γ, G, can be expressed, respectively, as follows [42,43] where σ µν = [γ µ , γ ν ] and P L,R are the usual chiral projectors.Here Σ γ L,R (Σ G L,R ) are model-dependent coefficients with dimension mass −1 (mass 0 ).
Taking into account that m e /m µ ≪ 1, the leading contribution to the first process can be approximated to where x a = (m fa /m η + ) 2 , α em is the fine-structure constant, and m µ ≃ 105.658MeV is the muon mass.Here, is the total decay width of the muon, with Γ new µ representing the new contributions to the muon decay width calculated in this section.Similarly, for the process involving the Goldstone boson emission, we have where ⟨G|φ 2 ⟩ is the Goldstone projection along φ 2 .
We would like to point out also that, despite the presence of several new neutral scalars as well as an extra neutral gauge boson (Z ′ ), flavour-changing neutral currents involving charged leptons only take place at loop level.Indeed, notice that, at tree level, two charged leptons can only couple to a scalar via the first term in Eq. ( 3), governed by the Y H matrix.The very same term leads to the charged lepton mass matrix: M e = Y H ⟨H⟩. Therefore, by diagonalising M e , the charged lepton couplings to neutral scalars are also automatically diagonalised.Lastly, since all the charged leptons couple to Z ′ with the same B − L charge (−1), and there is no non-standard charged lepton, the tree-level couplings between the charged leptons and Z ′ are also diagonal.
B. Numerical results
Having discussed the main cLFV processes, as well as the relevant bounds, we now proceed to calculate numerically the relevant contributions.Note that the Yukawa couplings controlling the cLFV processes are also involved in the generation of neutrino masses.Thus, in order to estimate the cLFV contributions, we need to choose as inputs benchmarks satisfying neutrino oscillation data.Although we perform this task numerically, it is useful to develop an analytical form for extracting the Yukawa couplings from measured parameters that satisfy neutrino oscillation restrictions, à la Casas-Ibarra [51], as this will optimise our subsequent scanning procedure.For this purpose, we write the sum of the tree-level (seesaw) and the one-loop (scotogenic) neutrino masses in Eqs. ( 26) and ( 27) in the following compact form with M c defined in Eq. ( 27).This way the ansatz below can be used to obtain our benchmarks where U ν is the lepton mixing matrix, m ν is a diagonal matrix containing the active neutrino masses, and ρ is a generic 3 × 3 orthogonal matrix.For the oscillation parameters that enter U ν and m ν , i.e., the neutrino mixing angles, CP phase and neutrino mass splittings, we take the values obtained in Ref. [13] within a 3σ range, for the case of normal neutrino mass ordering.Moreover, we fix some of the free parameters present in X M by assuming that all the relevant dimensionless scalar couplings are λ i = 0.1, the bare masses for the dark scalars µ η = µ σ = 1 GeV, while µ 1 = 5 TeV, and the Yukawa couplings are Y N = Y f a = 0.5.As for the lightest neutrino mass, m 1 , we vary it in the range [0.005, 0.03] eV.The upper value comes from taking the upper limit on the sum of the neutrino masses: Σ i m i < 0.12 eV [52] together with the normal ordering assumption [13].
On the other hand, the lower bound comes from the radiative origin of the lighter neutrino masses, i.e., m 1 , m 2 .In general, we expect no strong m 2 /m 1 hierarchy or, if present, such a hierarchy would be typically milder than the m 3 /m 2 hierarchy, since m 3 has a tree-level origin.This is seen from Fig. 3, where we show the m 3 /m 2 and m 2 /m 1 mass ratios versus m 1 for normal ordering.One sees that this is indeed the case, as long as m 1 ≥ 0.005 eV.
In what follows, we consider two scenarios.In the first, the induced VEV v Φ , defined in Eq. ( 9), varies with varying v φ ≡ v φ1,2,3 , which describes the B − L-breaking scale, whereas in the second case, v Φ remains constant although v φ varies.The other parameters -such as the scalar potential couplings (λ i , µ i ) and the Yukawa couplings (Y f a ) -are assumed constant throughout the scans, as explained below Eq. ( 36).This choice helps isolating the impact of v φ on the dark mediator masses and the cLFV branching ratios.Our results for the first scenario are presented in Fig. 4. In this case, v φ varies randomly between 1 and 15 TeV, and the induced VEV v Φ varies with v φ , according to Eq. ( 9), indicate the experimental bounds for the branching ratios of µ → eG [17,48,49] and µ → eγ [47], respectively, so that the points within the shaded red areas are excluded.Moreover, the gray points are excluded by the astrophysical constraints on the flavour-diagonal Goldstone couplings Σ G ee and/or Σ G µµ .The dashed orange lines represent future experimental sensitivities for both processes, i.e., BR(µ → eγ) ≲ 6 × 10 −14 from the MEG II experiment [50] and BR(µ → eG) ≲ O(10 −8 ) from the COMET experiment [40], as derived in Ref. [18].The remaining points are coloured according to the "size" of the corresponding Yukawas, defined as Y eff = | a Y η ea Y η * µa |, as shown in the colour bar on the right.One sees from the left panel of Fig. 4 that µ → eG is a very promising cLFV channel.Indeed, in order to comply with the existing µ → eG limit, BR(µ → eγ) must be in the 10 −14 range or less, which will be somewhat challenging for the upcoming round of experiments.As can be seen from the right panel, this implies lower bounds for the dark mediator masses.We now turn to a scenario in which the induced VEV v Φ is a constant, i.e., v Φ = 50 MeV, but v φ is allowed to vary, as previously, between 1 and 15 TeV.The plots in Fig. 5 show very different behaviours when compared to those in the first scenario.Here, the branching ratios increase as the mediator masses, m η + , m fa , increase, whereas the opposite happens in the first scenario.
To understand this, notice first that the branching ratios in Eqs. ( 33) and ( 34) vary roughly as µa | contribute to (scotogenic) solar neutrino mass generation which depend on v 2 Φ ; see Fig. 1.Thus, once we impose neutrino oscillation data, via Eqs.( 35) and ( 36), Y eff can also be described as a function of the induced VEV v Φ .In the first scenario (Fig. 4), as v Φ increases with v φ , following Eq.( 9), the scotogenic loop functions M c in Eq. ( 27) also increase.Consequently, to "neutralise" this increase in M c , so as to satisfy neutrino oscillation data, the relevant Yukawas should decrease accordingly.Therefore, since both Y eff and (m µ /v φ ) decrease with increasing v φ , the branching ratios also decrease, as shown in Fig. 4. In contrast, in the scenario where v Φ is constant, the scotogenic loop functions M c decrease with increasing v φ , and, in turn, the Yukawas increase to satisfy oscillation data.Therefore, while (m µ /v φ ) decreases, Y eff ends up increasing, with v φ .
Nevertheless, it turns out that the dependence of Y eff on v φ is stronger than linear so that the branching ratios, given as functions of Y 4 eff (m µ /v φ ) 4 and Y 4 eff (m µ /v φ ) 2 , end up increasing with increasing v φ , as shown in Fig. 5.
In short, we have seen that our model naturally leads to detectable cLFV rates involving Goldstone boson emission.
Although we have mainly focused on the implications of Goldstone boson emission in cLFV processes, there is a rich phenomenology associated to Goldstone boson processes that can affect other sectors.For instance, the SM Higgs boson can decay invisibly h ≡ S 1 → GG [53,54] and also the effective number of relativistic degrees of freedom could be affected.Our model is elastic enough to satisfy constraints such as the invisible Higgs decays [55,56] by adequately choosing the Higgs potential parameters.For a recent discussion on similar models see [36].Likewise, by properly choosing the Higgs potential parameters, Z ′ mass and couplings one can suppress excessive contributions to the effective number of relativistic degrees of freedom [57] so that it can fulfill the observational constraint [52].Therefore, the model offers sufficient flexibility to accommodate constraints arising from the Goldstone boson interactions without compromising the promising prospects of detecting cLFV through Goldstone boson emission.A detailed analysis of the parameter space falls outside the scope of the present work.
VIII. COLLIDER SIGNATURES
A. pp → Z ′ → N N Drell-Yann Production In the simplest scoto-seesaw model [7,8], the neutrino mass mediators are unlikely to be produced at colliders.
For example, the atmospheric neutrino mass mediator, i.e., the heavy neutral lepton N (right-handed neutrino), is inaccessible through the Drell-Yan mechanism.Indeed, N is typically very heavy.Even if (artificially) taken to lie in the TeV scale, its production would be suppressed by a tiny doublet-singlet ν − N mixing.
In the present case, however, both the neutral lepton N and the vector boson Z ′ can lie in the TeV-scale.The existence of the Z ′ provides a Drell-Yan pair-production portal for the heavy neutrino.Its subsequent decays into the SM states could potentially lead to verifiable signals.Due to the enhanced gauge charge ( 5) of the right-handed fermion N , its Z ′ phenomenology can differ from that of other U (1) B−L models.
We remark that the neutrino mass suppression mechanism of our scoto-seesaw model involving the induced VEV of the leptophilic doublet Φ, Figs. 1, requires a rich scalar sector compared to a vanilla U (1) B−L model.We have an extended visible scalar sector with four more CP-even scalars apart from the Higgs boson.Among them, the scalar predominantly coming from the leptophilic doublet Φ is relatively heavy and beyond the LHC reach, while the others are lighter, possibly around the TeV scale.Moreover, there are two massive CP-odd scalars.While one of them is heavy, as it comes mostly from Φ, the other can potentially lie at the TeV scale.
The pp → N N collision process can also proceed through the exchange of neutral scalars.These couple to the initial-state quarks and final-state neutral fermion pair through the mixing of the Higgs boson with other CP even scalars.Nonetheless, for simplicity, we assume that this mixing is not significant so that such decays (S i → N N ) are not kinematically allowed, and these contributions can be neglected.In Table II, we give estimates of the expected Z ′ mediated production cross section.II.Drell-Yan neutral lepton pair production through the Z ′ portal.The cross section is computed with λii ∼ 0.1, λij ∼ 0.01, Yi ∼ 0.5, g ′ = 0.1.The dominant contribution is through the Note that the neutral fermion mass is taken around 700 GeV, allowing the Z ′ boson to decay to it in on shell.
For a Z ′ mass of 1.5 TeV, the branching ratio to Z ′ → N N is relatively low, while it peaks near 2 TeV Z ′ mass, and saturates at around 25%.Therefore, σ(pp → Z ′ ) × BR(Z ′ → N N ) increases initially up to 2 TeV, and then it gradually decreases with increasing Z ′ mass.Notice that the heavy neutrino N will decay to SM states, such as leptons, through ν − N mixing in the electroweak currents.This leads to leptonic final-state signatures such as dileptons plus missing energy 8 .It is not in our scope here to present a dedicated numerical simulation of these signatures.A detailed collider analysis of the pp → Z ′ → N N mode in a slightly different scenario is performed in Ref. [58].
8 Dark sector fermions can also be produced through Drell-Yann processes like σ(pp → f i f i ), with less striking signatures.
B. Z ′ Phenomenology: Dilepton search ATLAS and CMS experiments are looking for a heavy BSM gauge boson in different channels.The leading LHC production mode for a Z ′ with couplings to quarks (as is the case for all U (1) B−L models) is q q fusion.In our model, we have many decay modes of Z ′ → S i A j and Z ′ → S i Z, where S i and A j are the CP-even and CP-odd scalars, in both visible and dark scalar (s i , a i ) final states.As a consequence, our Z ′ decay rates to the SM channels can be relatively small.
Despite such a smaller branching ratio to the SM channels, the Z ′ decay to a pair of charged leptons places strong constraints on the Z ′ mass and coupling; see Fig. 6.The exclusion bounds on the parameters of our model at the LHC are derived by comparing our theoretical predictions with the experimental limits obtained from the non-observation of any BSM excess in a particular mode.Here, we fix the gauge coupling at a value g ′ = 0.1 and compare the theoretical cross section σ(pp → Z ′ ) × BR(Z ′ → ℓℓ) in our model with the upper bound obtained from the LHC dilepton search.This way we provide limits on the Z ′ mass, M Z ′ .The Z ′ limit in the sequential standard model gauged lepton number charges [62], in the presence of higher dimensional operators, the lower bound on Z ′ hovers around m Z ′ ≈ 2.5 − 3 TeV.
Considering the left plot, the blue line gives the strongest constraint, excluding any Z ′ mass up to m Z ′ ∼ 3.9 TeV.This is the case where all scalar quartic couplings, except for λ H , chosen to fit the SM Higgs boson mass, are taken to be large λ i ∼ 0.75, leading to a heavier scalar spectrum which restricts Z ′ decays to BSM scalar modes.Likewise all BSM fermion Yukawa couplings are taken large Y a = 1.5 so that also the N, f a fermions are too heavy to be produced by Z ′ decays, leading to a smaller branching ratio for Z ′ to charged leptons, i.e., BR(Z ′ → ℓ i ℓ i ) ∼ 15%.Our second benchmark corresponding to the red line leads to weaker constraints.In this case, the scalar quartics take values λ i ∼ 0.10, making the scalar sector relatively light so that Z ′ can decay to some dark scalar-pseudoscalar pair.Even in this limit, Z ′ decays to visible sector scalars are not favored, mainly due to phase space.As for the BSM Yukawa couplings, we assume Y a = 0.5, which opens up decay modes like Z ′ → N N, f a f a with BRs in the order of 30% and 20% each, respectively.These new decay modes suppress the dilepton BR to BR(Z ′ → ℓ i ℓ i ) ∼ 2 − 4%, leading to a weaker limit on the Z ′ mass, m Z ′ ≥ 2.7 TeV. Finally, in the right plot, we compare our "aggressive" model prediction to that of the U (1) χ model, an E6-motivated Grand Unification construction [63].We observe that the lower bound for the Z ′ mass of the U (1) χ model lies around 4.8 TeV, roughly 1 TeV above the limit for m Z ′ in our model.For details regarding the collider aspects of Z ′ phenomenology, we use Ref. [62,64].
IX. DARK MATTER
In our gauged scoto-seesaw model, dark matter can be either the lightest "dark" scalar or fermion, stabilized through the residual matter parity symmetry defined in Eq. ( 4).Before embarking on our dark matter discussion we note that, for the particular benchmarks adopted in the previous sections, the dark matter candidate is always a scalar field.Radiative exchange of the "dark" particles generates the solar neutrino mass scale, as illustrated in the right panel of Fig. 1.The dark sector includes scalar dark matter candidates that are mixtures of neutral scalars of the SU (2) L doublet η and the singlet σ, plus the charged scalar η ± .On the other hand, the fermionic dark sector contains two neutral Majorana fermions f a .In addition to the standard Higgs portal mediated DM annihilation, we can have a novel Z ′ mediated annihilation, along with some new scalar mediated annihilation.
• Scalar dark matter
The scalar dark matter sector of our scoto-seesaw model includes, besides the doublet η, an additional singlet σ.
The neutral components of the singlet S σ , A σ and doublet S η , A η mix so as to produce the CP-even (s 1 , s 2 ) and CP-odd (a 1 , a 2 ) DM candidates, respectively.(In contrast, the dark scalar sector of the simplest scoto-seesaw scenario, analysed in Ref. [7,8], consists of a single scalar doublet η.)As the doublet-singlet mixing is small, see Eq. ( 17), two of the scalar DM candidates, one CP-even s 1 and one CP-odd a 1 , are mainly doublet DM candidates.The neutral components of the doublet η resonantly annihilate through the Higgs portal.In our setup, we can have additional resonant annihilation through non-dark CP-even Higgs bosons (S ≡ S Φ , S φi ).
This may lead to enhanced annihilation and underabundant relic at different DM masses m DM ∼ m S /2.
In addition to the η mediated t-channel annihilation to SM states ℓℓ, W W, ZZ, hh, present in the simplest scotoseesaw (Ref.[7,8]), in our new model, there are other t-channel contributions to scalar DM annihilation, which can enhance the annihilation cross section, leading to more parameter space with an underabundant relic.
The dark singlet σ present in our model annihilates through similar channels as the doublet-like dark matter, though the relative dominance of different contributions will depend on the mixing of the CP-even scalars h, S i .
For example, singlet s 2 and a 2 annihilation through quartic couplings like SSV V with S ≡ s 2 , a 2 and V ≡ W ± , Z are absent, whereas same vertices with V ≡ Z ′ and the corresponding annihilation channel will be present.
Moreover, the scalar dark sector phenomenology of our model will be modified due to the presence of a Z ′ portal.
This happens because, if CP-even (s 1,2 ) and CP-odd (a 1,2 ) dark scalars are almost degenerate, co-annihilation happens through a Z ′ mediated s-channel process.This coannihilation adds up to the existing DM annihilation modes, opening up new parameter space consistent with measured relic density.
• Fermionic singlet dark matter In our gauged scoto-seesaw model, two fermionic dark matter candidates f a mediate the loop-induced solar mass scale, instead of a single dark fermion in the simplest scoto-seesaw [7,8].These dark fermions f a have a new Z ′ portal to annihilate.In contrast to the simplest non-gauged scoto-seesaw model in Ref. [7,8], this brings a significant enhancement in the DM annihilation, including a resonant dip in the relic abundance at Moreover, in the simplest scoto-seesaw mechanism, the dark fermion mass term is a bare mass term.As a result, there is no Higgs-mediated s-channel fermionic dark matter annihilation channel.In contrast, in our gauged scoto-seesaw model there is a singlet φ 2 that provides mass to the dark Majorana fermions f a , and can act as a portal of DM annihilation, creating a resonant dip in the relic density.
• Constraints and dark matter detection In the standard scoto-seesaw (Ref.[7,8]), light doublet-like dark matter is constrained from the LEP Z decay measurement.In vanilla scotogenic models, including the original one [5] and its simplest triplet extension [6], nearly degenerate neutral CP-even and CP-odd scalars from η is typically assumed in order to implement scotogenic neutrino mass generation.In our present model, however, instead of relying on the quasi-degeneracy of the scalar mediators, the smallness of neutrino masses follows mainly from their dependence on the (small) induced VEV v Φ , Fig. 1.Consequently, here neutrinos can be light enough even when one has a light CPeven DM with a heavier CP-odd counterpart, or vice versa, avoiding constraints from the Z → S η A η decay.
Therefore, in our gauged scoto-seesaw a light doublet-like scalar dark matter is not as constrained as in vanilla scotogenic models [5,6] or the simplest scoto-seesaw scenario [7,8].The LEP constraint can be completely avoided in our proposed scheme.
Our dynamical scoto-seesaw model can have light dark scalars, free from constraints that hold for a generic doublet-only dark sector, allowing it to harbor a viable light DM candidate.The latter can be probed through dedicated experiments like CDMS-lite and CRESST, along with existing experiments, such as LZ, Xenon-nT, and PandaX.
Note that in our model we have DM coannihilation through a Z ′ portal, the DM relic density being fixed by Z ′ related parameters and dark sector mass degeneracy.Concerning direct detection, the scalar DM-nucleon scattering dominantly happens through the Higgs portal, as in generic scotogenic models.This relaxes the interdependence between the DM annihilation and detection, as they are not controlled by the same interaction.
In the simplest scoto-seesaw scenario, for instance, there is no mediator for fermionic DM detection through DM-nucleon scattering at the tree level.This results in a possible loop-suppressed scattering cross section.
The presence of DM-nucleon scattering mediated at tree level by Z ′ provides a significant advantage to our gauged scoto-seesaw model in this regard, as it enhances the possibility of direct detection of fermionic DM.The fermionic DM-nucleon scattering cross section then may reach the current sensitivity of the DM direct detection experiments.
X. CONCLUSIONS
We have proposed a scheme where the scoto-seesaw mechanism has a dynamical origin, associated to a gauged In this appendix, we provide in full form some of the important, but longer, analytical expressions obtained while deriving the scalar spectrum.The 5 × 5 squared mass matrix for the CP-and M P -even fields (S H , S Φ , S φ1 , S φ2 , S φ3 ) is given by the symmetric matrix with Similarly, for the M P -even but CP-odd fields, in the basis (A H , A Φ , A φ1 , A φ2 , A φ3 ), we have the symmetric matrix Out of the five eigenvalues of M 2 A , only the two below are non-vanishing where Finally, the exact expression for the physical Nambu-Goldstone, G, is ) .
Appendix B: The accidental U (1) From the Lagrangian interactions, we can derive relations amongst generic Abelian charges (Q) of the different fields in the model.These relations can be expressed in terms of four independent charges.For instance, taking ) as a basis of independent charges, we find the following relations The four independent charges are associated with four independent global U (1) symmetries, three of which were imposed: U (1) Y , U (1) B and U (1) L .Thus, by substituting B Q = (1/6, 1/2, 0, 0), B Q = (1/3, 0, 0, 0), B Q = (0, 0, −10, −8) into Eq.(B1) we obtain, respectively, the hypercharge, baryon number and lepton number of all scalars and fermions.
On the other hand, the fourth symmetry, U (1) X , was not imposed and appears accidentally in the model.The presence of a massless pseudoscalar in the spectrum tells us about the existence of such a symmetry and that it is broken spontaneously.Note that, as an accidental symmetry, the X-transformation properties of the fields have not been previously defined.Nevertheless, we can obtain such charges a posteriori by making use of the Goldstone profile given in Eq. (A6) as well as Goldstone's theorem.
According to Goldstone's theorem, the spontaneous breaking of a continuous global symmetry leads to a massless field, the Goldstone boson, which can be identified through the associated Noether's current.In the case of a global Abelian symmetry, such as the accidental U (1) X , once it is broken by the vevs of a set of scalars ϕ j = (v j +S j +iA j )/ √ 2, the corresponding Goldstone boson can be written as where X j represent the charges of ϕ j under U (1) X .Comparing this expression with Eq. (A6), we can extract the U (1) X charges of the scalar fields in terms of one of the charges, say X H , and the scalar vevs, The charges of the other fields can be obtained by substituting Eq. (B3) with Eq. (B1).
Fig. 1 .
Fig. 1.Seesaw and scotogenic contributions to neutrino masses.The fields that are even under MP are depicted in blue, whereas the black legs represent the MP -odd (dark) fields.
Fig. 4 .
Fig. 4. Left: branching ratios for µ → eG vs µ → eγ.Right: branching ratio for µ → eG vs the masses of the dark mediators, m η + , m fa , given as functions of vφ.The shaded red bands are excluded by the current limits on the branching ratios BR(µ → eG) and BR(µ → eγ), whereas the gray points are excluded by the limits on the diagonal couplings (Σ G ee , Σ G µµ ).The dashed orange lines are the expected experimental sensitivities, as discussed in the text.The colour bar denotes the effective Yukawa defined as Y eff = | a Y η eaY
Fig. 6 .
Fig. 6.The left panel gives the expected σ ×BR of the dilepton signal in our dynamical scoto-seesaw model, obtained when all exotic channels are accessible (red line) or kinematically forbidden (blue line); see text.The black line represents the dilepton channel measurement from ATLAS[59] (with similar results also from CMS[60]), while the green and yellow bands give 1σ and 2σ regions, obtained from the LHC data repository HEPData[61].Exclusion limits on the Z ′ mass within these two extreme benchmarks are readily obtained from the blue and red curves.The right panel compares the aggressive constraints (blue line) obtained in our model and in a reference χ model.
B
− L symmetry."Dark" states mediate solar neutrino mass generation radiatively, while the atmospheric scale arises a la seesaw ; see Fig.1.Indeed the origin of the solar scale is scotogenic, its radiative nature explaining the solar-to-atmospheric scale ratio.The two dark fermions and the TeV-scale seesaw mediator carry different dynamical B − L charges.Dark matter stability follows from the residual matter parity that survives the breaking of B − L gauge symmetry.Apart from the possibility of being tested at colliders, see Fig.6, our scoto-seesaw model with gauged B − L has sizeable charged lepton flavour violating phenomena.These include also processes involving the emission of a Goldstone boson associated to an accidental global symmetry present in the theory, see Fig.
2. Rate estimatesfor muon number violating processes are given in Figs.4 and 5.They indicate that these processes lie within reach of present and upcoming searches.Likewise, we also expect sizeable tau number violating processes. | 10,529.4 | 2023-07-10T00:00:00.000 | [
"Physics"
] |
Linguistic processing and classification of semi structured bibliographic data on complementary medicine.
Complementary and alternative therapies and medicines (CAM) such as acupuncture or mistletoe treatment are much asked for by cancer patients. With a growing interest in such therapies, physicians need a simple tool with which to get an overview of the scientific publications on CAM, particularly those that are not listed in common bibliographic databases like MEDLINE. CAMbase is an XML-based bibliographical database on CAM which serves to address this need. A custom front end search engine performs semantic analysis of textual input enabling users to quickly find information relevant to the search queries. This article describes the technical background and the architecture behind CAMbase, a free online database on CAM (www.cambase.de). We give examples on its use, describe the underlying algorithms and present recent statistics for search terms related to complementary therapies in oncology.
Introduction
In 1999, Zollmann and Vickers defined complementary and alternative medicine (CAM) as "a group of therapeutic and diagnostic disciplines that exist largely outside the institutions where conventional health care is taught and provided"). 1 In the last decade however, complementary medicine has not only succeeded in establishing itself in the academic context, with an increasing number of institutions at European universities working on CAM, it has also expanded its activities in all fields of research. 2 The growing academic development of CAM accompanies the increasing demand for complementary therapies in the public domain. All surveys of this field show that, depending on the country, between 20% and 65% of people use CAM, 3 and that patients often request to be updated on CAM options. Although the family physician is in most cases regarded as the first choice for such inquires, a growing number of patients also use the internet as a resource for CAM information. In one survey, the Pew Internet and American Life Project found that 48% of health seekers have looked for information about CAM on the internet. 4 In particular, there is an increasing demand for information on CAM among cancer patients. 5 Patients are actively seeking treatments and are thus aiming at utilizing coping strategies which might be helpful to extend survival time, gain a better quality of life or to relieve pain. 6,7 While many patients have discovered the World Wide Web as a resource for information on CAM, 8 modest empirical research has been conducted into the kind of sources of CAM information and search tools used by cancer patients. Some observations suggest that cancer patients increasingly search professional sources of clinical evidence in the internet, such as patient guidelines or databases. 9 Our own experience with CAMbase, a bibliographic database on CAM, revealed a strong interest from cancer patients in evidence based information on complementary therapies. 10 The increasing interest of cancer patients in the internet and the diversity of information resources has led to difficulties in identifying, filtering and implementing new and appropriate medical knowledge. Figure 1 illustrates the growth of information in various bibliographic databases using the example of the search term "cancer". Medicine Database (AMED) 238% and "Cumulative Index to Nursing and Allied Health (CINAHL) 454%.
The expansion of the internet in the last ten years has made information retrieval challenging not only for patients but also for health care providers and physicians who wish to separate "important/meaningful" from "unimportant/meaningless" information. 11 Thus, with the growing interest in CAM, there is an urgent need by users of online databases for tools to assist in reviewing the scientific information on CAM in cancer.
Despite the development of CAM research and application in recent decades, a large proportion of the publications reporting on CAM cannot be easily found in digital bibliographic databases like MEDLINE. This is mostly due to two problems:
A broadly accepted thesaurus for CAM in total
does not exist: With respect to heterogeneity, CAM has not developed a strong tradition for a controlled vocabulary to classify CAM literature 12 and despite some promising efforts, 13 this is still an unsolved problem. In addition, the conventional MESH-Keywords of MEDLINE do not adequately map the contents of CAM-Literature. 2. Despite the existence of semi-structured bibliographic data on CAM in electronic databases, a researcher might use misleading or ambiguous keywords in his search strategy and hence cannot find the relevant data. This problem is also known as the so called "vocabulary problem". 14 While the first problem is a result of the diverse composition of CAM itself, the second also arises in other biomedical contexts and is a problem on the client's side.
In our work, we have sought to provide search options suitable for (i) experienced users who want to employ a specific search strategy, e.g. for systematic reviews, as well as for (ii) casual users who want to inform themselves on a topic which at that moment of query issue might not be paraphrased very precisely. 15 This article describes the linguistic processing and classification of semi-structured bibliographical data on complementary and alternative medicine in CAMbase, a free bibliographic database on CAM realized with the semantic web standard XML and accessible online at the URL www.cambase.de.
Technical Realization
In situations where there is no concise repertoire of controlled vocabulary, there is a necessity to build tools which guide the user through the bibliographical landscape. Based on the inverted file structure and the implementation of XML together with the linguistic algorithms, we are able to offer such a search-tool of CAM-landscapes, which are created for every single search query based on elementary bibliographic data (i.e. keywords or authors).
The basic idea behind this approach is to guide the user through the mass of resulting datasets, which in the case of the search query 'cancer' in conventional databases like MEDLINE might lead to a large volume of results as can be seen in Figure 1." In the following section we focus on the specific topics of processing and classification of semistructured bibliographical data.
search procedures and strategies
Search strategies in medical research, e.g. for systematic reviews, are nowadays targeted towards identifying relevant articles in bibliographical databases such as MEDLINE and are commonplace. 16 Users are primarily scientists or specially trained library professionals, who are consulted by domain experts. On the other hand, an increasing number of physicians are embracing the internet to gather information about the latest guidelines or to deepen their education in web-based medical training and seminars. 17 When the structure of CAMbase was initially planned, we intended to develop search options which suit both of these user-groups. Therefore, in addition to conventional search options (Author, title, keywords, publication year, etc.), we implemented a natural language interface with linguistic algorithms designed to facilitate enhanced query capture before issuing to the search process.
Apart from conventional techniques of natural language processing (NLP) like stemming, text segmentation and the analysis of punctuation and normalisation (see Brants 18 for a review of common methods of NLP), we have developed special features, which are quite unique with respect to their decomposition of search phrases into their linguistic and grammatical entities. In the following we illustrate these features using two examples: Example 1: Modification and restriction of a subject, explicitly formulated by the user In the search term "European mistletoe for the treatment of cancer", the phrase "mistletoe" is the subject modified by "treatment of cancer" and restricted by "Europ∼ean", whereas in the search term "mistletoe treatment of cancer in Europe", the term "cancer" is the central subject modified by "Europe" and restricted by "mistletoe treatment" (Fig. 2). Thus, even though the textual phrase consists of the same words, the ranking of the search results will be different 19-21 because of the grammatical relation between the words. Note that most of these language processing features are quite unique for German.
As can be seen in this example, linguistic procedures like automated morphological reduction and decomposition algorithms also help in the processing of search queries: the search query "european" will also lead to hits on "europe" and the German search query "Krebstherapie" (cancer therapy) also finds "Therapie des Krebses" (therapy of cancer) and vice versa. Table 1 gives an overview of the implemented NLP-techniques.
Ranking Algorithm
Based on a mathematical analysis of how the search phrase is conceptually represented in the dataset, a numerical distance value between 0 (=no relevance) to 100 (identical) is computed, which helps the user to decide which documents fit to his/her search query, i.e. datasets with a higher ranking will be placed in a higher position than those with a lower ranking. The underlying algorithm can be described as follows: The inclusion of words depends on the given record structure and the decision of the database owner. In our case we included all words of bibliographical fields (i.e. title, keywords, abstract) and excluded content notes such as page numbers and stopwords such as "the","an" or "and", which were filtered out prior to the processing. is defined where a 1 , a 2 , …, a k denote additional coded information about syntactical and morphological attributes of the word w i d (See Table 1 for an overview of these additional informations). ∈ also is fulfilled if S j is 'part' of a word w i d (i.e. the German term "Therap∼ie" in "Misteltherap∼ie" or the English term "therap∼y" in "therap∼eutic").
In the case of a normally distributed heterogeneity
Rule based analysis of German
Treat/-ing/-ed; assess/-ment Krebstherapie (engl.: cancer therapy) is decomposed into "Krebs" and "Therapie"; "Lektine" (engl.: lectins) also finds "Mistellektine" (engl.: "mistletoe lectins") Figure 3 illustrates the behaviour of the search algorithm for the threshold values Q t = 35 and Q t = 45 and a high (+) versus a low (-) sensitivity. Note that the threshold values Q t = 35 together with a low sensitivity results in a total of 54 hits whereas Q t = 45 with a high sensitivity delivers only 10 hits. The remaining two combinations Q t = 45 with low sensitivity and Q t = 35 with a high sensitivity each achieve about 20 hits. Also note that in these cases the decrease in quality of the hits develops slightly differently.
Inverted File structure
In our context, along with the linguistic processing of the bibliographical metadata of documents, data volume and query processing loads increase continually and thus it is of great importance to have an efficient information retrieval system which is able to process the search queries in an appropriate run time. We therefore decided to use an indexing mechanism based on inverted file structure. 22 All bibliographical datasets (including sentences if an abstract is available) are processed into an index by the above mentioned linguistic algorithms, and together with additional discriminating information (ADIs) for example, the logical subset that a bibliographical record belongs to, an inverted file structure is created (Fig. 4).
Within the context of linguistic indexing, one new feature has been developed to enhance the speed of search algorithms operating on conventional inverted file structures: words are no longer regarded as textual 'atoms', but are recognized in their syntacticsemantic interrelation. Based on a 6-byte coding, the above mentioned mathematical algorithms produce a 'word-map' of the documents which provides the information on whether or not a word appears in certain documents. This information is used as a filter when more than two search terms are entered in the query. Thus, matching-algorithms may abstain from processing a document from the outset if the wordmap provides this information, which increases the speed of processing the search query.
Example 2: Search-query 'mistletoe in oncology'
When searching for the phrase "adverse effects of mistletoe preparations", our matching algorithms, based on the mathematical calculation of Q (adverse, effects, mistletoe, preparations), together with the inverse file coding, are able to detect whether the term "preparation∼s" occurs in a dataset or not. Thus, documents found in the search for 'mistletoe' without the words "preparation∼s" or "adverse effect∼s" are not processed any further because no match for the combined search term emerges. This has quite a dramatic effect on the processing speed of the search query. In the search query example given above, the single term "adverse" results in a performance of 0.05 seconds based on the average of 20 subsequent entries on a conventional notebook with a DSL-6Mbit data rate. For the terms "adverse" and "effects", the database responds with a mean of 0.06 seconds. Adding the search term "mistletoe" increases the speed by a factor of 6 to 0.01 seconds and finally the complete phrase is processed in 0.005 seconds on average. As shown in the example, our method is more efficient when more search terms are entered in the search query, in contrast to conventional data structures which at first sight shows a paradoxical search time behaviour as illustrated in Figure 5. This procedure becomes more important with larger data volumes and is also used for keyword landscapes described below.
Semantic Web Standard: XML
As bibliographical data, in the case of CAM, does contain several heterogeneous elements arising from the different sources from which the datasets originate (i.e. publishers, libraries, authors entries, offline collections of datasets of research institutes), we decided to use the innovative technological web-standard XML (eXtended Markup Language) 23 to receive capture the output in a homogeneous structure. We constructed XML-based import interfaces to process incoming documents of varying document type definitions (DTDs), to extract structural and descriptive metadata from these documents and to deliver special document output styles on demand.
As an additional feature, we also implemented XMLinterfaces for the standards given by the Open-Archives-Initiative (OAI). With this XML-based document-management, CAMbase can be easily connected with national and international electronic databases and other digital libraries. 24,25 In addition, connections with other metadata, e.g. semantic web standard ontologies, can be established.
The basic idea behind this approach is to guide the user through the bibliographical landscape. The following figure illustrates this approach with the example of the search query for "Krebs" (German for "cancer").
Given the German language search term "Krebs", the search algorithm which is similar to other conventional databases displays a listing of articles which includes this search term or compositions like "Krebstherapie" (Engl: "cancer therapy"). Additionally, a keyword-landscape on the right side (German Title: "Themenlandkarte") is also given which describes the resulting dataset more precisely. Keywords occurring more often are displayed in the landscape with a bigger font size and at least one dataset behind every keyword displayed in the landscape is guaranteed.
By clicking on a term displayed in the landscape, the system again displays the results. In our case, clicking on the term "Hyperthermie" leads to 243 hits which are equivalent to a boolean AND-relation "Krebs AND Hyperthermie"). Naturally, the results of this search query are also linked to another landscape (see Fig. 6). With each navigation step, the user might choose another keyword for narrowing the result set even further. An additional feature is available after the nth step (n 2) offering the user permutations of the search terms of the (n-1)th step: For the query "Krebs AND Hyperthermie AND Immunstimulation" (2 hits), the system also asks Figure 4. Schematic description of an inverted file structure of a bibliographical dataset with linguistic processing information in the front and additional discriminant information (ADIs) in the back of the coding. "*stem$mr" denotes the decomposition of a search term i.e. "oncology" leads to "oncolog$∼y". Note that front truncation information (marked with a "*") and umlauting are features specially designed for German language (i.e. for composition terms like "Krebstherapie" (engl. "therapy of cancer").
whether the user might want to switch to the search query "Krebs AND Immunstimulation" without the term "Hyperthermie" (19 hits).
Discussion
The next generation of the Web is often characterized as the "Semantic Web" 26 in which information processing is no longer the task for human users, but rather handled by machines and algorithms giving rise to semantically empowered search engines such as the one described above. Semantic processing, however, requires standards for the syntactic form and for the semantic content of documents. One important technology for this purpose is already in place: eXtensible Markup Language (XML). There are other semantic web standards used for indexing bibliographic resources, like Web Ontology Language (OWL) and Resource Description Framework (RDF), that can facilitate data integration i.e. for genomic data 27 and applications that employ ontologies such as indexes and query models, e.g. GoPubMed, 28 Knowlegator 29 and Textpresso). 30,31 These systems show sophisticated use of of bibliographic metadata for indexing literature abstracts or full texts according to domain content. All of these approaches are domain specific and the approach presented here is likewise a first for the field of CAM.
A first evaluation of the vocabulary used in more than 28,000 search queries applied to CAMbase between 2003 and 2006 found that 12.2% were regarding authors publishing primarily in CAM, 10.9% on general terms, 30.3% on diseases, disorders and symptoms, 25.2% on therapies and procedures and 10.2% on plants and ingredients. 32 However, according to the recently conducted review of Kim et al 33 of tag ontologies for semantically linked data, a reorientation in evaluation strategies seems to be necessary to develop appropriate metrics, particularly if the results of this evaluation shall be used for the creation of a CAM-ontology.
In particular, a precise identification of entities is essential for the creation of an ontologydriven document indexing and data integration. 34 In addition, most of our data can be considered to be "old data", which means, there is only limited information for the creation of a valid ontology and changes in underlying ontologies might have far-reaching side effects for more recent data as the old scheme might not be compatible with the new results. 35 Nevertheless, if one looks closely at the CAM-landscapes created for every search query they implicitly indicate the presence of some underlying ontology. Thus, according to Vatant 36 a closer analysis of our CAM landscapes would be a first step to establish an ontological framework of CAM. Within this context the use of RDFs as interchange formats and the OWL as a formal description of the underlying concepts, terms, and relationships seem promising.
Taking into consideration the high amount of German language articles on CAM, our features and algorithms even today are of direct relevance. Whilst we have addressed the technical specifications in the construction of CAMbase, the issue of generating semantic ontologies in CAM is even more challenging than in other disciplines. This is because a catalogue of homogeneous synonyms for medical terms (which is valid for all CAM-disciplines) remains intangible (i.e. the term "liver" has a quite different meaning in traditional Chinese medicine than in homeopathy).
Mertz 37 also pointed out that one of the important future tasks for research in CAM is to build CAM vocabularies and make them available to the public domain. In doing so, one has to be aware that even if some values are shared with traditional biomedicine, axiological differences have to be considered as well. Although our examples are based on CAM bibliographical records in "CAMbase", the underlying algorithms we have developed are generic and work for other contexts as well. For example, a user might not only want to access the scientific literature, but also wants to look for a research institution or an expert involved in a distinct search topic (i.e. for a unique cancer treatment). In the same way as described above, this can be realized with an XML-based content management system, in which such information is stored 24 and a multi-dimensional web-portal is the next logical step.
Another example where deployments of semantic ontologies can be found within the field of CAM is the use of these features and tools for electronic lexicography. 38,39 In particular, machine readable lexicons or dictionaries with a codification of domain knowledge and literature metadata in accordance with a generic and extendible XML scheme model are more suitable in this context. One open problem in complementary medicine in this context is the semantic processing of the many resources of homeopathic remedies, Materia Medicia. Such repositories contain structured data on medical symptoms and the corresponding homeopathic remedies. Natural language processing of and ontology creation for such vast resources would be beneficial in the complex process of homeopathic prescribing. 40 conclusion Although the internet has brought a 'revolution to information technology', most of the current forms of web content remain incomprehensible for intelligent analysis by computers. The semantic web standard XML can be used to unite structural properties of databases, web requirements and the demands of end users. In this publication we were able to show how XML together with other algorithmic features like linguistic processing may aid the user to find his/her way through incomplete and often only semi-structured data in the field of CAM. On a long term, our approach may lead to a creation of a CAM ontology as currently described for other areas of biomedical research. 41
Disclosure
The authors report no conflicts of interest.
Publish with Libertas Academica and every scientist working in your field can read your article "I would like to say that this is the most author-friendly editing process I have experienced in over 150 publications. Thank you most sincerely." "The communication between your staff and me has been terrific. Whenever progress is made with the manuscript, I receive notice. Quite honestly, I've never had such complete communication with a journal." "LA is different, and hopefully represents a kind of scientific publication machinery that removes the hurdles from free flow of scientific thought." Your paper will be: • Available to your entire community free of charge • Fairly and quickly peer reviewed • Yours! You retain copyright http://www.la-press.com | 5,048.4 | 2009-01-01T00:00:00.000 | [
"Medicine",
"Linguistics",
"Computer Science"
] |
SILVA tree viewer: interactive web browsing of the SILVA phylogenetic guide trees
Background Phylogenetic trees are an important tool to study the evolutionary relationships among organisms. The huge amount of available taxa poses difficulties in their interactive visualization. This hampers the interaction with the users to provide feedback for the further improvement of the taxonomic framework. Results The SILVA Tree Viewer is a web application designed for visualizing large phylogenetic trees without requiring the download of any software tool or data files. The SILVA Tree Viewer is based on Web Geographic Information Systems (Web-GIS) technology with a PostgreSQL backend. It enables zoom and pan functionalities similar to Google Maps. The SILVA Tree Viewer enables access to two phylogenetic (guide) trees provided by the SILVA database: the SSU Ref NR99 inferred from high-quality, full-length small subunit sequences, clustered at 99% sequence identity and the LSU Ref inferred from high-quality, full-length large subunit sequences. Conclusions The Tree Viewer provides tree navigation, search and browse tools as well as an interactive feedback system to collect any kinds of requests ranging from taxonomy to data curation and improving the tool itself.
Background
Reconstructing phylogenetic trees is an important method for studying the evolutionary relations among organisms. In molecular phylogeny, genetic data are the basis for any kind of phylogenetic inferences. The ever growing amount of genetic data calls for approaches to dynamically visualize phylogenetic trees that comprise hundreds of thousands of sequences.
The SILVA project provides comprehensive, highquality datasets of small (16S/18S, SSU) and large (23S/ 28S, LSU) subunit ribosomal RNA (rRNA) gene sequences for all three domains of life (Bacteria, Archaea, and Eukaryota) [1]. These datasets include large phylogenetic guide trees which need to be interactively visualised for biologist seeking relationships and evolutionary information. For the current release (r128) of the SILVA dataset, the following trees are available: The SILVA trees include a large amount of metadata about each sequence, which requires several gigabytes of memory footprint, even in compressed form. Currently, browsing such large trees is only possible using desktop applications such as ARB [2], which is only available for Mac OS and Linux. Web-based solutions, like phylo.io [3] or iTOL [4], are not able to effectively handle trees of this dimension. The Open Tree of Life [5] relies on folding and allows the navigation of only one depth level at a time. OneZoom [6] supports large trees but relies on fractals to keep a manageable tree visualization and, hence, does not preserve meaningful branch lengths. In need of an effective, platform independent phylogenetic tree navigation tool, we developed the SILVA Tree Viewer as a web application.
Implementation
The SILVA Tree Viewer has been developed to provide access to large scale (e.g. 38,585 taxa with 645,151 sequences for the SSU tree) trees in a web browser. The SILVA website hosts the viewer (version 1.1) to provide access to the SSU RefNr99 and LSU Ref phylogenetic guide trees of the SILVA database.
The viewer is based on Web Geographic Information Systems (Web-GIS) technology, in particular, it leverages the PostGIS extension [7] to the PostgreSQL database [8] for the back end, which stores the tree "map". The Leaflet JavaScript library [9] is used for the front-end, which enables visualization and navigation of the tree as if it was a map. Parsing of the input tree uses the ETE 2 Python library [10]. Additional components have been written in Python to provide features required to satisfy the needs of the biologists. The use of Web-GIS technology enables the viewer to display large scale trees at the cost of not supporting folding.
Results and discussion
The phylogenetic tree is presented to the users with pan and zoom functions similar to those in web mapping technology. A search function, as well as browsing of phyla and major clades, is provided to allow users to quickly lookup sequences and taxa. Additionally, the viewer is equipped with an interactive feedback system which allows reporting of taxonomic problems and asking questions to the tree curators, thus supporting them in improving the quality of the trees. The core components of the SILVA Tree Viewer's main interface are shown in Fig. 1.
Tree navigation
The tree can be navigated and zoomed in analogy to Google Maps. The different zoom levels allow rapid drill-down navigation from the overview to any area of interest. To further facilitate navigation in large trees, the Navigation bar indicates the current view in relation to the whole tree. The taxonomic groups are shown as coloured rectangles whose colour depend on the taxonomic rank of the group (Fig. 2), enclosing all sequences belonging to them. Further details of any sequence, provided by the SILVA database, can be shown by clicking on its label.
Taxonomic context
Taxonomic context information is essential for the understanding of the tree. The viewer provides taxonomic context about the currently displayed tree portion in Fig. 1 Overview of the SILVA Tree Viewer interface. The following components are available: Main display: shows the SILVA Tree Viewer in the browser or in full-screen mode. Tree area shows the phylogenetic tree; Control buttons to access the main functions; Navigation bar indicates the current view with respect to the whole tree. Taxon bar provides details of a taxon; Taxon brackets: show taxonomic information; Feedback provides access to the feedback system, where support requests -including reports about taxonomic data -can be filed; Scale: provides a reference for interpretation of the horizontal branch length at the current zoom level two ways. The Taxon bar, at the bottom of the screen, shows the full path of the taxon under the mouse pointer, along with sequence counts. The Taxonomy Brackets, shown to the right of the visible tree, are vertical square brackets according to the vertical extension of taxonomic groups "crossing" the current view of the tree. They are shown from right to left according to their rank (domain being the rightmost). In case the brackets are too short for the corresponding label to be shown, no label is shown and the taxonomic information remains available via tooltip.
Search and browse tools
A search function is provided which allows searching for distinct elements of the tree. The search result is displayed in the viewer (Fig. 3).
Search for sequences
The user can search for sequences by accession number, returning any sequence whose accession number starts with the query. In this case, the search function can handle start and stop positions, separated by full stops (i.e. the complete accession is composed of <ACCES-SION NUMBER > . < START>. < STOP>). Furthermore, a search by full name can be performed, returning all sequences whose species name contains the query string.
Search for taxa
The user can search for taxonomic groups by taxon name, returning all taxa whose name contains the query. The search by taxonomic path returns all taxa having the query as part of their full taxonomic path is also supported. In this case, since all taxa within a lower rank will match the search condition, only the highest (more generic) rank result will be returned. This search marks taxonomic groups as rectangular features which are loaded to the results navigator for browsing by the user.
Browse phyla
All taxonomic groups which are direct children of any domain (phyla for Bacteria and Archaea, kingdom/major clades for Eukaryota) are listed for direct browsing. A click on the taxon name loads all taxonomic groups comprising it, zooms the view to the first result and allows navigation through results as for the other search types.
Feedback system
The feedback system is an important part of the SILVA Tree viewer since the SILVA guide trees are manually curated. The feedback system includes the possibility to report a so-called "data problem": wrong classifications, Fig. 2 Colour legend of taxonomic rank. Each rank is assigned a different colour to allow clear distinction of taxonomic groups Fig. 3 Result set for a search for sequences. All sequences matching the search criteria are marked with an orange dot. The current result is highlighted with a red circle. The top right corner shows the result browser panel, allowing to show all results, navigate to the next/previous result and to leave the results navigator errors in the taxonomy and any other problem with the phylogenetic data. This eases the curation process, thus supporting the quality improvement of the dataset in a cooperative effort with the users. Reporting a data problem is a two-step process: first, the data affected has to be selected by picking it interactively on the SILVA Tree Viewer, second, a report form has to be filled to provide details about the problem.
The feedback system can also be used to report on technical problems and to propose improvements to the viewer, as well as for asking general questions to the SILVA team.
Known limitations and future developments
In its current version (1.1), the SILVA Tree Viewer implements only visualization tools, thus no editing of the trees is supported. Due to technical constraints of the Web-GIS technology there is also no support for group folding, thus the tree is always shown as fully expanded.
Integration with the SILVA search and download system is planned to allow extended search functionality and the direct download of sequence data. The integrated feedback system will provide further guidance for the SILVA team in providing new features, taking into account user suggestions and recommendations.
Conclusions
The SILVA Tree Viewer enables users worldwide to navigate the large phylogenetic trees provided by the SILVA database. Hundreds of thousands of sequences can be searched in the tree and browsed in their taxonomic context. An interactive feedback system is provided to collect requests for improving the tool itself and guide data curation.
Availability and requirements
Project name: SILVA Tree Viewer.
Any restrictions to use by non-academics: none. | 2,276.4 | 2017-09-30T00:00:00.000 | [
"Biology",
"Computer Science",
"Environmental Science"
] |
Selection for small body size favours contrasting sex-specific life histories, boldness and feeding in medaka, Oryzias latipes
Background Studying variation in life-history traits and correlated behaviours, such as boldness and foraging (i.e., pace-of-life syndrome), allows us to better understand how these traits evolve in a changing environment. In fish, it is particularly relevant studying the interplay of resource abundance and size-selection. These are two environmental stressors affecting fish in natural conditions, but also associated with human-induced environmental change. For instance, fishing, one of the most important threats for freshwater and marine populations, results in both higher mortality on large-sized fish and reduced population density. Results Medaka, Oryzias latipes, from lines selected for large or small size over ten generations, were exposed individually to high or low food availability from birth to adulthood. Maturation schedules, reproductive investment, growth, boldness and feeding were assessed to evaluate the effect of size-selection on the pace of life, and whether it differed between food contexts (high and low). Different food abundance and size-selection resulted in diverse life histories associated with different feeding and boldness behaviour, thus showing different pace-of-life-syndromes. High availability of food favoured faster growth, earlier maturation and increased shyness. Selection for small size led to slower growth in both males and females. But, the life-history trajectory to reach such growth was sex- and food-specific. Under low food conditions, females selected for small size showed earlier maturation, which led to slower adult growth and subsequent low willingness to feed, compared to females selected for large size. No line differences were found in females at high food conditions. In contrast, males exposed to selection for small size grew slower both as juvenile and adult, and were bolder under both feeding regimes. Therefore, the response to size-selection was more sensitive to food availability in females than in males. Conclusions We showed that size-selection (over ten generations) and resource abundance (over developmental time) led to changes in life history and behaviour. However, the effect of size-selection was sex- and context-specific, calling for precaution when drawing general conclusions on the population-level effects (or lack of them) of size-selective fishing. Conservation and management plans should consider this sex- and context-specificity. Electronic supplementary material The online version of this article (10.1186/s12862-019-1460-x) contains supplementary material, which is available to authorized users.
Background
Variation in life histories arises from different trajectories of survival, growth and fecundity that maximises fitness in different environments [1]. Somatic growth, time of maturation and reproductive investment are key factors in an individual's life cycle and hence represent a life-history strategy [2][3][4]. Numerous environmental factors (e.g., resource abundance, temperature, inter-and intraspecific interactions) can affect growth and maturation [5,6]. Studying how these factors affect variation in life-history strategies is particularly relevant to better understand the potential evolutionary adaptation of wild populations [7,8].
In the particular case of fish, size-selective mortality and resource availability are two environmental stressors of interest. Selection on body size in fish can result from natural mortality, as predators normally target small individuals [9][10][11] or from fishing-induced mortality that is commonly higher on large individuals [12][13][14]. Food abundance is affected by many factors such as intraspecific competition, climate change and predation [15]. Increased predation (by predators or fishing) not only leads to selection on size, but further results in lower abundance in the population and thus increases food availability for the survivors [16]. This interplay between size-selection and resource availability on life-history traits is still poorly understood [17].
Size-selective mortality on large individuals, as that induced by fishing, leads to decreased life span. And hence, according to life-history theory, fast life histories are favoured i.e., earlier maturation and increased reproductive investment [5,14,18,19]. However, how size-selective fishing affects juvenile and adult growth remains unclear [5,14]. Most empirical evidence shows no direct effect of size-selective fishing on growth [20], or an indirect effect resulting in reduced growth only after maturation [14]. Yet, such a response may not appear if resource availability is high [21]. Moreover, populations exposed to fishing would experience fast or slow growth, depending on the selectivity of the gear and the minimum size imposed [5,22,23].
Life-history diversity often entails co-variation with behaviour, as behaviour and physiology are the basis for the tradeoff between current and future reproduction (referred as pace-of-life syndrome) [18,24]. Fast life history is expected to be linked with behaviours that favour energy acquisition (e.g., foraging behaviour) and reproduction over survival (e.g., boldness), resulting in a fast pace of life [18,24]. However, the few studies that have evaluated how behaviour was affected by size-selective mortality showed that fish exposed to positive size selection were less bold and less eager to forage [25,26]. Behavioural and life-history co-variation are often sex-and context-dependent [24,27]. Thus, knowledge on how the suit of life-history and behavioural traits changes due to size-selective mortality is still limited. Particularly, little is known about whether such changes are affected by the release from density-dependent food limitation following the reduction in abundance in harvested populations, and whether both sexes despite their different investment in reproduction are responding in a similar way.
Here we tested how two different size-selective mortality regimes (on large or small size) affected the pace-of-life syndrome in medaka (Oryzias latipes). In addition, we evaluated whether the result of size-selection differed under different food availability conditions. Wild medaka were used to produce two laboratory-reared lineages over ten generations with selection for either large or small standard body length at 75 days post-hatching (dph); here referred to as large-selected and small-selected lines, respectively [28]. The line selected for small size mimics the size-selective pressure induced by fishing, where large individuals are harvested and mainly the smaller individuals can reproduce. The large-selected line mimics natural size-selection, where predators feed on small individuals. Lines were selected under common-garden conditions with abundant food supply. Individuals from the 11th generation were reared in isolation under two different feeding regimes (high and low food availability). We fitted data on age and length to a biphasic (juvenile and adult) growth model that allowed studying trade-offs in energy allocation between growth, maturation and reproduction [29][30][31]. Females were measured every 2 weeks, while males every three, thus models were fitted separately for each sex. Once fish became mature, we assessed the feeding behaviour and boldness of each individual. Both behaviours have ecological validity, as they are respectively linked to foraging rate and risk of predation and hence related to fitness [32].
We expected individuals from the small-selected line to express 1) fast life histories (i.e., early maturation and high investment in reproduction), with growth only slowing down after maturation, and 2) behaviours associated with a fast pace of life, i.e., higher boldness and feeding rate compared to large-selected line [14,18]. We also predicted that 3) individuals under low food availability would present slower pace of life relative to under high food [2,33]. The effects of size-selection on behavioural and life-history traits would hold true at differing food abundances and for both sexes if food availability did not play a role in the selection process [17,34], as expected under controlled equal and abundant food availability in the selection experiment. However, different reproductive investments between sexes may affect this expectation [27].
Life-history traits
Both male and female medaka showed large inter-individual variation in growth and life-history parameters (age at maturation and maximal potential growth; Figs. 1a and b). Lifehistory traits differed between the two laboratory-reared lines (i.e., small-selected vs. large-selected) and the two food availability conditions applied during development (high vs. low; see Methods for details). Both low food availability and selection for small size led to smaller size, but the life-history trajectory leading there was context-specific (different between treatments) and different between sexes.
At the end of experiment (180 dph), fish in the high food availability treatment were larger compared to those fed low food in both sexes (Figs. 1a and b). In addition, fish from the large-selected line were larger at 180 dph compared to those originating from the small-selected line, although this was food-dependent in females (Fig. 1a). Overall, the effect of sizeselection in females was absent in the juvenile phase of the growth curve (Fig. 1a). It was only observable in adult growth under low food conditions, when the earlier age at maturation of small-selected females resulted in a slower adult growth and hence smaller length at 180 dph ( Fig. 1a; Table 1). Thus, the effect of line on female growth only occurred indirectly through its effect on maturation. The effect of size-selective mortality in males was evident, with juvenile and adult growth always being slower for small-selected males, and direct, as it was not affected through any other parameter.
Average age at maturation (a mat ) was 68 dph (SD = 24.5) in females and 71 dph (SD = 10.6) in males, which is within the range commonly observed for medaka (60-90 days) [35]. It was affected by the interaction of food and line in females, but not in males. As expected, females delayed maturation under low food conditions, but the delay differed between small-and large-selected females (Tables 1 and 2). At low food conditions, small-selected females matured earlier (81 dph) than large-selected females (97 dph), while at high food both lines matured at 48 dph (Tables 1 and 2). In contrast, maturation in males was not affected by size-selection, and only delayed under low food (Tables 3 and 4). Investment in reproduction was not affected by size-selection in either sex, as lower food led to a lower investment in reproduction in both sexes (Tables 2 and 4).
The maximal potential growth rate (related to β; see Methods) describes the growth rate before the investment in reproduction is accounted for [29]. Higher values of β correspond to faster juvenile somatic growth rate, i.e., before any investment in reproduction takes a b Age at maturation, a mat , reproductive investment, r, and growth allometric exponent, β. See Table 2 for treatment effect and statistics on parameter estimates place. Specifically, β is the exponent of the allometric relationship between growth rate and weight. It describes the allometric scaling of energy readily available to growth (i.e., net energy after expenditure on metabolism has been accounted for) [29]. Our average estimated value for β was 0.58 (SD = 0.06), which is similar to the 0.66 assumed in many growth models [29,36]. For both males and females, β was lowerresulting in a slower juvenile growthat low food conditions than at high food conditions, (Tables 1 and 3). In males there was also a direct effect of line on β, which was additive to the effect of food on β, but lower in magnitude. Large-selected males at high food conditions presented the highest value of β and hence the fastest juvenile growth rate, while small-selected males at low food had the lowest β and slowest juvenile growth rate (Table 3; Fig. 1b).
Behavioural traits
Two different behaviours were assessed once the test fish reached maturity. Feeding behaviour was measured as the total number of bites the fish took at the supplied food, while boldness was related to time spent freezing at the bottom of the tankthe shorter freezing time the bolder an individual is [32]. Both behaviours were repeatable over time and across contexts (Additional file 1: Table S1). Feeding behaviour and boldness were evaluated as both total amount (bites on food or freezing time) and probability (of zero bites or zero freezing) due to the nature of the data (see Methods for details).
Food availability and size-selection affected the behaviours and this was sex-specific (Table 5).
Feeding rate was higher under low food conditions in both sexes (Fig. 2), but this effect was stronger in females than in males. Females reared at low food fed more than those at high food, as indicated by both their higher number of total food bites ( Fig. 2b; low-high food comparison, estimate ± SE = 0.27 ± 0.10 bites, z = 2.72, P = 0.007) and lower probability to remain without eating (low-high food comparison: 3 times lower odds, z = − 4.48, P < 0.0001). Males at low food had only lower probability of no biting at all relative to high food males (low-high food comparison: 1.6 times less odds, z = − 2.81, P = 0.005). In addition, small-selected females presented reduced feeding activity ( Fig. 2a-c) and they had higher probability of not eating at all compared to large-selected ones (small-large-selected comparison: 2.4 higher odds, z = 3.48, P < 0.001).
Finally, total freezing time (used as a proxy for boldness) was affected by size-selection in males and by food availability in females (Fig. 3). Females at low food conditions were bolder ( Fig. 3a-c) than females fed high food quantities, as the former had 1.6 times higher probability of not freezing at all (estimate ± SE = 0.48 ± 0.21 log(odds), z = 2.24, P = 0.025). Small-selected males were Age at maturation, a mat , reproductive investment, r, and growth allometric exponent, β. See Table 4 for treatment effects and statistics on parameter estimates
Discussion
Size-selection over ten generations affected life-history traits, boldness, and feeding behaviour. These traits were also affected by exposure to different food availabilities during development of each fish to a somewhat larger extent. Both the selection for small size (experienced by the parental generation) of fish and exposure to low food availability during development, led to shorter length at adulthood. However, the life-history traits (age at maturation, maximal potential growth and investment in reproduction) that led to such outcome were context-specific and differed between food availability and size-selection. Moreover, the effect of size-selection, but not the effect of food availability, was sexspecific.
It should be noted that comparisons between the two drivers evaluated here (i.e. size-selection and food availability) are challenging, as they were acting on different ecological processes (i.e. plasticity vs. evolution). Specifically, food availability affected the environment of the studied fish during their development from juvenile to adulthood, while size-selection was the stressor experienced by ten ancestral generations. Food availability and size-selection can lead also to evolutionary and plastic changes, respectively [33,37], but here the experimental design limited those effects. Indeed, food availability was a proximate driver of plastic change in life-history traits, while size-selection could affect life history through genetic change (evolution), as it was performed under common garden conditions over ten generations [38][39][40]. Epigenetic inheritance cannot be ruled out, but it is likely not related to size-selection [40]. We warrant this simplification to better disentangle the effect of each treatment, before more complex interactions between treatments, through intraspecific competition should be assessed. In addition, sex differences could be due to differences in handling the sexes during the experiment due to time and space limitations. Males were measured every 3 weeks and water quality was maintained with partial water changes (for half of the males) and constant flow-through (on the other half), while females were measured every 2 weeks and water quality was maintained by constant flow-through of water. Even though these differences should not affect the effect of experimental treatments, but the statistical power, its effect cannot be completely ruled out.
Effect of food
As expected, low food availability led to smaller size at adulthood and slower growth relative to high food availability in both sexes [2,33,41] . Low food availability led to higher willingness to feed and increased boldness (less freezing time) compared to high food availability, although the latter only happened in females. Thus, the hunger level might be the driver of increased boldness and increased willingness to eat, as seen elsewhere [42,43]. Specifically, Magnuson [43] found that under limited food medaka became more aggressive to ensure access to food and maximal growth rate. Fast growth is generally associated with boldness and willingness to forage through evolutionary trade-offs [18], but these links often depend on environmental conditions [44]. In the present study, fast-growing fish at high food condition might have been feeding at a maximum ratio. Indeed, they showed lower immediate willingness or need to feed (no hunger and low appetite) during our observational assays, and lower boldness (at least in females). The links here between growth and behaviour are driven by hunger levels and are thus probably plastic.
Effect of size-selection
Fish selected for small body size grew slower relative to those selected for large body size, as normally observed in other size-selective experiments on fish [25,38,39,45,46]. However, the overall shorter length at adulthood in small-selected fish, were achieved through different growth trajectories between sexes. Here, growth depended on the combination of the exponent of the maximal potential growth rate (related to β) and age at maturation. Size-selection affected growth rates in males throughout their lives, but only during adulthood in femalesthis was due to size-selection effects on different parameters. Males experienced a direct effect of size-selection on growth, affecting only βmales selected for small size had lower values of β and thus grew slower. This, together with the lack of differences in maturation and reproductive investment between lines, led to different growth curves throughout the life of the fish. Females experienced a direct effect of size-selection on age at maturation under low food conditions. This led to an indirect effect of size-selection on growth, which resulted in equal juvenile growth between lines followed by a slowed down adult growth in small-selected females reared at low food. At high food conditions, females of both lines presented equal growth curves. Reduced age at maturation after selection for small size is expected from theory [14,19] and observed in laboratory experiments [25,34,39,45], but only occurred for females reared under low food conditions in our experiment.
The interaction between size-selection and food suggests that food availability played a role during the selection process in females [17,34]. The higher investment in reproduction in females relative to in males may have resulted in females perceiving their food environment, during the selection process, as quantitatively low. This higher sensitivity to food in females also became apparent in our results on behaviourthe feeding and boldness of females were affected more by food than those of males. This is in concordance with earlier studies in medaka showing that females are more sensitive to fasting than males, resulting in reduced gonadosomatic index and fecundity [47]. Thus, it seems that females were only selected for small size under low food availability, and hence the response to selection is stronger under conditions similar to those experienced under the selection process [48]. Similar resource-dependent response to size-selection has been seen in other fish species [17,49]. In the case of the killifish, Rivulus hartii, the females response to size-selection were also more sensitive to food availability than the males [17]. The context-specificity of life-history trajectories, which differed between sexes here, but also between species [25,38,39,45,46], indicate that the life history response to size-selection is more complex than often assumed.
Medaka selected for small size were bolder than medaka selected for large size. This was evident at least in males, which showed reduced juvenile and adult growth. Medaka females fed less when selected for small size, which presented early maturation and slow adult growth. Size-selection experiments with Atlantic silversides, Menidia menidia, have shown a reduced food consumption in fish selected for small size [26]. However, similar selection experiments have indicated reduced boldness in fish selected for small size [25,26]. As noted earlier, the evolutionary link between fast growth and boldness is common mainly when predation is high and resources are limited [44]. For instance, in medaka exposed to low food, higher aggression (commonly correlated with boldness [32]) was linked with higher growth rate, but this link disappeared when food supply was high [43]. When resources are abundant, low activity and boldness leads to higher growth as predicted by the allocation model [50]. Overall, we observed that changes in life-history parameters due to size-selection also led to changes in behaviour, which were consistent over time and among context [51].
Medaka do not present drastic morphological sexual dimorphism [35], but males and females present behavioural and physiological differences. Both sexes present courtship and competitive behaviours, but these are more evident in males [43,52], while females invest more in reproductive tissues [53]. Moreover, external factors (e.g., food, temperature, pollutants) seem to alter reproductive investment in females (but not behaviour) and aggressiveness in males (but not reproduction or to a lower extent) [47,[54][55][56][57]. Overall, this sexual dimorphism in sensitivity to external stressors could explain the sexual differences observed in the present study. Moreover, our results highlight the need of assessing sexual differences while evaluating life-history and behavioural traits.
Here we show that the effect of size-selection, such as the one induced by fishing, on life history can entail behavioural changes. Boldness and foraging are at the core of predator-prey interactions, as they determine the effects of consumers on their prey and are affected by the presence of predators [58]. Therefore, changes in behaviour due to human-induced size-selection can in turn affect the resource community and ecosystem processes through different pathways, such as alteration of the strength of the trophic cascade [59]. Moreover, the links between life history and behaviour are more complex than often assumed and dependent on sex and environmental conditions. A better understanding on how sizeselective mortality affects this suit of traits is not only relevant for management and conservation of exploited species, but allow us to predict further ecosystem consequences. We suggest that future mesocosm-based experiments should assess whether differences in these sizedependent correlated traits can translate to changes in the trophic cascade to better evaluate the ecosystem impacts of size-dependent mortality. Moreover, such experiments could allow the interplay of food availability and size-selection and hence assess the context-dependency of such ecological variations.
Conclusions
In the present study, both sexes displayed smaller size and slower growth when exposed to selection for small size, similar to the size-selection induced by fishing. However, life-history strategy and pace-of-life syndromes were sex-and context-specific in the present study.
Small-selected females showed a fast pace of life under low food conditions, i.e., early maturation and fast juvenile growth only slowed down after maturation. Slower adult growth was linked to reduced feeding rate. However, small-selected males grew slower throughout life linked with a higher boldness relative to fish selected for large size. Sex differences may be due to differences in investment in reproduction and food requirements during the selection process.
Conservation plans concerned with size-selectivity (e.g., fishing or introduction of novel predators) should consider that a suit of behavioural and life-history traits, rather than only size, are responding to the new selection. In addition, the interplay between size-selection and resource availability should be evaluated to better account for the impact on ecosystem functioning and services.
Selection lines
Individuals used in this study were the offspring of the 10th generation (F10) produced by a size-selection experiment performed in the laboratory (see [28] for details on experimental protocol). Briefly, the selection for small standard length (SL ± 1 mm) (referred as Small-selected line) mimicked the selection imposed by fishing, where large individuals are removed and only small individuals are allowed to breed. Selection for large individuals (i.e., large SL, referred as Large-selected line) represents natural mortality in the wild. Specifically, at 60 days post hatching (dph), the ten brother-sister families with the shortest or largest average length were selected for the small-selected and large-selected lines, respectively. At 75 dph, individual size-selection took place and the breeders (two males and two females per family) for the next generation were chosen. The largest and the smallest individuals per family were chosen for the line selected for large size and selected for small size, respectively.
Fish rearing and feeding experiment
At generation F10, we randomly chose five families from each line to produce at least 64 fish per line and per feeding treatment (N = 256). F10 breeders were 90 dph when eggs were collected and kept in an incubator until the larvae hatched. The incubator was checked for daily hatchlings and hatching day was recorded for all larvae. Larvae hatched on the same day were kept together within family in 3 L tanks for 2 weeks. At 14 dph, each surviving individual larva was randomly assigned to one of two feeding regimes and housed in isolation in 1 L tanks.
Every 2 weeks females were measured for SL with a measuring board and weighted (W ± 0.001 g), while males were only measured and weighted every 3 weeks due to time limitation. The first two measurements (at 14 and 28 dph) were obtained with ImageJ (version 1.51 s; [60]) from photographs of the larvae placed in a petri dish filled with water (no weight was taken), as the larvae were too small to be handled otherwise. Later, individuals were anaesthetised (Metacaine; Sigma; following [35] protocols) to minimise stress during handling. At the end of the experiment 259 fish were included in the growth analysis as they had at least four measurements of size. Total number was not balanced among sexes and treatments. Specifically, we obtained growth trajectories from a total of 143 females (39 from large-selected line in high food, 36 from small-selected line in high food, 28 from large-selected line in low food and 40 from small-selected line in low food) and 116 males (19 from large-selected line in high food, 41 from small-selected line in high food, 20 from large-selected line in low food, and 36 from small-selected line in low food).
Fish were fed once a day (morning) with quantified amounts of newly hatched Artemia salina based on previous experiments [61,62] During the first 2 weeks, from birth until the initiation of the treatments, all fish received the same food quantities (0-14 dph: 0.01 ml of filtered, undiluted Artemia per fish). Then, high and low food levels were applied and increased every 2 weeks. These quantities of food delivery were chosen to sustain two different growth rates, with the high level being double the low level [61,62]. Specifically, high food level consisted of 0.01 mL of Artemia at 15-28 dph and reached 0.05 mL from 143 dph until the end of the experiment.
In the 1-L tanks, females were housed in a flowthrough system, while only half of the tanks holding males where housed in the same flow-through system. Due to space limitations, the other half of male tanks was housed outside of the flow-through system (partial water changes every 2 weeks), but still inside the same lab (same temperature of 26 degrees Celsius). These males were only outside of the flow-through system during the growth part of the experiment, and not during the behavioural assessment. In addition, to minimise this effect, males were randomly rotated every 3 weeks between inside and outside the flow-through system. Finally, measurements were more frequent in females than males and, because they were also exposed to constant flow-through, data sets are analysed separately. At the end of the study all individuals were anaesthetised with Metacaine (Sigma; following [35] protocols) and later euthanised with an overdose of Metacaine as they were included in a study on pituitary gene expression.
Estimation of life history traits
The life-history parameters, age at maturation, investment in reproduction and growth rate were estimated using the Quince-Boukal biphasic growth model [29,30]. This growth model fits better juvenile and adult growth curves compared to other commonly used growth models for fish, and has proven useful for generating management advice [36]. The model follows a continuous function with a smooth transition between the juvenile and adult growth phase. This transition is due to allocation of energy to reproduction. It assumes that the maximal potential growth rate scales allometrically with body size.
The formulation assumes that juveniles allocate all surplus energy into growth (i.e., the investment in reproduction r a = 0). Juvenile growth curve for length, at age, a, L a , follows: The adult growth rate considers the investment in reproduction, r, of the mature individuals, whose age is larger than their age at maturation (a > a mat ) and the weight-age growth curve follows: , assuming the conversion factor between somatic and gonadic investment, q, in [29] to be 1 as in [36,63]. L 0 is length at birth, c and β are the intercept and exponent in the allometric relationship of growth rate with weight, dW/dt = c W βwhich is also referred as maximal potential growth rate. b and α are the intercept and exponent of the allometric relationship between weight and length, W = b L α . See [29] for all details in formulation. The coefficient, b, and exponent, α, of the allometric relationship of weight, W, with length, L, were estimated with the data prior running the growth model. The values obtained from a regression model with log-weight and loglength were α = 2.7 and b = 0.04 mg mm -2.7 , which were used for both males and females. Length at birth, L 0 , was obtained from a subsample of the individuals (25 females and 22 males) that were photographed and measured at 0 dph. Length at birth did not differ between sexes (F 44, 1 = 0.67, P = 0.5) or lines (F 44, 1 = 0.36, P = 0.7). The mean length at 0 dph was 3.9 ± 0.4 mm, thus L 0 = 4.0 mm was used in the model for both males and females. The linear models were performed with the "stats" R package [64].
Growth curves were estimated separately for males and females, as sexes normally differ in their life-history optima and considering together might impede the study of pace-of life syndrome [27], but also due to the differences in experimental handling between sexes. Thus, here we aimed at studying the effect of size-selection on growth in both sexes without directly comparing sexes. For each sex, the growth model estimated age at maturation a mat , reproductive investment r, and the exponent in the allometric growth rate-weight relationship β. To improve model convergence the coefficient in the allometric growth rate-weight relationship was fixed to c = 0.15 mg 1-β day − 1 . Initial exploration of the data showed that this value was the most appropriate for our data and changes in this value with an increase or decrease of 10% did not qualitatively change the results. This scaling coefficient is species-specific [65], it has been estimated for another small freshwater fish, guppy Poecilia reticulata (c = 9-14 mg 1-β day − 1 ; [45]).
All statistical analyses were performed with the R software (version 3.5.0; [64]). The parameters were estimated from a non-linear mixed effect model in the R package "nlme" (version 3.1.137; [66]) with fish identity as random effect on r and β for males and on β only for females. The random structure was chosen following recommendations from [67]. Models included a residual autocorrelation structure ARMA (0,2), chosen according to guidelines in [67]. Line, food treatment and their interaction were tested as fixed effects on a mat , r, and β for both sexes separately. All possible models were run and model selection was done by comparing them with AIC (Akaike information criterion). The model yielding the lowest AIC was considered the best-ranked model in the Kullback-Leibler information [68]. However, for males this model was further simplified through hypothesis testing and non-significant predictors were dropped one by one to obtain a more parsimonious model [67].
Behavioural traits
Behavioural observations took place on mature fish (mean age = 124 dph ± 33 SD for females and 131 dph ± 32 sd for males) in three different settings: 1) Control conditions where fish were fed Artemia undisturbed, 2) Novel conditions where fish were fed a novel food source (4 pellets; JBL NovoGrano Mix) undisturbed, and 3) Threatening conditions where fish were fed Artemia immediately after being netted out of the water for 2 s as threat stimuli. Each fish was exposed to the three conditions inside their tank in a randomized order, each exposition replicated twice. Only one condition was tested per day and the second replicate took place 1 week after the first one. The six measurements (three conditions repeated twice) of each behaviour (i.e. feeding and boldness; see details below) ensured that we considered consistent intrinsic individual variation in behaviour [51]. However, the effect of time and experimental condition was not the main focus of the study, thus details on the experimental set up and acrosscontext repeatability can be found in the supplementary material (Additional file 1: Tables S1-S3).
During 5 min of observations we counted 1) number of bites to the supplied food (Artemia or pellets) as a measurement of willingness to feed, and thus is associated with foraging on prey [26], and 2) total time frozen at the bottom of the tank, as a proxy of boldness, which is related to predation risk and survival in natural conditions [32]. A fish that spends more time immobile is considered less bold than one that spends little time immobile [32]. Both behaviours are ecologically relevant, as they are linked to interactions with prey and predators [58]. Behaviour tests were performed on 80 females and 107 males, as only these were available at the time.
The two behaviours were analysed with generalized mixed effect models using the R package "glmmTMB" [69] with measurement nested within fish identity and experimental condition included as two separate random effect on model intercept. Full models contained size-selection (Large-selected vs. Small-selected), food (High food vs. Low food) and their interaction as fixed effects. The analyses of number of bites and freezing time (measured as an integer count of seconds) were performed following negative binomial distribution and allowing for zero inflation, as 33 and 49% of the data for number of bites were zeros for females and males respectively, and 52% of the values for freezing time were zeros for both males and females. Negative binomial distribution was used due to overdispersion observed with Poisson distribution. The final model was the one ranked with the lowest AIC. The residuals from all the final models were evaluated following [67] and fulfilled all the requirements. In addition, simulations showed that the zero-inflated models represented well the data; particularly the models estimated an equivalent number of zeros as in the original data on average in 40% of the simulations [70]. It should be noted that the analyses with negative binomial distribution and zero inflation evaluate each behaviour in two ways: 1) total amount of bites or seconds frozen, and 2) probability of zero bites or zero seconds frozen. Thus, the effects of sizeselection and food are evaluated for both cases.
Additional file
Additional file 1: Detailed experimental set up for the behavioural assessment. Behavioural repeatability (Table S1) and effect of experimental conditions on male (Table S2) and female (Table S3) behaviours. (DOCX 32 kb) Abbreviations AIC: Akaike information criterion; a mat : age at maturation; b: Intercept of the allometric relationship between weight and length; c: Intercept of the allometric relationship between growth rate and weight; dph: Days posthatching; F10: 10th generation; L: Length; L 0 : Length at birth; r: reproductive investment; SL: Standard length; W: Weight; α: Exponent of the allometric relationship between weight and length; β: Exponent of the allometric relationship between growth rate and weight | 8,192.8 | 2019-06-19T00:00:00.000 | [
"Biology",
"Environmental Science"
] |
Accounting for cell lineage and sex effects in the identification of cell-specific DNA methylation using a Bayesian model selection algorithm
Cell- and sex-specific differences in DNA methylation are major sources of epigenetic variation in whole blood. Heterogeneity attributable to cell type has motivated the identification of cell-specific methylation at the CpG level, however statistical methods for this purpose have been limited to pairwise comparisons between cell types or between the cell type of interest and whole blood. We developed a Bayesian model selection algorithm for the identification of cell-specific methylation profiles that incorporates knowledge of shared cell lineage and allows for the identification of differential methylation profiles in one or more cell types simultaneously. Under the proposed methodology, sex-specific differences in methylation by cell type are also assessed. Using publicly available, cell-sorted methylation data, we show that 51.3% of female CpG markers and 61.4% of male CpG markers identified were associated with differential methylation in more than one cell type. The impact of cell lineage on differential methylation was also highlighted. An evaluation of sex-specific differences revealed differences in CD56+NK methylation, within both single and multi- cell dependent methylation patterns. Our findings demonstrate the need to account for cell lineage in studies of differential methylation and associated sex effects.
Introduction
DNA methylation is a widely studied epigenetic modification that plays an essential role in the regulation of gene expression [1], cell differentiation [2] and the maintenance of chromatin structure [3]. Advances in
Materials and methods Data
Publicly available Illumina 450K methylation data were obtained from six healthy males subjects [13]. The data contained methylation β-values on seven isolated cell populations: CD19B+ cells, CD14+ Monocytes, CD4 + T cells, CD8 + T cells , CD56 + Natural Killers (NK), Neutrophils (Neu) and Eosinophils (Eos). Samples of the same cell types excluding Eosinophils were also obtained for five healthy female subjects [18]. In the absence of Eosinophils, The methodology described in this paper was therefore only applied to the six cell types common to both datasets. For both sexes, 445,603 CpG probes were available for analysis.
Consistent with previous findings [13], differences in DNA methylation levels across CpG sites were primarily driven by cell type differences, as opposed to being an artefact of between-subject variation (Fig 1).
A hierarchical clustering of cell-sorted samples revealed that differences in methylation levels were closely aligned with expected cell lineage in whole blood and were sex-independent. Additional cell sorted 450K methylation data on female and male subjects [18], aged 32-50 years, were obtained from public databases, for validation of model results. Female data consisted of cell-sorted samples on six healthy subjects downloaded from ArrayExpress (Accession number: E-ERAD-179) for CD19B + B cells, CD14 + Monocytes, CD4 + T and CD8 + T cells. Methylation data on six healthy male subjects from the same study were downloaded from Gene Expression Ominibus (Accession number: GSE71245). Following filtering and normalisation, 436,067 and 445,307 CpG sites were available for validation for females and males respectively, based on their correspondence with available sites in the training data. Table 1: Description of cell-specific models based on cell lineage from Fig 1. Each candidate model corresponds to a partition of cell types into non-overlapping groups. Cell types within each set of parentheses, {}, belong to the same partition and were assumed to have the same level of methylation. For each cell-specific model excluding 'All', the partition annotated by an asterisk ( * ) denotes the reference partition.
Model Formulation
The methodology presented in this paper was motivated by the identification of cell-specific methylation at the individual CpG level. A cell-specific CpG was defined as any CpG probe where the expected methylation level was different for one (single cell-specific) or more (multi-cell-specific) cell types, relative to others observed. The true cell-specific methylation pattern at each CpG probe was assumed unknown and treated as a model selection problem, with key details outlined in this Section. Information about cell lineage inferred from available samples was used to define candidate models for selection at each CpG. Using the results of hierarchical clustering (Fig 1), eleven candidate models were defined (Table 1) where each model represented a dendrogram split or node. Each candidate model therefore proposed a different cell-specific methylation pattern, characterised by a unique partition of cell types into non-overlapping groups. For a given candidate model, cell types assigned to the same group were assumed to share the same level of methylation. Two additional models, corresponding to the null and saturated cases, were also considered.
where N J (a, A) defines a J−dimensional Normal distribution with mean vector a and variance-covariance matrix A. The unknown variance was assumed to be common across all cell types, denoted by σ 2(m) ks I J×J where I is the identity matrix. This variance-covariance structure was chosen for identifiability reasons given sample sizes available.
A Bayesian approach to model selection was adopted, which allowed for probabilistic statements to be made about the relative fit of each candidate model. Under this approach, the posterior probability of each model conditional on the observed data was calculated for all candidate models. The posterior probability of model m compared to other candidates m ′ was calculated as, P r (Model m|y ks ) = p (y ks |Model m) P r (Model m) where the sum of probabilities over all candidate models was equal to 1. The term p (y ks |Model m) was obtained by integrating over all unknown parameters from the likelihood and prior distributions. Prior probabilities of each model, P r (Model m), were assumed to be equal to reflect a lack of model preference a priori.
Prior distributions for remaining parameters were selected such that Eq 1 could be derived analytically.
A g-prior distribution [19] was adopted for each µ The g-prior is a popular choice in linear model selection settings, as it allows the experimenter to introduce information on the scale of X m . Expected methylation levels were centered around the overall methylation level, b 0 , with variance proportional the standard error of each partition. The scaling factor g s > 0 is interpreted as a relative weighting of the prior versus the observed data. Here, g s was assumed common over all CpG probes and estimated by maximising the marginal likelihood averaged over candidate models. In this paper, b 0 was set to the global methylation mean, averaged over CpGs and cell types. A conjugate prior distribution for σ 2(m) ks of form p σ 2(m) ks ∝ 1/σ 2(m) ks completed model specification [19]. The expectation-maximisation (EM) algorithm was used to obtain a global Empirical Bayes estimate for g s [20]. This computational approach provided significant computational benefits over sampling-based approaches, namely Markov chain Monte Carlo (MCMC), and was possible given closed-form solutions for p (y ks |Model m) conditional on g s . In addition, the desired posterior model probabilities from Eq 1 were available upon convergence of the EM algorithm. The proposed methodology was implemented in R, with code available as Supporting Information (File S1).
Making the most of the Bayesian approach for model based inference
The accommodation of model and parameter uncertainty under the Bayesian approach formed the basis of subsequent inference, namely the identification of cell-specific CpGs, the estimation of differential methylation by cell type and the assessment of sex-specific differences by cell type. Brief details of each inference are provided in this Section.
CpG marker identification
CpG probes or 'markers' associated with each cell-specific pattern were identified by comparing posterior model probabilities at each CpG probe. For each candidate model and sex, markers were identified by reviewing the set of K model probabilities: P r (Model m|y 1s ) , . . . , P r (Model m|y Ks ). A 5% Bayes' False Discovery Rate (FDR) [21] was applied to each set of probabilities to control the expected number of false discoveries. Common CpG markers were defined as CpGs that were identified in both female and male samples, for the same candidate model. The compilation of common marker panels allowed for the assessment of sex-specific differences by cell type, within each candidate model.
Estimation of differential methylation
Posterior inference about each marker panel focused on the estimation of differential methylation by cell type relative to its corresponding reference partition (Table 1). Under model m, the posterior distribution of differential methylation for cell-specific partition p and sex s is, where c p is the number of cell types in partition p and µ The posterior distribution in Eq 2 was also used to validate selected CpG markers. Validation of common CpG markers was limited to CD19 + B, CD4 + T, CD8 + T and Pan T models, in light of validation samples available. Using all validation samples for each sex, the average β−value for each cell type and CpG was computed. The difference between each average β−value and reference partition was then compared with the corresponding 95% CI from Eq 2 for the appropriate candidate model. Concordance rates with respect to the predicted direction of methylation (hypomethylated, hypermethylated) for validation samples were also calculated. This joint approach was motivated by the limitation that validation based on coverage of 95% CIs relied on a posterior estimate of σ 2(m) ks . When this estimate is small and/or underestimated, proportions of validated markers based on CI coverage only were likely to be low, even if differences in methylation between training and validation samples were small. Finally, it is noted that not all cell types observed in the training data were available in the validation samples. Given the properties of the multivariate Normal distribution, this discrepancy was addressed by using the appropriate marginal distributions for the available cell types.
Evaluation of sex effects
A similar expression to Eq 2 was derived to evaluate sex-specific methylation differences within common CpG marker panels. In this case, focus was on the comparison of female and male methylation estimates for differentially methylated cell types, as defined by the corresponding candidate model. The posterior distribution of this difference across model partitions was Multivariate Normal, To assess evidence for sex effects, the posterior probability that the difference in methylation between sexes using Eq 3 was at least 0.10 was calculated for each cellspecific partition. This calculation again relied on estimates of the residual variance, σ 2(m) ks for each sex which were set to their posterior mean estimate. A sex-specific difference for a given partition was declared if the posterior probability of a difference greater than 0.10 exceeded 0.95.
CpG markers to genes: Assessment of SNP effects, genomic features and pathway enrichment
To provide additional evidence to support our method of marker identification, a pathways enrichment analysis was performed to explore the underlying biology of gene sets derived from common CpG marker panels. Common CpG markers were mapped to genes using Illumina Human Methylation 450k annotation data available from Bioconductor [22]. SNP information was also collated to infer the percentage of SNP associated markers with associations limited to SNPs located directly on the CpG loci. Using KEGG functional analysis in WebGestalt [23], a hypergeometric test was applied to each marker panel. A 5% FDR [24] was applied to resulting p-values to identify significant pathway enrichment for derived gene lists.
Cell lineage impacts the identification of differentially methylated CpGs
Across all cell-specific models, 83,449 and 97,747 CpG markers were identified for females and males, respectively ( Table 2). Among female samples, 42,834 markers (51.3%) were associated with differential methylation in multiple cell types, which included a three-fold increase in Pan T markers compared with males. A larger proportion of multi-cell dependent markers among males was observed (64.8%), due to larger numbers identified under Lymphocyte-II and Myeloid models. Among single cell dependent models, CD19 + B was the most frequently observed marker type for both sexes, with 25,611 in females and 18,271 in males.
A total of 42,452 CpGs were identified under the same cell-specific methylation pattern for both sexes, corresponding to 9.5% of the observed methylome. Within this subset, 23,551 (55.5%) were defined by differential methylation in more than one cell type. Over all candidate models, smaller frequencies of common markers were associated with T-cell dependent markers (CD4 + , CD8 + , Pan T).
Differential methylation is affected by cell lineage among common CpG markers
Common CpG markers associated with differential methylation in Myeloid cell types (CD14 + Mono, CD16 + Neu) were consistently hypomethylated across all relevant marker panels (Table 3). Among Lymphocytes, CD8 + T markers were the least likely to be hypomethylated (35.99%). Smaller proportions of hypomethylation among Lymphocyte I and II panels were indicative of greater methylation among Lymphocyte versus Myeloid cell subtypes.
The impact of cell lineage on differential methylation was greatest among marker panels related to Lymphocytes (Fig 2). Lower levels of differential methylation (<0.10) were concentrated within single cell The distribution of differential methylation by sex among common Pan T markers revealed considerable variation for both hypermethylated and hypomethylated states compared to CD4 + T and CD8 + T panels (Fig 3). Differential methylation levels among Pan T markers tended to be greater for CD4 + T cells; 25 markers in this subset showed strong evidence of differential methylation greater than 0.5 for both sexes ( Fig S1).
Validation of common CpG markers for B-and T-lymphocytes
The validation of common markers with respect to mean differences was generally higher in males than females, across all immune cell subtypes (Table 4). Whilst validation based on coverage of credible intervals was low, differences between training and validation estimates were relatively small across all markers tested; for CD4 + T markers, approximately 80% of absolute differences were less than 10% (Fig S2). Furthermore, comparisons with respect to inferred methylation state showed very high concordance rates for all marker panels and these findings were consistent for both sexes.
Genomic feature distributions and enrichment analysis
The effects of cell lineage were evident in the comparison of genomic features, with higher proportions of markers residing in Transcription Start Sites (TSS) for Lymphocyte cell subtypes compared with Myeloids (Table 5). Common markers were concentrated in the gene body and intergenic regions, with 54.73% of CD56 + NK markers located in the gene body. Among Lymphocyte cell subtypes, TSS proportions were highest for T cells, with 18.28% and 16.48% for CD4 + T and CD8 + T cells, respectively. Moderate proportions of markers were directly associated with SNPs and these levels were maintained between sex-specific and common marker panels, averaging 30.7% in common markers (Table S1). Higher SNP proportions were observed in marker panels related to the lineage of Myeloid cells, except for CD56 + NK markers of which 36.02% were SNP associated. For CpG probes not assigned to any common marker panel, a similar degree of SNP association was observed (26.61%). Significant pathways enrichment for common markers were associated with immune cell subtypes (Table 6) all including biologically relevant pathways.
Assessment of common marker profiles shows sex effects in CD56 + NK, CD16 + Neutrophils The majority of common CpG markers did not exhibit sex-specific differences in methylation profile (Table 7). markers exhibited sex effects, with the majority corresponding to autosomal CpGs. No sex-specific differences were identified among common CD8 + T markers.
Sex-specific differences identified among single cell-specific markers were uniquely mapped to 215 genes (Table S2). The majority of identified cases were concentrated in the CD56 + NK common marker panel.
Four CD4 + T markers were associated with greater methylation observed in females and were mapped to a single gene, CD40LG, located on the X chromosome. Annotation information for the full list of sex-specific Table 7: Summary of sex-specific differences by common CpG marker panel. A difference between male and female methylation estimates ≥0.10 was the outcome of interest. A marker was declared sex-specific if the posterior probability for this outcome exceeded 0.95, for one or more model-based partitions. For each marker panel, total numbers of sex-specific markers and autosomal sex-specific markers are given.
Non (File S2). sex effects in CD56 + NK methylation was prominent in Lymphocyte-I and Lymphocyte-II common marker panels, in addition to the CD56 + NK panel (Fig 5). Within the Lymphocyte-I panel, 1698 common CpG markers were identified as sex-specific, of which 1627 showed differences in CD56 + NK methylation only. Similarly, 442 common Lymphocyte-II markers showed differences in CD56 + NK methylation between sexes. The tendency for CD56 + NK methylation to be higher in males was common across all three panels. 684 common Myeloid markers were associated with differences in CD14 + Monocytes or CD16 + Neutrophils only. Differences with respect to CD14 + Monocytes tended to be higher in females, compared with higher male methylation in CD16 + Neutrophils (Fig 6).
Discussion
This paper has proposed new statistical methodology for the discovery of cell-specific methylation profiles in whole blood, applying principles of Bayesian model selection. The characterisation of CpGs by differential methylation in one or more cell types builds significantly upon existing work that has been restricted to univariate analyses of differential methlyation by cell type. Sex-specific differences in both the prevalence of cell-specific profiles and methylation signal for select cell types were also demonstrated. For immune cell subtypes, validation of common markers using external cell-sorted samples produced favourable results. Enrichment analyses of common marker panels provided additional support for the proposed methodology, where it was demonstrated that the detection of cell-specific methylation at the individual CpG level was also biologically meaningful.
The incorporation of knowledge about hematopoietic lineage into model specification represents a semisupervised approach that reflects relevant cell biology. Consistent with other model selection strategies, our approach assumes that the defined set of candidate models is exhaustive and true patterns beyond this set are not of primary interest. In the event that the true methylation pattern is not commensurate with any of the candidate models specified, two outcomes are likely. In the first instance, corresponding probes are It is well documented that there are sex-specific differences in the proportions of circulating white blood cells [25,26]. The application of the proposed methodology to female and male samples has highlighted the importance of accounting for sex effects in DNA methylation analyses. Greater numbers of CD19 + B and T-cell dependent markers in females are consistent with previous findings and are possibly indicative of higher levels of cell activation [27]. The association between sex-specific differences in select CD4 + T markers and the CD40LG gene have also been identified previously. Previous studies have pointed to alelle specific methylation for this gene [28,29] where CD4 + T hypomethylation is observed in healthy males compared with healthy women who carry one methylated and one hypomethylated alelle. One of our major findings was large differences of methylation between males and females in markers defined as CD56 + NK specific. This is interesting when considered alongside the observation that males show an increase in circulatory NK cells compared to females [27], which adds further support for the accuracy of the approach. Additionally there is some evidence of sex-specific methylation differences in CD + 56 NK, as well as CD + 8 T-cells [30]. Under the proposed approach, we have provided a potential solution to accurately account for potential bias introduced by sex effects at the marker level.
The presence of cell-specific methylation CpG markers highlights the need to account for cellular composition prior to conducting Epigenome Wide Association Studies (EWAS), in whole blood. Methods for this purpose have been developed [31,32] based on the assembly of methylation 'signatures' from cell-sorted data which are then projected onto heterogeneous samples to predict cell type proportions. A comparison of common markers with the top 500 CpG probes identified by the cell mixture methodology of [31] revealed 75% concordance between panels (data not shown), of which the majority were associated with Lymphocyte-I/II and Myeloid maker panels. The inclusion of other marker panels in these algorithms may lead to further improvement in cell mixture estimation, in particular for immune cell subtypes that may be present in low proportions. Furthermore, the performance of these algorithms rely on consistent cell-type effects across cohorts [33]. Given the sex-specific methylation differences we have identified in this study, failure to account for sex effects may also impact upon the quality of cell mixture estimation and should therefore be given due consideration.
It is common practice in array-based methylation studies to exclude CpG sites which contain SNPs both within the probe and on the CpG site. While this is a valid approach to filtering before analysis, it will often lead to dramatic reduction of overall data. As a result, it is likely that sites of potential interest may be lost before any association can be made. By mapping hg38 annotated SNPs to all 450K CpG loci, we were able to ascertain the overall proportion of cell marker sites which have a SNP present; on average, across the common set of markers, this was approximately 30.7% of markers. In light of these results, we suggest that deconvolution studies and methods should account for SNP events at cell marker sites, noting the proportion that are present. For the filtering stage, we recommend that the overall rarity of the SNP variant be taken into account, for example, retaining CpGs which also have a 'rare' (MAF < 0.01) variant mapping. This approach is likely to be beneficial to the overall study design and outcome.
Supporting information
Supplementary files S1 and S2 are available from the GitHub respository https://github.com/nicolemwhite/BayesMS. S1 File R code for Bayesian model selection algorithm. Core R functions required to prepare data and compute posterior model probabilities via the EM algorithm, across all listed candidate models.
S2 File List of common CpG markers associated with a sex-specific difference of ≥0.10. Illumina Human Methylation 450k annotation data are also included for each CpG marker. Table S2: Distribution of sex-specific common markers over chromosomes, for single cell-dependent makers only. A sex-specific marker was declared if the posterior probability from Eq 3 was greater than 0.95 for at least one differentially methylated cell type. | 5,131 | 2017-04-06T00:00:00.000 | [
"Biology",
"Computer Science"
] |
A p-Adic Model of Quantum States and the p-Adic Qubit
We propose a model of a quantum N-dimensional system (quNit) based on a quadratic extension of the non-Archimedean field of p-adic numbers. As in the standard complex setting, states and observables of a p-adic quantum system are implemented by suitable linear operators in a p-adic Hilbert space. In particular, owing to the distinguishing features of p-adic probability theory, the states of an N-dimensional p-adic quantum system are implemented by p-adic statistical operators, i.e., trace-one selfadjoint operators in the carrier Hilbert space. Accordingly, we introduce the notion of selfadjoint-operator-valued measure (SOVM)—a suitable p-adic counterpart of a POVM in a complex Hilbert space—as a convenient mathematical tool describing the physical observables of a p-adic quantum system. Eventually, we focus on the special case where N=2, thus providing a description of p-adic qubit states and 2-dimensional SOVMs. The analogies—but also the non-trivial differences—with respect to the qubit states of standard quantum mechanics are then analyzed.
Introduction
The field of p-adic numbers Q p was introduced by K. Hensel at the end of the XIX century, mainly in connection with pure mathematical problems. The peculiarity of this field, in sharp contrast with the fields of real and complex numbers R and C, is its natural ultrametric structure, that entails a non-Archimedean character of this field. It came as a complete surprise when, at the end of the past century, some concrete applications of p-adic numbers to physical theories began to appear. Indeed, in the late 1980s, Vladimirov, Volovich and Zelenov [1,2] argued that the existence of a smallest measurable length-i.e., the so-called Planck length l P ≈ 10 −35 m, predicted in quantum gravity and string theory, see [3] and references therein-forces one to adopt a non-Riemannian model, that is, a model in which the Archimedean property is no more valid at very small distances. In particular, they proposed a model of quantum mechanics based on the non-Archimedean field of p-adic numbers. Later on, different p-adic quantum mechanical models were studied [4][5][6][7][8][9][10][11][12], and several applications to quantum field theory were proposed [13][14][15][16][17][18]. More or less in the same years, other unexpected connections between p-adic numbers and theoretical physics were revealed. e.g., it was argued that the natural fractal-like structure of this field makes it suitable for the description of the dynamics of chaotic and disordered systems. In particular, it was proved that the ground state of spin glasses exhibits a natural (non-Archimedean) ultrametric structure [19][20][21].
More recently, new and interesting applications of p-adic numbers, not necessarily related to foundational physics, have begun to appear. Indeed, p-adic numbers have found a fertile ground of application in the context of algebraic dynamical systems, also in connection with problems from computer science, image analysis, compression of information, image
In the mathematical literature, property (2) is usually referred to as ultrametricity and, accordingly, a metric function satisfying it is called an ultrametric.
Example 1 ([1,32-34]). Recall that, according to the unique factorization theorem, every rational number x ∈ Q can be expressed as x = p k q, where p ∈ N is a fixed prime number, k some integer in Z, and q a rational number whose numerator and denominator are not divisible by p [32,35]. The p-adic absolute value is then defined as the map | · | p : Q → R, such that |0| p ≡ 0, and It is easily shown that | · | p is a non-Archimedean valuation on Q, since it is strictly positive on Q * ≡ Q \ {0}, it factorizes under the product of two elements in Q, and verifies the strong triangle inequality (V4). Therefore, if we consider the associated metric function we obtain an ultrametric on Q.
Consider the pair (Q, d |·| p ), where d |·| p is the ultrametric associated with the p-adic valuation (see Example 1). It is a metric space that, by means of a standard procedure, can be completed [36]. The resulting complete field is usually called the field of p-adic numbers Q p . This is a standard (through rather abstract) way to define p-adic numbers. A more concrete characterization is given as follows. Let x ∈ Q * p ≡ Q p \ {0}. It is possible to prove that x admits a unique decomposition of the form and, conversely, every series of this form converges to some non-zero element of Q p [32]. Therefore, we see that the decomposition (5) provides a representation of any p-adic number by means of a suitable converging series. In particular, this is reminiscent, to some extent, of the usual decimal expansion of a real number x ∈ R, namely, The p-adic valuation on Q can be extended-in a unique way-to a non-Archimedean valuation on Q p which, still using the same symbol (with a slight abuse of notation), is given by Clearly, by a similar reasoning, also the ultrametric (4) can be extended to an ultrametric on Q p . The ultrametricity condition satisfied by this ultrametric reflects in some topological peculiarities of Q p that, ultimately, justify the use of p-adic numbers when describing physics on length scales comparable to Planck's length l P [1,2,4,5,10,11]. Just to mention the most relevant ones [37,38], we list the following points: Two open (closed) balls are either disjoint or one is contained in the other.
(P3)
Every ball in Q p is both closed and open (in short, clopen) in the ultrametric topology of Q p .
(P4)
All triangles are isosceles in Q p .
As a topological space, Q p is completely regular (being a metric space) and totally disconnected; namely, the only connected subsets of Q p are the singletons [37]. We devote the last part of this section to a brief discussion of the quadratic extensions of Q p . The opportunity of switching to a quadratic extension is related to the lack of a nontrivial involution on Q p [25,33]. This is analogous to the formulation of standard quantum mechanics relying on the field C of complex numbers, with C regarded as a quadratic extension of the reals, and endowed with its natural involution (the complex conjugation).
The definition of a quadratic extension of Q p closely mimics the one given for the field of complex numbers C. Indeed, let µ ∈ Q p be a non-quadratic element in Q p , i.e., µ / ∈ (Q * p ) 2 . Introducing the symbol √ µ (which plays a role analogous to the one played by the imaginary unit in C), the quadratic extension Q p,µ of Q p induced by µ is defined as the set It is easily verified that Q p,µ is a field extension of Q p . Indeed, Q p,µ is a two-dimensional vector space on Q p , its elements can be added and multiplied following the usual rules, and any non-null element admits a unique inverse, which is given by where the denominator x 2 − µy 2 is not zero (otherwise µ should be a square in Q p ). On the field Q p,µ , it is possible to define a conjugation, namely, the mapping so that Moreover, the p-adic absolute value | · | p can be extended-in a unique way-to a non-Archimedean valuation | · | p,µ on Q p,µ , which is given by For the sake of conciseness, henceforth we will simply denote this valuation by | · |. However, differently from the real case, there exist various inequivalent quadratic extensions of Q p . In fact, we have [1,31]: (1) If p = 2, there are precisely three non-isomorphic quadratic extensions of Q p , i.e., Q p,µ , with µ ∈ {η, p, η p}, and where η ∈ Q p is a non-quadratic unit, i.e., η / ∈ (Q * p ) 2 , and |η| p = 1; (2) if p = 2, there are precisely seven non-isomorphic quadratic extensions of Q p , i.e., Q p,µ , with µ ∈ {2, 3, 5, 6, 7, 10, 14}.
p-Adic Hilbert Spaces and Operators
This section is devoted to introduce a suitable notion of a p-adic Hilbert space and the associated p-adic linear operators [31] (compare with [39,40], where different notions of non-Archimedean Hilbert spaces are introduced, and with [41], where orthogonal and symmetric operators in the non-Archimedean setting are studied).
p-Adic Hilbert Spaces
As is well known, complex Hilbert spaces are defined as (complex) Banach spaces endowed with a suitable inner product, namely, the one inducing the relevant norm. It turns out that this familiar picture keeps some of its main futures-but also requires some essential modification-when switching to the field of p-adic numbers. We start by setting the following: Definition 1. Let Q p,µ be a quadratic extension of the field of p-adic numbers Q p . By a p-adic normed space, we mean a pair (X, · ), where X is a vector space over Q p,µ , while · : X → R + is an ultrametric norm, i.e., a map satisfying the following conditions: x + y ≤ max{ x , y }, for all x, y ∈ X and α ∈ Q p,µ . A p-adic normed space which is complete w.r.t. the ultrametric associated with · , is called a p-adic Banach space.
Remark 1.
The explicit definition of a p-adic Banach space is motivated by the fact that the strong triangle inequality (N3) differs significantly w.r.t. the standard (real or complex) case, where the usual triangle inequality holds.
Let (X, · ) be a p-adic normed space. Our first concern is to provide a suitable notion of a basis for this space [31,34,37,42]. To this end, let us start by recalling that two vectors x, y in a p-adic normed space X are said to be (mutually) norm-orthogonal if, for any α, β ∈ Q p,µ , we have that αx + βy = max{ αx , βy }. Moreover, an arbitrary subset B of X is norm-orthogonal if any finite subset of B is such; namely, if for every set {x 1 , . . . , x n } of elements in B, and every set {α 1 , . . . , α n } in Q p,µ , we have that We say that a subset B of X is normal, if it is norm-orthogonal and, additionally, . . , N}, for some N ∈ N, in the case where this set is finite; otherwise, I = N). We say that B is a norm-orthogonal (normal) basis, if B is a norm-orthogonal (normal) set, and every x ∈ X can be expressed-in a unique way-as In such a case, we define the dimension of X-in symbols, dim(X)-to be the (countable) cardinality of any norm-orthogonal basis in X, i.e., we set dim(X) = card(I). In the following, we call a p-adic Banach space X admitting a normal basis a normal p-adic Banach space.
Example 3.
Let us consider the space c 0 (I, Q p,µ ), of zero-converging sequences in Q p,µ : . This set is a vector space over Q p,µ , and it becomes a p-adic Banach space once it is endowed with the so-called 'sup-norm', which is defined as A normal basis for c 0 (I, Q p,µ ) is given by the set B = {b i } i∈I (the so-called standard basis of c 0 (I, Q p,µ )), where As in the standard complex case, also in the p-adic setting an essential step in the definition of a p-adic Hilbert space is the introduction of a suitable notion of inner product. In particular, we set the following: Definition 2. Let (X, · ) be a p-adic Banach space over Q p,µ . By a non-Archimedean inner product we mean a map · , · : X × X → Q p,µ such that, for all x, y ∈ X and α, β ∈ Q p,µ , (a) x, αy + βz = α x, y + β x, z (linearity in the second argument); We call the triple (X, · , · , · ) where · , · is a non-Archimedean inner product, an innerproduct p-adic Banach space.
From conditions (a) and (b) of Definition 2, it is clear that the inner product · , · is conjugate-linear in its first argument, i.e., it is a sesquilinear form. Also note that, from the Hermitianity condition (b), and the sesquilinearity of · , · , it follows that 0, x = 0 = x, 0 , for all x ∈ X; in particular, 0, 0 = 0. We also say that the inner product · , · is non-degenerate if the condition x, y = 0, for all y ∈ X, implies that x = 0.
Example 4. Let (X, · ) be a normal p-adic Banach space, and let B ≡ {b i } i∈I be a normal basis in X. The canonical inner product associated with B is defined as the-non-degenerate, Hermitian-sesquilinear form where x = ∑ i∈I x i b i and y = ∑ i∈I y i b i are the (norm converging) expansions of the vectors x and y w.r.t. the fixed normal basis B. One can easily check that this sesquilinear product verifies all the defining conditions of a non-Archimedean inner product.
Remark 2.
The notion of non-Archimedean inner product naturally leads us to a notion of innerproduct orthogonality, which is distinct from the-previously introduced-norm orthogonality. Explicitly, we say that two vectors x, y, in an inner-product p-adic Banach space X, are innerproduct orthogonal (IP-orthogonal, in short) if x, y = 0.
The notion of inner-product orthogonality, introduced in Remark 2, entails the following natural extension of the notion of normal basis: Definition 3. Let (X, · , · , · ) be a normal inner-product p-adic Banach space. We say that a (finite or countable) sequence of vectors Ψ ≡ {ψ i } i∈I in X is an orthonormal basis, if Ψ is a normal basis in X, and its elements are mutually IP-orthogonal, namely, ψ i , ψ j = δ i,j , ∀i, j ∈ I.
We stress that the existence of an orthonormal basis in an inner-product p-adic Banach space X is, in general, not guaranteed. On the other hand, when X is a normal p-adic Banach space-where the existence of a normal basis is assumed-it is always possible to turn any given normal basis into an orthonormal one by making a suitable choice of the inner product. Indeed, it suffices to consider the canonical inner product associated with this normal basis in X (recall Example 4). Therefore, we have the following natural definition of Hilbert space in the p-adic setting: is a normal p-adic Banach space, and · , · ≡ · , · B is the canonical inner product associated with the normal basis B in X.
From the previous definition, it is clear that a p-adic Hilbert space may be thought of as a normal p-adic Banach space endowed with a distinguished normal basis and with the associated canonical inner product. It is then not difficult to check the following two properties of a p-adic Hilbert space: Every vector x ∈ H can be uniquely expanded w.r.t. any orthonormal basis Ψ ≡ {ψ i } i∈I in H, namely, (H2) The non-Archimedean Parseval identity holds true: Example 5. Let us consider the set c 0 (I, Q p,µ ) introduced in Example 3. We have already observed that it is a normal p-adic Banach space once endowed with the sup-norm · ∞ and with the standard basis (17). Then, introducing the canonical inner product in c 0 (I, Q p,µ ) of Example 4, we obtain a p-adic Hilbert space. In the literature [41,43], this Hilbert space is sometimes called coordinate p-adic Hilbert space, and denoted by H(I). It plays a role analogous to the role played by 2 (I) for (separable) complex Hilbert spaces. There exists an isomorphism of p-adic Banach spaces between H and H(I) (dim(H) = card(I)); see [34].
As in the complex setting, also in the p-adic case one can define a convenient notion of isomorphism of Hilbert spaces (or unitary operator, defined as a bounded operator mapping an orthonormal basis into another) [31]. Let us briefly outline this notion. Let (H, · , · , · B ) be a p-adic Hilbert space, where · , · B is the canonical inner product associated with a given normal basis B. Denote by N (H) the collection of all the normal bases in H and by N (H, B) ⊂ N (H) the class of all normal bases that are orthonormal w.r.t. · , · B . A Hilbert space automorphism of (H, · , · , · B ) is a bounded linear map transforming a basis in N (H, B) into another normal basis in the same set; equivalently, a surjective norm-isometry of H onto itself that preserves the inner product · , · B . This notion admits a straightforward generalization to a notion of isomorphism relating two Hilbert spaces over Q p,µ (of the same dimension). Interestingly, if C ∈ N (H) is such that C / ∈ N (H, B)i.e., N (H, B) = N (H, C)-then (H, · , · , · B ) and (H, · , · , · C ) are different, but mutually isomorphic, p-adic Hilbert spaces. The p-adic Hilbert spaces stemming from the same p-adic Banach space (H, · ) are in a natural one-to-one correspondence with the classes of normal bases of the type N (H, B), that form a partition of the set N (H).
Remark 3.
It is worth stressing that the analogies between complex and p-adic Hilbert spaces cannot be pursued too far. Indeed, quite generally, in a p-adic Hilbert space, H, the norm does not stem directly from the inner product; i.e., in general, x = | x, x |. Moreover, note that a p-adic Hilbert space may contain isotropic vectors, i.e., non-zero vectors x such that x, x = 0. e.g., for p ≡ 1 (mod 4), taking into account the fact that −1 is a square in Q p [1], let x be a vector in the p-adic Hilbert space H (dim(H) ≥ 2), and let {ψ 1 , ψ 2 , . . .} be an orthonormal basis in H. Then, setting x = ψ 1 + √ −1ψ 2 , we have that x, x = 0.
Hereafter, borrowing the terminology from the standard (complex) quantum mechanics, we shall call a quantum system with associated p-adic Hilbert space H of finite dimension N a p-adic quNit.
Linear Operators
In [31], it is demonstrated that some fundamental classes of operators used in the standard formulation of quantum mechanics-e.g., bounded and trace class operators in a complex Hilbert space-can be suitably introduced in the p-adic framework as well, with some non-trivial differences w.r.t. the standard complex setting.
Since our main concern is to consider applications to quantum information theory, we will actually focus our attention to linear operators acting in a finite-dimensional padic Hilbert space. In this case, we only need to consider the space L(H) of all linear operators in H, and the distinction between the various classes of operators mentioned above becomes irrelevant.
Then, let H be a finite-dimensional p-adic Hilbert space, with dim(H) = N, and let Ψ ≡ {ψ i } N i=1 be an orthonormal basis in H. Every L ∈ L(H) can be represented-w.r.t.
-as a matrix operator, namely, where (L ij := ψ i , Lψ j ) ∈ M N (Q p,µ ) is the matrix associated with the operator L and the fixed orthonormal basis Ψ (here M N (Q p,µ ) denotes the set of N × N matrices on Q p,µ ). Conversely, every matrix (M ij ) ∈ M N (Q p,µ ) defines a linear operator M ∈ L(H) by putting On the space L(H), we can define a (ultrametric) norm-namely, the operator norm-which is given by Then, by means of a standard argument (cf. Theorem 6.2.1 in [33]), it is not difficult to show that the space (L(H), · ) is complete w.r.t. the (ultra-)metric associated with (23); i.e., (L(H), · ) is a p-adic Banach space.
Remark 4.
Let us explicitly note that, by using the Dirac notation, the operator |ψ i ψ j | appearing in the matrix representation of L ∈ L(H) should be understood as the linear operator ψ i , · |ψ j , whose action on a generic element φ ∈ H is given by For every L ∈ L(H), the adjoint L * of L is given by i.e., L * is the operator in L(H) with matrix coefficients given by L * ij := ψ i , L * ψ j = ψ j , Lψ i = L ji . As in the standard complex setting, the adjoining operation so defined is easily seen to be an involutive automorphism of L(H); that is, the map L(H) L → L * ∈ L(H) verifies the following conditions: for all A, B ∈ L(H), α, β ∈ Q p,µ . Therefore, we get to the conclusion that the p-adic Banach space (L(H), · ), equipped with the adjoining operation (24), has a natural structure of a p-adic Banach * -algebra. In fact, in the next section, the set L(H) will be regarded as the Banach * -algebra of physical observables of a (finite-dimensional) p-adic quantum system.
Remark 5.
As in the complex setting, also in the p-adic case it is possible to single out the subset L(H) sa ⊂ L(H) of selfadjoint elements of L(H), namely, the linear operators for which the additional condition is verified.
To conclude this section, we will now argue that L(H) turns out to be a p-adic Hilbert space. Indeed, let us first observe that given L ∈ L(H), we can define its trace, tr(L)-w.r.t. any fixed orthonormal basis Ψ ≡ {ψ i } N i=1 in H-in the usual way as We call this Hermitian sesquilinear form the p-adic Hilbert-Schmidt product. Next, note that, for all A, B ∈ L(H), we have: i.e., · , · HS satisfies the Cauchy-Schwarz inequality. Hence, we conclude that · , · HS is a non-Archimedean inner product, and L(H), endowed with this sesquilinear form, is an inner-product p-adic Banach space.
be an orthonormal basis in H. We can introduce a family of namely, in the usual Dirac notation, ij E Ψ = |ψ i ψ j |.
We now prove that the set { ij E Ψ } N i,j=1 is an orthonormal basis in L(H). To this end, first note that { ij E Ψ } N i,j=1 in a normal set of vectors in L(H). In fact, consider that, for every finite subset {α jk } N j,k=1 in Q p,µ , we have: Moreover, we also have that ij E Ψ , rs E Ψ HS = tr(|ψ j ψ i | |ψ r ψ s |) = ψ s , ψ i ψ j , ψ r = δ si δ jr ; i.e., { ij E Ψ } N i,j=1 is an IP-orthogonal set w.r.t. the Hilbert-Schmidt product. Finally, by noting that any L ∈ L(H) is written-w.r.t. the orthonormal basis Ψ ≡ {ψ i } N i=1 -as (cf. (21)) we see that { ij E Ψ } N i,j=1 is an orthonormal basis in L(H). Summarizing, we have the following result: Theorem 1. Given an N-dimensional p-adic Hilbert space H, the p-adic Banach space L(H)endowed with the p-adic Hilbert-Schmidt product · , · HS -becomes an inner-product p-adic Banach space. In particular, the triple (L(H), · , · , · HS ) is a p-adic Hilbert space and, for every is an orthonormal basis in this space.
Physical States and Observables
As is well known, the most general and abstract description of quantum mechanics is provided by the so-called algebraic formulation. This formulation essentially relies on two fundamental assumptions; namely, that every quantum system can be described by means of two main classes of objects-i.e., states and observables-mutually related by means of a natural pairing map. Specifically, the observables can be identified with the selfadjoint elements of an abstract non-commutative unital C * -algebra, whereas the states are normalized positive functionals on the C * -algebra. In particular, in the case of ordinary quantum mechanics, the C * -algebra of observables is realized by the noncommutative unital C * -algebra of bounded operators B(K) in a complex Hilbert space K. The associated states are realized by trace-one positive trace class operators, the so-called density or statistical operators [44][45][46][47][48][49].
Following the same route, it has been recently argued that, in p-adic quantum mechanics, physical states should be defined as (normalized) involution-preserving bounded functionals on the unital Banach * -algebra B(H) of bounded operators [31]. It has been further shown that the role played by the density operators in the complex case is played, in the p-adic setting, by the class of the so-called p-adic statistical operators, defined as a suitable subclass of the selfadjoint trace class operators in a p-adic Hilbert space H [31]. The properties, as well as the conditions, that must be satisfied by these states are ultimately ruled by a suitable model of p-adic probability theory [6,7,50,51].
Remark 7.
Since p-adic probability theory is a rather non standard topic, for reader's convenience, we now briefly sketch its main features. Let us first observe that both p-adic and classical probability theory arise in a natural way from a common conceptual background. In fact, in both theories, one starts by considering the set O Q = {q ∈ Q | 0 ≤ q ≤ 1} ⊂ Q of the (relative) frequencies of experimental outcomes. Then, the set where all experimental statistical distributions take their values should coincide with the closure cl(O Q ) of O Q , where the closure is relative to some suitable topology. In classical probability theory, one assumes that this topology is the one induced by the standard valuation on Q, so obtaining cl(O Q ) = [0, 1] ⊂ R. In the p-adic case, instead, we should consider the topology induced by the p-adic valuation, which now yields cl(O Q ) = Q p . This means that, all possible normalized (i.e., summing up to 1) sequences in Q p provide legitimate (discrete) p-adic probability distributions [6,7,50,51]. The consequences of this fact are noteworthy. Just to mention the most relevant ones, we observe that p-adic probability theory naturally involves affinerather than convex-structures. Moreover, certain (say, rational) values of p-adic probability that, when considered in the standard real setting, would be greater than 1 or negative (therefore, inconsistent), are actually allowed in this model. e.g., the set {1, −1, −6, 7} is a legitimate p-adic probability distribution, even if it is not a standard probability distribution. Definition 5. Let H be an N-dimensional p-adic Hilbert space. By a p-adic statistical operator we mean a linear operator ρ ∈ L(H), such that ρ = ρ * and tr(ρ) = 1. Equivalently, ρ is a trace-one selfadjoint linear operator in H. We denote by the set of p-adic statistical operators in H.
It is convenient, at this point, to better clarify the statistical interpretation of p-adic quantum mechanics. To begin with, we need to introduce a suitable notion of observable in the p-adic setting. In particular, as argued in [31], a convenient mathematical tool for the description of a quantum measurement in the p-adic setting is provided by the so-called (discrete) selfadjoint operator valued measures (SOVMs) in H. A SOVM may be regarded as a suitable p-adic counterpart of a POVM in a complex Hilbert space. In the finite-dimensional setting we are considering, a discrete SOVM can be defined as a family M ≡ {M i } i∈I -where I is a finite index set-of selfadjoint operators in H such that ∑ i∈I M i = Id.
Remark 8.
We remark that, as for a discrete POVM, a discrete SOVM should actually be defined as an additive measure on the algebra of subsets of the index set I (the power set of I), by putting, e.g., for j = k, M {j,k} = M j + M k . The definition of SOVM on a general measurable space is beyond the aims of the present contribution.
Remark 9.
We stress that, in p-adic quantum mechanics, there is no straightforward counterpart of a POVM of standard (complex) quantum mechanics. In fact, the field of p-adic numbers Q p is not ordered. As a consequence, there is no natural notion of positivity in Q p . Accordingly, there is no natural way to define positive operators in a p-adic Hilbert space.
As a further step, we need to specify the pairing between states (i.e., p-adic statistical operators) and observables. To this end, similarly to the standard complex case, by means of the trace functional tr(·) : L(H) → Q p,µ we can associate, with any fixed ρ ∈ S(H), the linear functional ω ρ on L(H) defined by Now, taking into account the defining conditions ρ = ρ * and tr(ρ) = 1, one can easily check that the following two conditions of ω ρ hold true: That is, for every ρ ∈ S(H), ω ρ is a normalized involution-preserving linear functional on L(H). Then, we reach the following two conclusions: • The trace functional tr(·) provides a well defined pairing between p-adic statistical operators and observables. • For every ρ ∈ S(H), and every SOVM M ≡ {M i } i∈I ⊂ L(H) sa , relations (39) guarantee that the sequence {ω ρ (M i ) = tr(ρM i )} i∈I is a p-adic probability distribution. We have then obtained a complete description of the statistical content of the theory.
We next turn our attention to the characterization of a suitable p-adic counterpart of the complex qubit. To this end, let us first note that, since H is finite-dimensional, by considering the matrix representation-w.r.t. a fixed orthonormal basis Ψ ≡ {ψ i } N i=1 in H-of any linear operator in H, it is clear that L(H) ∼ = M N (Q p,µ ); i.e., one can identify the set L(H) with the set M N (Q p,µ ) of N-dimensional matrices in Q p,µ . In particular, the set S(H) can be identified with the following set of N × N matrices: Namely, we identify S(H) with the set S N (Q p,µ ) of trace-one p-adic Hermitian N × N matrices. Let us focus on the particular case where N = 2. We can give a complete characterization of S(H) ≡ S 2 (Q p,µ ) as follows. Let be a matrix in M 2 (Q p,µ ). We first consider the most general form of a traceless p-adic Hermitian matrix. In particular, the conditions tr(Q) = 0 and Q = Q * immediately yield the following relations for Q: From these conditions, we deduce that a two-dimensional p-adic Hermitian matrix, with zero trace, is given by: Next, let us introduce the p-adic Pauli matrices σ 1 , σ 2 , σ 3 ∈ M 2 (Q p,µ ), defined by Exploiting these matrices, we can rewrite Q in a more compact form: (Here we have set y 2 ≡ x 3 ). It is then clear that -where Id 2 denotes the identity matrix in M 2 (Q p,µ )-gives the most general form a twodimensional trace-one p-adic Hermitian matrix in M 2 (Q p,µ ). Therefore, we conclude that the set of all states of a two-dimensional p-adic quantum systems is where σ 1 , σ 2 , σ 3 are the p-adic Pauli matrices defined in (44). In particular, we find out that a qubit state can be represented, in the p-adic setting, as From the matrix representation (48), we observe that the p-adic qubit shares some analogies with the qubit states of standard quantum mechanics. In particular, the matrix representation of a p-adic qubit is essentially the same as in the complex case, the main formal difference consisting in the presence, in the p-adic case, of √ µ. However, there are two substantial differences between the p-adic and the complex case.
As a first point, note that S 2 (Q p,µ ) is a norm-unbounded subset of M 2 (Q p,µ ). Moreover, let us compute the eigenvalues of the p-adic qubit (48). As is easily verified, they are given by where one should require that x 2 1 + x 2 2 − µx 2 3 is a quadratic element of Q p,µ . On the other hand, it is a well known fact that the field of p-adic numbers (and its quadratic extensions) is not algebraically closed, namely, not every non-constant polynomial admits a root in Q p .
Let us clarify this point by means of an explicit example. Take p = 2, and consider the quadratic extension of Q 2 by √ 2, i.e., the field Q 2,2 (see Section 2). Now, consider the qubit state associated with the parameters (x 1 , x 2 , x 3 ) = (4, 4, 3). Then, from its characteristic polynomial, we obtain the following two formal eigenvalues However, 7 in not a quadratic element of Q 2 ; that is, the characteristic polynomial of the matrix (48) does not admit any root in the quadratic extension of Q 2 just considered. Otherwise stated, we have constructed an example of a 2-adic qubit state that is not diagonalizable. Actually, it is not difficult to construct examples of non-diagonalizable qubit states also for all other quadratic extensions of Q 2 (as classified in Section 2). The same fact holds true also for p > 2. Namely, for suitable values of x 1 , x 2 , x 3 ∈ Q p , it is possible to construct p-adic qubit states that cannot be diagonalized, for every quadratic extension Q p,µ of Q p .
Conclusions
As a first step towards a quantum information theory based on a quadratic extension of the non-Archimedean field of p-adic numbers, we have proposed a model of QuNit on the field Q p,µ , where µ is a non-square element of Q p .
We started by introducing a notion of p-adic Hilbert space and, restricting to the case where H is finite-dimensional, the associated space of linear operators L(H). Then, we have described various properties of the ultrametric Banach space L(H). We have argued that L(H), endowed with the operator norm and the adjoining operation, turns out to be a p-adic Banach * -algebra. Then, we have proved that the linear space L(H) itself has a natural structure of a p-adic Hilbert space, once it is endowed with the p-adic Hilbert-Schmidt inner product.
Owing to the distinguishing features of p-adic probability theory, we have argued that the states of an N-dimensional p-adic quantum system are implemented by p-adic statistical operators, i.e., trace-one selfadjoint operators in the carrier Hilbert space. In particular, it turns out that the set of p-adic statistical operators, S(H), is a Q p -affine subset of L(H)coherently with the affine structure of p-adic probability theory-hence, it is an unbounded subset of L(H).
We have next introduced the notion of (discrete) selfadjoint-operator-valued measure (SOVM)-a suitable p-adic counterpart of a POVM in a complex Hilbert space-as a convenient mathematical tool describing the physical observables of a p-adic quantum system.
Eventually, focusing on the special case where N = 2, we have provided a description of p-adic qubit states and of two-dimensional SOVMs.
We close by outlining some potential extensions of this work, especially focusing on those ones that are relevant for our (final) program aimed at developing a p-adic model of quantum information theory. Tensor products and entanglement play a central role in quantum information theory, and we expect that they will play a central role in the p-adic setting too. Therefore, as a first step, we plan to investigate tensor products of p-adic Hilbert spaces and the associated classes of entangled states. Our next concern is the description of dynamical maps and dynamical (semi-)groups in p-adic quantum mechanics. This will provide a suitable p-adic counterpart to quantum channels. Another interesting prospect concerns the possibility of defining typical entropic quantities, such as the von Neumann and the Rényi entropies-which are relevant in standard quantum information theory-in the p-adic framework too. | 8,551.2 | 2022-12-31T00:00:00.000 | [
"Mathematics"
] |
Deep Learning Applications with Practical Measured Results in Electronics Industries
: This editorial introduces the Special Issue, entitled “Deep Learning Applications with Practical Measured Results in Electronics Industries”, of Electronics . Topics covered in this issue include four main parts: (I) environmental information analyses and predictions, (II) unmanned aerial vehicle (UAV) and object tracking applications, (III) measurement and denoising techniques, and (IV) recommendation systems and education systems. Four papers on environmental information analyses and predictions are as follows: (1) “A Data-Driven Short-Term Forecasting Model for O ff shore Wind Speed Prediction Based on Computational Intelligence” by Panapakidis et al.; (2) “Multivariate Temporal Convolutional Network: A Deep Neural Networks Approach for Multivariate Time Series Forecasting” by Wan et al.; (3) “Modeling and Analysis of Adaptive Temperature Compensation for Humidity Sensors” by Xu et al.; (4) “An Image Compression Method for Video Surveillance System in Underground Mines Based on Residual Networks and Discrete Wavelet Transform” by Zhang et al. Three papers on UAV and object tracking applications are as follows: (1) “Trajectory Planning Algorithm of UAV Based on System Positioning Accuracy Constraints” by Zhou et al.; (2) “OTL-Classifier: Towards Imaging Processing for Future Unmanned Overhead Transmission Line Maintenance” by Zhang et al.; (3) “Model Update Strategies about Object Tracking: A State of the Art Review” by Wang et al. Five papers on measurement and denoising techniques are as follows: (1) “Characterization and Correction of the Geometric Errors in Using Confocal Microscope for Extended Topography Measurement. Part I: Models, Algorithms Development and Validation” by Wang et al.; (2) “Characterization and Correction of the Geometric Errors Using a Confocal Microscope for Extended Topography Measurement, Part II: Experimental Study and Uncertainty Evaluation” by Wang et al.; (3) “Deep Transfer HSI Classification Method Based on Information Measure and Optimal Neighborhood Noise Reduction” by Lin et al.; (4) “Quality Assessment of Tire Shearography Images via Ensemble Hybrid Faster Region-Based ConvNets” by Chang et al.; (5) “High-Resolution Image Inpainting Based on Multi-Scale Neural Network” by Sun et al. Two papers on recommendation systems and education systems are as follows: (1) “Deep Learning-Enhanced Framework for Performance Evaluation of a Recommending Interface with Varied Recommendation Position and Intensity Based on Eye-Tracking Equipment Data Processing” by Sulikowski et al. and (2) “Generative Adversarial Network Based Neural Audio Caption Model for Oral Evaluation” by Zhang et al.
This Special Issue had received a total of 45 submitted papers with only 14 papers accepted. A high rejection rate of 68.89% of this issue from the review process is to ensure that high-quality papers with significant results are selected and published. The statistics of the Special Issue are presented as follows. Topics covered in this issue include the following four main parts: (I) environmental information analyses and predictions, (II) unmanned aerial vehicle (UAV) and object tracking applications, (III) measurement and denoising techniques, and (IV) recommendation systems and education systems. Four topics with accepted papers are briefly described below.
Environmental Information Analyses and Predictions
Four papers on environmental information analyses and predictions are as follows: (1) "A Data-Driven Short-Term Forecasting Model for Offshore Wind Speed Prediction Based on Computational Intelligence" by Panapakidis et al. [43]; (2) Intelligence" considered that the time series data of wind speed has the characters of high nonlinearity and volatilities. Therefore, an adaptive neuro-fuzzy inference system (ANFIS) and a feed-forward neural network (FFNN) were constructed to analyze the nonlinearity and volatilities of wind speed for short-term wind speed prediction. In their experiments, five cases were selected to predict the wind speeds of the 1-min-ahead and 10-min-ahead prediction horizons for the evaluation of the proposed method. The results show that all of mean absolute range normalized errors (MARNEs) of each case by the proposed method were lower than the MARNEs of each case by other methods (e.g., regression neural network, regression trees, support vector regression, etc.) [43].
Wan et al. from China in "Multivariate Temporal Convolutional Network: A Deep Neural
Networks Approach for Multivariate Time Series Forecasting" considered that the long-term multivariate dependencies of time series data are hard to be captured. Therefore, a multivariate temporal convolution network (M-TCN) was proposed to combine convolutional layers and residual block for extracting the spatio-temporal features of environmental data. In the experiments, two benchmark datasets including a Beijing PM2.5 dataset and an ISO-NE Dataset were used to compare the M-TCN with other methods for evaluating the proposed method. The results show that the root mean squared errors (RMSEs) of each case by the M-TCN were lower than the RMSEs of each case with other methods (i.e., long short term memory (LSTM), convolutional LSTM (ConvLSTM), Temporal Convolution Network (TCN) and Multivariate Attention LSTM-FCN (MALSTM-FCN)) [44].
Xu et al. from China in "Modeling and Analysis of Adaptive Temperature Compensation for Humidity Sensors" considered that the nonlinear compensation of sensing data is required because the humidity sensitive materials may be sensitive to temperature with nonlinear relationships. Therefore, a genetic simulated annealing algorithm (GSA) was proposed and adopted into a back propagation neural network (BPNN)-based nonlinear compensation model to compensate the sensing data of different temperature ranges. In their experiments, 150 practical datasets were collected by a humidity sensor and used to train the proposed nonlinear compensation model; furthermore, 15 practical datasets were collected and analyzed to test the trained nonlinear compensation model for the performance evaluation of the proposed method. The results show that the errors the proposed method were lower than the errors of other methods (i.e., genetic algorithm-BPNN (GA-BPNN) and artificial fish-swarm algorithm-BPNN (AFSA-BPNN)) [45].
Zhang et al. from China in "An Image Compression Method for Video Surveillance System in Underground Mines Based on Residual Networks and Discrete Wavelet Transform" considered that the image compression can be used to transfer a large number of digital images through lower bandwidth underground channels for the applications of underground mines. Therefore, a neural network containing an encoder module and a decoder module with residual units was constructed, and a metric termed discrete wavelet structural similarity (DW-SSIM) was proposed for the loss function of the neural network. In the experiments, this study collected the images from the COCO 2014 dataset and the images of underground mines for training and testing. The results show that the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) of the proposed method were higher than the PSNR and the SSIM of other methods (e.g., denoising-based approximate message passing (D-AMP), ReconNet and total variation augmented Lagrangian alternating direction algorithm (TVAL3)) [46].
UAV and Object Tracking Applications
Three papers on UAV and object tracking applications are as follows: (1) "Trajectory Planning Algorithm of UAV Based on System Positioning Accuracy Constraints" by Zhou et al. [47]; (2) "OTL-Classifier: Towards Imaging Processing for Future Unmanned Overhead Transmission Line Maintenance" by Zhang et al. [48]; (3) "Model Update Strategies about Object Tracking: A State of the Art Review" by Wang et al. [49].
Zhou et al. from China in "Trajectory Planning Algorithm of UAV Based on System Positioning Accuracy Constraints" considered that the location information cannot be accurately determined by UAVs with the limitation of system structure. Therefore, this study considered multi-constraints (e.g., vertical errors, horizontal errors, and flight distance) and proposed an improved genetic algorithm and an improved sparse A* algorithm to find the shortest trajectory length. In their experiments, two practical case studies were selected to evaluate the improved genetic algorithm and the improved sparse A* algorithm. The results show that the trajectory length could be reduced by 57.79% by the proposed methods [47].
Zhang et al. from China in "OTL-Classifier: Towards Imaging Processing for Future Unmanned Overhead Transmission Line Maintenance" considered that the transmission line-based robots equipped with cameras can only travel a line to inspect for maintenance. Therefore, an overhead transmission line classifier based on ResNet (deep residual network) and Faster-RCNN (faster regions with convolutional neural network) was proposed to analyze the images from robots for classification and inspection.
In the experiments, 1558 images, which include 406 positive samples and 1152 negative samples, were collected for evaluating the proposed classification method. The results show that the area under curve (AUC) of the proposed classification method was higher than support vector machine (SVM). Furthermore, the precision-recall (PR) curve of the proposed classification method (i.e., ResNet) was also higher than the PR curve of the combination of VGG and Faster-RCNN [48].
Wang et al. from China in "Model Update Strategies about Object Tracking: A State of the Art Review" considered that tracking model update strategies were important factors for the robustness of image recognition. Therefore, the study conducted the literature review of target model update occasions, target model update strategies, and background model updates. Four update strategy types, which include (1) update strategies based on correlation filters, (2) update strategies based on dictionary learning and sparse coding, (3) update strategies based on bag-of-words, and (4) update strategies based on neural network models, were summarized and presented. The experimental results of different update strategies from recent publications were discussed, and it was concluded that the local representation, target re-detection, and background models were important factors for the improvement of object tracking [49]. Lin et al. from China in "Deep Transfer HSI Classification Method Based on Information Measure and Optimal Neighborhood Noise Reduction" considered that high redundant spectral information in the hyperspectral images (HSIs) may interfere with the accuracy of image classification. Therefore, a deep learning method based on a dimensionality reduction method and convolutional neural networks was proposed to improve the accuracy of HIS classification. In the experiments, the dataset of Indian Pines and Salinas which were obtained by Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) sensors were collected for evaluating the proposed method. The results show that the accuracy of the proposed method was higher than that of other methods (e.g., principal component analysis (PCA)) [52].
Chang et al. from Taiwan and India in "Quality Assessment of Tire Shearography Images via
Ensemble Hybrid Faster Region-Based ConvNets" considered that the bubble defect detection is an important issue to filter out defective tires for the improvement of driving safety. Therefore, the combination of ensemble convolutional neural network and Faster-RCNN was proposed to detect bubble defects in the shearography images of tires. In their experiments, for the evaluation of the proposed method, 3279 tire images were selected as training data; 797 tire images were selected as testing data. The results show that the accuracy, sensitivity and specificity of the proposed method were higher than those of other methods (e.g., SVM, random forest, Haar-like AdaBoost, etc.) [53].
Sun et al. from China in "High-Resolution Image Inpainting Based on Multi-Scale Neural Network" considered that the blurred textures and the unpleasant boundaries may be obtained by the image inpainting method based on GAN in the cases of high resolution images. Therefore, this study applied the super-resolution using a generative adversarial network (SRGAN) to inpaint image and extract the features of textures for the improvement of image recognition. In the experiments, COCO and VOC datasets which included 135,414 images as training data and 200 images as testing data were selected to evaluate the proposed method. The results show that the PSNR and SSIM of the proposed method were higher than the PSNR and SSIM of other methods [54].
Recommendation Systems and Education Systems
Two papers on recommendation systems and education systems are as follows: (1) Eye-Tracking Equipment Data Processing" considered that high correlations may exist between users' gaze data and interests in human-computer interaction for recommendation inferences. Therefore, this study collected eye-tracking data to train a deep learning neural network model for building an e-commerce recommendation system. In the experiments, 15,922 fixation records were generated by eye-tracking devices from 52 participants. The results show that the accuracies of training dataset and testing dataset were 98.4% and 98.2%, respectively [55].
Zhang et al. from China in "Generative Adversarial Network Based Neural Audio Caption Model for Oral Evaluation" considered that the massive human work is required by oral evaluation for testing children's language learning. Therefore, an automated expert comment generation method based on gated recurrent units (GRUs), LSTM networks and GANs was proposed to extract the features of orals and generate expert comments. In their experiments, the proposed neural audio caption model (NACM) and the proposed GAN-based NACM (GNACM) were implemented and compared; several oral audios from the children of 5-6 years old were collected for evaluating the proposed models. The results show that scores of GNACM were higher than the scores of NACM; furthermore, the average response time of GNACM was lower than that of NACM [56].
Conclusions and Future Work
Four main parts, including (I) environmental information analyses and predictions, (II) UAV and object tracking applications, (III) measurement and denoising techniques, and (IV) recommendation systems and education systems, are collected and discussed in this Special Issue. These articles utilized and improved the deep learning techniques (e.g., ResNet, Fast-RCNN, LSTM, ConvLSTM, GAN, etc.) to analyze and denoise measured data in a variety of applications and services (e.g., wind speed prediction, air quality prediction, underground mine applications, neural audio caption, etc.). Several practical experiments were given in these articles, and the results indicated that the performance of the improved deep learning methods could be higher than the performance of conventional machine learning methods [43][44][45][46][47][48][49][50][51][52][53][54][55][56].
In the future, the federated learning techniques can be considered to train deep learning and machine learning models across multiple decentralized servers for data privacy and data security in electronics industries. Furthermore, the optimization techniques (e.g., gradient descent algorithm, Adam optimization algorithm, particle swarm optimization algorithm [57,58], etc.) can be improved for finding the global optimal solution. | 3,140.2 | 2020-03-19T00:00:00.000 | [
"Computer Science"
] |
Study the E ff ect of Various Sulfonation Methods on Catalytic Activity of Carbohydrate-Derived Catalysts for Ester Production
: In the present study, four types of sulfonation method, including thermal treatment with concentrated sulfuric acid (H 2 SO 4 ) , thermal decomposition of ammonium sulphate (NHSO 4 ), thermal treatment with chlorosulfonic in chloroform (HSO 3 Cl), and in situ polymerization of poly(sodium4-styrenesulfonate) (PSS), were employed to convert incomplete carbonized glucose (ICG) to sulfonated heterogeneous catalysts for the fatty acid methyl ester (FAME) production. The characteristics of synthesized catalysts were further examined using Raman spectroscopy, Fourier transformation infrared (FT-IR), ammonia temperature programmed desorption (NH 3 -TPD), Brunauer–Emmett–Teller (BET), thermogravimetric analysis (TGA), scanning electron microscopy (SEM), and energy dispersive X-ray (EDX). According to experiments, the sulfonic acid density was varied in a range from 4.408 to 14.643 mmol g − 1 over various sulfonation methods. The catalytic activity of synthesized catalysts over di ff erent sulfonation methods was determined by performing the conversion of palm fatty acid distillate (PFAD) to ester synthesis in a batch-system reactor. The findings reveal that using PSS-ICG resulted in the highest FAME yield of 96.3% followed by HSO 3 Cl-ICG of 94.8%, NHSO 4 -ICG of 84.2%; and H 2 SO 4 -ICG of 77.2%. According to results, the ICG sulfonated by PSS method with the highest acid density (14.643 mmol g − 1 ) gave the highest catalytic activity over PFAD conversion to biodiesel. According to experiment results, acid density played a crucial role over FAME yield percentage. Besides acid density, it is also worth mentioning that various sulfonation methods including di ff erent mechanisms, chemicals and sulfonating agents played crucial roles in the FAME yield percentage. These results were in accordance with the FFA conversion rate, esterification of PFAD over ICG sulfonated by in situ polymerization of poly(sodium4-styrene also o ff ered the highest FFA conversion.
Introduction
It is not deniable that conventional fossil fuels are being swiftly depleted which has raised a global concern. Moreover, the consumption of the current fuels has negatively affected the global climate due to the emission of the hazardous particulates such as carbon monoxide, sulphur dioxide, etc., which lead to the greenhouse gases (GHG) emissions [1][2][3][4]. The harmful consequences of GHG emission on the wellbeing of individuals has expanded public awareness to employ renewable resources of energy.
Biodiesel, as a promising alternative for diesel fuels, has substantially attracted the awareness of scientists. Moreover, biodiesel is considered as an environmentally friendly source of energy that is able to lessen the illness possibility by up to 90%. Biodiesel, fatty acid methyl ester (FAME), is generally derived from esterification of the free fatty acids (FFAs) or transesterification of triglycerides (TGs) over a suitable catalyst [5][6][7].
There are several important factors that highly influence the general expense of ester production: (i) accessibility of the feedstock, (ii) FFA content, (iii) synthesis route, and (iv) type of catalyst. The refined edible vegetable oils such as sunflower, rapeseed, cottonseed, soybean, palm oil, and canola have all been employed as raw feedstocks [8][9][10]. Nevertheless, the consumption of these edible oils always threatens the food crop sources, resulting in high costs of fuel production. To minimize the expense of analysis, there are demands to switch over non-eatable feedstocks. Castor oil [11], rubber seed [11], jojoba oil [12], Jatropha curcas [13], Cynara cardunculus oil [13], waste oils (trap grease, frying oil) [14] and microalgae [15] are some examples of non-edible oil used for biodiesel production. Between all the nonedible oil sources, palm fatty acid distillate (PFAD), as a by-product from the palm oil milling process, has drawn current attention due to its low-grade oil and its availability in Malaysia (which is recognized as the second major palm oil producer in the world). However, because of the high content of FFA in PFAD (85%-95%) [16], a highly stable acid catalyst is required rather than a base catalyst in order to avoid soap formation.
In contrast with homogeneous catalysts, heterogeneous acid catalysts are preferable through the process of ester generation due to the remarkable advantages that they possess such as the easy and inexpensive separation process, higher recyclability, highly intensive to FFA and aquatic content in raw materials and being environmentally friendly. Conversely, homogeneous acid catalysts are capable of concurrently catalyzing together esterification-transesterification reactions due to the presence of hydrophobic and strong acid sites. Recently, sulfonated carbon-based solid catalysts [17][18][19][20] have been significantly studied regarding their catalysis activities through biodiesel production due to possessing unique characteristics like a hydrophobic surface area, excellent thermal and mechanical stability, and low cost [21]. In a specific study, D-glucose was initially carbonized to obtain rigid graphite and then post-sulfonation treatment took place using concentrated sulfuric acid (H 2 SO 4 ) [18]. The synthesized starch derived solid acid catalyst successfully converted PFAD to biodiesel using conventional batch system. The utmost FAME yield and FFA transformation gained 90.4% and 94.6%, respectively, in presence of 2 wt.% of catalyst at 75 • C applying 10:1 methanol:PFAD molar ratio within 3 h. In another study, a carbohydrate-derived mesoporous ZnO-TiO 2 hollow sphere [22] and a mesoporous NiO core-shell solid sphere [23] were initially produced over incomplete carbonized glucose (ICG) as a template agent via the one-pot hydrothermal technique. The as-prepared composites were later immobilized by thermal treatment with chlorosulfonic acid in chloroform and thermal decomposition of ammonium sulphate, respectively, to add -SO 3 H functional species to the surface of the catalysts as well as the mesopore walls. The SO 3 H-ZnO-TiO 2 -ICG and SO 3 H-NiO-ICG catalysts showed promising FAME yield of 96.1% and 95.6%, respectively, by using an autoclave reactor.
The selection of sulfonation route depends on several considerations. One of the most crucial factors is the expected characteristics of the desired products. Certain methods are incredibly adaptable, whilst others generate a small variety of products. Every single procedure makes somewhat unique products. For instance, the air/SO 3 method has the potential to sulfonate a broad range of feedstocks and synthesize exceptional property products. The second important element to choose a sulfonation process is the capacity of the expected fabrication. Some batch procedures suit the fabrication of minor capacities of material and some are capable of large-scale continuous processes to produce tons per hour of product. Some processes such as chlorosulfuric acid can be operated as either continuous or batch processes. Chemical expense, initial equipment cost and required safety systems may possibly Catalysts 2020, 10, 638 3 of 13 have a notable effect on picking a sulfonation procedure. The apparatus cost is practically the opposite of the reactants cost. Here, the air/SO 3 route is the top in cost, whilst the uncomplicated batch sulfamic method is the lowest. The last aspect to consider in the choosing of sulfonation procedures is the cost of waste discarding. The chlorosulfuric acid route generates significant by-product streams of either sulfuric acid or hydrochloric acid where the vital apparatus is expensive, and the dumping costs would be ridiculous.
In the current research, the principal aim was to fabricate various sort of carbohydrate-derived solid acid catalyst using different sulfonation methods to maximize the FAME yield over esterification of PFAD. In this regard, ICG was initially fabricated over pyrolysis of D-glucose and further sulfonated over various sulfonating agents and conditions. The synthesized catalysts were characterized via state-of-the-art analytical methods. Moreover, the produced catalysts were later employed at a fixed esterification conditions to examine their catalytic activity over biodiesel production yield. According to our knowledge, no detailed study has been conducted on the characterization and comparison of different sulfonation methods to convert ICG to sulfonated catalysts for ester production using high FFA feedstock.
Fourier Transformation Infrared (FT-IR) Analysis
The FT-IR spectroscopy was applied for analyzing the nature of functional species linked to the active sites. From the spectrum shown in Figure 1a, all sulfonated catalysts had clear and strong vibration at around 1597.31 and 1686.02 cm −1 , which corresponded to the C=C aromatic ring vibration mode and C=O stretching mode of the carboxyl spices in turn [16]. The presence of sulfonic groups was confirmed by clear peak at 1036.70 and 1227.21 cm −1 , which were attributed to symmetric and asymmetric O=S=O stretching vibration [17]. From the spectra, the synthesized HSO 3 Cl-ICG (C) and PSS-ICG (D) catalysts had stronger vibration at peak 1036.70 cm -1 as compared to those H 2 SO 4 -ICG (A) and NHSO 4 -ICG (B). Moreover, all four catalysts showed a peak at 750.10 cm −1 which proved attachment of S−O group to the active sites. The spectra showed two distinct bands occurring at wavelength 1580 and 1370 cm −1 that are referred to as G and D-bands, respectively [25,26]. The presence of these two signals proved that polycyclic aromatic carbon sheets
Ammonia Temperature Programmed Desorption (NH 3 -TPD) and Brunauer-Emmett-Teller (BET) Analysis
The acid density of the sulfonated ICG can highly manipulate the catalytic performance through esterification. According to the previous literature, an excess of acid density increased the catalytic performance and subsequently enhanced the conversion rate of biodiesel production [24]. The NH 3 -TPD was used to determine the intensity and dispersal of the acid sites. TPD profiles (Figure 1b) revealed that all the catalysts except HSO 3 Cl-ICG (C) possessed two unique desorption peaks from 150, 350, and 400 to 850 • C, which were attributed to weak and strong brönsted acids on the surface in turn. However, it was observed that HSO 3 Cl-ICG (C) had only one peak, which was a broad peak as of 400 to 850 • C and maximized at 600 o C. The presence of these peaks ensured that the synthesized catalysts in this research were chemically and thermally steady up to 300 • C prior to the decomposition of -SO 3 H species. According to the results indicated in Table 1, the sulfonic acid density was varied between 4.408 and 14.643 mmol g −1 over various sulfonation methods. The specific surface areas (S BET ) of the synthesized sulfonated catalysts were analyzed from the N 2 adsorption/desorption isotherm. The synthesized sulfonated catalysts possessed specific surface areas within the range of 4.27-8.70 m 2 g −1 , as summarized in Table 1. As expected, the specific surface area of the synthesized catalysts was very low and had minor effect on catalysis process, hence much more attention was paid to studying the differentiation of the sulfonic methods and their effects on the catalytic activity. *Esterification was performed over following reaction condition: operating temperature of 80 • C, catalysts concentration of 2.5 wt.%, methanol to oil ratio of 10:1, operating time of 4 hrs.
Raman Spectroscopy Analysis
Figure 2a depicts Raman spectra for all sulfonated ICG catalysts. The spectra showed two distinct bands occurring at wavelength 1580 and 1370 cm −1 that are referred to as G and D-bands, respectively [25,26]. The presence of these two signals proved that polycyclic aromatic carbon sheets were successfully formed. In detail, the D-band with A 1g D breathing mode corresponded to the deficiencies in the carbon sheets or presence of amorphous carbon which confirmed that the carbonization process had already taken place [27]. The G band was assigned to carbon atoms with a single crystal that vibrated in the opposite direction, which confirmed the formation of graphitic structures. The degree of structural order changes of graphite and defect was determined using the intensity ratio of amorphous carbon I(D)/I(G) [28]. The intensity ratio for all the synthesized catalysts-H 2 SO 4 -ICG (A); NHSO 4 -ICG (B); HSO 3 Cl-ICG (C); and PSS-ICG (D)-were calculated to be 0.714, 0.711, 0.704, and 0.615, respectively. It disclosed that the sulfonation treatment through inserting −SO 3 species on the imperfect carbonized sp 2 caused the imperfections of the carbon structure [26].
of 13
Catalysts 2020, 10, x FOR PEER REVIEW 5 of 13 sulfonation method which affected the stability of the acid site through the attachment of the -SO3H group to the carbon sheets.
Scanning Election Microscopy (SEM) and Energy Dispersive X-ray (EDX) Analysis
The SEM images and elemental arrangement for all the sulfonated products are shown in Figure 3; Figure 4, respectively. As expected, the carbon sheets composed predominantly of polycyclic aromatic species coupling with enormous quantities of SO3H groups, which were similar to images depicted by other researchers [28]. As outlined in Table 1 from the EDX data, PSS-ICG possessed the highest binding content of S element (7.87%) and O (10.85%), followed by HSO3Cl-ICG with content of S element (6.26%) and O (9.82%) ; NHSO4-ICG; S element (4.51%), O element (36.16%), H2SO4-ICG with S element (1.91 %), O element (33.83%).
Thermogravimetric Analysis (TGA)
The thermal stability of each prepared catalyst was examined using the TGA technique ( Figure 2b). The thermal disintegration of the synthesized catalysts demonstrated that slight and gradual weight losses of about 5%-10% in the temperature up to 150 • C could be assigned to dihydroxylation. This was followed by substantial weight losses between 150 and 360 • C that could be assigned to decarboxylation. The third strong desorption was distinguished in the array of 360-600 • C that could be corresponded to the de-sulfonation. This implied that the attached sulfonic group could remain highly stable at 360 • C [29][30][31].
From the analysis, it was notified that the ICG sulfonated by thermal treatment with chlorosulfonic acid in chloroform (C) and ICG sulfonated by in situ polymerization of poly(sodium4-styrene sulfonate) (D) had quite a similar pattern of TGA as compared to ICG sulfonated by thermal treatment with concentrated sulfuric acid (A) and ICG sulfonated by thermal treatment with ammonium sulphate (A), which were slightly different. This may be because of the different type of sulfonation method which affected the stability of the acid site through the attachment of the -SO 3 H group to the carbon sheets.
Scanning Election Microscopy (SEM) and Energy Dispersive X-ray (EDX) Analysis
The SEM images and elemental arrangement for all the sulfonated products are shown in Figure 3; Figure 4, respectively. As expected, the carbon sheets composed predominantly of polycyclic aromatic species coupling with enormous quantities of SO 3 H groups, which were similar to images depicted by other researchers [28]. As outlined in Table 1 The SEM images and elemental arrangement for all the sulfonated products are shown in Figure 3; Figure 4, respectively. As expected, the carbon sheets composed predominantly of polycyclic aromatic species coupling with enormous quantities of SO3H groups, which were similar to images depicted by other researchers [28]. As outlined in Table 1 from the EDX data, PSS-ICG possessed the highest binding content of S element (7.87%) and O (10.85%), followed by HSO3Cl-ICG with content of S element (6.26%) and O (9.82%) ; NHSO4-ICG; S element (4.51%), O element (36.16%), H2SO4-ICG with S element (1.91 %), O element (33.83%).
Figure 3
Scanning electron microscopy (SEM) images of ICG sulfonated by thermal treatment with concentrated sulfuric acid (a); ICG sulfonated by thermal treatment with ammonium sulphate (b); ICG sulfonated by thermal treatment with chlorosulfonic acid in chloroform (c); ICG sulfonated by in situ polymerization of poly(sodium4-styrene sulfonate) (d) at 500× and 5000× magnification
Comparisons of the Catalytic Activities using Different Sulfonation Methods
The catalytic performance of sulfonated catalysts was examined through the esterification of PFAD. The FFA conversions for each catalyst were assessed via titration route and later the values were calculated using Equation (2). The esterification reaction conditions were used as follows; operating temperature of 80 • C, operating time of 4 h, methanol to oil molar ratio of 10:1, and catalyst amount of 2.5%. According to the obtained results shown in Figure 5, the presence of PSS-ICG (D) gave the highest FFA conversion of 97.6% followed by HSO 3 Cl-ICG (C) of 95.2%, NHSO 4 -ICG (B) of 89.8%; and H 2 SO 4 -ICG (A) of 85.3%. It can be concluded that catalysts with a superior total acidity yielded higher FFA conversion in turn.
Subsequently, the percentage of total FAME yield produced was determined via Equation (1) and illustrated in Figure 5. From the calculation, ICG sulfonated by in situ polymerization of poly (sodium4-styrene sulfonate) (D) offered the highest FAME yield of 96.3% followed by ICG sulfonated via thermal treatment with chlorosulfonic acid in chloroform (C) of 94.8%, ICG sulfonated by thermal treatment with ammonium sulphate (B) of 84.2%, and ICG sulfonated by thermal treatment with concentrated sulfuric acid (A) of 77.8%. These results were in accordance with the FFA conversion rate, whereas esterification of PFAD over ICG sulfonated by in situ polymerization of poly(sodium4-styrene sulfonate) also offered the highest FFA conversion. The summary of FAME yield and total acidity over the four unique sulfonation methods is shown in Table 1. From all the results obtained, it was clearly notified that acid density played a crucial role in the FFA conversion rate and FAME yield percentage. FAME yield increased with the increase in total acid density. It was observed that PSS-ICG (D) gave the highest FAME yield among other prepared catalysts with acidity of 14.463 mmol g -1 . This may result from the significant forms caused by the benzenesulfonic acid species suppressed in poly(sodium4-styrenesulfonate). The SO3H sulfonic group became stronger in the existence of the three electronegative oxygen atoms and also increased the properties of an electron-withdrawing group. It is also worth mentioning that using sulfonated ICG as a catalyst gave higher FAME yield (96.3%) as compared to sulfonated multi-walled carbon nanotubes and SO3H-ICS catalyst with the ester yield of 93% [32] and 90.4% [18], respectively.
On the other hand, not only different type of sulfonation methods may influence the FAME yield percentage, but also different mechanisms and chemicals involved sulfonating agents and conditions may influence sulfonation process and subsequently FAME yield percentage. Sulfonation via poly (sodium-4-styrenesulfonate) method at lower temperature may influence the polymerization process and bonding of SO3H species to the active sites. This was supported by Kanokwan et al. [33] study, whereas with rising the sulfonation temperature, the overall acidity decreased. It may result in increasing of oxygen to carbon ratio because of the fact that oxidation, dehydrogenation, or condensation may take place. As indicated in Table 1, the oxygen to carbon ratios of PSS-ICG (D) and HSO3Cl-ICG (C) were much lower as compared to those of NHSO4-ICG (B) and H2SO4-ICG (A) where both of them (B and A) were treated at a higher sulfonation temperature. A higher oxygen:carbon ratio may increase the number of weak acid groups on the ICG, resulting in a lower total acid density [34]. Although post-sulfonation treatment of ICG sulfonated by thermal treatment with concentrated sulfuric acid (A) was conducted at extremely elevated temperatures (150 °C), it possessed the lowest acidity of 4.408 mmol g -1 . However, for the ICG by thermal treatment with ammonium sulphate (B), The summary of FAME yield and total acidity over the four unique sulfonation methods is shown in Table 1. From all the results obtained, it was clearly notified that acid density played a crucial role in the FFA conversion rate and FAME yield percentage. FAME yield increased with the increase in total acid density. It was observed that PSS-ICG (D) gave the highest FAME yield among other prepared catalysts with acidity of 14.463 mmol g -1 . This may result from the significant forms caused by the benzenesulfonic acid species suppressed in poly(sodium4-styrenesulfonate). The SO 3 H sulfonic group became stronger in the existence of the three electronegative oxygen atoms and also increased the properties of an electron-withdrawing group. It is also worth mentioning that using sulfonated ICG as a catalyst gave higher FAME yield (96.3%) as compared to sulfonated multi-walled carbon nanotubes and SO 3 H-ICS catalyst with the ester yield of 93% [32] and 90.4% [18], respectively.
On the other hand, not only different type of sulfonation methods may influence the FAME yield percentage, but also different mechanisms and chemicals involved sulfonating agents and conditions may influence sulfonation process and subsequently FAME yield percentage. Sulfonation via poly (sodium-4-styrenesulfonate) method at lower temperature may influence the polymerization process and bonding of SO 3 H species to the active sites. This was supported by Kanokwan et al. [33] study, whereas with rising the sulfonation temperature, the overall acidity decreased. It may result in increasing of oxygen to carbon ratio because of the fact that oxidation, dehydrogenation, or condensation may take place. As indicated in Table 1, the oxygen to carbon ratios of PSS-ICG (D) and HSO 3 Cl-ICG (C) were much lower as compared to those of NHSO 4 -ICG (B) and H 2 SO 4 -ICG (A) where both of them (B and A) were treated at a higher sulfonation temperature. A higher oxygen:carbon ratio may increase the number of weak acid groups on the ICG, resulting in a lower total acid density [34]. Although post-sulfonation treatment of ICG sulfonated by thermal treatment with concentrated sulfuric acid (A) was conducted at extremely elevated temperatures (150 • C), it possessed the lowest acidity of 4.408 mmol g -1 . However, for the ICG by thermal treatment with ammonium sulphate (B), the low point of acidity of 5.299 mmol g -1 might correspond to the fact that post-sulfonation over a closed system may reduce the effect of oxidation, dehydrogenation, or condensation.
Materials
The initial material, D-glucose, was purchased from R&M Chemicals. All the required acidic agents were supplied accordingly; concentrated sulfuric acid [
Synthesis of Incomplete Carbonized Glucose (ICG)
To produce ICS, the pyrolysis procedure was adopted using a furnace tube, where 10 g of D-glucose was filled in the calcination boat and put into the furnace under the following conditions: temperature of 400 • C for 12 h with nitrogen (N 2 ) gas flow rate at 1 mL/min to provide the inert atmosphere. After the calcination, the calcined sample was cooled down to room temperature and later was processed via milling at 1000 rpm for 60 min to get the fine black powder.
Sulfonation by Thermal Treatment with Concentrated Sulfuric Acid
Conventionally, acid treatment is thermally carried out using concentrated sulfuric acid (H 2 SO 4 ) in order to produce carbon solid acid catalysts to synthesize biodiesel [7,20]. In this regard, 2.0 g of ICG was interspersed together with 50 mL of H 2 SO 4 and then sonicated for half an hour. The blend was later stirred and refluxed for 12 h at 150 • C, over N 2 flow at 100 mL/min to give an inert condition. The calcined sample was cooled down to room temperature then filtered and washed with double distilled water (DW) until the pH of the residual water became neutral and dried out at 120 • C, overnight. The final produced catalyst sample was labeled as H 2 SO 4 -ICG.
Sulfonation by Thermal Decomposition of Ammonium Sulphate
This sulfonation method was implemented and amended as proposed by Siew et al. [32], where 0.4 g of ICG was interspersed with 30 mL ammonium sulphate (NH 4 ) 2 SO 4 and sonicated for 10 min. Then, the mixture was flowed into an autoclave and sealed at 200 • C for half an hour under autogenous pressure. As it cooled down, the mixture was then filtered using a vacuum pump and washed ruinously using DW to eliminate dissipation of (NH 4 ) 2 SO 4 . The mixture was later dried in the oven at 120 • C, overnight. The final sample was labeled as NHSO 4 -ICG.
Sulfonation by Thermal Treatment with Chlorosulfonic Acid in Chloroform
This method was referred from Pravin et al. [35], where 2.0 g of ICG was dissolved into 50 mL of chloroform for 60 min in a sonicator, resulted in a black distributed ICG suspension. Then, 5 mL of chlorosulfonic acid was cautiously poured into the mixture, stirred and refluxed at 70 • C for 4 h. As the mixture cooled down, it was then filtered and washed with a mixture of DW/Ethanol to eliminate organic moieties till the pH of the filtrate turn into neutral. The sample was then dried at 120 • C for 12 h in the oven. The final product was labeled as HSO 3 Cl-ICG.
Sulfonation by in situ Polymerization of Poly(Sodium4-styrene Sulfonate)
This sulfonation system was adopted as proposed by Shuit et al. [32], where 0.4 g of ICG was stirred in a mixture of 0.8 g of poly(sodium 4-styrene sulfonate) and 100 mL deionized (DI) water in a round bottom flask at room temperature for 10 h. Then, ammonium persulphate (NH 4 ) 2 S 2 O 8 was poured into the blend, stirred and warmed up to 65 • C for 48 h, using a hot plate stirrer. As the mixture cooled down, 100 mL of DI was poured into the cooled blend, sonicated for 60 min, then filtered and washed frequently by DW. Next, the filtrate was added to 500 mL of 4 M H 2 SO 4 and stirred at lab temperature for 24 h. After this, the blend was filtered using a vacuum pump and washed again with DW until the pH of the filtrate turn into neutral. The sample was finally dried out at 120 • C in the oven for 12 h. The final product was labeled as PSS-ICG.
Characterization Methods
The morphology of the synthesized catalysts was examined via Raman spectroscopy (WITec; Alpha 300R). Fourier transformation Infrared (FT-IR; Perkin Elmer -model GX) was employed to determine type of functional spices connected to the active site of the catalyst. In addition, ammonia temperature programmed desorption (NH 3 -TPD; Thermo Finnigan, model TPDRO. 1100. series) was utilized to investigate the acid density of the synthesized catalysts. The specific surface area (S BET ) of the synthesized sulfonated catalyst were analyzed using the Thermo Finnigan Sorptomatic 1990 series apparatus. Thermogravimetric analysis (TGA) was applied to determine the mass losses of the functional groups from room temperature to 1000 • C in an air flow of 200 mL min -1 at a rate of 10 • C min -1 using Mettler Toledo, 990. The morphology of the synthesized catalysts was studied by scanning electron microscopy (SEM; JEOL, JSM-6400) fitted with the energy dispersive X-ray (EDX) spectroscopy analysis system.
Catalytic Activity of the Catalysts
The catalytic performance of each prepared sulfonated catalyst was examined through the esterification of PFAD. The esterification reaction was performed with a reflux system where the round-bottom flask was connected to a condenser to re-condense the vaporized methanol to minimize the loss amount of methanol. An important point that needs to be taken into consideration is that high loss of evaporated methanol can significantly affect the FAME yield production during esterification reaction [7]. The esterification was done at a reaction temperature of 80 • C, methanol to PFAD molar ratio of 10:1, with catalyst concentration of 2.5 wt.%, in the presence of 5 g of PFAD feedstock for 4 h, as per our previous research [7]. During the esterification reaction, the blend was stirred at 600 rpm to have homogenized dispersed of catalyst in the mixture along the operating time. By the completion of the reaction, the catalyst was removed from the mixture by centrifuging at 5000 rpm for 20 min. After phase separation overnight, the methanol was withdrawn and recovered from the blend. The final FAME was collected for more evaluation.
FAME Analysis and FFA Conversion
The characteristics of the synthesized ester were examined utilizing a GC outfitted with a flame ionization detector (FID; HP, 6880). In addition, to separate the FAME compound, an extremely polar vessel column (BPX 70, SGE Company) was employed. N-hexane was introduced as the solvent for the weakening of the samples while helium gas as the carrier gas. In this process, 500 ppm of each standard methyl myristate, methyl palmitate, methyl oleate, methyl linoleate, and methyl stearate were used as the reference standard. Methyl heptadecanoate was mixed with the prepared ester to be used as an internal standard. Later, 1 µL of the sample solution was inserted into injector port. The oven temperature was initially adjusted at 100 • C and risen to 230 • C with the heating rate of 10 • C/min. The indicator temperature was put at 270 • C. The product yield was further calculated using Equation (1) as depicted below [36]: where C is the calculation of ester yield, A is the amount of maximum zone for the FA summits, A meh is peak zone of methyl heptadecanoate, C meh is the internal standard strength, V meh is the volume utilized of internal standard, and W t is the FAME's mass. The FFA conversion rate (%) from PFAD feedstock to FAME was determined using titration method. Equation (2) shows the calculation for acid value determination [37]: where AV f and AV p are the acid values of the feedstock and the product, respectively.
Conclusions
In the present investigation, incomplete carbon glucose (ICG) was initially fabricated using D-glucose as carbon resources. Various sulfonation methods were introduced in order to convert the ICG to sulfonated heterogeneous catalysts for FAME generation. The esterification of PFAD was conducted over four different synthesized sulfonated catalysts upon using the batch technique. The present study showed that acidity played a crucial role in catalytic activity, where FAME yield improved with the rise in acidity. In addition, different types of sulfonation method along with different mechanisms and chemicals involved sulfonating agents and conditions influence the FAME yield percentage. The in situ polymerization of poly(sodium 4-styrene sulfonate) method resulted in the highest FFA conversion and FAME yield of 97.6% and 96.3%, respectively, as compared to other studied methods.
Conflicts of Interest:
The authors declare no conflict of interest. | 6,896.4 | 2020-06-08T00:00:00.000 | [
"Chemistry"
] |
Spatial Planning of Electric Vehicle Infrastructure for Belo Horizonte , Brazil
1The Environmental Research Institute, University College Cork, Ireland 2Faculty of Science and Technology, Nova University of Lisbon, Caparica, Portugal 3Department of Cartographic Engineering, University of the State of Rio de Janeiro, Rio de Janeiro, Brazil 4Center for Environmental and Sustainability Research, Faculty of Science and Technology, Nova University of Lisbon, Caparica, Portugal 5Engineering Department, Military Engineering Institute (IME), Rio de Janeiro, Brazil 6Center for Innovation, Faculty of Science and Technology, Instituto Superior Técnico, University of Lisbon, Lisbon, Portugal 7Energy Policy and Modeling Group School of Engineering, University College Cork (UCC), Ireland
Introduction
There is a large mobility demand in urban centers.These urban centers hold about 80% of global Gross Domestic Product (GDP) and account for two-thirds of all primary energy consumption and 70% of carbon dioxide (CO 2 ) emissions [1,2].Global transport is responsible for almost a quarter of energy-related emissions [1,2].In Brazil, more than 32% of final energy consumption is from the transport sector with road transportation accounting for more than 92% of this amount [3].Ethanol has an important role for low carbon transportation in Brazil.A Comparative Life Cycle Assessment (LCA) study shows that when replacing gasoline with ethanol fuels, Greenhouse Gas (GHG) emissions can decrease by around 81% [4].On the other hand, several studies highlight that the agricultural process for producing ethanol can be responsible for considerable problems such as ecotoxicity, soil acidification, and human toxicity among others [5,6].Some even studies that identify benefits in the production of ethanol also highlight the negative environmental impacts [7].Other studies suggest that the competition from the production of ethanol with food is another substantial problem [8][9][10].
A study for Sao Paulo (with adopted emissions by car of 220 gCO 2 /km) revealed that EV could reduce around 11,0 TgCO 2 , if 20% of its fleet of gasoline cars is replaced by 2030 [19].This is equivalent to removing the emissions of about 140,000 medium gasoline-powered cars in one year when considering an LCA methodology-European study indicates emissions of 250 gCO 2 / km [20].Therefore, the EV can eliminate the emission of local pollutants (namely, HC, CO, NOx, and PM) and significantly contribute to reducing energy consumption and the mitigation of climate change, especially when the electricity is produced from a renewable energy mix [21,22].
Emissions from the transport sector have grown 2.5% annually between 2010 and 2015; however, OECD (Organization for Economic Cooperation and Development) countries aim to reduce emissions by 2.1% annually between 2015 and 2025 [1,2].The development of EV technology and corresponding recharging infrastructure can be one potential way to reduce emissions as well as increase energy efficiency [23].
However, the lack of expansion of suitable electric vehicle supply equipment (EVSE) networks provides a barrier for EV, e.g.range anxiety effect (the drivers fear that the electric vehicle will not have enough range to reach its destination) [24], even though features such as Safe-Range-Inventory (SRI) is capable of overcoming the obstacle of range anxiety.In some developing countries, other problems also need to be considered when planning for the development of EVSE.These include greater social inequality, public security problems, and lack of logistic support, as developing countries also do not have the same level of experience when compared to developed countries.China is the only developing country that is prioritizing EV.However, this Asian country may not be a good benchmark, since China and Brazil have different public safety standards; i.e., there is no significant record of vandalism in Chinese EVSE.
Consequently, the aim of this study is to measure and identify the optimal locations for the installation for EVSE in the municipality of Belo Horizonte (BH), Brazil, to meet a penetration of 1% of light-duty electric vehicles (LDEV) by 2025.This study answers the following questions: (i) what attributes should be highlighted for the installation of an EVSE network in urban areas in developing countries with substantial risk areas?And (ii) where is the optimal location for a network of EVSE network in the municipality of Belo Horizonte?
The literature review is presented in the next section.The third section is dedicated to the methodology, the fourth section presents the results, the fifth section presents the discussion, and the sixth section presents the conclusions, limitations, and suggestions for upcoming studies.
Characterization of the Studied Region.
A large share of the total GHG emissions in Minas Gerais (MG), whose capital is BH, is due to transportation activities.It ranks second in all of Brazil's 26 states for GHG emissions.The transport sector is responsible for around 36% of the state's total emissions with the road transport sector accounting for more than 96% of these emissions [25].
The estimated population in BH is around 2.5 million inhabitants, the highest population density in the MG state.The municipality of BH has 17% of the total road fleet in the state of MG which is estimated to be 1.8 million units-70% of these are light-duty vehicles (LDV).
BH is divided into 9 administrative regions, 8 districts, and 487 neighborhoods [26].Some of BH neighborhoods, such as Sion, Lourdes, and Belvedere, have a significant population concentration with an average household income higher than US $ 2.2 thousand monthly.The income square meter price in the area is higher than the average within the BH municipality [27].
Other neighborhoods such as Savassi, Santo Agostinho, and Funcionários have a high population concentration.The BH South-Central region (Figure 1) is characterized as a metropolitan center with a diversity large quantity of services and a concentration of economic activity [28].
From an environmental and energy point of view, BH reveals a favorable potential for the expansion of electric mobility since Brazil has a predominantly clean energy mix.In 2016, almost half the energy consumed in Brazil was from renewable sources [29].The Brazilian electricity matrix uses 82% of renewables, due to hydroelectric generation that accounts for 68% of the total electric generation [29].The use of renewable energy in recharging of EV is crucial to guarantee a reduction in CO 2 life-cycle emissions.
On the other hand, there are some challenges that need to be overcome in the case of large-scale implementation of an EVSE network in Brazil's urban areas.These include irregular urbanization with large areas of subnormal agglomerates (cluttered and dense poor settlements, most lacking basic urban utilities and essential services).Often it is necessary to cross-areas of irregular urbanization to have access to airports, railway stations, bus terminals, highways, financial zones of the city, and upper middle class neighborhoods.
In 2010, Brazil registered more than 6,300 subnormal agglomerates-with more than 50% of them in the Southeast of the country-containing around 1.6 million households and a population of around 5.5 million inhabitants [30].These areas are often associated with urban violence, due to the difficulties in accessing essential public services in these regions.
In BH, the situation is not different; in 2009, there were more than 364,000 living in subnormal agglomerates, corresponding to 23% of the population and 10% of the territory of the municipality [31].Areas of geological hazards and areas subject to flooding should also receive special attention.BH experiences recurrent floods, mainly in the regions of the municipality involving the basins of the Arruda (South and South-Central) and Onc ¸a (North) due to problems involving irregular soil patterns and steep relief [32].were 2 million EVs worldwide, with more than 750,000 EVs sold in the same year.In addition, there are more than 200 million two-wheeled EVs and 345,000 electric buses mainly concentrated in the Chinese market.This highlights the increase of EV penetration that can occur over a relatively short period of time; as in 2005, there were only a few hundred EVs [1,2].A global goal of 100 million electric cars and 400 million electric 2-and 3-wheelers by 2030 was defined in the Paris Declaration on Electro-Mobility and Climate Change and Call to Action [33].
Electric Vehicle
Restrictive policies on the use of ICE vehicles adopted by some countries may favor the penetration of EV.Some European countries, for example, are already restricting the circulation of ICE vehicles in some areas and allowing access only to low emission vehicles such as EV [34].Europe has set a goal of reducing transport emissions by 20% by 2030 and by 70% in 2050 compared to 2008 levels [34].In Brazil, electric mobility is beginning to develop but the number of EVs is still insignificant.From 2011 to 2016, around 3,500 units of LDEV (most of them for demonstration purposes) have been licensed in the country (in this period BH does not have any EV registered), while in this period more than 15.3 million LDV (ethanol, gasoline, and fuel flex) have been licensed in Brazil [35].
Importance of EVSE for the Success of EV.
In 2016, the EVSE infrastructure grew substantially and reached around 2.3 million charging points globally.However, most EV drivers depend on public charging points to recharge their vehicles [1,2].The lack of charge points near the place of residence, roads, and workplace is a potential barrier to EV market penetration due to the range anxiety barrier [36].The EVSE network must be adequately planned considering both volume and location [37].
The characteristics of the EVSE, especially those related to the charging level, are relevant for the creation of an infrastructure network.The EVSE levels (Table A2) considered in this study are as follows [38,39]: (i) EVSE L1 (level 1) is equivalent to a slow recharging station corresponding to 8 h or more for a full charge (5): [43,55,57,61] * Contemplated by other studies.[40,41].It is recommended for the area where the driver lives.
(ii) EVSE L2 (level 2) corresponds to the fast charging station accounting for 3 h to 8 h for a full charge [40,41].It is recommended for workplaces, public transportation connection areas, and shopping centers.
(iii) EVSE L3 (level 3) is characterized by super-fast recharge of up to 30 minutes for a full charge [40,41].It is recommended for shopping malls, roads, corridors and main avenues.
The optimization of public and private investments in the expansion of EVSE requires research in order to gain knowledge on the main attributes capable of calibrating the supply versus demand equation [42].These attributes include the availability of garages in residential areas, points of public transport connections, high traffic roads, shopping centers, large parking lots, and other areas with a considerable concentration of vehicles.In addition, another important factor is the EV range, which continues to improve.For example, in 2011, only three models of pure EV were available for the mass-market in the USA, with a range between 100 and 150 kilometers.In 2017, 15 pure EV models were available in the USA, with a minimum range of 92 kilometers for the Smart Fortwo Electric Drive Coupe model and a maximum range of 540 kilometers for the Tesla Model S 100D [41].
Additionally, the availability of physical space to install EVSE must also be accounted for, as well as the impact on the electricity generation network [43][44][45][46][47], the general attitude towards EV [44], and consumer attitude [48] and psychological aspects [49] as well.The installation of an EVSE network should include technical and geographic resources in order to meet consumer's expectations of charging the vehicle's battery in the least amount of time and with short displacements [43], thus avoiding unnecessary energy consumption.
Among the most important innovations pointed out by the research are the areas of restriction since they are significantly important in projects that contemplate the development of EV infrastructure.A detailed literature review reveals the main attributes for the expansion of the EVSE network in urban areas, as shown in Table 1.
Materials and Methods
The study is divided into two parts.The first focuses on establishing the methods, defining geographic criteria, identifying the demand, and completing a survey with specialists.The second part of the study aims at parameterizing and processing the data in the GIS tool.
Method Definition.
GIS was used because it is a spatial information system capable of processing different types of data and accurately indicating the desired spatial location.Similarly to other studies (Church 2002), GIS was used to carry out this study and to manage geospatial data analysis supporting the criteria and indicators for the optimal location on EVSE [67].Furthermore, the Multi-Criteria Decision Making (MCDM) method for locating EVSE is important and requires maximum assertiveness [68], as facilities of this type are meant to endure for large periods (i.e., decades).
In the context of applying the MCDM, a multidisciplinary group of specialists were invited to evaluate-through the assignment of the importance of different attributes-the best location for the EVSE network.The MCDM allows the consistent analysis of different quantitative and qualitative elements in a decision-making process.The MCDM method has been widely used for decision-making processes applied to the environment, energy, business, and infrastructure for EV [42,[69][70][71].Additionally, MCDM was considered adequate for this study, since Brazil is an incipient EV market, with the additional problem of the limitation and complexity of the available processing data [72], justifying the use of MCDM for the localization of infrastructure for electric vehicles.
The Weighted Linear Combination (WLC) method was applied, as it is an analytical method for dealing with multiattributes or when more than one attribute must be taken into consideration.The WLC method has been largely used in similar studies [73,74].Due to the complexity of the problem, the Analytical Hierarchy Process (AHP) was also considered.AHP helps to solve problems using a comprehensive and rational procedure.The AHP technique is also widely used in studies linked to geographic analysis [73][74][75].
Demand Identification.
The demand identified in this study corresponds to an estimated LDEV penetration of 1% in BH by 2025.This estimate was based on a report on energy demand for the Brazilian automotive sector [29].This is a conservative scenario; however it is justified as after more than half a decade of EV availability, which is comparable with the market share of less than 1% in most countries [1,2].Among developing countries, only China shows a high EV penetration; nevertheless, the EV market penetration in 2016 was only around 1% [1,2].In Brazil, the rate of EV expansion may be affected by the lack of public policies for electric mobility development and the existing governmental ethanol protection policy.
In the absence of a reference for the ideal number of EVSE, as well as the ideal ratio between EVSE level 2 (L2) and EVSE level 3 (L3), the experience of other urban regions was used as a reference.We defined the proportion of one EVSE for each ten electric vehicles for use in this study.This is comparable to a case study of the main Chinese cities that have the same target [39].Regarding the EVSE by level-(L2) and (L3)-the proportion of 80% for EVSE L2 and 20% for EVSE L3 was used.This was based on the average number of EVSE for sixteen countries that have a more intensive electric mobility use [21,22,29,33].This study also considers that each EV owner has a home charger (EVSE L1).
Survey to Specialists.
The electrified mobility is in very incipient phase in Brazil.There are a limited number of experts in the country.To identify specialists in this topic, we made email and phone contact was made with Brazil electric vehicles associations, car manufacturing association, car dealers federation, automakers, car dealerships, suppliers of EV equipment, union of taxi companies working in EV test projects, specialized media, and research by Internet.
After preparing a list containing the names of the specialists, we sent 67-approach email to clarify the research and invite them to participate by answering a questionnaire that would be sent later.
The main attributes identified based on the literature review (Table 1) used to create the questionnaire that was classified into three groups: economic, sociodemographic, and geographic.Before submitting the questionnaire, we carried out a 40-day trial with a group of 17 people formed by researchers and friends linked to the automotive sector.The test proved to be valuable in adjusting some issues.
The survey was conducted by email from 09/05/2016 to 10/09/2016, and the sample covered 51 specialists with an interest in electric mobility.
The survey targeted EVSE L2 and EVSE L3, and specialists were required to assign a value for each attribute, according to the criteria explained in the introduction of each question.The comparison between two elements using AHP can be performed in different ways [76].However, the scale of relative importance between two alternatives proposed by Saaty [77] is the most widely used.Therefore, the specialists assigned values from 1 to 9 for each attribute, where 1 meant that the attribute had no importance to the location of the EVSE and 9 was linked to highest importance.The questionnaire had ten questions with one space for the respondent to identify new attributes.
Although the number of respondents was not expressive, it does not compromise the study's results, because the specialists limited themselves to evaluating the attributes-elements of greater relevance-that were identified in publications covering studies and experiences with EVSE implementation in other countries.Most of the respondents were in executive direction (57%), worked in the private sector (79%), in the automotive industry (37%), and lived in Southeast Brazil, as shown in Table 2.
GIS Analyses.
The use of the GIS tool can be considered as the backbone of the study.It is composed of three steps: preprocessing, processing, and generation of maps, as described in the next sections.
First Processing
Step.The first stage of the GIS analysis called preprocessing was divided into two phases.The first one focused on creating parameters and obtaining and preparing the processing data and the second phase on management of the attributes.(i) Preprocessing and normalization consist of the preprocessing of the geographic data in Flooding areas [65] order to structure and parameterize the study.The attributes were established in vector format, as presented in Table 3.
In the evaluation of the spatial analysis and location of EVSE, operational characteristics for the different levels of EVSE were considered: (i) EVSE L1: the attributes considered were household income (equivalent to US $ 2.2 thousand or higher) and population density of people aged 18 years or older (Table 1).(ii) EVSE L2: it is suitable to be located near shopping centers, workplaces, or public transportation connections with parking lots with a minimum of 50 parking spaces, among other attributes shown in Table 1.(iii) EVSE L3: the places with the greatest potential for installation are roads, shopping centers, and public transportation connections among other attributes revealed in Table 1.The attributes were grouped and processed according to the EVSE levels (L1, L2, and L3), as shown in Table 4.
The Euclidean distance was adopted and defined as a grouping from the delimitation of the average distance between the attributes, i.e., if the average distance between shopping malls is 20 km, 10 km is considered to be the maximum radius (limited to 30 km) from each shopping mall in which the EVSE should be installed.The attributes are standardized by the minimum and maximum values of classification.The Boolean method was applied and consists of the logical combination of binary values, where each attribute was evaluated and standardized at "0" as an unsatisfactory hypothesis and "1" as a satisfactory hypothesis [3].Spatial inference is a step in the MCDM that aims to integrate the data involved.For this, we perform the standardization of attributes (Table 3) into two categories: benefits and cost [78].From the distance condition of the attributes, the benefits criteria are the maximization indicators whose values are always higher (value 1).The cost criteria are related to distance minimization (value 0).
(ii) Management of attributes defines the importance of each attribute from the peer-to-peer comparison between attributes.
The AHP that is created under MCDM is composed of techniques that are suitable for ranking of critical management problems.The AHP is a ranking process that is used in making group decision and is widely used around the world in a variety of fields.
Through the AHP process and the multicriteria decisionmaking technique, the quantitative and qualitative aspects were combined generating weights for the attributes [79] that were compared in pairs from an importance definition scale in relation to EVSE installation [80].The weights of comparison respected the scale of Saaty, as shown in Table 5.
The weights of the attributes were determined in four stages and the AHP technique was performed using the Easy AHP tool integrated with the QGIS software [81]: Establishing the hierarchical structure of the factors influence.
(ii) 2nd Stage.Assembling the judgment matrix, based on the AHP technique consisting of comparative notes between the attributes (from 1 to 9) indicated by the experts.
(iii) 3rd Stage.Determining the uniformity of the matrix applied to the normalization process.The eigenvector max (the highest value of the matrix) was adopted in order to verify the consistency index (CI), the judgment, and consistency ratio (CR), thus indicating the degree of randomness of the matrix [82] and checking the consistency of decisions, as shown in (1) and Table 6: and = where CI is the Consistency Index, Λmax is the highest matched array value, CR is the consistency ratio, N is the number of attributes, RI is the Random Index.
Although there are cases in the literature that consider CR > 0.1 because of the low knowledge of the parameters [83], in this project we consider that the RC reached the conditions of execution of the model, since only about 12% of cases CR was > 0.1 provoking a rediscussion of the project with the experts.
(iv) 4th Stage.Determination and analysis of the weights that are the basis for the MCDM method of linear combination weighted, according to the scenarios modeling: (a) Modeling for the L1 Scenario.For the L1 scenario, since these two attributes have the same degree of importance, a comparative matrix was not required.To refine the model results, a spatial grouping of the average income data by households was applied, considering only the regions with a minimum of 40 households (respecting the principle of equal US $ 2.2 or higher) and people 18 years old or older.
(b) Modeling for the L2 Scenario.In the L2 scenario, all attributes were confronted and judged according to the degree of inherent importance.Attributes shopping and workplace were characterized as strong and very strong importance.Table 7 shows the application of the weights according to the AHP technique.
(c) Modeling for the L3 Scenario.To ensure the consistency of the judgments and the degree of randomness of the matrix, the eigenvalue max (9.57) and the CI (0.072) and CR (0.05) indexes were calculated.Then, the value of the weights for each attribute was determined in order to meet the WLC method.
In the modeling for EVSE L3, the roads, avenues and corridors, shopping centers, and public transportation connection were characterized as strong and very strong importance.Table 8 shows the importance levels for the L3 scenario.The matrix was consistent and with a low degree of randomness, with the following values being reached: Λmax (9.24), CI (0.031), and CR (0.021).
Second Processing
Step.The weights for each attribute in the model were applied according to the given scenario.The WLC method was used in the linear combination of the weights generated by the comparison matrix in each attribute [84] and considering the three steps: (i) Weight Application.The WLC method was adopted for linear combination and matrix application.
The WLC approach is a widely used GIS-based decision rule technique [85].Its most common applications are land use analysis, site matching and selection, and resource assessment [60].In this work, the WLC method is implemented in the GIS environment with the support of the reclassification and map algebra tools.The AHP technique provided the weights of the attributes and the WLC performed the weights calculation in relation to the attributes in a georeferenced way.The calculation for the identification of the most suitable areas for the EVSE network was performed according to (2), : (ii) Grouping of Areas.To identify the areas with the greatest potential for EVSE network, the technique of Natural Break (Jenks) was adopted [58].
(iii) Subtraction of Restriction Areas.Inappropriate areas for the implementation of any type of infrastructure linked to road transport, i.e., water bodies, irregular settlements, environmental conservation units, etc.Therefore, all restriction areas were unified and subtracted from the global map [86].The term unsafe area was used to refer to subnormal agglomerates considered with a high degree of violence that endangers physical and property integrity, thus resulting in serious material and personal damage.Risk areas refer to the flooding areas that are vulnerable to major floods that can cause damage to materials and endanger personal safety.
Third Processing
Step.This stage highlights the appropriate areas by neighborhood, taking into account spatial and cartographic issues.Spatial intersection processing was used in the relation of the maps (municipality and neighborhoods) and the cartographic questions are characterized by mapping indicating the main locations for the installation of the EVSE network.Finally, the cartographic presentation is designed to support decision-makers in the areas with the highest potential areas (percentage) for EVSE and their overall geographic location.
Results
The considered scenarios account for 12,000 units of LDEVs by 2025 in BH, corresponding to a total of 1,200 EVSEs.The study revealed new attributes in this type of research, such as subnormal agglomerates-a patrimonial and physical risk in the region studied because it deals with areas with high violence rates, and flooding areas-in the case of high rainfall event.These findings were fundamental for this and future studies.
Survey Results.
Around 30% of the response rate was obtained from the 51 expert surveys.The sectors with the highest percentage of participation were taxi companies and EV associations.The sector with the lowest return was the industrial segment.The experts evaluated that, for the EVSE L2, the most important attributes were workplace, shopping centers, public transportation connections, and household income.When compared to EVSE L3, the household income and workplace attributes had higher weights.For EVSE L3, the best-evaluated attributes were roads, public transportation connections, shopping centers, and population density.When compared to EVSE L2, the roads, shopping centers, public transportation connections, population density, and geohazards were considered the most important attributes, revealing experts' preference for super-fast charging stations (L3), as shown in Figure 2.
The experts also revealed the further away the EVSE are from the violence zones such as subnormal agglomerates and risk zones such as flooding areas the better.The result of the analysis also revealed a wide standard deviation, especially in attributes as flooding areas and slopes.The most plausible explanation for this deviation is that attributes characteristics directly affect those who regularly live or access the region and these respondents usually give higher value to these attributes.On the other hand, the roads attribute presented the lowest standard deviation, which revealed a more balanced evaluation due to the fact that it shows a common and widespread demand in the diffusion process of the electric mobility infrastructure, as shown in Figure 3.
Regarding the GIS analysis, the optimum location of Level 1 installations of the EVSE network is largely affected by attributes like household income and population density.However, proximity of attributes, like shopping centers, public transportation connections, workplaces and roads, corridors and avenues, has positive impacts on Level 2 and Level 3 installations of the EVSE network, while attributes like flooding areas, slopes, and geohazards areas have negative impacts.
GIS Analysis Results of Recommended Area for EVSE L1.
The study showed that 65% of the recommended EVSE L1 installations were located in 10 neighborhoods located in the South-Central region, as shown in Figure 4.This is due to higher household income concentration and people aged 18 years or older in the targeted regions.
GIS Analysis Results of Recommended Area for EVSE L2.
The analysis revealed that the suggested areas for installation of the EVSE L2 network are contained in 45% of the BH neighborhoods, covering the South-Central and East regions (Figure 5) mainly due to the greater concentration of companies (workplace), shopping centers, and public transport connections, attributes that have been well evaluated by the specialists for this type of infrastructure.
GIS Analysis Results of Recommended Area for EVSE L3.
Around 36 neighborhoods of the municipality of BH hold about 40% of the location recommended areas for EVSE L3, as shown in Figure 6.These areas are distributed in the Northeast, Northwest, West, South-Central, and Pampulha regions.
They are located near expressways, major avenues, and highways such as the Celso Mello Azevedo ring road that connects the South-Central (economic pole) to the other regions of BH and neighboring municipalities, as well as the Cristiano Machado avenue, Dom Pedro Primeiro Avenues, access to Pampulha Airport and Confins International Airport, Contorno Avenue, among others.This region has some important shopping malls in BH, public transportation connections, and a high population density as well as access to major highways.
The distances between the extreme areas in the municipality of BH are short and compatible with the use of EVs.For example, the connection from extreme North to extreme South is around 37 km and from East to West it is less than 20 km.From the central area in the municipality to the farthest point (from extreme South to extreme North) the distance is close to 17 km.
Discussion
The identification of unsuitable areas for infrastructure installation in developing countries has been considered in other investigations.The study to analyze the best locations for new filling stations in Malaysia, Asia, Khahro indicated that 24% of the planned areas for EVSE installation were in inadequate areas [87].In the aforementioned study, the adaptations of the areas were carried out based on environmental risk factors and not as inadequate areas due to risks of public security.In the present study, this problem did not occur, since we excluded the risk areas from the recommended areas for EVSE installation.However, the problem is not completely solved because it is necessary to take into account the areas close to the risk areas; i.e., many access routes in the studied region cross long areas of subnormal agglomerate, which are not recommended for EVSE installation, due to the risks of violence during the loading time.Therefore, our recommendation is for the EVSE to be installed as far as possible from unsafe areas, in order to avoid serious occurrences with risks of physical integrity and damage.
Other studies proposed an assignment model to distribute charging infrastructure in Beijing, China, and to investigate the development of optimal EV charging station assignments in Seattle, USA.Hao and Chen [88,89] also did not address the issues of risk areas with a high rate of urban violence.Actually, almost all of the projects on the expansion of electric mobility are restricted to a few developed countries and public safety issues are focused on the safety of people and the vehicle with regard to the potential for battery accidents that can cause fires or explosions.However, the expansion of electric mobility in urban areas of some developing countries will probably require concern with other factors, such as unsafe urban areas highlighted in this study.
Conclusions, Limitations, and Recommendations
This Brazilian pioneering study aimed at identifying the optimal locations to implement EVSE in the municipality of BH and identify what attributes should be prioritized for the installation of an EVSE network in urban areas.Some attributes such as subnormal agglomerates and flooding areas were identified for the first time in this study and support the identification of the best locations for EVSE (L2 and L3) in BH, and in developing countries with substantial risk areas.Therefore, projects for the development of infrastructure for electric vehicles in regions with a high risk of violence and therefore subject to material and personal damage should include projects risk attributes such as those identified in this study.
In this context, the study reveals that EVSE L3 should not be located in public places, only in places that offer 24/7 personal security, as they do with gas stations currently in Brazil, and places as hotels, airports, shopping malls, etc.Besides that, the flooding areas may be obstacles to the expansion of its infrastructure network in Brazil.Similar regions around the world, especially in developing countries, are likely to present the same limitations for EVSE network expansion and require the development of business models capable of overcoming this issue.
Most of EVSE should be concentrated in small and specific areas of the municipality of BH.This can be a facilitating factor for optimizing the investment potential and maximizing the offer of user services by the expansion of the EV.The installation of an EVSE network will depend on government regulation and potential financial incentives for the users.Participation of the public sector and other stakeholders will likely create sustainable business models capable of attracting investments to EVSE.In this context, the involvement of utilities, automakers, and companies that own large fleet and various level of government federal, state, and municipal government) could promising start for the development of electric mobility and an incentive to support the EV infrastructure network.
The short distances that connect the main regions to the central region of the BH municipality may be favorable for the adoption of EV, as with a full charge the range of most EVs is much higher than the distances needed to cross the BH municipality (maximum distance of 37 km from extreme North to extreme South).The results of this research can contribute to better understand the diffusion of the charging infrastructure for EV and guide stakeholders and public policies in Brazil and other regions (not restricted to South America) with similar characteristics as BH in electric mobility projects.
Although it was not the objective of this study, it was possible to observe that the specification of the types of EVs during the vehicle registration process (up to now, the Brazilian authorities have not started this process) will be of great importance for the stakeholders and for future studies.This type of practice and information on various types of EVs might reveal significant differences in energy consumption and have an impact on the grid.
This study was limited due to the difficulty of locating specialists in electric mobility in Brazil, and further studies are still required to complement the current one.These include incorporating the EVSE financial aspect, evaluating the impacts of the EVSE network growth on the grid, incorporating the drivers' psychological aspects, and consumer attitudes towards e-mobility.
3. 2 .
Territorial Divisional.The geographical area considered was based on the territorial division of the municipality BH, based on the Brazilian Constitution of 1988 and the Brazilian Institute of Geography and Statistics or IBGE (Portuguese: Instituto Brasileiro de Geografia e Estatística).It considers 487 neighborhoods [26, 63].
Figure 2 :
Figure 2: Replies received from the specialists.
Figure 3 :
Figure 3: Average rating of expert rated attributes and standard deviation.
Penetration.Due partially to public policy support in some nations, the global number of EVs has almost doubled every year since 2010.In 2016, there
Table 3 :
Data source of the attributes in vector format.
Table
Number of attributes and random index.
Meaning of acronyms in Supplement 1. | 8,050 | 2018-12-19T00:00:00.000 | [
"Economics"
] |
Power laws in pressure-induced structural change of glasses
Many glasses exhibit fractional power law (FPL) between the mean atomic volume va and the first diffraction peak position q1, i.e. va∝q1−d\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{\mathrm{a}} \propto q_1^{ - d}$$\end{document} with d ≃ 2.5 deviating from the space dimension D = 3, under compression or composition change. What structural change causes such FPL and whether the FPL and d are universal remain controversial. Here our simulations show that the FPL holds in both two- and three-dimensional glasses under compression when the particle interaction has two length scales which can induce nonuniform local deformations. The exponent d is not universal, but varies linearly with the deformable part of soft particles. In particular, we reveal an unexpected crossover regime with d > D from crystal behavior (d = D) to glass behavior (d < D). The results are explained by two types of bond deformation. We further discover FPLs in real space from the radial distribution functions, which correspond to the FPLs in reciprocal space. A puzzle in metallic glass research is the existence of the fractional power law in reciprocal space, whilst its origin remains controversial. Zhang et al. show that nonuniform local deformations under compression induce this phenomenon and quantify the power law exponent at both two and three dimensions.
G lasses are amorphous solids and ubiquitous in our daily life and in industry, but the understanding of glasses remains a major challenge in science 1,2 . In particular, microscopic structural changes in response to mechanical deformation is poorly understood [3][4][5] . A well-known puzzle is the fractional power law (FPL) in the reciprocal space of many metallic glasses 3,6 , whose mechanism and generality remain controversial 3,[6][7][8][9] .
For crystals, the position of Bragg diffraction peaks is inversely proportional to the lattice plane distances in real space, i.e. q j ∝ 1/ a, where q j is the position of the jth peak of structure factor S(q) and a is the lattice constant. Therefore the volume per atom v a / a D / q ÀD 1 must hold for a D-dimensional crystal. Surprisingly, diffraction experiments for many metallic glasses show an FPL in three dimensions (3D), v a / q Àd with a fractional exponent d ≃ 2.5 < D = 3 under composition change 6,8 or compression 3,7,8 . Recently, power laws with large fluctuation of d were observed in glasses under compression and composition change, which raises questions about the generality of the FPL 9 . Here we summarize five open questions: (1) Does the FPL generally hold in glasses? (2) Which factors affect the value of d? (3) What is the origin of the FPL? The anomalous FPLs have been attributed to atomic-scale fractal packing 8 and mediumrange order 6 , but both explanations are derived from a single state, not from a series of states as the FPL arises. Moreover, further studies of metallic glasses did not reveal a fractal structure 9 . S(q 1 ) contains structural information spanning broad length scales in real space. It is therefore difficult to connect the FPL concerning q 1 in reciprocal space to certain structure changes in real space. (4) How does the FPL change from crystal behavior (d = D) to glass behavior (d < D)? It was not measureable before because glasses were usually produced from supercooled liquids instead of crystals 2 . (5) Does the FPL exist in two dimensions (2D)? This question has not been explored. Dimensionality strongly affects material properties and phase behavior. Recent studies showed that 2D and 3D glasses are fundamentally similar, but also differ in the local dynamics [10][11][12] . Here we try to answer these five questions by systematically changing the parameters in six systems in both 2D and 3D. To deepen our understanding of the FPLs, we measure not only the glass regime, but also the crossover to crystals. Besides the FPL in reciprocal space, other structural power laws have also been observed in real space, e.g. based on the distances between neighbors in a granular glass 13 and the correlations of structural order parameters in supercooled liquids 14 . Whether the FPLs in reciprocal space relate to certain power laws in real space has not been explored. Here, we discover a new set of FPLs in real space that correlates with the FPLs in reciprocal space.
Results
Six model systems. We perform simulations with three types of binary particles in 2D and 3D, i.e. a total of six systems: hard/hard spheres (Fig. 1a) in 2D (2DHH) and 3D (3DHH), spheres with the Weeks-Chandler-Andersen (WCA) potential 15 (Fig. 1c) in 2D (2DWCA) and 3D (3DWCA), and soft/hard spheres (Fig. 1e) in 2D (2DSH) and 3D (3DSH) 16 . The three types of pair interactions exhibit distinct deformation behavior, thus can help to identify which type of structural change gives rise to the FPL. The shoulder potential has been widely used to model metals, water, silica, micelles and colloids [17][18][19][20][21] . The mixtures of soft/hard particles can mimic materials whose components have different compressibilities such as alloy Ce 75 Al 25 with soft Ce and hard Al atoms 22 . We define the packing fraction ϕ as the volume fraction of hard spheres and the hard cores of soft spheres. Generally, when soft and hard particles are of the same size at low pressures, they can form crystals 16 . As the pressure increases, an increasing number of soft particles are compressed, resulting in finergrained polycrystals and eventually glasses ( Supplementary Fig. 1). Previously we resolved a sharp polycrystal-glass transition which distinguishes fine-grained polycrystals and glasses in 2DSH and 3DSH systems in ref. 16 Here we explore the crossover of the power law from crystal behavior (d = D) to glass behavior (d < D) for the first time. We systematically study the FPL by continuously tuning the fraction of soft particles η and the softness λ.
The systems contain N = 12,800 disks in a square box for the 2D case and 10,000 spheres in a cubic box for the 3D case. Each state is directly compressed from a low-density liquid (see Methods). Hence states under different pressures are uncorrelated and the structural FPLs are not related to affine or non-affine deformation. As a comparison, we also compressed the 2DSH system step by step and obtained the same FPLs. The step by step compression in 3DSH system yields similar FPLs in the crystal and glass regimes, but different behaviors in the crossover regime.
Power laws in 2D systems. We calculate the structure factor from the positions of particles, r. q 1 is measured from the Lorentzian fit of the first peak 23 . Figure 1b, d, f shows that v a / q Àd 1 holds in the 2DHH, 2DWCA and 2DSH glasses. d ≃ D = 2 for the 2DHH and 2DWCA systems (Fig. 1b, d), indicating a uniform deformation at all length scales like a crystal under compression. As hard disks cannot overlap, the uniform deformation in the 2DHH glass arises from squeezing the free volume of the gaps between particles. The uniform deformation in the 2DWCA glass arises from both the free-volume squeezing and the compression of WCA particles. As small and large WCA particles have the same softness characterized by the repulsive potential U(r)~r -12 , their size changes are proportional to each other, resulting in a uniform deformation. By contrast, the shoulder potential has two length scales which causes nonuniform deformation at the singleparticle length scale and gives rise to two distinct FPLs in crossover and glass regimes (Figs. 1f and 2).
Unlike binary HH and WCA systems, SH systems can form crystals at low pressures and glasses at high pressures. Figure 1f shows three regimes of v a / q Àd 1 in the 2DSH system. In the crystal regime, d = 1.95 ± 0.05 ≃ D = 2, as expected. In the glass regime, d = 1.36 < 2, similar to the FPLs with d < 3 in 3D metallic glasses 3,6 . We find that the crossover regime can also be fitted by an FPL with d = 4.6 (Fig. 1f).
The FPL in Fig. 1f is replotted in Fig. 2 as a function of packing fraction ϕ, in order to compare the three power-law regimes with the five regimes observed in refs. 16 : 1 polycrystals at ϕ < 0.66 featuring Hall-Petch behavior, i.e. the mechanical strength increases as the crystalline grains become finer; 2 ultrafine-grained polycrystals at 0.66 < ϕ < 0.70 featuring inverse-Hall-Petch behavior 24 ; 3 shadow glass at 0.70 < ϕ < 0.76 featuring strong dynamics 25 ; 4 low-density glass at 0.76 < ϕ < 0.80; and 5 high-density glass at ϕ > 0.80. The boundaries of these five regimes were identified from various structural, dynamic, mechanical and thermodynamic quantities 16 .
Here the FPL provides new features at their boundaries: the three power laws in Fig. 2 intersect at the boundary between the Hall-Petch and inverse-Hall-Petch regimes and the boundary between the shadow and low-density glasses. In addition, the minimum slope of q 1 (ϕ) (black star in Fig. 2) coincides with the boundary between the polycrystal and glass regimes identified via other methods 16 . These results generally hold in other 2DSH systems with different values of (η,λ). Hence the q 1 (ϕ) curve could provide empirical criteria to distinguish between Hall-Petch and inverse-Hall-Petch regimes, and between ultrafine-grained polycrystal and glass regimes, at least in 2DSH systems.
The FPL in the crossover corresponds to a regime with abnormally large compressibility (dashed curve in Fig. 2). The low-and high-density glasses in Fig. 2 have been observed 16,26 when the particle interaction has two length scales such as the square-shoulder potential. Here we find that low-and highdensity glasses have the same d = 1.36 (Fig. 2), which is consistent with the observation that both the low-and high-density metallic glasses of Ce 68 Al 10 Cu 20 Co 2 have the same FPL with d = 2.5 3 .
Fraction of soft shells governing the FPL. 2DSH systems with different values of η or λ similarly exhibit three power laws at the crystal, crossover and glass regimes as shown in Fig. 3a, b, respectively. The exponent d in the glass regime varies with η and λ ( Supplementary Fig. 2a, b). Interestingly, d decreases linearly with the area fraction of the soft deformable part in the total area of all the particles: X = η(λ 2 − 1)/λ 2 (Fig. 3c). X describes the amount of size mismatch available in the 2DSH system under compression, which determines the amount of defects that can be produced in the crystal and reflects the glass-forming ability of 2DSH crystals. We further measure three other systems, and all of their d values lie on the linear d(X) as shown in Fig. 3c. Therefore, we conclude that the soft deformable part governs d. A larger area fraction of the soft part can produce more nonuniform deformation under compression, thus d deviates more from D (Fig. 3c). Compression-induced FPLs have been measured in two types of metallic glasses based on La and Ce, e.g. La 62 Al 14 -Cu 11.7 Ag 2.3 Ni 5 Co 5 and Ce 68 Al 10 Cu 20 Co 2 . They both yield d ≃ 2.5 3,7 , indicating that they have similar fractions of soft compressible parts. In metallic glasses, Al, Cu, Ag, Ni and Co are known to be hard-sphere-like atoms, while La and Ce are much softer due to their localized electrons 22,27 . In fact, Ce can be described by the square-shoulder potential 17 .
U(r) U(r) U(r) U(r)
In(q 1 ) 1.78 Fig. 1. Red, green and blue lines in the log-log plot denote the fitted q 1 ∝ ϕ 1/d in crystal, crossover and glass regimes, respectively. The intersection of the three lines coincide with the boundary between Hall-Petch and inverse-Hall-Petch behaviors (red star) and the boundary between the shadow glass and normal glass regimes (blue star) 16 . The slope of q 1 (ϕ) reaches the minimum at ϕ = 0.70 (black star), coinciding with the polycrystal-glass boundary identified in ref. 16 The five regimes are identified in ref. 16 and can be roughly seen from the compressibility (dashed curve). NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-15583-4 ARTICLE NATURE COMMUNICATIONS | (2020) 11:2005 | https://doi.org/10.1038/s41467-020-15583-4 | www.nature.com/naturecommunications η = 0.5), which leads to phase separation 28 . Consequently, systems beyond these regimes can only form large-grained polycrystals instead of glasses even under the highest pressure 16 , and the corresponding FPL does not exist in the glass regime ( Supplementary Fig. 3a, b).
FPLs near random-close packing. At the high-pressure limit, almost all soft particles would be compressed so that the 2DSH system becomes a 2DHH system with the same (η,λ) at the random-close packing (RCP) point ϕ RCP 29 . This is confirmed in systems with various (η,λ) values. For example, the extrapolations of q 1 (ϕ) of 2DSH and 2DHH systems with the same (η, λ) = (0.5,1.3) intersect at ϕ = 0.850 ≃ ϕ RCP ≃ 0.848 of binary hard disks with diameter ratio 1.4 29 (Fig. 3d). To approach ϕ RCP , the rapid increase in q 1 (ϕ) (i.e. d < D) in the glass regime needs to be compensated by a slow increase in q 1 (ϕ) (i.e. d > D) in the crossover regime. More compressible parts (i.e. larger η or λ) create a broader crossover regime with a stronger deviation from the line of d = 2, hence a steeper q 1 (ϕ) (i.e. smaller d) in the glass regime is needed for the compensation as shown in Fig. 3a.
Power laws in 3D systems. Similar power laws are observed in 3D systems (Fig. 4), which further confirms that the FPL, d < D, requires two length scales in the potential. 3DHH and 3DWCA glasses exhibit the normal power laws with d = 3.0 = D (Fig. 4a, b) similar to their 2D counterparts (Fig. 1b, d). Similar to its 2D counterpart, the 3DSH system also exhibits the crystal regime with d = 3.0 = D, the crossover regime with d = 4.03 > D and the glass regime with d = 2.48 < D as shown in Fig. 4c. In contrast to the continuous v a (q 1 ) curve in the 2DSH system (Fig. 1f), v a (q 1 ) in the 3DSH system abruptly jumps at ϕ = 0.5 in Fig. 4c, coinciding with the crystal-glass transition point (Fig. 4d). This is in accordance with the observations in ref. 16 that the crystal-glass transition is like first order in 3D and more continuous in 2D. If the 3DSH system is compressed step by step, the crystal behavior of the power law extends to the crossover regime and exhibits a jump at the onset of the glass regime. Such protocol dependence in the crossover regime should be due to the first-order-like polycrystal-glass transition in 3D.
Theoretical explanation for the FPL. At thermal equilibrium, the Helmholtz free energy F = U − TS is minimized. U is the internal energy and S is the entropy related to the free volume 30 . For SH systems, when a SS or SH-bond is compressed (Fig. 5a), ΔU = U 0 and TΔS ∝ TA SS or TA SH . Note that the configurational entropy is neglected because we focus on the entropy change instead of entropy when a bond is compressed in a given glassy configuration. Apparently, the compressed area A SS > A SH as shown in Fig. 5a. Therefore, SS bonds will be compressed first as they reduce F more effectively. Consequently, we expect three regimes. At low pressures, the compressed volumes come from the gaps between particles. Few SS or SH bonds have been compressed so that the structure remains a crystal. At medium pressures, the volume change mainly arises from the compressed SS bonds, resulting in a more disordered structure. The nonuniform spatial distribution of the compressed volume ( Supplementary Fig. 4a) causes d to deviate from D. At high pressures, almost all SS bonds have been compressed so that further volume shrinkage is caused solely by the compression of SH bonds. These compressions do not occur randomly in space, but only in the previously uncompressed areas (Supplementary Fig. 4b). This compensates for the deviation from D in crossover regime, consistent with the result based on the RCP point in Fig. 3d. The three stages are confirmed in Fig. 5b for the 2D case and Fig. 5c for the 3D case. Interestingly, the three stages in Fig. 5b, c coincide well with the crystal, crossover and glass regimes in Figs. 2 and 4c, suggesting that the compressions of SS and SH bonds are responsible for the crossover and glass regimes, respectively. Figure 5b for the 2DSH system shows that the crossover regime is dominated by the compression of SS bonds, and the glass regime is dominated by the compression of SH bonds. Figure 5c for the 3DSH system shows an abrupt increase in the number of compressed SS bonds at ϕ = 0.5, which coincides well with the sharp first-order-like crystal-glass transition identified in Fig. 4d. In the crossover regime (0.50 < ϕ < 0.58), both SS and SH bonds are compressed (Fig. 5c), and their mixing effects result in a nearly power law in Fig. 4c. In the glass regime (ϕ > 0.58), the number of compressed SS bonds is saturated so that the deformation solely arises from the compression of SH bonds (Fig. 5c). Therefore, glass and crossover regimes have distinct FPLs.
As the crossover regime is dominated by the compression of SS bonds (Fig. 5b), systems with the same λ and different values of η should have the same d because the compression of one SS bond has the same impact on Δϕ and Δq 1 . This is confirmed in Fig. 3a. For systems with different values of λ, compressing a SS bond changes ϕ and q 1 differently. Hence, the power laws in the crossover regime show a different d, as confirmed in Fig. 3b. Systems with a larger η have more SS bonds which gives rise to a broader crossover regime (Fig. 3a). The large exponent, i.e. d > D, implies that q 1 (ϕ) increases much more slowly than it does in crystals. This can be understood as the compressed SS bonds reducing ϕ effectively, but affecting q 1 much less as q 1 is from the average of all bonds. Note that the compressed SS bonds at the highest pressure represent <20% of all bonds. In addition, a larger λ gives a larger A SS , i.e. more change in ϕ than q 1 , therefore yielding a larger d (Fig. 3b) in the crossover regime. For example, suppose that compressing one large-λ SS bond and two small-λ SS bonds results in the same Δϕ, but the structural deformation in the former case is more localized in real space and thus less effective in changing q 1 . As this effect gives d > D in the crossover regime, the glass regime should have d < D for the compensation to be able to reach the RCP point.
The above arguments about the SS-and SH-bond compressions without using any simulation results have qualitatively explained d > D in the crossover regime and d < D in the glass regime, and d(λ) behaviors. The simulation results in this section just provide consistency checks and are not necessary for the theoretical explanations. Beside the above approach, next we introduce another approach based on Eqs. 4 and 5 below. This second approach needs simulation results to fully explain the observations. Thus the first approach above provides a full qualitative explanation, while the second approach below is just a consistency check.
We further estimate the effects of SS and SH bonds on the FPL. The FPL indicates that 1/d is the slope of lnq 1 =lnφ.q 1 ¼ q 1 =q 0 1 , ϕ ¼ ϕ=ϕ 0 and q 0 1 and ϕ 0 are values of the initial single crystal (Fig. 3a, b). Under compression,q 1 is a function of the volume change. In 2DSH systems, the volume change arises from the squeezing of the free volume characterized byφ, and the compression of the SS and SH bonds. Note that the compressed volume from a SS or a SH-bond, i.e. A SS or A SH , is a constant at a fixed λ (Fig. 5a), and thus their numbers, n SS and n SH as functions ofφ, determine the amount of volume change. Consequently, q 1 ¼q 1 ðφ; n SS ; n SH Þ, and the FPL becomes The constant n SS and n SH in the first term describe the fixed numbers of SS and SH bonds. Hence the compression solely occurs from the free-volume change, which is similar to the uniform compression of SH crystals or HH glasses. Thus, The second and third terms denote the contributions from SS and SH bonds, respectively. lnq 1 is proportional to n SS and n SH in Supplementary Fig. 6a, and we find a similar relationship in Cu c Zr 1−c using the data in ref. 31 ( Supplementary Fig. 6b) Thus, the integration of Eq. 4 yields Fitting lnq 1 ðlnφÞ curves with Eq. 5 ( Supplementary Fig. 7) yields C SS < 0 and C SH > 0 (Fig. 5d), indicating that the compressed SS and SH bonds increase and decrease d relative to D, respectively. This is consistent with d > D in the crossover regime dominated by compressing SS bonds and d < D in the glasses regime dominated by compressing SH bonds (Figs. 3b and 5b). Moreover, C SS and C SH vanish at λ ≃ 1.14 ( Fig. 5d), indicating d → D as λ decreases toward 1.14. This is consistent with the fact that the 2DSH system cannot be compressed into a glass at λ ≤ 1.16 16 .
Power laws in real-space g(r). Power laws have been observed in real-space structures of amorphous states, e.g. the jth pair distance r j (ϕ) in a granular system 13 and correlation functions of structural order parameters in supercooled liquids 14 . These power laws cast important light on the disordered structures, but are not directly related to the radial distribution function where V is the volume. g(r) is usually derived from the Fourier transformation of the measured S(q) in scattering experiments 6 . Thus it has been used to explore the structural origin of the FPL in reciprocal space 6,8 . The FPL in reciprocal space has been attributed to the fractal structures at the length scale of the nearest neighbors, i.e. the first peak of g(r) 8 , but ref. 9 pointed out that the fractal structure is absent at the atomic scale. Ref. 6 suggested that the FPL arises from the medium-range order from the fit of the envelop of gðrÞ À 1 j j$ r Àγ expðÀr=ξÞ. However, this fit only captures the structure at a fixed ϕ rather than the structural change at a series of ϕ values as the way that the FPL in reciprocal space is derived.
Here we discover FPLs from the g(r) peaks of glasses with a series of ϕ values. For binary systems, the first peak of g(r) splits into three subpeaks because the first-layer neighbors have three typical separations corresponding to S-S, S-H and H-H bonds ( Supplementary Fig. 5c). These three subpeak positions cannot deviate much from their corresponding bond lengths, thus r 1 changes very little, resulting exceptionally large exponents in the power law ( Supplementary Fig. 8). The second-and third-layer peaks of g(r) split into more subpeaks which interferences with each other, thus their ambiguous peak positions are not measured. We focus on the positions of the unimodal peaks, i.e. fourth to eighth peaks for 2DSH glasses (Fig. 6e), third to seventh peaks for 2DHH glasses (Fig. 6a), and second to sixth peaks for 2DWCA glasses (Fig. 6c). We find that with d j = D for 2DHH and 2DWCA glasses (Fig. 6b, d) and d j < D for 2DSH glasses (Fig. 6f). Such a shift of the jth peak position, r j , demonstrates that the medium-range pair distance changes uniformly with ϕ for 2DHH and 2DWCA glasses and nonuniformly for 2DSH glasses, in accordance with the FPLs in reciprocal space. These results suggest that the power laws in real space and reciprocal space have the same structural origin. Whether d j from real space and d from reciprocal space have a quantitative relation is worth to explore in the future.
Discussion
From the simulations of the six model systems, we found answers to the five questions about the FPL in reciprocal space raised in the Introduction section as follows: Does the FPL generally hold in glasses? Yes, but the exponent d is not a universal constant. We observed the compressioninduced FPLs with d < D in 2DSH and 3DSH glasses, and power laws with d = D in 2DHH, 3DHH, 2DWCA and 3DWCA glasses.
Compression or composition change can produce the power law v a / q Àd 1 in our simulations and in the literature, but they have different impacts on the value of d. Therefore the data points with both pressure and composition change do not exhibit a good power law 9 .
Which factor affects the value of d? The answer was previously unclear because the responsible parameters were not systematically varied. We found that the compression-induced FPL with d ≠ D requires a mixture of soft and hard particles so that the local structure can be deformed nonuniformly. By systematically adjusting η and λ of 2DSH systems, we found that d is linearly governed by the fraction of the soft shells X (Fig. 3c).
For particles of different sizes but with the same softness (e.g. binary WCA spheres or binary hard spheres), the interparticle distances change uniformly under compression, resulting in the trivial d = D. This is consistent with the observed d ≃ 3.0 = D when Zr 46 Cu 54 is compressed 9 because both Zr and Cu atoms are like hard spheres 31,32 . Note that this does not conflict with d ≃ 2.3 < D = 3 in Zr x Cu 1−x metallic glasses when the mixing ratio x is varied 6 because changing x is analogous to compressing a SH system rather than an HH system.
The FPL was mainly observed in metallic glasses when the mixing ratio of different types of atoms was varied 6 . A few studies showed that compression can also induce the FPL with d ≃ 2.5 in La-or Ce-based metallic glasses, e.g. Ce 68 Al 10 Cu 20 Co 2 3,7 , but these coincidences at 2.5 do not mean that changing the pressure or changing the mixing ratio would have the same effect on the FPL. Ce atoms are much softer than other atoms 22,27 and can be described as spheres with a square-shoulder potential 17 because its 4f-electron orbit is localized at low pressure and delocalized at high pressure, resulting in atomic volume collapse 3,22 , thus the metallic glasses in ref. 3 are similar to our SH systems and exhibit the FPL. However, our simulation suggests that d would not be constant at 2.5 when metallic glasses with different fractions of soft atoms like Ce are compressed.
What is the origin of the FPL? We found that the FPL and its exponent d are determined by different types of volume changes. For HH systems, the compression arises solely from the squeezing of the free volume, i.e. the gaps between particles. Its d = D indicates that such a volume change is uniform. For WCA systems, the volume change arises from both the squeezing of the free volume and the compression of the particles. The latter is also uniform as both large and small WCA particles have similar softness, hence d = D in WCA systems. For 2DSH and 3DSH systems, the volume change arises from the squeezing of the free volume and the compression of SS and SH bonds. In the crystal regime, the uniform free-volume change dominates and thus d = D. In the crossover and glass regimes, the volume change is dominated by the compression of SS and SH bonds, respectively (Fig. 5). Compressing a large-λ SS bond is equivalent to compressing multiple small-λ SS bonds in changing Δϕ, but the former deformation is more localized in space which is less effective at changing S(q) at a small q 1 . Therefore, d is larger for a large-λ system in the crossover regime. When λ = 1, the SH system reduces to the HH system where d = D. Hence, d > D for λ > 1 in the crossover regime reflecting a nonuniform local structural change.
The volume change in the glass regime must occur in previously uncompressed local regions rather than at random positions. In other words, the structural change in the glass regime will compensate for the nonuniform structural built up in the crossover regime. Consequently, d > D in the crossover regime is accompanied by a d < D in the glass regime, in accordance with the same RCP structure of 2DSH and 2DHH systems. This explanation for d < D and d > D in different regimes does not need the assumption of any fractal structure. In fact, fractal structures with dimension d > D cannot exist in a Ddimensional space.
Note that the FPL is not related to affine or non-affine deformation because each solid state is directly compressed from a low-density liquid, i.e. structures at different values of ϕ are uncorrelated. This is also supported by the existence of the FPL in metallic glasses when their compositions change, which is not related to any non-affine deformation.
How does the FPL change from crystal behavior (d = D) to glass behavior (d < D)? We studied this question for the first time by creating a novel crystal-to-glass transformation. We found that the crossover regime between a crystal and a glass can also be fitted by a power law, but its d > D does not sit between the d values of the crystal and the glass. This is because the volume change in the crossover regime mainly comes from the compression of SS bonds, which is more effective at changing ϕ than changing q 1 as explained in the answer to question 2. We discovered that the onset of the crossover FPL regime coincides with the boundary between Hall-Petch and inverse-Hall-Petch regimes, while the minimum slope coincides with the polycrystalglass transition in 2DSH systems (Fig. 2). In 3DSH systems, however, the crystal regime with d = D terminates at the crystal-glass transition (Fig. 4c). These coincidences generally hold for systems with different values of η and λ, which cast new light on the poorly understood Hall-Petch-to-inverse-Hall-Petch transition 16,24 and the polycrystal-to-glass transition 16 .
Does the FPL exist in 2D? Yes, the FPL in 2D is similar to that in 3D and has been explained above. Low-dimensional systems are much softer because there are more long-wavelength fluctuations 33 and particles have fewer neighbors providing constraints. Consequently, the space dimension could strongly affect the nature of phase transition 33,34 . For example, 3D crystal melting is a first order phase transition, while 2D melting often exhibits two continuous transitions 33 . Similarly, here we found that the v a (q 1 ) curve at the crystal-to-glass transition is continuous in 2D (Fig. 1f), but makes an abrupt jump in 3D (Fig. 4c), which is consistent with the behaviors of other quantities in ref. 16 Besides the FPLs in reciprocal space, we discovered power laws about the medium-range g(r) peaks. Although g(r) has been intensively studied in liquids, crystals and glasses, the shift of the jth peak has rarely been explored because (1) the positions of medium-range peaks are difficult to measure accurately from the Fourier transform of S(q) in scattering experiments, and (2) the trivial relation, v a / 1=ϕ / r D j , is expected to hold. Surprisingly, we find that d j can deviate from D (Fig. 6f) when the pair interaction has two length scales. d j < D in SH systems and d j = D in HH and WCA systems. These real-space results are similar to those in reciprocal space, suggesting that they have the same structural origin.
The results bring new insights on material fabrication. For example, how particle interaction affects material properties is a key question in materials science, but the understanding is limited. For instance, soft solvent particles are empirically argued to be responsible for the elastic modulus of metallic glasses 35 . Here we found that two length scales in the pair potential result in d ≠ D, indicating that the soft particles play a key role in the FPL and the structural change in glasses. Our results predict that compressing metallic glasses composed of hard-sphere-like atoms will result in d = D, and the higher the fraction of soft atoms like Ce, the more d deviates from D. Fabricating ultrafine-grained polycrystals is another important challenge in materials science as they are unstable and tend to coalesce into larger grains 36 . We found that a large η, i.e. more soft particles and SS bonds, causes a broad crossover regime, which corresponds to a broader regime of ultrafine-grained polycrystals with inverse-Hall-Petch behavior and abnormally large compressibility. The fraction of the soft shells of soft particles determines d. 2DSH glasses can only form at 1 < d < 2. Beyond this range, the systems can only form largegrained polycrystals. These results provide guidance for fabricating ultrafine-grained polycrystals and glasses with different degrees of nonuniform deformation under compression.
Methods
Simulation methods. We performed Brownian dynamics simulations for 2DWCA and 3DWCA systems using LAMMPS 37 and event-driven molecular dynamics simulations 38 for the other four types of systems. All the simulations were performed under periodic boundary conditions in NVT ensembles. Samples were relaxed for long enough at each ϕ.
Simulations of soft/hard and hard/hard mixtures. 2DSH: The simulations contained N = 12,800 disks with the mixing ratio η = N S /N, where N S is the number of soft disks. The soft particles had square-shoulder potential (Fig. 1e) where σ and λσ are the diameters of the inner core and outer shell, respectively. σ serves as the length unit. U 0 is the height of the shoulder. The pair potential of the hard particles The packing fraction was calculated as where A is the area of the box. Particles were randomly distributed in the box at ϕ = 0.62, and then relaxed at T = 2.0 U 0 /k B for a time period of 10 5 t 0 and finally equilibrated at T = 0.133U 0 /k B for 10 5 is the amount of time a disk takes to move a distance σ, where m is the unit of mass. To facilitate the equilibration, initial velocities of particles were reassigned every 10 4 t 0 with a Gaussian distribution. Starting from ϕ = 0.62, crystals were compressed into higher packing fractions using the Lubachevsky-Stillinger algorithm 39 . All the results were measured at T = 0.133U 0 /k B , a low temperature used to thermalize the system 16 .
The results of the 2DSH system with (η, λ) = (0.5, 1.3) are shown in Figs. 1f, 2, 5b and 6c. Other values of (η,λ) were explored and some of them are shown in Fig. 3. η varies from 0.35 to 0.60 in Fig. 3a, in which range the system can be compressed to the glass state 16 . During compression, defects steadily accumulated through collapse of shoulders in soft particles (Fig. 5a), which caused the crystal to transform into glass ( Supplementary Fig. 1). As a nonequilibrium state, glass depends not only on state parameters such as temperature, density and pressure, but also on its fabrication history 2 . We compared two glass states compressed from a liquid (Fig. 3a, b) and from a crystal (Fig. 1f) at the same (η, λ) = (0.5, 1.3), and found similar FPLs: d = 1.36 in Fig. 1f and d = 1.35 in Fig. 3a, indicating that the FPL is insensitive to the glass formation pathway.
3DSH: The simulations were performed in a cubic box containing 5000 soft and 5000 hard spheres. The initial state was set to a fluid with ϕ = 0.3 and relaxed at T = 2.0U 0 /k B to obtain different configurations across trials. Then it was directly compressed into the target ϕ (Fig. 4d) and relaxed at T = 0.2U 0 /k B for 10 5 t 0 . The packing fraction is defined as ϕ ¼ N V πσ 3 6 ð1 À ηÞ λ 3 þ η 1 3 À Á : The system with (η, λ) = (0.4, 1.25) exhibits similar features at the crystal-glass transition to those of the 2DSH systems 16 .
The simulations of 2DHH and 3DHH systems are the same as those of 2DSH and 3DSH systems, except that the binary HH spheres cannot form crystals at a low ϕ. The time unit for HH systems t 0 ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi mσ 2 =ðk B TÞ p .
Simulations of WCA systems. WCA potential 15 where σ = 1.3, 1.15 and 1.0 for large-large, large-small and small-small particle interactions, respectively. U 0 = 1.0 is the unit of energy. WCA potential is a wellknown short-range repulsive potential which has often been used to model colloidal interactions 14,15,40 . At each ϕ, particles were randomly distributed in the box. After energy minimization using the FIRE algorithm 41 , the system was relaxed at T = 0.002ε/k B for 10 8 steps with the time step δt = 0.001. The packing fraction was calculated using Eq. 10 for the 2D case and Eq. 11 for the 3D case with the effective diameter 2 1/6 σ.
Structural identification. To characterize the local crystalline order of each particle, we used the modified orientational order parameter 16,42 . This parameter is more accurate than the conventional bond-orientational order parameter because each neighbor is properly weighed by the corresponding edge of the Voronoi cell 42 . For 2D systems, the modified orientational order parameter where θ jk is the orientational angle of the bond between particle j and its neighbor k. The Voronoi polygon 43 has N j edges with perimeter l tot , and the length of the edge between j and k is l jk . A higher jψ 6j j represents a higher crystalline order. Particles with three or more crystalline bonds were defined as crystalline, where a crystalline bond (between particles j and k) is one that satisfies jψ 6j Á ψ * 6k j > 0:6 16 . Two neighboring crystalline particles belong to the same grain if the difference between their orientational angles jArgðψ 6j Þ À Argðψ 6k Þj ≤ 5:0 . Noncrystalline particles and single isolated crystalline particles are defined as disordered 16 .
For 3D systems, the modified orientational order parameter where θ ij and ϕ ij are the spherical angles of the vector from particle i to its jth nearest neighbor. A j is the area of the Voronoi facet to the jth neighbor. A is the total surface area of the Voronoi cell. Y lm is a spherical harmonic function of degree l and order m. q l=6 ≤ 0.4 are disordered particles; and q l=6 > 0.4 are crystalline particles 44 . Crystalline particles with q l=4 > 0.143 are defined as having an facecentered cubic (FCC) structure and the rest are hexagonal close-packed (HCP) structures (Fig. 4d) 44 .
Data availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Code availability
The codes that are used to generate results in the paper are available from the corresponding author upon reasonable request. | 9,389.8 | 2020-04-24T00:00:00.000 | [
"Materials Science",
"Physics"
] |
Assembly-free genome comparison based on next-generation sequencing reads and variable length patterns
Background With the advent of Next-Generation Sequencing technologies (NGS), a large amount of short read data has been generated. If a reference genome is not available, the assembly of a template sequence is usually challenging because of repeats and the short length of reads. When NGS reads cannot be mapped onto a reference genome alignment-based methods are not applicable. However it is still possible to study the evolutionary relationship of unassembled genomes based on NGS data. Results We present a parameter-free alignment-free method, called Under2¯, based on variable-length patterns, for the direct comparison of sets of NGS reads. We define a similarity measure using variable-length patterns, as well as reverses and reverse-complements, along with their statistical and syntactical properties. We evaluate several alignment-free statistics on the comparison of NGS reads coming from simulated and real genomes. In almost all simulations our method Under2¯ outperforms all other statistics. The performance gain becomes more evident when real genomes are used. Conclusion The new alignment-free statistic is highly successful in discriminating related genomes based on NGS reads data. In almost all experiments, it outperforms traditional alignment-free statistics that are based on fixed length patterns.
Introduction
The comparison of sequences is fundamental for the analysis of many biological processes. The use of alignment tools like BLAST [1] to assess the degree of similarity between two sequences is a dominant approach. Alignment-based methods produce good results only if the biological sequences under investigation share a reliable alignment. However there are cases where traditional alignment based methods cannot be applied, for example, when the sequences being compared do not share any statistical significant alignment. This is the case when the sequences come from distant related organisms, or they are functionally related but not orthologous. Another drawback is that alignment methods are usually time consuming, thus they cannot be applied to large-scale sequence data produced by NGS technologies.
With the advent of NGS, a large amount of short read data has been generated. These data are used to study many biological problems, such as transcription factor binding sites identification, de novo sequencing, alternative splicing, etc. The first step of most studies is to map the reads onto known genomes. However, if a reference genome is not available, the assembly of a template sequence is usually challenging because there may be a large number of repeats within a genome and the short length of reads.
When the NGS reads cannot be mapped onto a reference genome alignment-based methods are not applicable.
Moreover the size of NGS data demands the use of very efficient algorithms. For these reasons the comparison of genomes based on the direct comparison of NGS reads has been investigated only recently using alignment-free methods [2].
The use of alignment-free methods for comparing sequences has proved useful in different applications. Some alignment-free measures use the patterns distribution to study evolutionary relationships among different organisms [3][4][5]. In [6], researchers have shown that the use of k-mers frequencies can improve the construction of phylogenetic trees traditionally based on a multiplesequence alignment, especially for distant related species. The efficiency of alignment-free measures also allows the reconstruction of phylogenies for whole genomes [4,7,8]. Several alignment-free methods have been devised for the detection of enhancers in ChIP-Seq data [9][10][11][12] and also of entropic profiles [13,14]. Another application is the classification of protein remotely related, which can be addressed with sophisticated word counting procedures [15,16]. For a comprehensive review of alignment-free measures and applications we refer the reader to [17].
To the best of our knowledge, so far only one group of researchers have compared sets of NGS reads using alignment-free measures based on k-mers counting [2]. Here we intend to follow the same approach by adapting our alignment-free pairwise dissimilarity, called U nder 2 [8], for the comparison of two sets of NGS reads. The current study differs from our previous studies [7,8] in the following aspects. First U nder 2 was originally developed to compare pairs of genomic sequences, here we extend it to compare pairs of reads set. Another important aspect is the way patterns are weighted in our similarity score, where we need to consider the expected number of occurrences of a pattern in a set of reads.
Almost all other methods are based on statistics of patterns with a fixed-length k, where the performance depends dramatically on the choice of the resolution k [4]. Finally, one the most important contributions is the use of reverse and reverse-complement patterns, as well as variable-length patterns to mimic the exchange of genetic material. In summary, in this paper we present a parameter-free alignment-free method, called Under 2 , based on variable-length patterns. We will define a similarity measure using variable-length patterns along with their statistical and syntactical properties, so that "uninformative" patterns will be discarded.
The paper is organized as follows. In the next section we review alignment-free methods and their applications. Then we present our contributions, the Under 2 statistic. In the result section we test the performance of several alignment-free measures with both synthetic and real NGS data. In the last section, the conclusions and future work are discussed.
Previous work
Historically, one of the first papers that introduces an alignment-free method is due to Blaisdell in 1986 [18]. He proposed a statistic called D 2 , to study the correlation between two sequences. The D 2 similarity is the correlation between the number of occurrences of all k-mers appearing in two sequences. Let A and B be two sequences from an alphabet Σ. The value A w is the number of times w appears in A, with possible overlaps. Then the D 2 statistic is: This is the inner product of the word vectors A w and B w , each one representing the number of occurrences of words of length k, i.e. k-mers, in the two sequences. However, it was shown by Lippert et al. [19] that the D 2 statistic can be biased by the stochastic noise in each sequence. To address this issue two other popular statistics, called D S 2 and D * 2 , were introduced respectively in [11] and [20]. This measures were proposed to standardize the D 2 in the following manner. Let à w = A w − (n − k + 1) * p w and B w = B w − (n − k + 1) * p w where p w is the probability of w under the null model and n is the length of the strings A and B. Then D S 2 and D * 2 can be defined as follows: and, These similarity measures respond to the need of normalization of D 2 . All these statistics have been studied by Reinert et al. [20] and Wan et al. [21] for the detection of regulatory sequences. In [2] the authors extend these statistics for genome comparison based on NGS data, and define d 2 , d s 2 and d * 2 . The major difficulties are the random sampling of reads from the genomes and the consideration of double strands of the genome. They tested the performance of d2, d s 2 and d * 2 on synthetic and real datasets. In particular, the common motif model, introduced by [20], is used to mimic the exchange of genetic material between two genomes, and MetaSim [22] is used to simulate the sequencing. We describe the common motif model in the next sections and propose a more realistic formulation. In this paper we will follow the same experimental setup of [2] and compare our results with these statistics.
Under 2 an assembly-free genome comparison based on next-generation sequencing reads and variable length patterns In this section we describe our parameter-free alignmentfree dissimilarity measure, called Under 2 , which extends our previous work [8] to the case of NGS reads. The dissimilarity Under 2 is based on two concepts: irredundancy and underlying positioning.
Let's consider two sets of reads R 1 and R 2 that are sampled from two genomes. Every set is composed by M reads of length b in the alphabet Σ = {A, C, G, T}. We say that a pattern in Σ * is shared between the two sets of reads if it appears at least once in some read of R 1 and once in some other read of R 2 . The notion of irredundancy is meant to remove the redundant patterns, i.e. those patterns that do not convey extra information for the similarity measure. The second driving principle is the fact that, in previous approaches, every position of a read contributes a multiple number of times to the final score.
In the following we address these two issues separately. The goal is to build a similarity measure between the two sets of reads R 1 and R 2 using all exact patterns of any length, Σ * , that are shared between the two sets.
Removing redundant patterns
One can easily show that most sequences share an unusually large number of common patterns that do not convey extra information about the input. To keep the article self-contained, here we summarize the basic facts already proved in [16] and extend the notion of irredundant common pattern to the case of two sets of reads. If the occurrence of a pattern in a read completely overlaps with the occurrence of another longer pattern, we say that the occurrence of the first pattern is covered by the second one.
Definition 1 (Irredundant/Redundant common patterns) A pattern w is irredundant if and only if at least an occurrence of w in R 1 or R 2 is not covered by other patterns. A pattern that does not satisfy this condition is called a redundant common pattern.
We observe again that the set of irredundant common patterns I R 1 ,R 2 is a subset of the well-known linear set of maximal patterns [23]; therefore the number of irredundant common patterns is bounded by A simple algorithm that can discover all such patterns has already been described in [8] and it employs a generalized suffix tree of two sequences. To extend this algorithm to the new input R 1 and R 2 , it is sufficient to use the two sets of reads, while maintaining separated the occurrences that belong to the two sets. The construction of the generalized suffix tree and the subsequent extraction of the irredundant common patterns can be completed in time and space linear in the size of sequences [8]. In summary, the notion of irredundancy is useful for removing noninformative patterns, and thus for drastically reducing the number of candidates to be analyzed to estimate the sequence similarity between R 1 and R 2 .
Selecting underlying patterns
The basic idea behind our approach is that a position on the sequences should contribute only once to the final similarity. Traditionally alignment-free statistics fail to comply with this simple rule. In fact, every position, apart from the borders, belongs to k different k-mers and thus contributes k times to the similarity.
In previous works on whole-genome comparison, to solve this problem we used the notions of pattern priority and of underlying pattern [8]. The pattern priority rule is mainly based on the idea of selecting, for each position, those patterns that represent the largest number of matching sites between sequences, and thus that are more likely to be conserved patterns. Here we recall the definition of pattern priority and of underlying pattern from [8], and adapt these concepts to the new settings.
Let's consider the set of irredundant common patterns I R 1 ,R 2 as input. Given two patterns w and w′, we say that w has priority over w′, denoted w w′, if and only if either |w| > |w′|, or |w| = |w′| and w is less likely to appear in the sequences than w′, or w and w′ have the same length and probability to appear, but the first occurrence of w appears before the first occurrence of w′. We say that an occurrence l of w is tied to an occurrence l′ of another pattern w′, if these occurrences (partially) overlap to each other, [l, l + |w| − 1] ∩ [l′, l′ + |w′| − 1]) ≠ ∅, and w′ w. Otherwise, we say that l is untied from l′.
Definition 2 (Underlying patterns) A set of patterns U R 1 ,R 2 ⊆ I R 1 ,R 2 is said to be the Underlying set of {R 1 , R 2 } if and only if: (i) every pattern w in U R 1 ,R 2 , called underlying pattern, has at least one occurrence in both sets of reads that is untied from all the untied occurrences of other patterns in U R 1 ,R 2 \w , and (ii) there does not exist a pattern w ∈ I R 1 ,R 2 \U R 1 ,R 2 such that w has at least two untied occurrences, one per set of reads, from all the untied occurrences of patterns in U R 1 ,R 2 .
The objective of this definition is to select the most important patterns in I R 1 ,R 2 for each location of the reads in the two sets, according to the pattern priority rule. If a pattern w is selected, we filter out all occurrences of patterns with less priority than w that lay on the untied locations of w, in a simple combinatorial fashion. The complete procedure to discover the set U R 1 ,R 2 can be found in [8]. Here below we give an overview of the algorithm.
Underlying pattern extraction (Input: R 1 , R 2 ; Output: Compute the set of Irredundant common patterns Rank all patterns in I R 1 ,R 2 using the pattern priority rule.
for Select the top pattern, w, from I R 1 ,R 2 : do if Check in Γ if w has at least one untied occurrence per sequence that is not covered by some other patterns already in U R 1 ,R 2 then Add w to U R 1 ,R 2 and update the location vector, Γ, in which w appears as untied.
else Discard w. end if end for An auxiliary vector Γ, of length L, is used to represent all locations of R 1 and R 2 . For a pattern w in I R 1 ,R 2 , we can check whether its occurrences are tied to other patterns by looking at the vector Γ. If some untied occurrences are found, then we can add the new underlying pattern w to U R 1 ,R 2 , and update the vector Γ accordingly using all the untied occurrences of w. In total the extraction of all underlying patterns, using this scheme, takes O(L 2 ) time. A more advanced algorithm with a better complexity, O(L log L log log L) time and O(L) space, can be found in [8].
Building the Under 2 similarity measure Our similarity is inspired by the Average Common Subword approach (ACS) [24], where the scores of common patterns found are averaged over the length of sequences. Here we follow the same approach, but, instead of counting all common patterns, we use just the untied occurrences of the underlying patterns, which by definition do not overlap [8]. We can note that the set of underlying patterns U R 1 ,R 2 is not symmetric, in general U R 1 ,R 2 = U R 2 ,R 1 . Thus, in order to build a symmetric measure, we need to consider both sets.
In ACS the contribution of each position is given by the length of the pattern covering that position. In our approach we use instead the ratio of the number of occurrences for an underlying pattern w, and the expected number of occurrences for that pattern. Let's define occ w as the number of occurrences of w, and untied 1 w as the number of untied occurrences of w in R 1 . First we compute the score: Recalling that the untied occurrences do not overlap with each other, we notice that the term |w| * untied 1 w counts the positions where w appears without over-lapping any other pattern. For each such position we sum the score occ w E[occ w ] , where E[occ w ] is the expected number of occurrences. Note that the expectation of this ratio is exactly 1. This sum is then averaged over the length of the first sequence under examination, R 1 . This score is large when the two sequences are similar, therefore we take its inverse. Then, since the total number of occurrences of an underlying pattern w present in R 1 is expected to logarithmically increase with the length of R 2 , we consider the measure log 4 (|s2|)/Score(s1, s2), where a base-4 logarithm is used to represent the four DNA bases.
To center the formula, such that it goes to zero when R 1 = R 2 , we subtract the term log 4 |R 1 |. If R 1 = R 2 there will be just one underlying pattern that is equal to the sequence itself. In this case, Score(R 1 , R 1 ) will be 1 and the term log 4 |R 1 | makes sure that Under 2 (R 1 , R 1 ) = 0 . These observations are implemented in the general formula of Under 2 (R 1 , R 2 ) .
Finally, to correct the asymmetry, our similarity measure called Under 2 is the average of the two statistics Under 2 (R 1 , R 2 ) and Under 2 (R 2 , R 1 ) .
An important aspect in this formula is the computation of the expected number of occurrences of a pattern w. A Markov model usually outperforms the Bernoulli model on biological sequences. In our case the length of reads is relatively short and thus, to avoid overfitting, we will rely on a first order Markov model. In summary, the expectation is computed as E[occ w ] = p w M (b −|w|+1), where pw is the probability of w using the Markov model, M is the number of reads and (b − |w| + 1) are the possible occurrences of w. Finally, we extend our approach to account for untied occurrences that are present in the reverse, complement, and reverse-complement of each sequence, in order to simulate the DNA strand and the evolution of sequences. For more details about this extension, we refer to [8].
Experimental results on synthetic and real data
To compare the performance of Under 2 and all d-type statistics proposed in [2], we performed several experiments using both simulated and real data. The common motif model revised We start from a background sequence which can be either synthetic or a real genomic reference, we call such sequence negative to indicate that no correlation exists between any two of them. For each negative sequence we created a positive one using three different correlation models. The first is the Common Motif (CM ) model introduced in [20]. In the Common Motif model a pattern of length five is inserted at position j with probability l while the background is left unchanged with probability 1 − l, we chose the same pattern and the same length used in [20,2]. In the CM model the pattern inserted is always the same. The second model we adopted is the Simple Multiple Motifs (SMM ), in this model five patterns with length varying from four to six bases are considered. Note that the five patterns are all different now, moreover we consider also their reverse complement in this model. For each position j a pattern is inserted with probability l, the pattern to be inserted is chosen so that all five patterns and their reverse complements are inserted with the same probability. The last model introduced is the Full Multiple Motifs (FMM ) model which is a slight variation of SMM where, for each pattern, not only the reverse complement is considered, but also the reverse is inserted. The introduction of these two models SMM and FMM try to mimic the exchange of genetic material between genomes, where regions of variable lengths as well as reverse and reverse complements are important.
Experimental setup
We test the performance of the different statistics by assessing if sequences from the positive set score higher than those from the negative set. We compute the similarity scores for all pairs of sequences in the positive set and all pairs of sequences in the negative set. Then we sort all scores in one combined list. We consider as positive predictive value (PPV) the percentage of pairs from the positive set that are in the top half of this list, PPV of 1 means perfect separation between positive and negative sequences, while a PPV of 0.5 means no statistical power.
Following the experimental setup of [2], during all the experiments we maintained a constant pattern intensity l = 0.001. For each sequence (either positive or negative) we used MetaSim (http://ab.inf.uni-tuebingen.de/software/ metasim/) [22] to generate M reads with length b = 200 and with standard deviation 0 (i.e. all reads have length exactly b ), in order to obtain an overall coverage g = 5.
We will use these parameters for most of the experiments. Except where indicated, exact (i.e. no errors) sequencing has been simulated, when errors are considered, the MetaSim preset for 454 model is used with all parameters set to their default values.
For each experimental setup we compute the average score over five runs of Under 2 and of all d-type statistics (http://www-rcf.usc.edu/~fsun/Programs/D2_NGS/ D2NGSmain.html). During all simulations, parameters of different algorithms have been maintained fixed, more specifically we used k = 5 for d-type statistics because this is the best value measured in [2] as well as the best value we observed in a set of preliminary tests.
Simulations with random background
In this first test we use random sequences as background. Although real datasets are always more desirable than simulations, the use of random sequences is very useful to establish the behavior of alignment-free statistics. Moreover random background sequences can be used to formally prove the statistical power of the d-type statistics (see [20,21]).
To simulate data we used the same setup of [2], we considere two different i.i.d. models for negative sequences, uniform background with p A = p C = p G = p T = 1/4 and GC-rich background with p A = p T = 1/6, p C = p G = 1/3, we measure the PPV of 40 sequences, 20 positive and 20 negative, as the sequence length N varies from 500 to 10000 bases.
Results for the CM model are shown in Figure 1 with both uniform background (a) and GC-rich background (b). Using this setup we observed no significant improvement as N grows (recall of 0.5 means no statistical power). All measures are almost aligned around PPV of 0.5 and only for higher values of N (4000 or more) d * 2 and d s 2 show a slight improvement of their performance. This is explained by the fact that the number of patterns inserted grows with length of the sequence, thus longer sequences from the positive set will have more chance to obtain an higher similarity score. However all methods perform poorly on this dataset, as can be seen from the scale of Figure 1. In general d-type statistics need longer sequences or an higher pattern intensity l to improve their predictive power.
In Figures 2 and 3 are shown results for the SMM and FMM models, respectively, with uniform (a) and GC-rich (b) backgrounds. The introduction of multiple motifs does not lead to significant performance improvements for d-type statistics, even if these statistics consider also the reverse complement. On the other hand we see a slight improvement of Under 2 for the SMM model and a significant improvement for the FMM model, this is due to the fact that introducing the reverse complement (SMM) and also the reverse (FMM) gives better results as the Under 2 statistic explicitly considers them.
By comparing subfigures (a) and (b) of all Figures 1, 2 and 3, we can note that changing the background from uniform to GC-rich produces worse PPV values. However such effect becomes significant only for small values of N and when the FMM model is used, while all d-type statistics and all other cases of Under 2 are almost immune from such effect, probably because performance in these cases are already poor. Finally in Figure 3(c) we double the coverage, g = 10. If we compare this plot with Figure 3(a) we can note a moderate improvement, especially for longer sequences. Thus, for random backgrounds, increasing the coverage will produce a small performance improvement.
Simulations with Drosophila genome
To assess the performance in a more realistic scenario in this section we use as background real genomic sequences from Drosophila. We first downloaded all the intergenic sequences of the Drosophila genome from FlyBase (http://flybase.org, dmel-all-intergenic-r5.49. fasta) and then we created the negative backgrounds by picking at random 10 sequences for each length varying from 1000 to 10000. We then generated positive sequences using the foreground models CM and FMM described above. To test the impact of sequencing error, we also performed a set of experiments using the 454 error model provided by MetaSim [22] with the FMM foreground, all results are shown in Figure 4 and Figure 5(b).
We observed a consistent trend among all the experiments with Under 2 always outperforming d-type statistics. Our measure, in fact, always gives better PPVs at all the tested lengths and for all models. As we introduced sequencing errors results degrade, however this effect is more relevant for short sequences where errors become more important and their effect are, therefore, more Comin and Schimd BMC Bioinformatics 2014, 15(Suppl 9):S1 http://www.biomedcentral.com/1471-2105/15/S9/S1 visible while at higher lengths the impact of sequencing errors become less significant.
Starting from this latest and more realistic setup, i.e. using Drosophila genome as background for the FMM model with 454 sequencing errors, we further evaluate how the different parameters affect the performance. Thus we will compare the next plots with Figure 4(c) that has been obtained with the following parameters (g = 5, l = 0.001 and b = 200). In Figure 5 we report the PPV values while changing only one parameter at a time. If we double the coverage (g = 10), subfigure (a), the recall values do not improve; only with random backgrounds we see a small improvement (see Figure 3 (c)). If we increase the probability to insert a pattern (l) in the FMM model, subfigure (b), as expected, all statistics improve and Under 2 quickly converges to 1. Finally the use of shorter reads (b = 100), subfigure (c), does not degrade the recall rates of Under 2 that remains around 0.7.
Phylogeny of genomes based on NGS data
In this section we test the ability of alignment-free statistics on the reconstruction of whole-genome phylogenies of different organisms. We first selected 12 prokaryotic organisms among the species in [24] for DNA phylogenomic inference. The organisms come from both the major prokaryotic domains: Archaea, 6 organisms (Accession No. BA000002, AE000782, AE009439, AE009441, AL096836, AE000520), and Bacteria, 6 organisms (Accession No. AE013218, AL111168, AE002160, AM884176, AE016828, L42023). The reference taxonomy is interred using the 16S rDNA Comin and Schimd BMC Bioinformatics 2014, 15(Suppl 9):S1 http://www.biomedcentral.com/1471-2105/15/S9/S1 sequences and the multiple alignment of these sequences available from the Ribosomal Database Project [25]. Then we perform a maximum likelihood estimation on the aligned set of sequences using Dnaml from PHYLIP [26] in order to compute a reference tree.
We simulate the sequencing process with MetaSim following the same setup as above and then we compute the distance matrices using all statistics. From these distance matrices we derive the taxonomies with the PHYLIP [26] software using neighbor joining (NJ) and the unweighted pair group method with arithmetic mean (UPGMA). We compare the resulting trees with the reference taxonomy using the Robinson and Foulds (R-F) distance. For two unrooted binary trees with n ≥ 3 leaves, the R-F score is in the range [0, 2n − 6]. A score equal to 0 means that the two trees are isomorphic, while 2n − 6 means that all non-trivial bipartitions are different.
The R-F distance between the reference taxonomy and the resulting phylogenetic trees, for all statistics and the two reconstruction methods, are summarized in Table 1.
In general Under 2 outperforms all d-type statistics obtaining the lower value with both reconstruction methods NJ and UPGMA. We can also observe that d S 2 and d * 2 obtain comparable results and, in some cases the former outperforms the latter confirming a similar observation in [2]. This latter experiment confirms that Under 2 is able to detect the genetic signal between unassembled NGS data.
Conclusion and future work
In this paper we introduced a parameter-free alignmentfree method called Under 2 that is designed around the use of variable-length words combined with specific statistical and syntactical properties. This alignment-free statistic was used to compare sets of NGS reads, in order to detect the evolutionary relationship of unassembled genomes. We evaluate the performance of several alignment-free methods on both synthetic and real data. In almost all simulations our method Under 2 outperforms all other statistics. The performance gain becomes more evident when real genomes are used. As a future direction of investigation, we will try to create a linear time linear space alignment-free measure based also on read quality values. | 6,962.2 | 2014-09-01T00:00:00.000 | [
"Biology",
"Computer Science",
"Environmental Science"
] |
Ageism and sexism amongst young computer scientists
. A study was undertaken with 189 young computer science students to assess whether as future developers of technologies for older people, they have ageist and sexist attitudes about people as users of technology. They were shown a picture of either a young or old woman or man and asked to assess the likelihood that this person would use a desktop computer, laptop computer and a smartphone, and their level of expertise in each of these technologies. The results showed that the students did have negative perceptions of the older people in comparison to young people. They although thought that women were less expert with the technologies than men, although there was not difference in the likelihood of them using the technology. However, there was no evidence of a “double standard” of older women being perceived particularly negatively.
1! Introduction
To create technologies that are useful, usable and acceptable to older users, developers need to be able to understand and empathise with the needs and wishes of those users. Yet it is well-known that young people tend to have negative attitudes and beliefs about older people. There has been a considerable amount of research exploring different parameters of these attitudes [e.g. 4] and attitudes by different types of young people [7,8,13,14], particularly those who will interact with older people in their professional lives such as doctors, nurses, and social workers [e.g. 3, 6]. Given that there is an increasing imbalance towards women in cohorts of older people, it is also relevant that there appears to be a Òdouble standardÓ in attitudes and beliefs about older people, with older women being more negatively viewed than older men [9,12]. In response to these issues, there has been interesting research on how to overcome such negative attitudes and beliefs [e.g. 1, 2, 5].
Very little work on the attitudes and beliefs about older people amongst those creating technologies for older people has been conducted. This work explores the attitudes of ageism and sexism, and the potential double standard of these two parameters in a group of young professionals who in the future might be asked to develop technologies for older people Ð young computer scientists at the beginning of their professional education. It builds on preliminary work [11] by expanding on the sample of young people participating, which allows for more detailed and robust analyses of both ageism and sexism.
2! Method
This study investigated the perceptions of young university students studying computer science of younger and older men and women as users and experts of smartphones and related technologies.
Two classes of first year computer science students at the University of York in the United Kingdom completed a very short survey for the study as part of one of their courses. Students who completed the survey were entered into a prize draw for five Amazon gift vouchers worth £5 (approximately USD 7.50) each.
The survey comprised a photograph of either an old or young man or woman (see Figure 1). Eight different versions of the survey were created, each with a different photograph. Four of the photographs were of older people, four were of younger people. Photographs were chosen carefully so that the person looked to be in their 70s for the older people, and in their late 20s/early 30s for the younger people (so a little older than the target respondents for the survey, but people they would still consider young). Within each group two images were of women and two were of men. All the photographs were chosen to be close up shots of a person reading a book. All the photographs were copyright free images from the Internet. The survey asked the following nine questions about the person in the photograph.
Firstly, three questions about the age of the person and old age in genera: How old do you think the person is? Would you call this person old? What is the minimum age you would think of someone as old?
Three questions about the person's use of technology: How likely do you think it is that this person uses a desktop computer regularly (rated on a scale from 1 = not at all likely to 7 = very likely)? How likely do you think it is that this person uses a laptop computer regularly (same rating as above)? How likely do you think it is that this person uses a smartphone regularly (same rating as above)?
Three questions about the person's expertise with technology: How expert do you think this person would be with a desktop computer/(rated on a scale from 1 = not at all expert to 7 = very expert)? How expert do you think this person would be with a laptop computer (same rating as above)? How expert do you think this person would be with a smartphone (same rating as above)?
Finally, respondents were asked their age and gender.
189 students completed the survey, 162 (85.7%) were men, 24 (12.7%) were women and 3 (1.6%) preferred not to identify their gender. The imbalance between women and men respondents unfortunately reflects the strong male bias in our undergraduate computer science community.
Because of the small number of women, no analyses could be attempted on differences due to the gender of the respondents, which would have been interesting to investigate. Respondents ages ranged from 18 to 28 years, with a median age of 18 years.
3! Results
In response to the question on when old age begins, on average respondents estimated that old age begins at 53.2 years (SD: 11.95), with a very wide range of answers, from 18 to 78 years. However, somewhat less than half the respondents (40.9%) felt that old age begins between 60 and 65 years, which are the typical ages for retirement and also those used in demographics and aging research [2,15]. In response to the likelihood that the people in the photos would use a desktop computer/laptop computer/smartphone regularly, a three way multivariate analysis of variance was conducted: Device (desktop/laptop/smartphone) x Age of person in the photograph (Young or Old) x Gender of person in the photograph (woman or man) This showed a significant main effect for device (F (2, 360) = 48.95, p < .000) with smartphone being rated as the most likely to be used, followed by laptop, with desktop the least likely to be used. There was also a main effect for Age (F (1, 185) = 427.98, p < .000) with young people rated more likely to use all the devices than older people (mean young people: 5.47; mean older people: 2.74). There was no main effect for Gender (F (1, 57) = 3.02, n.s.). There was no significant interaction between Age and Gender (which might suggest the double standard in ageism) (F (1, 185) = 0.64, n.s.).
The results for the expertise questions were similar to those for the likelihood of use question, with one interesting difference in relation to gender. The three way multivariate analysis of variance showed a significant main effect for device (F (2, 370) = 28.12, p < .000) with smartphone being rated as the device with which people with have the most expertise, followed by laptop, and desktop the device with which people would have the least expertise. There was also a significant main effect for Age (F (1, 185) = 266.88, p < .000) with young people rated more likely to use all the devices than older people (mean young people: 4.75; mean older people: 2.44). In this instance, there was a significant main effect for Gender (F (1, 185) = 6.81, p < 0.01), with women seen as less expert than men. There was no significant interaction between Age and Gender (which might suggest the double standard in ageism) (F (1, 185) = 1.31, n.s.).
4! Discussion and Conclusions
This paper reported on the results of an investigation into the perceptions of older people as users of technology, particularly desktop computers, laptop computers and smartphones by young, predominantly male, British computer science students. The results showed that the students perceived older people as both less likely to use these technologies and less expert in using them. However, there was less evidence of sexism, with no significant differences in the likelihood of women and men as users of technology, although there were significant differences in the perception of expertise, with women being seen as less expert in the technologies than men. However, there was no evidence of a double standard in ageism, in which older women are perceived less positively than older men. It was interesting that for younger people, they were seen as most likely to use and be more expert in smartphones in comparison to laptop computers and least likely to use and be expert in desktop computers. This reflects the move away from desktop machines to mobile devices and computing.
These results agree with numerous previous studies which have shown that young people hold negative attitudes and beliefs about older people (see Introduction). While the uptake of computing technologies by older people is still a lower that of younger people, older people in the UK are currently the fastest group adopting mobile technologies, especially smartphones and tablet computers [10]. Indeed the usage of portable devices such as laptop or tablet computers amongst older people has grown, in 2016 43% of 65 to 74 year olds now use a laptop or netbook (20% of those 75 and older), 31% use a tablet computer (15% of those 75 and older) and 83% use a mobile or smartphone (50% of those 75 and older) [10]. Undoubtedly these figures will continue to grow as the "baby boomer" generation of those born after the World War II ages. And with the decreasing number of younger people to care for them in old age, they will rely much more on technology than previous generations of older people. Thus, it is particularly important that the younger generations of computer scientists appreciate that older people are users of computing technologies. Clearly awareness of the issues around older computer users is needed. | 2,357.6 | 2018-07-11T00:00:00.000 | [
"Computer Science",
"Sociology"
] |
ion production from collisions between antiprotons and excited positronium: cross sections calculations in the framework of the GBAR experiment
. In the framework of the gravitational behaviour of antihydrogen at rest (GBAR) experiment, cross sections for the successive formation of ¯ H and ¯ H + from collisions between positronium (Ps) and antiprotons ( ¯ p) have been computed in the range 0–30 keV ¯ p energy, using the continuum distorted wave-final state theoretical model in its three-body and four-body formulations. The effect of the electronic correlations in ¯ H + on the total cross sections of ¯ H + production has been studied using three different wave functions for H − (the matter equivalent of ¯ H + ). Ps excited states up to n p = 3, as well as ¯ H excited states up to n h = 4, have been investigated. The results suggest that the production of ¯ H + can be efficiently enhanced by using either a fraction of Ps(2p) and a 2 keV ( ¯ p) beam or a fraction Ps(3d) and antiprotons with kinetic energy below 1 keV.
Introduction
The formation of positronium (Ps) in collisions between a positron and an atomic hydrogen target (equation (1)) has already been widely studied for it is the prototype of a three-body charge exchange reaction, where the three particles involved are distinguishable. However, even more than providing a testing ground for atomic collision theories, this reaction, and more precisely its charge conjugated inverse (equation (2)), stirred interest for antimatter experiment very early: e + + H(n h , l h , m h ) → Ps(n p , l p , m p ) + p, Ps(n p , l p , m p ) +p → e − +H(n h , l h , m h ).
Indeed, in the 1980s, a sub-GeV beam of antiprotons was available at CERN's LEAR facility and the possibility of efficiently producing antihydrogen (H) atoms using ground state Ps had already been discussed [1]. The importance of using low-energy antiprotons was stressed; a note added by the authors of [1] also suggested investigating the production of excited states ofH. These cross sections were computed, for instance by Igarashi et al [2] using a hyperspherical coupled-channel calculation and by Mitroy [3] using the unitarized Born approximation. In the latter, the production ofH states up to n h = 7 was considered for Ps energies between 0 and 4 eV and, furthermore, for Ps being itself excited (from n p = 1 to 4). Mitroy thus showed that high-n h antihydrogen levels provide the main contribution to the totalH formation cross section. This was later confirmed by a more accurate close coupling (CC) calculation [4]. Mitroy also pointed out the interest of having the Ps excited to a state n p = 3 or 4, since these states lead to the highest cross sections below 1 eV centre of mass energy. Around the same period, experimental values became available for both direct and reverse reactions of (1): (i) Weber et al [5] and Zhou et al [6] performed scattering experiments of positrons on, respectively, a mixture of H/H 2 and H 2 ; they deduced from their measurements the total cross section for Ps formation from ground state hydrogen (in the range 0 to ≃ 200 eV positron energy). Their experimental measurements are in good agreement with each other and with available two-centre CC calculations [7][8][9]. The maximum of the cross section is found to be about 3.5 πa 2 0 around 1 Ryd positron energy; and (ii) three experimental values of hydrogen formation from protons and ground state Ps are also available in the range 11-16 keV (proton energy) [10]. The CC calculations of Mitroy and Ryzhikh [4] give lower values however, they are almost within the error bars.
If reaction (1) is a fundamental three-body charge exchange reaction, then the reaction of Ps formation from collisions between positrons and H − ions (equation (3)) would be the four-body equivalent. However, the literature is much less abundant on that process, certainly because of its complexity and the extra care required to describe correctly the highly correlated system that is H − . Usually, the main motivation in studying equation (3) lies in astrophysics, where it is supposed to be a major contribution of the 511 keV annihilation line observed: Ps(n p , l p , m p ) +H(n h , l h , m h ) → e − +H + .
Straton and Drachman used the Coulomb Born approximation (CBA) with two wave functions for H − to compute the cross sections of equation (3) at four different positron energies (0.1, 0.5, 1 and 100 eV) when Ps and H are in their ground states [16]. Chaudhuri [17] highlighted the importance of the choice of the H − wave function and demonstrated the influence of the correlation description on the total cross section above 50 eV positron energy. Biswas [18] used the two-channel exchange coupled-channel theory to compute the cross section of the reverse process studied in [16], but took a plane wave for the exit channel and thus did not take into account the long range Coulomb interaction between e + and H − . This resulted in a total cross section going to zero close to the threshold when a finite value is expected. The most complete studies currently available are those of McAlinden et al [19] and Roy and Sinha [20], who both considered the formation of H − from H(1s) + Ps(1s-2p), using respectively a coupled pseudostates approach and a Coulomb modified eikonal approximation (CMEA). The both found the predominance of Ps(2p) at a low energy. Currently, no experimental result for this four-body reaction can be found in the literature. So far, the two sets of equations have not been studied as an ensemble of two consecutive reactions. Taken as a whole, they describe the production of H − ions from protons interacting with a gas of Ps or, from the antimatter point of view, the formation ofH + from antiprotons and Ps. The GBAR experiment precisely relies on these two processes to obtainH + ions.
GBAR, which stands for gravitational behaviour of antihydrogen at rest, will be a future experiment at CERN's antiproton decelerator [21,22]. The aim of the experiment is to perform a free fall of an antihydrogen atom in order to make a direct measurement ofḡ, the acceleration constant of antimatter on Earth. To observe this free fall and to reduce the uncertainty due to the initial velocity of the antiatom, the latter has to be cooled to a few neV (about 10 µK). The key idea of the GBAR experiment is to useH + ions that can then be trapped and sympathetically cooled in an ultracold cloud of Be + . Using this regular technique of cold atoms science, the required energy can be reached. The production of theH + ions therefore has to be optimized. In particular, for equation (4), a classical energetic argument suggests that the reaction could be nearly resonant close to the threshold (which is 4 meV antihydrogen energy) in the case n h = 1 and n p = 3. To our knowledge, cross sections for the four-body reaction are not available for Ps in a state n p = 3.
The main objective of this work is thus to provide a complete set of cross sections for the second reaction upon which an estimation ofH + production can be made for GBAR. Since both reactions (2) and (4) are of interest for the experiment, it was also decided to compute cross sections for the first reaction, with the same theoretical model, at the same level of approximation. The model chosen is called CDW-final state (CDW-FS) and has both a threebody and a four-body formulation. CDW-FS for light particles was first introduced by Fojón et al [23] to study Ps formation from positron capture by light hydrogen-like ions. Asymptotic states are described by Coulomb wave functions giving the exact boundary conditions (hence the relation to CDW methods). In the prior formulation of the theory, if the target is charged in the entrance channel, extra care must be taken so that the perturbative potential remains short ranged. The model also includes distortions in the final channel due to the Coulomb field of the residual target; this special treatment gives its name to the theory. On the same basis, this model has later been adapted to four-body reactions, in particular the case of positron capture by metastable helium [24]. More details on CDW-FS can be found in the mentioned publications [23,24].
However convenient CDW-FS is, it holds several drawbacks inherent to this family of perturbative theories. Firstly, its usual validity domain is at intermediate and high energies and it is probably not valid towards the threshold, energy region which is of primary importance for GBAR. Secondly, it has been demonstrated that the post/prior discrepancy of CDW theories increases towards low energies [25]. Despite the lack of reliability close to the threshold energies, useful informations can be extracted from the relative behaviours of the computed cross sections. Also, the use of this theoretical framework allows us to investigate in full detail, for the first time, the role played by the electron correlations of H − in four-body scattering processes (equation (4)). Furthermore, a comparison with available experimental (in the case of the three-body reaction) and theoretical results should give hints on the validity of CDW-FS at low energy.
In the following section, the analytical expressions of the cross sections calculated within the framework of CDW-FS theory and using a partial wave technique are detailed for very general cases (any state of hydrogen and Ps). Particular cases of l h = 0 and/or l p = 0 are included in appendices. In the next section, the results of the numerical computation are then presented and the importance of the correlations in H − is discussed. Differential cross sections are not considered. The cross sections have been computed for reactions (1) and (3) but, assuming invariance by charge conjugation and micro-reversibility, they are related to cross sections of reactions (2) and (4) by a simple kinetic factor. Throughout this paper, atomic units have been used, unless otherwise specified.
Three-body reaction
2.1.1. General case: l h , l p 0. The capture of an electron with Ps formation in e + -H collisions is considered (reaction (1)). Figure 1 describes the coordinates used in this section; for further discussion and details about the CDW-FS model of this reaction and on the choice of the coordinates, see [23]. The wave function corresponding to the entrance channel is where is the wave function of the electron in the hydrogen atom in the (n h l h m h ) state and F (+) k α (R) is the Coulomb wave function of the positron in the continuum of the hydrogen target. The final state is described by the wave function where ψ n p l p m p (ρ) =R n p l p (ρ)Y l p m p (ρ) is the wave function of the electron bound in the Ps atom, formed in the (n p l p m p ) state, F (−) k − (r) and F (−) k + (R) are, respectively, the Coulomb wave functions of the outgoing electron and the positron in the continuum of the residual proton of charge Z T = 1. The Coulomb wave functions write as with α + = (Z T − 1) µ α k α = 0 since the charge Z T − 1 of the hydrogen target is 0 (in which case, the Coulomb wave function actually reduces to a plane wave) and N α+ = Ŵ(1 + iα + ) exp − π 2 α + . k α is the wavevector of the reduced positron in the entrance channel and µ α is the reduced mass of the positron-target system (µ α = m(M+m) M+2m ≃ m); with β + = Z T µ β k + and N β+ = Ŵ(1 − iβ + ) exp − π 2 β + . k + is the wavevector of the reduced positron in the exit channel and µ β is the reduced mass of the Ps-residual target system (µ β = 2m×M M+2m ≃ 2m); with β − = Z T µ β k − and N β− = Ŵ(1 + iβ − ) exp + π 2 β − . k − is the wavevector of the reduced electron in the exit channel. k + ≃ k − ≃ k β µ β , where k β is the wavevector of the reduced Ps in the exit channel. Even if the exit channel of the reaction (1) is not Coulombic as it contains two species with only one having the overall charge, the CDW-FS model includes distortions in the final channel which are related to the Coulomb continuum states of the positron and the electron (of the Ps atom) in the field of the residual target (the proton). Therefore, the continuum wave functions (8) and (9) do not reduce to plane waves. If no distortions are included in the final channel, the CBA is obtained. This situation is similar to the one described in [1,26]. By artificially fixing the charges β − and β + to zero, we have got back almost the same results, which are quoted in [1]. The latter will be discussed in section 3.1.2.
In order to compute the transition matrix element, the following partial wave expansion of the Coulomb wave functions has been used: where the functions F l (kr ) are the Coulomb radial functions with the Sommerfeld parameters η = β ± or α + and δ l are the usual Coulomb phase shifts δ l = arg Ŵ(l + 1 + iη). The spherical harmonic function Y l p m p (ρ) has been treated using the development [27] where 0 λ L and −λ µ λ. The perturbative potential in the entrance channel (initial state) is The transition matrix element is then with In the above expression, the variable u is defined as ρ = √ r 2 + R 2 + 2r Ru andl indicateŝ l = 2l + 1. P l ′ is the Legendre polynomial of degree l ′ , coming from the multipole expansion The CDW-FS total cross section for Ps(n p , l p ) formation from H(n h , l h ) (reaction (1)) is given by These expressions can be simplified in the cases when either l h or l p , or both, are taken to be 0; details can be found in appendix A.
Four-body reaction
2.2.1. General case: l p , l h 0. We develop here the very general case where both l h and l p are different from 0 (the particular cases when l h = 0, l p = 0 and l h = l p = 0 are described in appendix B) and the wave function for H − is of the form where r 12 is defined as r 12 = |r 1 − r 2 |. This wave function has been treated using a partial wave expansion, which can be written as follows: with where r 12 = r 2 1 + r 2 2 + 2r 1 r 2 u. Three different wave functions have been used for modelling H − . The first one is the 'uncorrelated' Chandrasekhar wave function (labelled UC in the following) [28]. This is a simple and convenient wave function which takes into account the radial correlations between the two electrons of H − but no angular correlations. It has been used in several papers already cited and is usually considered as a sufficient level of description for where a uc = 1.0392, b uc = 0.2831 and the normalization constant N 1 = 0.3948. The binding energy of H − computed with this function is E H − = −0.5133, which has to be compared with the exact theoretical value, −0.5277 [29]. This wave function can be modified by introducing g = 0, so that now it takes into account angular correlations. This 'correlated' Chandrasekhar wave function CC α (r 1 , r 2 ) is defined by [28] where a cc = 1.0748, b cc = 0.4776, D = 0.3121 and the normalization constant N 2 = 0.3942. The binding energy is then E H − = −0.5259. Note that a misprint turned the sign of D into negative in several papers (in [17], for instance) whereas it is positive in the original Chandrasekhar paper, as it should be. The last wave function chosen to describe H − is the Le Sech wave function (labelled LS in the following) [30]. LS α (r 1 , r 2 ) decomposes as follows: where α = 0.05, β = 0.04, γ = 0.50, λ = 0.57 and the normalization constant N 3 = 0.03615 [31]. The binding energy calculated with the LS wave function is E H − = −0.5270. This wave function has been emphasized as a very accurate description of H − [32]. Table 1 shows the contribution of the different angular components in the H − wave function Coordinates used for reaction (3). and the convergence of the normalization when using the partial wave expansion (18); in this table, a l t is defined by a l t = (4π) 2 dr 1 dr 2 r 2 1 r 2 2 |h l t (r 1 , r 2 )| 2 . Figure 2 describes the coordinates used for the four-body reaction, similarly to [24] and derived from [33]. The wave function of the initial state is with the Coulomb wave function F (+) k α (R) describing the positron in the continuum of H − (see equation (7)). k α is the wavevector of the reduced positron in the entrance channel with µ α = m(M+2m) M+3m ≃ m the reduced mass of the positron-H − target system. The wave function for the final state is where ψ β (ρ i )F (−) k − (r i ) describes electron i captured in the outgoing Ps while electron j remains in the residual hydrogen atom (ϕ n h l h m h (r j )); F (−) k + (R) is the Coulomb wave function of the outgoing positron (see equation (8)). k β is the wavevector of the reduced Ps in the exit channel; we also define µ β = 2m(M+m) M+3m ≃ 2m, the reduced mass of the Ps-residual target system. As defined in the previous section, k + Again, a partial wave expansion of the Coulomb wave functions has been used (see equation (10)). The Sommerfeld parameters are we have β ± = 0: therefore, it should be noted that the Coulomb wave functions for the electron and the positron in the exit channel are actually plane waves. The Coulomb phase shifts δ l are defined as in the previous section. The 1 S symmetry of the initial state of H − imposes the symmetry in the final state, hence the choice of expression (22) for ξ (−) β . The chosen short-range perturbative potential is and the four-body CDW-FS matrix element T (−) αβ reads as with αβ can be thus written as a sum of two terms, t cap and t exc . After lengthy calculations, the former term may be expressed as where Similarly the excitation transition matrix element is given by with r < (respectively r > ) is defined as min(r 2 , R) (respectively max(r 2 , R)). The total cross section for a given state (n h , l h ) of H and a given state (n p , l p ) of Ps (reaction (3)) is given by with
Uncorrelated Chandrasekhar wave function.
Because of the absence of angular correlations in the uncorrelated Chandrasekhar wave function, further simplifications can be applied to the transition matrix element. The H − wave function can be treated slightly differently. Indeed, one notices that the UC wave function can also be written as Now, in T − αβ , terms depending on r 1 and r 2 are fully separable and the transition matrix element can be written as where a uc is shortened into a and b uc in b. t i j describes the capture into the Ps of the electron initially in state i in H − while, simultaneously, the electron initially in state j is excited to a final state H(n h , l h , m h ). Since both electrons in H − are indistinguishable, T − αβ is the coherent sum of two terms. We have where According to this, the matrix element t i j may be expressed as The matrix elements t i j cap and t i j exc may be obtained by replacing V i j in equation (34) with V i j cap and V i j exc , respectively. The term t i j cap mostly describes the capture of the electron in state i in H − into the Ps; the process is pondered by the overlap between the wave function of the electron in the hydrogen atom and the wave function of the H − electron initially in state j. Due to the scalar product n h l h m h | j , it is worthwhile noticing that the term of 'capture' is equal to zero for l h = 0. The term t i j exc encloses the excitation of the electron that remains bound to the proton, from state j in H − to state (n h , l h , m h ) in the hydrogen atom. Indeed, K j n h l h m h (R) can be written as The total cross section with the uncorrelated Chandrasekhar wave function can be written as If m h or l h are different from zero, most of these terms are null. The expression of the pure capture terms is Pure excitation terms can be written as follows: Finally, the cross terms can be expressed with the previously introduced quantities. For instance
Three-body continuum distorted wave-final state (CDW-FS).
Cross sections for reaction (1) have been computed for 6 states of the Ps atom, from Ps(1s) to Ps(3d) and for 13 states of hydrogen, from H(1s) to H(5d) (and up to H(5g) for ground-state Ps). Higher states have not been investigated yet since the calculation for each (n p , l p , m p )-(n h , l h , m h ) pair is very time-consuming. To obtain the cross sections for the reverse reaction (2), the following kinetic transformation is applied, assuming microreversibility and invariance by charge conjugation [1]: production of s-state antihydrogen is always the lowest contribution, whereas the formation of states with non-zero angular momentum quantum numbers gets more and more favoured as l h increases. For the highest l h state, this hierarchy can be slightly perturbed towards the threshold and towards the intermediate energy region, as can be seen in figure 6 for theH(4f).
In the prospect of the GBAR experiment for which the highest production rate of antihydrogen atoms is sought, figure 7 compares the different states of Ps, when all the contributions ofH states are summed from (1s) to (5d). The maximum ofH production occurs around 6 keV antiproton energy when the Ps is in its ground state and, because of the threshold constraints for n h > 1,H formation from Ps(1s) appears to be completely negligible below 5 keV. For excited states of Ps, the production of antihydrogen atoms peaks in the region between 2 and 3 keV. The highest cross section is obtained for Ps(2p) at 2 keV: this maximum dominates the others by at least a factor of 2.
Comparison with experimental data.
As already explained in the introduction, cross sections for reaction (2) can only be compared to one experiment performed by Merrison [10]. The solid lines are our CDW-FS results and the dotted lines are the results from Mitroy and Ryzhikh [4]. In light grey, the production of hydrogen in its ground state; in darker grey, the sum of the hydrogen production cross sections up to state H(3d). The solid black line corresponds to the CDW-FS results taking into account the hydrogen states up to H(5d) whereas the dotted black line from [4] is an estimation of the total hydrogen production using the 1 n 3 scaling for the states above (3d). Also represented: the CBA results for H(1s) production (light grey triangles) compared to the cross section given in [1] (light grey circles).
et al [10] for ground state Ps in the energy range 11-15 keV. In these inclusive measurements, no distinction between the hydrogen states produced could be done. The comparison, shown in figure 8 will thus take place between these experimental data and the sum of our cross sections over the (anti)hydrogen states. Despite a slight over-estimation from the summed cross sections, the theoretical calculations and the experimental data are in rather good agreement. To give another perspective, the CC (13,8) and CC (28,3) calculations by Mitroy and Ryzhikh [4] are also included in figure 8. As mentioned in the introduction, these calculations and the experimental data are in good agreement, except for the value at 11.3 keV. In that region, the CDW-FS results seem to better reproduce the behaviour of the total hydrogen production. On the other hand, CDW-FS largely underestimates the production of ground state hydrogen, which nonetheless remains small compared to the production of excited states. In the absence of other experimental results for reaction (2) (or its matter counterpart), in particular in the case of excited Ps, this positive comparison validates the use of CDW-FS to address the needs of the GBAR experiment.
The predictions of the CBA model (obtained by setting β − = β + = 0) are also depicted in figure 8, along with the cross section given in [1]. As expected, the two models exhibit very similar results.
Four-body CDW-FS.
Uncorrelated Chandrasekhar wave function The cross sections for reaction (3) were first computed using the uncorrelated Chandrasekhar wave function. Ps in states (1s)-(3d) were (3) and (4), assuming microreversibility and invariance by charge conjugation, are presented in figures 9-11, for energies between 0 and 30 keV antihydrogen energy and the contribution of the antihydrogen excited states have been summed over l h . As an example of the cross sections dependence in the orbital quantum number l h , more detailed results are shown in figure 13 for Table 2. Thresholds (keV) for reaction (4) whenH is in its ground state, given for each H − (H + ) wave function used in this work.
The general behaviour of the four-body reaction cross sections is a dramatic increase of theH + production towards the thresholds (see table 2) when the antihydrogen is in its ground state. This had already been noted by McAlinden et al (who gave a 1 E law to estimate the cross section of reaction (3) at very low energies) and Roy et al for Ps(1s) to (2p) [19,20]; it is now also demonstrated for n p = 3. The other tendency, over all the energy range considered, is a shift of the cross sections towards lower values when n h increases. This decrease of the cross sections with n h is amplified as l p goes up. Two cases can be distinguished: l p = 0 and l p = 0. Indeed, for s-states of Ps, values of the cross sections forH + formation fromH(1s) are very similar: they are all of the order of 10 πa 2 0 at the threshold and decrease with the impact energy, until, around 10-15 keV, they become comparable to the cross sections whenH is in the n h = 2 states in the entrance channel. For the l p = 0 states of Ps, the channel where the antihydrogen is in its ground state is always dominant; below 5 keV, these cross sections are one or two orders of magnitude higher than the ones with s-state Ps, as can be seen in figure 12.
The cross sections presented in figure 13 are representative of the behaviour for n h 3 and l p > 0. TheH + ions are preferentially produced from s-state and p-state antihydrogen atoms in close competition and then, to a lower extent, from d-states. For s-state antihydrogen atoms, the cross sections always exhibit the same structure with a maximum in the region between 10 and 15 keV. In the case of s-state Ps (l p = 0), not presented here,H + production will be favoured for l h = 0, but when n h = 4, it is now the p-state antihydrogen channel which dominates the one with s-state antihydrogen. For any state of Ps, whenH is in a state n h = 2 in the entrance channel, it isH(2s) that leads preferentially toH + .
Using equation (39) given for the UC wave function, it is possible to investigate the respective contributions of the capture and the excitation terms to the total cross section by taking either the t i j exc or t i j cap terms equal to zero. This is shown in figure 14(a) in the case of Ps(2p) andH(4s). The maximum observed in the cross section is explained by the preponderance on the capture process at these energies. The behaviour of the capture cross section is typical and the energy value at the maximum (11 keV) corresponds, as expected, to a projectile velocity twice larger than the velocity of the positron in the n p = 2 level of the Ps atom. The excitation is largely dominant below 5 keV. The cross terms, not presented here, are completely negligible above 10 keV. Similarly, the contributions of the processes 'ab' and 'ba' can be investigated by taking respectively t ba or t ab equal to zero. The results are presented in figure 14(b) and show that the major contribution to the total cross section ofH + is the 'ba' process, which corresponds to the de-excitation of the positron initially in the antihydrogen atom towards the lower level a ofH + while, simultaneously, the positron in the Ps is captured in the outer level b ofH + .
Correlated Chandrasekhar and Le Sech wave functions.
From the previous results with the uncorrelated Chandrasekhar wave function, cases of interest for GBAR have been selected to be investigated with the correlated Chandrasekhar wave function and the LS wave function. Since the formation ofH + ions from excited states of antihydrogen atoms is negligible, onlyH(1s) and (2s) have been considered. Other cases can of course be calculated using the formulas given in section 2.2 and in appendix B but at the cost of a long computational time. The results are presented in figures 15-17, where they are also compared to the ones obtained with the uncorrelated Chandrasekhar wave function. In the case of the LS wave function, figure 18 compares, for each state of Ps investigated, the cross sections ofH + production whenH is in its ground state in the entrance channel. The first observation is that the angular correlations taken into account in the CC and LS wave functions have an important effect on the total cross sections in all the energy ranges investigated and thus cannot be overlooked as a small correction. However, there is little difference between the CC and the LS wave functions: forH(1s) in the entrance channel, the CC and LS wave functions give the same results within 1% below 10 keV. The most notable difference is observed forH(2s) above, roughly, 5 keV; below, the cross sections obtained with the CC and the LS wave functions converge towards the threshold. This means that the collisional model is not sensitive to the level of description of these angular correlations at low energy, close to the reaction thresholds. The observed effect of the angular correlations is, when H is in its ground state, to increase the cross section for ground state Ps by almost a factor of 2, to slightly increase the cross section above 3 keV for Ps(2s) and (3s) (and slightly decrease them below 3 keV), to decrease it for Ps(2p) and (3p) below, roughly, 10 keV and finally to largely decrease that cross section for Ps(3d), losing up to one order of magnitude close to the threshold. ForH(2s) in the entrance channel, the use of the correlated wave functions leads to a decrease of the total cross sections compared to the UC results. The predominance of thē H(1s) channel below 20 keV is confirmed. The expected nearly resonant behaviour for n p = 3 [20] is for the CMEA results of Roy and Sinha [20] and the pseudostate approach results of McAlinden et al [19] correspond to the (green) line labelled [19]; the (blue) line with label [34] for Ps(1s) is the calculation at the threshold done by Blackwood et al [34].
is indeed observed, but, in the prospect of the GBAR experiment, does not lead to a very sharp increase of theH + production, unless using Ps(3p) or (3d) well below 2 keV antiproton energy. However, as can be remarked in figure 18,H(1s) with the (3p) or, above all, the (2p) states of Ps dominate all the other processes below 6 keV. Above 10 keV, Ps(1s) is the dominant channel.
Comparison with available theoretical results.
Since no experimental results are yet available for the four-body reaction, only a comparison with other theoretical models can be undertaken. The coupled pseudo-states computations of McAlinden et al, using an approached wave function for H − , and the CMEA calculations of Roy and Sinha, who chose the uncorrelated Chandrasekhar wave function, are thus compared to our four-body CDW-FS results with both uncorrelated Chandrasekhar and LS wave functions. So far, others authors kept the hydrogen atom in its ground state. The CC calculation at the threshold for the Ps(1s) + H(1s) channel performed by Blackwood et al [34] is also included. Figure 19 details the cases of Ps(1s), Ps(2s) and Ps(2p).
In most of the cases, although the general behaviour of the cross sections is similar, the CDW-FS results are in disagreement with the other theories' calculations, giving a much higher cross section. The notable exception is for Ps(2s) below 5 keV, where all the models seem to agree towards the threshold. Otherwise, the discrepancy can be higher than one order of magnitude. It could be expected, since CDW-FS, like the CMEA model used by Roy and Sinha, is not a low energy theory and is subdued to the post/prior discrepancy (which is emphasized towards the low energy region). In the medium energy region above 20 keV, both CDW-FS and CMEA should be valid but still disagree. The main difference between the two models is the treatment of the asymptotic states, which is exact in the case of CDW-FS and could be a reasonable explanation. However, the latter argument cannot be used when comparing to the CC calculation, which is intended to be accurate in the low energy region. Nonetheless, the model developed by McAlinden et al is itself not free from approximations: in particular, the wave function they used for H − (a split shell function involving one electron in the H(1s) state and the other in a linear combination of s-orbitals) gives a binding energy of −0.513, similar to the uncorrelated Chandrasekhar wave function. These observations stress that our results should be handled with particular care when used to predict experimental behaviours, either for GBAR or for other future projects.
Consequences for GBAR
As has been already underlined, no quantitative conclusions should be drawn. Concerning the energy of the antiprotons, two regions have been identified: below 2 keV and between 6 and 7 keV. In the 6 keV region, the use of Ps(1s) only (and this is the state in which Ps is produced) appears to be sufficient to produce sequentiallyH andH + ; theH production can be enhanced using a fraction of Ps excited in the (3d) state, whereas the second reaction can be slightly enhanced by a fraction of Ps(2p). Below 2 keV, Ps(1s) is almost of no use (below thresholds or too low cross sections), with the exception of the case whenH is in the (2s) state. Then, the best solution is to excite as much as possible the Ps to the (2p) state; the alternative is Ps(3p) or (3d) around 1 keV. Ps(3d) has the advantage to be a longer-lived state compared to Ps(2p) and (3p). To illustrate this discussion, figure 20 presents a rough approximation of the cross sections for the two reactions combined (using the cross sections obtained with the LS wave function). They have been calculated using equation (47). It is assumed that the proportions of Ps in the ground state (1 − f ) and in the (n p , l p ) excited state ( f ) are fixed during the whole process and that 20% (ε) of the excitedH produced in the first reaction had time to de-excite to the ground state before undergoing the second reaction. These cross sections are given for 100% Ps(1s) ( f = 0), 20% Ps(2p) ( f = 0.2; n p = 2, l p = 1) and 40% Ps(3d) ( f = 0.4; n p = 3, l p = 2), values that are thought to be experimentally feasible: 4 10;10 + f σ 4B,4 10;n p l p .
As suggested above, the interaction region of the GBAR experiment will also have to take into account the constraints on theH states. Indeed, the first reaction will mainly, if not exclusively produce excited states of antihydrogen, while the second reaction requires groundstate antihydrogen. Some time must be given to the antihydrogen atoms for them to de-excite towards the ground state or the (2s) state, which will give a small contribution, and still be able to interact with (excited) Ps atoms. This implies that one needs to have a long but still dense enough Ps cloud (or, ideally, two separated Ps clouds, which can be excited into different states at different times).
A simulation taking into account these different options is currently being developed.
Conclusions
In the framework of the GBAR experiment, three-body and four-body CDW-FS models have been adapted to compute the cross sections of positronium (Ps) formation by charge exchange reactions between a positron and, firstly, a hydrogen atom (equation (1)) and secondly, a negatively charged hydrogen ion (equation (3)). From the results, total cross sections for antihydrogen and antihydrogen ion production using Ps have been deduced (equations (2) and (4)). The different contributions of the lower excited states of both H (H) and Ps have been systematically investigated (up to n h = 4-5 in the case of the (anti)hydrogen atom and up to n p = 3 for the Ps). Three approximated wave functions have been used to describe H − (H + ), from the simple and useful uncorrelated Chandrasekhar wave function, to the more refined, with respect to the angular correlations, correlated Chandrasekhar and LS wave functions. To our knowledge, this is the largest study available concerning reactions (3) and (4). Results on the three-body reaction draw the same conclusions as previous studies (production of excited (anti)hydrogen; gain with Ps excitation) and compares quite well to the experiment of Merrison et al in the range 10-15 keV. Three-body CDW-FS has thus been proven a good tool to estimate the cross sections of antihydrogen formation from antiprotons and Ps. In the absence of experimental data and due to the few works available on the fourbody reaction, it is more difficult to put this model to the test. Despite the discrepancy between CDS-FS and the other two theories that were used for the reaction (3), it is worth noticing that, when the comparison is available, the relative behaviour of the cross sections is the same. This represents the information that can be extracted from these results, which can be used to guide experimental choices. It has been thus demonstrated that ground state antihydrogen is required in the entrance channel to have high cross sections, that Ps(2p) is the most interesting Ps state to enhance theH + production and that antiproton energy close to the reaction threshold should be aimed at. The expected nearly-resonant behaviour for Ps in a state n p = 3 has been observed, but would require ultra-low energy antiprotons (a few hundred eV) to be competitive with Ps(2p). Finally, it has also been shown that theH + formation fromH(2s) and Ps(1s) is a non-negligible channel. du R n p 0 (ρ) The total cross section is then given by σ 3B,1 n h 0;n p 0 = 1 4π 2 k β k α µ α µ β dk β T (−) with L n h l h (r 1 ) = ∞ 0 dr 2 r 2 2 R n h l h (r 2 )h l h (r 1 , r 2 ), The pure excitation term is given by m h dk β t * exc,l p =0 × t exc,l p =0 = 2 dk β t * cap,l p =0 × t exc,l p =0 = 2 The cross section is thus σ 4B,3 n h l h ;n p 0 = 1 4π 2 k β k α µ α µ β dk β T (−) αβ 2 = 1 4π 2 k β k α µ α µ β 2(4π) 5 (k + k − k α ) 2l h l ill il P * l il × P l il +P * l il ×P l il + P * l il × P l il + c.c. . (B.12) and the cross terms are of the form This leads to the cross section σ 4B,3 n h 0;n p 0 = 1 4π 2 k β k α µ α µ β dk β T (−) | 9,870 | 2013-09-01T00:00:00.000 | [
"Physics"
] |
Using Sentinel-2 Multispectral Images to Map the Occurrence of the Cossid Moth ( Coryphodema tristis ) in Eucalyptus Nitens Plantations of Mpumalanga , South Africa
Coryphodema tristis is a wood-boring insect, indigenous to South Africa, that has recently been identified as an emerging pest feeding on Eucalyptus nitens, resulting in extensive damage and economic loss. Eucalyptus plantations contributes over 9% to the total exported manufactured goods of South Africa which contributes significantly to the gross domestic product. Currently, the distribution extent of the Coryphodema tristis is unknown and estimated to infest Eucalyptus nitens compartments from less than 1% to nearly 80%, which is certainly a concern for the forestry sector related to the quantity and quality of yield produced. Therefore, the study sought to model the probability of occurrence of Coryphodema tristis on Eucalyptus nitens plantations in Mpumalanga, South Africa, using data from the Sentinel-2 multispectral instrument (MSI). Traditional field surveys were carried out through mass trapping in all compartments (n = 878) of Eucalyptus nitens plantations. Only 371 Eucalyptus nitens compartments were positively identified as infested and were used to generate the Coryphodema tristis presence data. Presence data and spectral features from the area were analysed using the Maxent algorithm. Model performance was evaluated using the receiver operating characteristics (ROC) curve showing the area under the curve (AUC) and True Skill Statistic (TSS) while the performance of predictors was analysed with the jack-knife. Validation of results were conducted using the test data. Using only the occurrence data and Sentinel-2 bands and derived vegetation indices, the Maxent model provided successful results, exhibiting an area under the curve (AUC) of 0.890. The Photosynthetic vigour ratio, Band 5 (Red edge 1), Band 4 (Red), Green NDVI hyper, Band 3 (Green) and Band 12 (SWIR 2) were identified as the most influential predictor variables. Results of this study suggest that remotely sensed derived vegetation indices from cost-effective platforms could play a crucial role in supporting forest pest management strategies and infestation control.
Introduction
In South Africa, emerging forest pests have caused extensive damage to Eucalyptus plantations [1].Approximately 1.3 million hectares of South African land is composed of both hard and softwoods with the majority located in the eastern parts of the country; primarily in Mpumalanga (40.8%),KwaZulu-Natal (39.5%) and the Eastern Cape (11.1%) [2].These plantations contribute annually to South Africa's gross domestic product with Eucalyptus plantations contributing over 9% to the total of exported manufactured goods [3].These species are the most productive planted exotics that mostly offer timber, pulp and paper in South Africa [4][5][6].Therefore, a robust mechanism needs to be established to prevent excessive damage, as numerous investments have been injected into the forestry sector, particularly the Mpumalanga province [7].Since 2004, Coryphodema tristis, commonly known as Cossid moth, has been the major cause of damage to Eucalyptus nitens resources across Mpumalanga, with forest managers requiring up-to-date information to support their forest protection interventions at ground level [8][9][10].
C. tristis is an indigenous wood-boring insect that commonly infests tree families, such as Ulmaceae (Elm Family), Vitaceae (Wild Grape family), Rosaceae (Rose family), Scrophulariaceae (Figwort family), Malvaceae (Mallow family) and Combretaceae (Indian almond family) [11,12].However, a sudden shift by the C. tristis to infest E. nitens in Southern Africa has been observed.According to Gebeyehu et al. [10], the shift of the C. tristis to infest E. nitens trees may be caused by a few to non-existent natural enemies in the area.As a result, the absence of natural enemies influences the increase of pests in the geographic area, due to less interspecific competition [13].This results in the moth breeding and multiplying at faster rates and increasing the intensities of E. nitens infestation.Adult female moths lay eggs on the bark of the E. nitens trees and the larvae feed on the bark damaging the cambium [10].The damage reduces the movement of water within the tree and also extends to the trunk and branches which turn black [8].Furthermore, as the larvae grow, it drills extensive tunnels into the sapwood and hardwood of the E. nitens which results in the trees producing resin on their trunks and branches and sawdust on the base of the forest floor [11].However, extensive tunnelling by the moth has resulted in severe damage to trees increasing the probability of tree mortality.Additionally, pupal casings are found protruding on the tunnelled bark or either at the base of the floor indicating the presence of C. tristis.
In recent years, researchers have attempted to use environmental variables to predict the spatial distribution of C. tristis [8,11].For example, Boreham [9] conducted a study that investigated the outbreak and impact of C. tristis on E. nitens in the Highveld of Mpumalanga, using environmental variables and the Residual Maximum Likelihood (REML) statistical method.The results showed that older E. nitens trees (above 8 years) and lower elevation sites less than 1600 m were the most susceptible to C. tristis infestations.Similarly, Adam et al. [8] used climatic and topographical variables to map the presence and extent of C. tristis infestations in E. nitens plantations of Mpumalanga.Using a random forest classifier, results indicated that with September and April's maximum temperatures; April's median rainfall and elevation played a crucial role in identifying conditions that are suitable for C. tristis occurrence.Their results furthermore predicted that areas with a maximum temperature greater than 23 • C in September and 22 • C in April were the most susceptible to infestation.While these studies have successfully utilised environmental and climatic variables to predict the presence of the moth, different studies have identified a number of limitations regarding traditional data collection methods to determine the presence or absence of pests.
Different studies stated that traditional methods such as field surveys are mostly time-consuming, costly, labour-intensive, spatially restrictive and likely unreliable as data collection is based on the knowledge of the surveyor [14,15].Hence, a direct detection approach that provides real-time information and can be repeated regularly for up-to-date decisions is required.Furthermore, utilizing environmental or climatic variables only for mapping the spatial distribution of pests can be challenging since these variables focus precisely on the surrounding factors and not the actual damage of plantations.For example, Germishuizen et al. [16] utilized environmental factors to determine the susceptibility of pine compartments to bark stripping by Chacma baboons (Papio ursinus).Results indicated that indirect variables such as altitude provide a challenge in explaining the complex relationship of baboon-damage risk.Moreover, Donatelli et al. [17] indicated that observed environmental datasets alone were no longer sufficient to predict the behaviour of pests due to climate change that has influenced the variability of temperature averages, rainfall means and distributions.Thus, requiring more traditional field surveys to confirm whether a particular area has been truly infested.Bouwer et al. [11] indicated that actual confirmation of infestation was certainly confirmed by tree felling which is impossible for large-scale assessments.Hence, the use of remotely sensed data with the ancillary data such as environmental and climatic variables would provide an up-to-date, repeatable source of information for forest assessment and inventory.
Remote sensing has improved the accuracy of predictions of forest-damaging pests using narrow and broad bands in the visible, near, shortwave-infrared and red edge regions [15,18,19].For example, Adelabu et al. [20] sought to discriminate the levels of change in forest canopy cover instigated by insect defoliation using hyperspectral data in mopane woodland.Results indicated that the overall accuracy of classification was 82.42% using a random forest algorithm and was 81.21% using ANOVA.In another study, Oumar and Mutanga [19] successfully assessed the potential of WorldView-2 bands, environmental variables, as well as vegetation indices which resulted in the prediction of Thaumastocoris peregrinus infestations on Eucalyptus trees.Results indicated that WorldView-2 sensor bands and indices predicted T. peregrinus damage with an R 2 value of 0.65 and a root mean square error of 3.62% in an independent test data set.Similarly, Lottering et al. [18] also found that vegetation indices derived from the red edge region correlated with Gonipterus scutellatus-induced vegetation defoliation using WorldView-2 satellite data.Furthermore, Pietrzykowski et al. [15] assessed the presence and severity of defoliation and necrosis caused by the Mycosphaerella fungus in a Eucalyptus globulus plantation, using multi-spectral imagery in north-western Tasmania, Australia.Their results indicated that high spatial resolution airborne digital imagery performed well, producing an accuracy of 71% for defoliation and 67% for necrosis.Therefore, despite the optimal modelling accuracies attained using multispectral remotely sensed data in these studies, these data sets are expensive and limited to a local scale.In this regard, there is an urgent need for testing and assessing the utility of other cheaper data sets that could capture the disease and pest incidences at landscape levels.
This study, therefore, sought to model the probability of the occurrence of the C. tristis on E. nitens plantations in Mpumalanga, South Africa using the cost-effective Sentinel-2 multispectral instrument and derived vegetation indices.Sentinel-2 images across the valuable red edge portion of the electromagnetic spectrum are suitable for forest health applications related to pest and disease damage detection [21,22].The large swath width and a 5-day temporal resolution make this sensor suitable for repeatable monitoring over forest plantations and detect pest-related damage continuously for effective management and control.Therefore, we used Maxent a robust machine-learning algorithm to predict the probability of the occurrence of the C. tristis using remotely sensed data.
Study Area
The research was conducted in the Mpumalanga province of South Africa in the Lothair village, also known as Silindile, and is located in the Msukaligwa Local Municipality (Figure 1).The study site is located between 26 • 26 25.08" S and 30 • 3 59.4"E in the Highveld of Mpumalanga.Elevation of the study area ranges from 1200 to 2100 m above sea level.
The area is associated with between 783-1200 mm of rainfall on average per year from November to March.The Highveld has a summer (October to February) to winter (April to August) temperature range of approximately 19 • C, with average temperatures ranging between 8 • C and 26 • C in the contrasting seasons.The Highveld is among South Africa's highly productive commercial plantation forests that consist of Pine and Eucalyptus plantations.Greater parts of the Highveld are comprised of sandstone and granite derived soils which the majority of commercial tree species are grown.
Image Acquisition
A cloud-free Sentinel-2A MSI image of the study area acquired on 19 August 2016, was downloaded from the United States Geological Survey website (www.earthexplorer.ugs.gov).
The MSI sensor has a revisit time of 5 days making the detection of pest damage to vegetation instantaneous [21,23,24].The MSI sensor covers a large area with a swath width of 290 km for increasing the spatial coverage of area of interest [22,24].Sentinel-2A has thirteen bands ranging from 443.9 nm to 2202.4 nm including four 10 m visible and near-infrared bands, six 20 m red edge, near infrared and shortwave infrared bands and three 60 m bands visible, near-infrared and shortwave infrared bands.The narrow red edge bands cover spectral regions of 703.9 nm, 740.2 nm and 782.5 nm that can be utilised for monitoring vegetation status [22,23,25].
Image Processing and Analysis
Atmospheric correction of the image was done using the Sentinel Application Platform (SNAP) software, which incorporates the plugin, Sen2Cor.In total, ten bands were derived for modelling the probability of the occurrence of the C. tristis as shown in Table 1.In this study, Sentinel-2A bands 1, 9 and 10 were excluded because of their sensitivity to aerosol, clouds and spatial resolution (60 m).Furthermore, these three bands are not used for vegetation mapping.Using the Index Database (https://www.indexdatabase.de/db/i.php),we selected vegetation indices with the best capacity to detect and map the occurrence of the C. tristis (see Table 2).Additionally, a number of published vegetation indices that have been effective in characterizing vegetation defoliation, many of which are sensitive to reflectance in the visible and NIR regions were derived.However, vegetation indices with wavelengths from the red edge region were given more emphasis based on their ability to identify stressed vegetation [18].[26]
Field Data Collection
On the 19 August 2016, a field visit was conducted in two South African Pulp and Paper Industries (SAPPI) plantations totalling 23,928 hectares to establish the presence/absence of the pest in the area.SAPPI plantations are divided into two blocks namely, Woodstock and Riverbend that contains 878 E. nitens compartments.Compartments are partitioned from the blocks that contain the E. nitens plantations and vary in size.Woodstock is located in the northern region of the SAPPI plantation and consists of 55 E. nitens plantations, whilst Riverbend located in the southern region comprises of 1145 plantations.Field crews from SAPPI were assigned different compartments to assist with field work in order to cover the whole study area.To determine the presence/absence of the C. tristis in E. nitens compartments, we used a quadrat sampling technique to carry out mass trapping of C. tristis.Mass trapping was carried out from 15 June to 19 August 2016 using a minimum of 19 and a maximum of 348 yellow bucket funnel traps with pheromone lures across all E. nitens compartments.Pheromones that match the chemical scent of a female adult moth was used to lure male moths into the traps that were located in the compartments [8].The number of traps used in the field varied with the size of the compartments where traps were placed at 50 m apart from each other hence, in bigger compartments there were more traps compared to smaller compartments.To determine the presence/absence, the sawdust and resin on the stem or the base of the tree were used as indicators of the presence/absence of the C. tristis.Locations of these indicators were then measured using a handheld Trimble GeoHX 6000 Global Positioning System (GPS) with a sub-meter accuracy (<10 cm).The dataset of pest damage indicators was then used to extract spectra from the Sentinel-2A image and develop training and testing datasets for statistical analysis.
Maxent Modelling Approach
The freely available Maxent approach (version 3.4.0) is developed for species distribution modelling (SDM) and was used in this study for modelling the probability of the occurrence of the C. tristis (http://biodiversityinformatics.amnh.org/open_source/maxent/)[35].Maxent is a machine learning technique that uses presence-only data to determine the potential spatial suitability preference of species [35,36].The model evaluates the probability of the occurrence from a number of spatial environmental variables [37][38][39].For Maxent to determine the probability of occurrence and reduce uncertainty, it requires more presence information of the target species [40].The background dataset definition contributes to the model's output significantly and requires the species full environmental distribution of those areas that have been searched [41].As a result, Maxent establishes a model with a maximum entropy in relation to the data of presence locations and variables to similar interactions with background locations [36,41].
In this study, a total of 20 predictor variables with a correlation −0.8 < r < 0.8 were considered for determining the probability of the occurrence of C. tristis.Bands and vegetation indices from Sentinel-2A MSI data were used to run four model scenarios in Maxent to determine the probability of the occurrence of the C. tristis (as shown in Table 3).These four model scenarios were carried out independently to identify which predictor variables were more robust in modelling the probability of the occurrence of the C. tristis.
Model Accuracy Assessment
For this study, presence data of the C. tristis infested locations (n = 371) within the compartments were randomly partitioned into two sets, 70% training data (n = 259) and 30% test data (n = 111).A sub-sample was used as the replicate run and iterations were fixed to 500.The regularization multiplier was maintained at 4 to avoid overfitting of the test data [36].The remaining model values were set to default values.A complementary log-log (clog log) output was utilised because it strongly predicts areas of moderately high output compared to the logistic output [34].To avoid bias of estimation, the study used a nonparametric method called the jack-knife to analyse the effects of environmental variables on model results to indicate influential variables.This method can estimate parameters and adjust the deviation without assumptions of distribution probability [42,43].Hence, during training, Maxent performs a jack-knife test that assesses the relative importance of each predictor variable which explain the spatial distribution of the species [41].Model performance was assessed using the area under the curve (AUC) of the receiver operating characteristics (ROC) [39,44,45].ROC is a graphical plot generated by the Maxent algorithm based on the AUC when model sensitivity is plotted against 1 minus model specificity [16,37].Hence, the model was characterized as more accurate when the curve followed the plot y-axis when compared to the x-axis because it attained a higher sensitivity value than a specificity value.Validation of results were carried out using test data.
In that regard, the AUC ranged from 0 to 1 and the accuracy was classified as poor between 0.5-0.70,while 0.70 and 0.80 are good and above 0.90 are termed high [46,47].Additionally, the jack-knife test was used to assess the contribution of each variable's to the model and highlighted the dominant variables [39,46].Furthermore, True Skill Statistic (TSS), also known as the Hanssen-Kuipers discriminant was utilized to assess the accuracy of the model.TSS accommodates both sensitivity and specificity errors and success as a result of random guessing [48,49].It ranges from −1 to +1, whereby +1 indicates perfect agreement whilst values of zero or less indicate random performance.The advantage of TSS compared to Kappa is that, TSS is not affected by prevalence making it a better accuracy assessment method [50,51].In terms of prevalence, Kappa may introduce bias with regards to the frequency of validation sites (field data) that is, a higher frequency of a specific species would result in higher prevalence rates, which would ultimately affect the classification accuracy [50].
Mapping C. tristis Occurrence
To determine the spatial distribution of the C. tristis, Maxent applies the maximum-entropy principle to fit the model and compares the interactions between the presence locations and variables to estimate the probability of species distribution [52,53].A complementary log-log (clog log) output was utilized as it strongly predicts areas of moderately high output [35].The regularization multiplier was set at 4 to avoid overfitting the test data [36].Model parameters were set to default replication of 1 with 500 iterations using cross-validation run type.Based on a threshold value, we used a 10-percentile threshold value in Maxent to generate model predictions using combined predictor variables (bands and vegetation indices).An estimate of probability of occurrence of C. tristis was exported to ArcGIS 10.4 from Maxent showing presence = 1 and absence = 0. Using ArcGIS 10.4, maps were generated to indicate presence/absence of C. tristis.
Maxent Modelling of C. tristis Occurrence
Table 4 shows the results attained after running the three models for determining the probability of the occurrence of the C. tristis.Using spectral bands, an overall accuracy of test data = 0.898 and training data = 0.891 with a TSS value of 0.282 was achieved while vegetation indices produced an overall accuracy of test data = 0.872 and training data = 0.875 with a TSS value of 0.324.When comparing the two models, the overall accuracy decreased by 0.026 test data and 0.04 training data.As a result, Sentinel-2 derived vegetation indices were outperformed by bands in detecting the probability of the occurrence of C. tristis.The results in Table 4 show that the overall integration of bands and vegetation indices produced higher prediction accuracy in this study.Using the combined data set, the model yielded a high overall accuracy of 0.890 test data and 0.900 training data with a TSS value of 0.344.Bands performed slightly weaker than vegetation indices.Based on the results, the models performed above the random prediction of 0.5 indicating good results.
Respectively in Figure 2, the Maxent model produced a test jack-knife that indicated the relative importance of each variable in the modelling process.In Figure 2a, the most influential bands in the model were Band 5 (Red edge 1), Band 4 (Red), Band 3 (Green), Band 12 (SWIR 3) and Band 2 (Blue) respectively.As illustrated in Figure 2b, PVR, GNDVIhyper, PSND, SR774/667 and NDSI respectively were the most influential variables in the vegetation indices model.Figure 3 is Jack-knife test variable importance graph of combined variables derived in modelling the spatial distribution of the C. tristis.
Band 5 (Red edge 1) contributed significantly to the probability of the occurrence of the C. tristis with a variable importance of 0.814 (Figure 2a).This shows the significance of the vegetation red edge band in discriminating healthy and unhealthy E. nitens trees.Moreover, Band 4 (Red) was the second highest variable with a contribution of 0.802.Band 4 (Red) recorded a decrease in the reflectance indicating the possibility of infested vegetation in the study area.Additionally, Figure 2a illustrates that bands in the VIS had the highest contribution as Band 3 (Green) was the third highest variable with a contribution of 0.793.Moreover, both Bands 11 (SWIR 2) and 12 (SWIR 3) performed well in the modelling of the C. tristis, with Band 12 (SWIR 3) contributing 0.784 as the fourth highest variable.Band 2 (Blue) also yielded a contribution of 0.757 and was the fifth highest variable in the model.In addition, Band 8 (NIR), Band 6 (Red edge 2), Band 7 (Red edge 3) and Band 8A (Narrow NIR) displayed a significant contribution above 0.65 each to the overall model.Sentinel-2 derived bands demonstrated the high potential of predicting the likely spatial distribution of the C. tristis.
As shown in Figure 2b, PVR was the most prominent variable in the model with a contribution of 0.818.The index has the potential to detect any changes in the chlorophyll content and identify weakly active vegetation affected by stress [29].The results showed that GNDVIhyper was the second highest important variable with a contribution of 0.797.The test jack-knife highlighted that the PSND was the third highest variable that performed well in the model with a contribution of 0.776.Both the NDSI and NDVI performed fairly equally with a contribution of 0.720.The remaining vegetation indices had a contribution above 0.500 in the model.The results obtained using Sentinel-2 derived vegetation indices alone produced slightly lower prediction accuracies when compared to those derived using the spectral bands.
Comparing the results attained in the model 1 and model 2 for each variable, it is evident that contribution accuracies did not significantly increase indicating similar strength in the prediction of the occurrence of C. tristis.Moreover, of all the three models, the results showed that PVR increased its contribution factor to 0.853 while Band 5 (Red edge 1) increased to 0.821 resulting in vegetation indices outperforming the spectral bands.Hence, results showed that vegetation indices (TSS = 0.324) outperformed bands (TSS = 0.282).However, model 3 produced a TSS value of 0.344, which is closer to +1 indicating a higher accuracy.Therefore, the results from model 3 using both bands and vegetation indices established a significant improvement on the overall contribution accuracies integrated into this study.Clearly, the results from the three models that surpassed the random prediction of 0.5 highlighted the great potential of the model to predict the probability of the occurrence of C. tristis.Based on the Jack-knife results obtained from using the bands alone and vegetation indices alone, we ran the final models using the most influential predictor variables and the results for each group are shown in Table 5 below displaying their performance metrics.As shown in Figure 2b, PVR was the most prominent variable in the model with a contribution of 0.818.The index has the potential to detect any changes in the chlorophyll content and identify weakly active vegetation affected by stress [29].The results showed that GNDVIhyper was the second highest important variable with a contribution of 0.797.The test jack-knife highlighted that the PSND was the third highest variable that performed well in the model with a contribution of 0.776.Both the NDSI and NDVI performed fairly equally with a contribution of 0.720.The remaining vegetation indices had a contribution above 0.500 in the model.The results obtained using Sentinel-2 derived vegetation indices alone produced slightly lower prediction accuracies when compared to those derived using the spectral bands.
Comparing the results attained in the model 1 and model 2 for each variable, it is evident that contribution accuracies did not significantly increase indicating similar strength in the prediction of the occurrence of C. tristis.Moreover, of all the three models, the results showed that PVR increased its contribution factor to 0.853 while Band 5 (Red edge 1) increased to 0.821 resulting in vegetation indices outperforming the spectral bands.Hence, results showed that vegetation indices (TSS = 0.324) outperformed bands (TSS = 0.282).However, model 3 produced a TSS value of 0.344, which is closer to +1 indicating a higher accuracy.Therefore, the results from model 3 using both bands and
C. tristis Spatial Distribution
Using the Maxent models, we used both bands and indices to determine the highest probability of C. tristis occurring across the study as illustrated in Figure 4.The highest probability of occurrence is detected in the upper northern parts of the boundary in the Woodstock area descending to the southern areas in the Riverbend area.In the middle of the Riverbend plantation, the highest probability of occurrence is expected, whilst minimum occurrence is anticipated at the lowest parts of the study area.Generally, the presence of the moth is spread across the plantation and observed from the northern parts to southern parts of the study area.
vegetation indices alone, we ran the final models using the most influential predictor variables and the results for each group are shown in Table 5 below displaying their performance metrics.
C. tristis Spatial Distribution
Using the Maxent models, we used both bands and indices to determine the highest probability of C. tristis occurring across the study as illustrated in Figure 4.The highest probability of occurrence is detected in the upper northern parts of the boundary in the Woodstock area descending to the southern areas in the Riverbend area.In the middle of the Riverbend plantation, the highest probability of occurrence is expected, whilst minimum occurrence is anticipated at the lowest parts of the study area.Generally, the presence of the moth is spread across the plantation and observed from the northern parts to southern parts of the study area.
Discussion
In this study, using remotely sensed data we modelled the probability of the occurrence of C. tristis on E. nitens through the application of Maxent.Derived Sentinel-2 vegetation indices and bands combined together performed well in modelling the probability the C. tristis occurring.However,
Discussion
In this study, using remotely sensed data we modelled the probability of the occurrence of C. tristis on E. nitens through the application of Maxent.Derived Sentinel-2 vegetation indices and bands combined together performed well in modelling the probability the C. tristis occurring.However, when testing both bands and vegetation indices individually, they did not perform as well as the combined variables.The significance of these vegetation indices compared to bands could be explained by their ability to detect the health status of vegetation.C. tristis damages the tree trunk and branches of E. nitens resulting in foliage turning black through chlorosis and then it ultimately dies.As a result, there is a reduction in the absorption rates of the visible light as there are fewer green pigments available, which cause changes in the spectral reflection.
Results obtained in this study regarding the significance of vegetation indices concurs with previous studies of Minařík and Langhammer [54], Metternicht [34] and Hart and Veblen [55].According to Gitelson and Merzlyak [30], they identified that healthy and unhealthy (stressed) vegetation is mostly observed in the green peak and vegetation red edge region, hence vegetation indices such as PVR and GNDVI yielded an outstanding performance in detecting the probability of C. tristis occurring.In addition, Metternicht [34] highlighted that PVR detects any changes in the reflective properties originating from changes in chlorophyll content and produce low values for photosynthetically weakly active vegetation.Moreover, Gitelson et al. [29] stated that new vegetation indices such as GNDVIhyper have an extensive dynamic range compared to NDVI, hence, they are more sensitive to chlorophyll changes.Therefore, this accounts for the high results yielded by GNDVIhyper in predicting the probability C. tristis occurring in this study.Sanchez-Azofeifa et al. [56] pointed out that SR and NDVI indices are used to estimate the chlorophyll concentration of vegetation as well as observing fundamental variations on leaf age, henceforth, these attributes boosts its performance.Findings from this study showed that SR800/500, SR 774/667 and NDVI performed exceptionally well and can be credited to the above-mentioned.In addition, a combination of two robust bands (NIR and Red) strengthens the probability of modelling and picking up vegetation characteristics that indicate the occurrence of pests.Therefore, different studies have stated that the integration of NIR and band 4 (NDVI) and vegetation indices derived from the red edge bands have enhanced the prediction of pests [19,21,54].For example, Hart and Veblen [56] illustrated that the vegetation indices were the most important predictors to detect tree mortality caused by spruce beetle (Dendroctonus rufipennis) at grey-stage.Therefore, future studies should seek to improve the detection of C. tristis and its associated impacts on E. nitens trees using powerful vegetation indices.
The results of this study also revealed that the band 5 (Red edge 1) was significant in determining the probability of C. tristis occurring.There is a high correlation between red edge bands and the chlorophyll content of the leaves, so that the spectral signature of E. nitens after chlorosis due to an attack by the C. tristis is easily detected on the red edge spectrum.Several studies that sought to detect and map the spatial distribution of insect pests affecting forest species confirmed that the red edge region played a significant role in predicting of the occurrence of such pests [19,20,57,58].In support of these results, Oumar and Mutanga [19], Murfitt et al. [59] and Pietrzykowski et al. [15] concluded that red edge bands perform slightly better than other bands in the detection of insect pests in forest damage.For example, Oumar and Mutanga [19] illustrated that the red-edge and NIR bands of WorldView-2 were sensitive to stress-induced changes in leaf chlorophyll content, therefore, improved the potential to detect T. peregrinus infestations.In this regard, the Sentinel-2's red edge bands demonstrated its great potential in monitoring the probability of C. tristis occurring, using its higher temporal and spatial resolution.
In determining the probability of the occurrence of C. tristis, results of this study further revealed the significant potential of the SWIR region.This region has the ability to map vegetation statues due to its sensitivity to changes in the water content of vegetation [54,55].Generally, the larva of C. tristis feeds on the cambium, which is responsible for providing layers of phloem and xylem in E. nitens plantations.Therefore, damage to the cambium affects both phloem and xylem which ultimately alters the movement cycle of water from the roots through the trunk to the leaves of E. nitens trees [54].This results in foliage and canopy water changes.It induces stress which leads to the reduction of the water content present in the main trunk and branches contributing to the change in colour to black.Subsequently, the variations are then detected effectively in the SWIR portion of the electromagnetic spectrum.This then explains the optimal influence of the band 11 (SWIR 2) and band 12 (SWIR 3) in detecting E. nitens compartments that are vulnerable to C. tristis.Similarly to this study, Senf et al. [60] accurately detected the infestations of bark beetle at the red and grey-attack stage using the SWIR bands which distinguished changes in the water content.In a similar study, Ismail et al. [61] indicated that infestation caused by the S. noctilio on pine trees altered the water balance of the tree and bands within the SWIR captured these changes and improved the overall prediction of the pests' distribution.Furthermore, Hart and Veblen [55] indicated that in the spruce beetle and mountain pine beetle-infested trees, reflection increased in the SWIR and decreased in NIR due to a decrease in the foliar moisture content.
As a species distribution model (SDM), the Maxent model developed a spatial distribution map that shows the probability of the occurrence of C. tristis across the study area.High levels of presence of the moth spread across from the upper (Riverbend plantation) to the lower (Woodstock plantation) portions of the study area while medium presence along the centre of the study area was recorded.The increase in the presence of the moth from the upper portions to the lower portions might be characterized by the absence of natural enemies and hence, could explain the higher level of infestations.The results were similar to Adam et al. [8] which illustrated that in the upper portion of the study area where there was a high presence of the C. tristis compared to the lower portions indicating that C. tristis is rapidly spreading.However, our results may be affected by the trap density and as a result, future studies should look at better sampling strategies.Hence, distribution maps of the C. tristis can help to formulate and improve on-going monitoring and management efforts to reduce the current infestation of E. nitens forests.
Conclusions
This study tested the utility of the new generation Sentinel-2 multispectral instrument in detecting and mapping the probability of the occurrence of C. tristis infestations on E. nitens plantations.Based on the findings of this study, we conclude that bands in the VIS, NIR and SWIR are significant in modelling the probability of the occurrence of C. tristis.These three regions measure the spectral reflectance of vegetation that results in determining the amount of healthy and unhealthy vegetation.Additionally, the red edge bands played a crucial role in the probability of occurrence of C. tristis.Consequently, vegetation indices derived from the VIS/NIR have demonstrated their influence in detecting changes in the chlorophyll concentrations and improving the overall modelling concept in this study.Overall, these results underscore the significance of the Sentinel-2 sensor in detecting C. tristis.The results are a platform towards the detection and mapping of the highest probability of occurrence of C. tristis, using different multispectral sensors and their spatial resolution.The utility of remotely sensed data will improve the monitoring and management strategies used in forecasting the prevalence of pests as well as their spread.Moreover, key stakeholders such as forest managers will be in a position to control the damage of pests and devise proactive measures that are seemingly appropriate.This information is critical for preventing extensive damage in the forestry sector.
Figure 1 .
Figure 1.(a) Map of South Africa showing the Mpumalanga Province, (b) the location of the study area within the Mpumalanga Province (c,d) show healthy and infested Eucalyptus nitens and (e) shows the sampled compartments over the Sentinel-2 image with a false-colour composite of R-NIR-B (B4, B8 & B2).
Figure 1 .
Figure 1.(a) Map of South Africa showing the Mpumalanga Province, (b) the location of the study area within the Mpumalanga Province (c,d) show healthy and infested Eucalyptus nitens and (e) shows the sampled compartments over the Sentinel-2 image with a false-colour composite of R-NIR-B (B4, B8 & B2).
17 Figure 2 .
Figure 2. Jack-knife test variable importance graph of (a) bands and (b) vegetation indices derived in modelling the spatial distribution of C. tristis.
Figure 2 .
Figure 2. Jack-knife test variable importance graph of (a) bands and (b) vegetation indices derived in modelling the spatial distribution of C. tristis.
Figure 3 .
Figure 3. Jack-knife test variable importance graph of combined variables derived in modelling the spatial distribution of the C. tristis.
Figure 3 .
Figure 3. Jack-knife test variable importance graph of combined variables derived in modelling the spatial distribution of the C. tristis.
Figure 4 .
Figure 4. Map of C. tristis occurrence predicted with Maxent models using bands and vegetation indices as predictors.
Figure 4 .
Figure 4. Map of C. tristis occurrence predicted with Maxent models using bands and vegetation indices as predictors.
Table 1 .
Sentinel-2 bands used in this study.
Table 2 .
Sentinel 2 vegetation indices tested in this study.
Table 3 .
Bands and indices from Sentinel-2A MSI data used as independent variables for predicting the probability of occurrence of the C. tristis with Maxent models.
Table 4 .
Evaluation results for all Maxent models used for predicting the probability of the occurrence of C. tristis.
Table 5 .
Evaluation results for all influential predictor variables used for predicting the probability of occurrence of the C. tristis.
Table 5 .
Evaluation results for all influential predictor variables used for predicting the probability of occurrence of the C. tristis. | 8,421 | 2019-01-31T00:00:00.000 | [
"Environmental Science",
"Biology"
] |
High-Barrier, Biodegradable Films with Polyvinyl Alcohol/Polylactic Acid + Wax Double Coatings: Influence of Relative Humidity on Transport Properties and Suitability for Modified Atmosphere Packaging Applications
Polyvinyl alcohol (PVOH) exhibits outstanding gas-barrier properties, which favor its use as a biodegradable, high-barrier coating on food-packaging films, possibly in combination with modified atmospheres. Nonetheless, its high sensitivity to water can result in a severe loss of barrier properties, significantly limiting its applications with fresh foods and in high-humidity conditions. In this work, the water vapor (PWV) and oxygen permeability (PO2) of high-barrier biodegradable films with PVOH/PLA + wax double coatings were extensively characterized in a wide range of relative humidity (from 30 to 90%), aimed at understanding the extent of the interaction of water with the wax and the polymer matrices and the impact of this on the permeation process. What is more, a mathematical model was applied to the PWV data set in order to assess its potential to predict the permeability of the multilayer films by varying storage/working relative humidity (RH) conditions. The carbon dioxide permeability (PCO2) of the films was further evaluated, and the corresponding permselectivity values were calculated. The study was finally augmented through modified atmosphere packaging (MAP) tests, which were carried out on double-coated films loaded with 0 and 5% wax, and UV-Vis analyses. The results pointed out the efficacy of the PLA + wax coating layer in hampering the permeation of water molecules, thus reducing PVOH swelling, as well as the UV-shielding ability of the multilayer structures. Moreover, the MAP tests underlined the suitability of the double-coated films for being used as a sustainable alternative for the preservation of foods under modified atmospheres.
Introduction
Plastic pollution is an issue of major concern in our society today.According to the 2022 OECD Global Plastic Outlook [1], about 22% of all plastic waste worldwide ends up in landfills or as litter, causing significant harm to land and marine environments.Biodegradable polymers are a viable alternative to conventional plastics from a sustainability and circular economy perspective [2]; however, their performance needs to be enhanced to meet the needs of the market, especially in those sectors where plastic consumption is very high [3].
PVOH is generally combined in multilayer structures realized by coating [8,17,18], coextrusion [19,20], and electrospinning [21,22] with the aim of joining gas barrier, heat resistance, printability, and antioxidant properties [23][24][25][26].The coating technology, in particular, is garnering a wide interest within scientific and industrial research because it allows one to functionalize biodegradable substrate films in a simple and effective way, conferring them an advanced performance without modifying their bulk properties nor their biodegradable and/or recyclability features [27,28].
Despite its excellent gas barrier, the main limitations of PVOH, like other hydrophilic biopolymers, are its poor water vapor barrier properties: indeed, water plasticizes the polymer matrix, significantly affecting the absorption and diffusion of even other gases and weakening the structural integrity.This precludes its use in high-humidity environments or in contact with moist foods [29][30][31].
Therefore, the study and mathematical modeling of the mechanisms governing mass diffusion in PVOH packages at different relative humidities becomes an essential task to predicting their final performance and determining their suitability for preserving the quality of packaged foods under real working conditions [32][33][34].To the best of our knowledge, only a few studies have described and modeled the effect of relative humidity (RH) on PVOH permeability.Abdullah et al. [35] studied the water-and gas-barrier properties of PVOH/starch/glycerol + halloysite nanotubes (HNT) nanocomposite films.The authors found that the water vapor permeability coefficient increased linearly as the RH gradient increased, following the typical trend of materials that obey Fick's law.Mo et al. [36] investigated the effects of RH and temperature on the barrier properties of bioriented polyvinyl alcohol (BOPVOH) films, finding that the relationship between water vapor and oxygen permeability and RH% was well expressed by an exponential trend.
In our previously published work [8], biodegradable films with a high gas-barrier performance were realized by spreading a double coating layer of modified PVOH (m-PVOH) and PLA + ethylene-bis-stereamide (EBS) wax (from 0 to 20%) on a poly(butyleneadipate co-terephthalate) (PBAT)/poly(lactide) (PLA) substrate.The films combined the ductility of the PBAT/PLA web layer with the high oxygen barrier of the m-PVOH and the heat sealability and water-resistance of the PLA + EBS wax layer, with a potentially positive effect on the water vapor barrier properties.
In this work, our goal was to investigate how ambient humidity affects the transport properties of these multilayer structures, and the impact of the films' configuration on the final barrier performance.
The water vapor and oxygen barrier properties were determined at ambient temperature (23 • C) over a wide range of relative humidities.Additionally, a mathematical model was applied to the data set to describe the trends of P WV evolution by increasing the RH and assess its potential for predicting the barrier performance of the multilayer systems at different humidities.The carbon dioxide permeabilities were determined and the suitability of the developed films for MAP were assessed.The optical properties of the films were finally measured to evaluate the UV-shielding ability, which plays a fundamental role in preserving sensitive foods from light-induced deterioration mechanisms.
Materials
A blown web layer made of a commercial PBAT/PLA blend (Biofilm) with the trade name Bioter TM (Euromaster S.p.a., Pistoia, Italy) was chosen as the substrate.
Exceval AQ-4104 is a polyvinyl alcohol modified with ethylenic insertions that exhibits a low gas barrier from dry conditions to approximately 55% relative humidity (RH%), according to the technical data provided by the manufacturer.It is totally hydrolyzed (98.0-99.0mol%), chlorine-free, and water-soluble, and it has received FDA approval for food-contact use.The polymer will be hereafter named m-PVOH.
Deurex (Elsterau, Germany) supplied the Ethylene-bis-Stearamide (EBS) wax X2010M (CAS 110-30-5), produced by revaluing sugar cane residues.The wax is approved for the production of commodities intended to come into contact with food, according to the EU Regulation 10/2011 and the FDA 21 CFR [37], and is classified as a no-hazard substance by the European Chemicals Agency (ECHA) [38].Analytical-grade solvents were utilized throughout.
Production of the Double-Coated Films
The double-coated films were realized with the same methodology and conditions reported by Apicella et al. [8].Briefly, the Biofilm substrate monolayer, having a thickness of 22 ± 2 µm, was realized using a GIMAC lab scale blown film plant, outfitted with a single-screw extruder (D = 12 mm, L/D = 24) with a thermal profile set between 190 • C and 160 • C, a screw speed of 50 rpm, and a collection speed of 3 m/min.The substrate was then double coated using a laboratory bar coater (RK, Printocoat Instruments Ltd., Litlington, UK), equipped with a stainless steel closed-wound rod with a wire diameter equal to 0.80 mm.Firstly, the m-PVOH/water solution (mass ratio 10:90) was prepared, spread on the web, and dried at 120 • C for 4 min in an oven.Then, the second layer made using a PLA/acetone solution (20:80) loaded with 0%, 5%, 10%, and 20% w/w PLA EBS wax was deposited and dried at room temperature.In the present study, unlike the previously published work by Apicella et al. [8], the thickness of the m-PVOH layer was increased from 5 ± 1 µm to 10 ± 1 µm to further improve the barrier performance of the films.Table 1 reports the list of the realized films and the relative thicknesses of the layers.
Water Vapor Permeability and Modeling
The water vapor transmission rate (WVTR) measurements of the films were performed through a Water Vapor Permeation Analyzer (Model 7002-Systech Illinois, Princeton, NJ, USA) at 23 • C and at 30%, 50%, 70%, 85%, and 90% relative humidity.All the measurements were carried out in triple according to the ASTM F 1249-90 standard.The area of the tested films was 5 cm 2 .The values of the water vapor permeability coefficients (P WV ) were calculated assuming that the double-coated films were homogeneous [39] and applying the following equation: where WVTR is the water vapor transmission rate (g/m 2 day) of the film, l is the average film thickness (mm), and ∆P is the partial water vapor pressure difference (bar) at the two sides of the film.
In order to describe and predict the P WV behavior as a function of relative humidity for the developed systems, the P WV data were fitted using an exponential mathematical model, using the software CurveExpert Professional © 2.7 to solve the equation.The fit was optimized based on the minimization of the sum of the squared residuals, and the coefficients of the determination R 2 were calculated as a measure of the goodness of the fit.
Gas Permeability Measurements and Permselectivity
The oxygen transmission rate (OTR) and carbon dioxide transmission rate (CO 2 TR) tests were conducted by means of a gas permeabilimeter (GTT, Brugger, Munich, Germany) following the ISO 15105-1 standard.The measurements were carried out in triple on film specimens with an area of 16 cm 2 ; the gases' flow rate was set at 80 mL/min.
The OTRs were evaluated at 23 • C and at three different levels of RH% (10%, 50%, and 85%) by conditioning the oxygen insufflated in the measuring chamber through paper filters soaked with different saturated salt solutions of known water activity (a w ), while the CO 2 TR measurements were performed at 23 • C and 50% RH.
The oxygen permeability coefficients (PO 2 ) and carbon dioxide permeability coefficients (PCO 2 ) were calculated by multiplying the measured OTRs and CO 2 TRs, respectively, by the thickness in mm of the samples, assuming that the double-coated films were homogeneous [39].
The CO 2 /O 2 permselectivity of the film samples was assessed according to the following formula:
Analysis of Packaging Headspace Gas Composition after MAP
Sample bags having 15 × 10 cm size of Bioter/m-PVOH/PLA and Bioter/m-PVOH/PLA + 5%wax were utilized to assess the suitability of the developed films for modified atmosphere packaging (MAP) applications.The bags were filled by gas flushing with a gas mixture consisting of 5% oxygen (O 2 ), 30% carbon dioxide (CO 2 ), and 65% nitrogen (N 2 ), after which they were hermetically sealed.The changes in gas composition within the headspace of the packages were monitored at specific time intervals (0, 7, and 14 days) using an O 2 /CO 2 gas analyzer (PBI Dansensor, Checkmate3, Ringsted, Denmark).The results were expressed as the average gas concentration (%) ± standard deviation (SD), based on ten replicated bags at each sampling time.
UV-Vis Spectroscopy
The ultraviolet-visible (UV-Vis) spectroscopic measurements were performed on the film samples using a Lambda 800 UV-VIS spectrophotometer (Perkin Elmer, Waltham, MA, USA) in accordance with the ASTM D1746.The transparency of the films was determined by measuring the transmittance at 560 nm.Three replicates of each film were tested.The percent transparency (T R %) was calculated as follows: where T r is the transmittance with the specimen in the beam and T 0 is the transmittance with no specimen in the beam.
The UV-barrier properties were quantified using the average transmittance of UVA (315-400 nm) and UVB (280-315 nm) calculated according to the following equations [40]: where T(λ) is the average transmittance of the film at the wavelength λ, and dλ is the film's bandwidth.
Evaluation and Modeling of the Effect of RH on the Water Vapor Barrier Properties of Coated Films
Figure 1a displays the values of P WV for the neat substrate and all the coated films plotted as function of the relative humidity, while Figure 1b provides the magnification between 65 and 95% RH.For a better comprehension, the numerical values of P WV at different RH% are also reported in Table 2.
Evaluation and Modeling of the Effect of RH on the Water Vapor Barrier Properties of Coated Films
Figure 1a displays the values of PWV for the neat substrate and all the coated films plotted as function of the relative humidity, while Figure 1b provides the magnification between 65 and 95% RH.For a better comprehension, the numerical values of PWV at different RH% are also reported in Table 2.As observable in Figure 1 and Table 2, for all the tested films, the water vapor permeabilities remained fairly constant until 50% RH and, after, they increase following an exponential growth curve, with some noticeable differences for films with different compositions.These trends can be interpreted by taking into account the fact that the increase in water vapor partial pressure leads to two opposite effects on the transport phenomena through the film: (i) a compaction effect on the polymer chains due to hydrostatic pressure, resulting in an increase in polymer density, which inhibits the diffusion process; and (ii) a plasticization effect as a direct consequence of the increased concentration, which enhances the segmental motions within the polymer and promotes the diffusion [41][42][43].This is a generally observed behavior in which these two effects are competing; the prevalence of one phenomenon over the other depends on the polymer-penetrant affinity as well as on the relative humidity, and determines the overall permeability of the film to moisture and other molecules.In particular, the Biofilm showed the highest P WV values in the whole RH interval investigated, ranging from 64.0 g mm/m 2 d bar at 30% RH to 124.0 g mm/m 2 d bar at 90% RH.These values categorize the substrate as a poor barrier-grade material, according to the classification proposed by Wang et al. [44], and are in line with the literature data reported for films based on PBAT/PLA blends [45,46].The further spreading of the m-PVOH layer was able to decrease the P WV of one order of magnitude with respect to the neat Biofilm: between 30 and 50% RH, the P WV of Biofilm/m-PVOH film was constant and equal to 3.5 g mm/m 2 d bar, among the range of high-barrier biodegradable polymers (P WV < 40 g mm/m 2 d bar) according to Wang et al. [44].The highwater vapor barrier performance of the m-PVOH coating was attributable to the insertion of ethylenic moieties on the polymer backbone, which were also responsible for the constant barrier properties up to 60% RH, as also reported in the producer's technical information [8].Above this critical value, the plasticization of the macromolecular chains prevailed, entailing an increase in the polymer-free volume, a faster diffusion, and, therefore, a steep increase in the P WV : indeed, the P WV rose at 10.3 g mm/m 2 d bar and 107.3 g mm/m 2 d bar at 70% and 90% RH, respectively.This outcome holds significant relevance in relation to other unmodified PVOH grades, for which the absence of water vapor resistance has been reported even under dry conditions [47], as well as a sharp increase in the P WV (>200%) at a relatively low RH (30-40%) as an effect of the rapid swelling [48,49].
The addition of the amorphous PLA-based second coating layer does not substantially impact the water vapor permeability of the films, since the P WV of the amorphous PLA (~1630 g mm/m 2 d bar at 23 • C and 50%RH) turned out to be approximately three orders of magnitude larger than that of the Biofilm/m-PVOH support film [50].Looking at the permeability curves of the Biofilm/m-PVOH/PLA and Biofilm/m-PVOH films, it is interesting to note a crossover point at 85% RH: at this RH, the P WV values of the two films were comparable (equal to 54.0 g mm/m 2 d bar and 52.8 g mm/m 2 d bar, respectively), while at 90% RH the Biofilm/m-PVOH/PLA sample exhibited a lower water vapor permeability value (equal to 82.2 g mm/m 2 d bar) with respect to the Biofilm/m-PVOH (equal to 107.3 g mm/m 2 d bar).This result can be explained by the fact that in glassy polymers, such as the amorphous PLA, the penetrant molecules occupy specific sites within the polymer's pre-existing micro voids up to the reaching of a saturation Polymers 2023, 15, 4002 7 of 14 level: at this point, all the available sites are occupied; the permeation phenomenon is no longer dependent on the penetrant-polymer interactions; and the permeation process slows down [51,52].
The inclusion of the EBS wax at different concentrations within the PLA matrix resulted in a reduction in the P WV of the double-coated films with respect to the Biofilm/m-PVOH/PLA sample, as underlined by the ∆P WV % values reported in Table 2.This is attributable to the hydrophobic effect of the EBS wax, which hinders the permeation of water vapor molecules, partially inhibiting the plasticization of the PVOH matrix [53].The maximum decrease in the P WV ensued after the inclusion of 5% of EBS wax in the PLA matrix and was equal to 37% at 70% RH.Further incorporation of higher percentages of EBS wax into the PLA layer did not substantially change the water vapor barrier' performance of the double-coated films with respect to the Biofilm/m-PVOH/PLA + 5%wax.These outcomes suggest that, among the developed structures, this latter configuration is the most promising for packaging applications, as it retains the best water vapor barrier performance over such a wide range of relative humidities, with the lowest concentration of wax.
In order to predict the performance of the developed films at specific storage/working RH conditions, it is useful to apply a mathematical model to the data set obtained.In the literature, several authors proposed different models for this purpose.For materials that obey Fick's law of diffusion, generally, the water vapor permeability increases linearly by increasing the RH.This behavior has been reported for PE and BOPP films [54].Unlike Fickian materials, the relationship between the RH and the P WV of moderate or high hydrophilic materials, such as most biopolymers, follows an exponential growth curve, according to the following equation [36]: where a, b, and c are the model parameters.This behavior has been reported for WPI [30], BOPVA [36], alginate, casein, chitosan, and zein films [32].This latter model was then selected to fit the values: the dotted lines in Figure 1a,b show the best fit of the equation to the P WV data set, while Table 3 gives an overview of the values of the model parameters (a, b, and c) obtained with the fitting procedure, and the R 2 value as an indicator of the goodness of the fit.As it can be observed above, the trends of P WV evolution by increasing the RH provided by the model's numerical solution are in close agreement with the empirical values.In particular, the calculated coefficient of determination R 2 is higher than 0.994 for all the samples investigated.
These outcomes underline the perspective to apply this simple mathematical model for predicting the water vapor permeability of the multilayer films over a wide range of relative humidity conditions, and its utility for the characterization, designing, and optimization of their barrier performance depending on the shelf-life requirements of the packaged foods.
Evaluation of the Effect of RH on the Oxygen Barrier Properties of Coated Films
The effect of the environmental relative humidity was also assessed on the oxygen barrier of the biodegradable films developed, as oxygen permeation into the package can be responsible for the harmful decay in the safety and quality of packed foods.Figure 2 illustrates the oxygen permeability (PO 2 ) values of all the coated films, measured at 10%, 50%, and 85% RH.
for predicting the water vapor permeability of the multilayer films over a wide range of relative humidity conditions, and its utility for the characterization, designing, and optimization of their barrier performance depending on the shelf-life requirements of the packaged foods.
Evaluation of the Effect of RH on the Oxygen Barrier Properties of Coated Films
The effect of the environmental relative humidity was also assessed on the oxygen barrier of the biodegradable films developed, as oxygen permeation into the package can be responsible for the harmful decay in the safety and quality of packed foods.Figure 2 illustrates the oxygen permeability (PO2) values of all the coated films, measured at 10%, 50%, and 85% RH.Our previously published research [8] highlighted the outstanding oxygen barrier performance obtained after the deposition of a 5 ± 1 µm thick m-PVOH layer, with PO 2 values comprised between 0.22 and ~0.30 cm 3 mm/m 2 d bar for the Biofilm/m-PVOH and the double-coated films, respectively.In the present study, a further improvement in the oxygen barrier performance has been achieved by doubling the thickness of the m-PVOH barrier layer, up to 10 ± 1 µm.In particular, a decrease in PO 2 values comprised between 42% and 66% for all the coated films was obtained with respect to the results measured by Apicella et al. in 2022 [8].
As already observed for the water vapor permeability coefficients, the oxygen permeabilities remained fairly constant up to 50% RH, thanks to the hydrophobic moieties on the m-PVOH chain structure.At 85% RH, the oxygen permeability values increased by one order of magnitude for all the coated films, compared to the values measured at 10% and 50% RH levels.In a similar fashion to what is reported for the P WV in Figure 1 and Table 2, at 85% RH, the measured PO 2 for the Biofilm/m-PVOH and Biofilm/m-PVOH/PLA samples were comparable and equal to 3.94 and 4.02 cm 3 mm/m 2 day bar, respectively.This outcome confirms that, at this relative humidity, the amorphous PLA coating layer reaches a saturation level which slows down the mass transport through the polymer matrix.
At 85% RH, it is also worth noting that, compared to the pristine Biofilm/m-PVOH/PLA, wax incorporation resulted in a PO 2 reduction comprised between 15%, for the Biofilm/m-PVOH/PLA + 5%wax, and 22%, for the Biofilm/m-PVOH/PLA + 10%wax.This result suggests that the hydrophobic effect of wax, which reduces the swelling of the m-PVOH layer by restricting the diffusion of water molecules, also allows to partially reduce the permeation of the oxygen molecules.
Carbon Dioxide Barrier Properties, Permselectivity, and Evaluation of Suitability for MAP Application
CO 2 barrier properties, along with oxygen permeability, are of paramount importance, especially in MAP systems where the levels of both oxygen and carbon dioxide must comply to specific values depending on the food type.CO 2 , when present at moderate concentrations, inhibits enzymatic reactions and delays spoilage, thereby extending the shelf-life of fresh fruits and vegetables [55].In the case of dry, fatty foods with a low water content (less than 12%), packaging should maintain high levels of carbon dioxide and oxygen levels below 2% in order to effectively inhibit oxidation and prevent rancidity [56].
For this reason, in the designing of MAP systems, the permselectivity (PCO 2 /PO 2 ratio) of the packaging material is a parameter that plays a crucial role, as it must be tailored to create an optimal gas composition within the package, thereby satisfying the specific preservation requirements and extending the shelf-life of packaged food products.
The CO 2 permeability and permselectivity values of the Biofilm substrate and all the multilayer films are provided in Table 4.The Biofilm substrate exhibited a remarkably high CO 2 permeability, equal to 305 cm 3 mm/m 2 day bar, which falls between the permeability values of the pristine PLA (~47.5 cm 3 mm/m 2 day bar) and PBAT (~390 cm 3 mm/m 2 day bar) films reported in the literature [57,58].The further spreading of the m-PVOH layer reduced the PCO 2 of the neat substrate by three orders of magnitude: the PCO 2 of the Biofilm/m-PVOH film came out to be equal to 0.37 cm 3 mm/m 2 day bar, highlighting the good barrier properties of m-PVOH also against CO 2 , as reported by other authors [59,60].Similarly to the water vapor and oxygen permeabilities, the addition of the amorphous PLA had no effect on the barrier against CO 2 , since its PCO 2 was ~47.5 cm 3 mm/m 2 day bar, three times larger than the Biofilm/m-PVOH support film [61].No variations were also appreciable due to the incorporation of the EBS wax, since all the double-coated films exhibited permeability coefficients ranging from 0.44 to 0.46 cm 3 mm/m 2 day bar.The calculated permselectivity values for the double-coated films ranged from 2.1, for the Biofilm /m-PVOH/PLA sample, to 4.9, for the Biofilm /m-PVOH/PLA + 10%wax.These values are comparable to those of conventional, fossil-based polymeric films [13,[62][63][64][65][66][67] stored under modified atmosphere packaging, with the possibility to tailor the optimal film layout on the basis of the target food's shelf-life needs.
To further assess the films' capability to retain modified atmosphere gas composition over time, MAP tests were carried out on bags made of Biofilm/m-PVOH/PLA and Biofilm/m-PVOH/PLA + 5%wax samples, flushed with a gas mixture consisting of 5% O 2 /30% CO 2 /65% N 2 , hermetically sealed, and stored for up to 14 days.The tests were conducted in the absence of foods that could alter the gas composition due to respiration, solubilization mechanisms in the food matrix, or spoilage effects.The headspace gas composition was monitored during that time and the results, expressed in terms of the measured O 2 and CO 2 percentages during that time, are depicted in Figure 3.As it can be observed, the O 2 concentration remains constant in both the Biofilm/m-PVOH/PLA and Biofilm/m-PVOH/PLA + 5%wax film samples throughout the investigated period.Although the m-PVOH exhibits a good CO 2 barrier under steady-state conditions, the CO 2 concentration slightly decreases during that time, from an initial level of 30%, on day 0 to a final value of ca.25% on day 14, for both samples.However, a slight CO 2 venting is a desirable condition for the preservation of some foods such as fruits and vegetables in order to avoid an excessive build-up of the gas produced by their respiration up to deleterious levels, which could result in cell membrane damage and physiological injuries to the product [68].
O2 concentration remains constant in both the Biofilm/m-PVOH/PLA and Biofilm/m-PVOH/PLA + 5%wax film samples throughout the investigated period.Although the m-PVOH exhibits a good CO2 barrier under steady-state conditions, the CO2 concentration slightly decreases during that time, from an initial level of 30%, on day 0 to a final value of ca.25% on day 14, for both samples.However, a slight CO2 venting is a desirable condition for the preservation of some foods such as fruits and vegetables in order to avoid an excessive build-up of the gas produced by their respiration up to deleterious levels, which could result in cell membrane damage and physiological injuries to the product [68]
Optical Analysis
The complete UV-Vis transmittance spectra and T R % for the neat Biofilm and all the produced multilayer films are depicted in Figure 4 and Table 5, respectively.As it can be observed, almost zero transmittance was registered in the UVC (200-280 nm) and UVB (280-320 nm) ranges for all the investigated samples, while transmittance remained limited in the UVA (320-400 nm) and visible range, with maximum values around 10-20% at 800 nm, which represents the upper limit of the visible range.The addition of the transparent m-PVOH and PLA layers yielded a slight improvement in the transparency, from 9.6% to 17% and 17.1% for the Biofilm substrate, and the Biofilm/m-PVOH and Biofilm/m-PVOH/PLA + 0%wax films, respectively.A drop in the films' transparency was observed by incorporating the wax in the PLA coating layer, the extent of which was more significant with increasing wax concentration.
Optical Analysis
The complete UV-Vis transmittance spectra and TR% for the neat Biofilm and all the produced multilayer films are depicted in Figure 4 and Table 5, respectively.As it can be observed, almost zero transmittance was registered in the UVC (200-280 nm) and UVB (280-320 nm) ranges for all the investigated samples, while transmittance remained limited in the UVA (320-400 nm) and visible range, with maximum values around 10-20% at 800 nm, which represents the upper limit of the visible range.The addition of the transparent m-PVOH and PLA layers yielded a slight improvement in the transparency, from 9.6% to 17% and 17.1% for the Biofilm substrate, and the Biofilm/m-PVOH and Biofilm/m-PVOH/PLA + 0%wax films, respectively.A drop in the films' transparency was observed by incorporating the wax in the PLA coating layer, the extent of which was more significant with increasing wax concentration.The films exhibited significant potential in shielding UV-light, as highlighted by the UVA-and UVB-blocking % values reported in Table 5, which are equal or higher than 92.0% and 99.2% for all the samples, respectively.The UV-light screening ability of these films plays a fundamental role in preserving the chemical, physical, and biological properties of food products, since UV radiation primarily induces light-induced oxidation in protein-rich and high-fat foods [69].The films exhibited significant potential in shielding UV-light, as highlighted by the UVA-and UVB-blocking % values reported in Table 5, which are equal or higher than 92.0% and 99.2% for all the samples, respectively.The UV-light screening ability of these films plays a fundamental role in preserving the chemical, physical, and biological properties of food products, since UV radiation primarily induces light-induced oxidation in protein-rich and high-fat foods [69].
Conclusions
This study aimed to investigate the transport properties of high-barrier biodegradable films made by the deposition of a double coating layer of m-PVOH and PLA + EBS wax on a PBAT/PLA (Biofilm) substrate, exploring the effect of relative humidity on barrier performance and the suitability of the structures for MAP applications.The deposition of the first coating of m-PVOH on the Biofilm web layer resulted, in the range of 30-50% RH, in a one-order-of-magnitude decrease in the water vapor permeability compared to the pure substrate, from ~67.0 to 3.5 g mm/ m 2 d bar for the Biofilm and Biofilm/m-PVOH samples, respectively.The water vapor barrier properties remained constant up to 60% RH, above which the plasticization of the m-PVOH chains prevailed, entailing a faster diffusion.This result is of particular significance with respect to the unmodified PVOH grades, which showed no water vapor resistance and a fast increase in the P WV already at RH ≥ 30%.
The films' barrier performances were not considerably affected by the addition of the second PLA layer up to 85% RH.Nonetheless, the incorporation of the wax in the PLA coating was effective in slowing down the permeation phenomenon by hindering the permeation of water vapor molecules and partially inhibiting the swelling of the sensitive PVOH matrix.The maximum decrease in P WV with respect to the Biofilm/m-PVOH/PLA film occurred for the Biofilm/m-PVOH/PLA + 5%wax sample and was equal to 37% at 70% RH.Further incorporation of higher percentages of EBS wax into the PLA layer did not significantly alter the water vapor barrier performance of the double-coated films.
The trends of P WV evolution by increasing RH were also fitted using an exponential mathematical model.The numerical results obtained agreed with the empirical data, highlighting the efficacy of the model in predicting the water vapor permeability of the multilayer films over a wide range of relative humidities, and its utility for the characterization, designing, and optimization of their barrier performance.
For what concerns PO 2 , a dramatic increase was observed at 85% RH compared to 50% RH.Also, in this case, the wax incorporation in the PLA coating was effective in decreasing the gas permeation; for the samples loaded with different wax concentrations, a PO 2 reduction comprised between 15% and 22%, with respect to the Biofilm/m-PVOH/PLA, was observed.
The permselectivity values of the biodegradable, double-coated films were found to be comparable to those of conventional, fossil-based polymeric films used for the MAP storage of several categories of foods, from fresh-cut agricultural products to cheese.
These results, coupled with those of the MAP tests, endorsed the potential of the films to be used as sustainable alternatives for the preservation of foods stored under modified atmosphere packaging, with the possibility to tailor the optimal film layout on the basis of the target food's shelf-life.
Figure 4 .
Figure 4. UV-Vis transmittance spectra of all the biodegradable films.
Figure 4 .
Figure 4. UV-Vis transmittance spectra of all the biodegradable films.
Table 1 .
List of film samples and thickness of coating layers.
Table 2 .
PWV at 30%, 50%, 70%, 85%, and 90% RH for all the biodegradable films and percentage decrease in PWV (ΔPWV%) for the double-coated films with wax with respect to the Biofilm/m-PVOH/PLA film, at the different relative humidities.
Table 2 .
P WV at 30%, 50%, 70%, 85%, and 90% RH for all the biodegradable films and percentage decrease in P WV (∆P WV %) for the double-coated films with wax with respect to the Biofilm/m-PVOH/PLA film, at the different relative humidities.
Table 3 .
Model parameters (a, b, and c) and coefficients of determination (R 2 ) for all the biodegradable films.
Table 4 .
CO 2 permeability and permselectivity values of all the biodegradable films.
Table 5 .
Percent transparency values and UVA-and UVB-blocking % values of all the biodegradable films. | 7,593.6 | 2023-10-01T00:00:00.000 | [
"Materials Science",
"Environmental Science"
] |
A Nasal Brush-based Classifier of Asthma Identified by Machine Learning Analysis of Nasal RNA Sequence Data
Asthma is a common, under-diagnosed disease affecting all ages. We sought to identify a nasal brush-based classifier of mild/moderate asthma. 190 subjects with mild/moderate asthma and controls underwent nasal brushing and RNA sequencing of nasal samples. A machine learning-based pipeline identified an asthma classifier consisting of 90 genes interpreted via an L2-regularized logistic regression classification model. This classifier performed with strong predictive value and sensitivity across eight test sets, including (1) a test set of independent asthmatic and control subjects profiled by RNA sequencing (positive and negative predictive values of 1.00 and 0.96, respectively; AUC of 0.994), (2) two independent case-control cohorts of asthma profiled by microarray, and (3) five cohorts with other respiratory conditions (allergic rhinitis, upper respiratory infection, cystic fibrosis, smoking), where the classifier had a low to zero misclassification rate. Following validation in large, prospective cohorts, this classifier could be developed into a nasal biomarker of asthma.
Supplementary
: Visual description of the machine learning pipeline used to select predictive features (genes) and develop classification models based on them in the RNAseq development set. By considering 100 splits of the development set into training and holdout sets (dotted box), many such models were evaluated for classification performance and then compared statistically using Friedman and Nemenyi tests. From this comparison, the best combination of predictive genes and global classification algorithms was determined, which was then executed on the development set to train the final asthma classifier model. This model was applied to an independent RNAseq test set and external microarray-derived cohorts with asthma and other respiratory conditions for final evaluation. Given a training set, this component used a 5x5 nested (outer and inner) cross-validation (CV) setup to select sets of predictive features (genes). The inner CV round was used to determine the optimal number of features to be selected, and the outer round was used to select the set of predictive genes based on this number, thus reducing the cumulative effect of potential sources of overfitting. The selection of features itself was performed using the Recursive Feature Elimination (RFE) algorithm in combination with wrapper Logistic Regression and SVM with Linear kernel classification algorithms. precision, recall, positive predictive value, and negative predictive value are summarized. F-measure, which is a harmonic (conservative) mean of precision and recall that is computed separately for each class, provides a more comprehensive and reliable assessment of model performance when classes are imbalanced, as is frequently the case in biomedical scenarios. Figure 6: Performance of permutation-based random classification models in test sets of independent subjects with asthma and controls. To determine the extent to which the performance of the classifier could have been due to chance, 100 permutation-based random models were obtained by randomly permuting the labels of the samples in the development set and executing each of the feature selection-global classification combinations on these randomized data sets in the same way as described above for the real development set. These random models were then applied to each of the asthma test sets considered in our study, and their performances were also evaluated in terms of the F-measure. Figure 7: Performance of permutation-based random classification models in test sets of independent subjects with non-asthma respiratory conditions and controls. 100 permutation-based random models were obtained by randomly permuting the labels of the samples in the development set and executing each of the feature selection-global classification combinations on these randomized data sets in the same way as described above for the real development set. These random models were then applied to these test sets, and their performances were also evaluated in terms of the F-measure.
Supplementary Figure 8: Distribution of DESeq2 FDR values of differentially expressed genes in the asthma classifier (blue bars) vs. other genes in the RNAseq development set (coral bars)
. The Y-axis shows the probability of a gene having a -log10(FDR) value in the corresponding bin. This plot shows that the genes in the asthma classifier were likely to be more differentially expressed, i.e., higher -log10(FDR) or lower differential expression FDRs, than other genes in the development set. Race Caucasian n/a n/a n/a 57% 82% 89%
Supplementary
African American n/a n/a n/a 17% 9% 0% Hispanic n/a n/a n/a 13% 0% 0% Asian/Other n/a n/a n/a 13% 9% 11% n/a n/a n/a n/a PC20 (mg/ml) n/a n/a n/a 4. *For Asthma2, data that the authors deposited in GEO GSE46171 are a subset of their published results [32]. GSE46171 has data for 16 of the 23 subjects with controlled asthma, 7 of the 11 subjects with uncontrolled asthma, and 5 of the 9 controls reported in the authors' publication [32]. We indicate the number of subjects with publically available data (GSE46171) that were used in our analyses. The summary statistics shown are drawn from the authors' publication on their reported sample. † Median (range)
Supplementary Table 4: Characteristics of the external cohorts with non-asthma respiratory conditions and controls used for testing the asthma classifier
Allergic Rhinitis [35] GEO GSE43523* URI Day 2 [32] GEO GSE46171^ URI Day 6 [32] GEO GSE46171^ Cystic Fibrosis [36] GEO GSE40445 Smoking [12] GEO GSE8987 Results are number (%) or mean (SD) unless otherwise indicated *Data that the authors deposited in GEO GSE43523 are a subset of their published results [35]. GSE43523 has data for 7 of the 15 subjects with allergic rhinitis, and 5 of the 13 controls reported in the authors' publication [35]. We indicate the number of subjects with publically available data (GSE43523) that were used in our analyses. The summary statistics shown are drawn from the authors' publication on their reported cohort. ^E ach subject provided a URI and control sample. The data that the authors deposited in GEO GSE46171 are a subset of their published results [32]. GSE46171 has data for 6 of the 9 healthy subjects reported in the authors' publication who provided samples during URI, and 5 of the 9 healthy subjects who provided samples after resolution of their URI [32]. We indicate the number of subjects with publically available data (GSE46171) that were used in our analyses. The summary statistics shown are drawn from the authors' publication on their reported cohort. † Median (range)
Gene
Annotation References ALOX15B Member of lipoxygenase family whose members can affect bronchiolar constriction, cytokine secretion, and immune cell migration. The ALOX15B isoform may regulate cytokine secretion by macrophages and macrophage differentiation.
C3
Central role in classical and alternative complement pathway system activation; downstream effects include smooth muscle contraction, vascular permeability, histamine release.
CD177
Glycoprotein expressed by neutrophils. Used as a cell surface marker in studies of IL-17RB granulocytes in asthma.
CDH26
One of eight genes targeted in a candidate gene study of asthma; negative results.
CDHR3
Variant in this gene associated with rhinovirus-induced wheezing and rhinovirus C illness. May function as a rhinovirus C receptor. 6, 7
CDKN1A
Mediator of microRNA-221-modulated airway smooth muscle hyperproliferation in cell culture studies of severe asthma. 8 CEBPD CEBPD gene expression in bronchial specimens from asthma subjects associated with asthma susceptibility and inhaled corticosteroid treatment. 9
CLEC7A
Expression on CD11b+ dendritic cells plays a role in house dust mite-induced allergic airway inflammation in murine models. 10
CPA3
Mast cell mediator whose gene expression in epithelial brushings is upregulated in mild asthma and suppressed by corticosteroids in moderate asthma.
CYFIP2
One of 237 candidate genes targeted in a candidate gene study of asthma in Mexicans.
CYP1B1
One of 25 candidate genes targeted in a candidate gene study of xenobiotic-metabolizing enzymes in asthma among Russians. 13
DEFB1
Protein level elevated in induced sputum from severe asthmatics vs. controls. Also studied as one of 44 candidate genes in a candidate gene study of innate immune pathways in asthma and eczema among children from Boston and Connecticut.
DUSP1
Expression in bronchial epithelial cell culture increased by dexamethasone, leading to suppression of p38 MAPK signaling and cytokine inhibition. 16
ESR1
SNPs in this gene associated with bronchial hyperresponsiveness and FEV1 decline, especially in females. miRNAs may impact pathogenesis of dust mite-induced asthma via regulation of ESR1. [16][17][18] FOS Encodes a transcription factor involved in anti-inflammatory activity of steroid action in asthma. 19, 20
GSTT1
Modifies the impact of air pollution exposure on asthma. 20
LPHN1
SNP in LPHN1 associated with asthma and found to regulate airway smooth muscle cell adhesion and proliferation in vitro 22
LTBP1
siRNA knockdown of LTBP1 inhibited TGFbeta1 release in airway fibroblasts from asthma subjects. 23
MMP9
Released by neutrophils in allergic asthma subjects and in murine models of asthma. 11,24,25 NMU Neuropeptide that amplifies Type 2 innate lymphoid cell-driven allergic lung inflammation in murine models. 26
S100A7
Antimicrobial peptide induced by IL-22 in T-cell lines derived from lung biopsy specimens of asthmatic subjects.
S100A8
Anti-apoptotic protein detected in supernatant of neutrophils treated with house dust mite extract and elevated in BAL from asthmatic vs. control subjects. 28
SCD
Inhibition of SCD in mice promoted airway hyperresponsiveness. SCD1 expression reduced in bronchial epithelial cells from asthma subjects vs. controls. 29 SCGB1A1 Levels in induced sputum higher in subjects with severe asthma vs. mild-moderate and healthy controls. BAL levels of SCGB1A1 correlated with epithelial detachment in bronchial biopsies. 30
SEMA5A
One of 11 genes mapped by 1000 of the top SNPs shared across European, African, and Hispanic populations in a rank-based analysis of shared genetic factors for asthma. 31 SERPINB3 In vitro-polarized Th2 cells from subjects with grass pollen allergy expressed higher mRNA levels of this serine protease inhibitor relative to CD27+CD4+ cells. Mediates mucus production in murine models of asthma. 32,33 SERPINE2 Selected SNPs in this gene were associated with asthma and related traits.
SLC26A4
Up-regulated in airway epithelial cells in association with mucus overproduction in murine models. 36, 37
SPRR1A
Intratracheal inoculation of mice with IL-13 induced more gene expression of SPRR1A than inhalation of IL-4. 38 TFPI TFPI level concentration studied in 17 subjects with asthma during early and late stages of reaction. 39
TPSAB1
Mast cell biomarker used to subtype sputum subtypes in a study of eosinophilia and corticosteroid response in asthma. 40
TPSB2
Encodes mMCP-6, which is required for airway hyperresponsiveness in murine models of asthma. | 2,373 | 2017-06-07T00:00:00.000 | [
"Computer Science",
"Medicine"
] |
Data Management Plan for Moore Investigator in Data Driven Discovery Grant
This Data Management Plan (DMP) was created for Ethan White's Moore Investigator in Data Driven Discovery award. It describes the management and sharing of all data and code associated with Gordon and Betty Moore Foundation grant GBMF4563 (White 2014). This includes raw data collected as part of the proposal, data compilations, and software. Research associated with this award is related to data-intensive approaches to studying ecological systems and the development of software for automating the cleaning, restructuring, and integration of heterogenous data sources.
Data Description
The majority of the data involved in this project will be pre-existing data collected by other individuals and organizations.Much of this data is already openly available.In cases where it is not we will work with the relevant stake-holders to make as much of that data available as possible.All pre-existing data that my group has collected have already been made openly available.
We will likely collect some additional data during this project by compiling existing data from the literature and other publicly available sources.This data will typically be tabular data involving information on ecological systems and the traits of individual organisms.It will be stored in standard data formats such as csv and HDF5.
In terms of data products more broadly (i.e., including databases, analyses, and software tools) this project will generate both project specific computational analyses and general use computational tools.The computational tools will be software designed to make working with the diversity dimension of big data easier by automating the cleanup and restructuring of data sets and the combination of multiple data sets for analysis.
Metadata will be provided for all data products.In the case of newly collected data this will include both well written documentation and Ecological Metadata Language files.In the case of software it will include documentation in the source code as well user focused documentation and tutorials.For data provided by other individuals or institutions we will help that individual or institution develop publicly available metadata, even in cases where the data itself is not public.
As described above, much of the actual data used in this project will be owned by other individuals and organizations.Data compilations that we build and software that we develop will be owned by the PI and the University of Florida, but be made publicly available under open source and open access licenses.
Data Management
During development the data, software, and associated metadata that we produce will be stored on GitHub or another Git-based hosting service.It will be automatically backed up nightly to both a lab server and to the University of Florida High Performance Computing Center's Replicated Long-Term Storage.Data and software access and distribution will be managed using Git and the centralized Git host.Following development data and software will be archived in more permanent locations (see Data Sharing).
Data obtained from other individuals and organizations will be stored in PostgreSQL databases on a lab server and backed up nightly to the University of Florida High Performance Computing Center's Replicated Long-Term Storage.Some of this data is proprietary (e.g., the Audubon Society's Christmas Bird Count, eBird, and the North American Butterfly Associations Butterfly Count).Permission to use this data is obtained through the use of Memorandums of Understanding or other data use agreements.These agreements typically prevent the redistribution of the raw data.There is also data that is publicly available but lacks fully open licenses.In many cases this data cannot be redistributed, but the software we will develop circumvents this issue by automating the downloading and installation of this data.As such, we can provide easy to use copies of assembled databases involving this type of data by distributing open source code that will download and assemble the datasets locally.
All personel on the project will be responsible for entering and maintaining data archives.The data and software archives will be archived indefinitely by using existing archives that are intended to be self-sustaining and are backed by CLOCKSS (see details in Data Sharing).
Newly collected data will be independently entered from by two separate individuals and then compared to identify any errors (i.e., double data entry or two pass verification) and then entered into the main database.Once in the database, the data will be checked again using an automated data validation step with constrained data values to flag errors (e.g., outside of range limits, inconsistent taxonomy, or incorrect data types).Quality control and assurance for software will be accomplished using a combination of code review, unit testing, and integration testing.Code review will be conducted prior to changes being accepted into the core repository and tests will be run automatically using Travis CI or a similar continuous integration service.
Data Sharing
Potential data users for the data we will compile include ecologists, environmental scientists, managers, and conservation planners.Potential users for our software include anyone working with long-tailed data.
Data and software will be developed in the open, with new data and changes to software being made public (typically via GitHub) in near-real-time.Formal releases of raw data and software products will be conducted as these products reach meaningful milestones.At each of these releases the relevant product will be archived in a long-term repository that includes meaningful assurances that the archive will be available even if the repository shuts down (e.g., CLOCKSS).Documentation, tutorials, and metadata for using both data and software will be archived with the relevant product.Examples of data repositories we will use include Dryad, Figshare, Ecological Archives and DataONE.Examples of software repositories include Dryad, Zenodo and Figshare.The original software and data sources will also be maintained on theGit-repository host to facilitate further development.The goal of these combined efforts is to maintain archives of all data and software products indefinitely.By using centralized archives hardware maintainence is not required.All software will be released under permissive open source licenses (MIT orBSD).All data will be published using the CC0 Public Domain Dedication if allowed by the repository, with CC-BY as a fallback if the repository does not allow CC0.The only exception to the use of these permissive licenses will be incases where data or code produced by other groups needs to be included that involves non-permissive licensing.If this is the case we may be required to use GPL or CC-BY-SA licenses.
In cases where we compile data from other sources we will cite all of the relevant sources and, where appropriate based on substantial contributions, include original data collectors as authors on relevant data products.
For major releases of major data products we anticipate publishing associated Data/ Software papers for referencing and to help advertise the availability ofthe data product as broadly as possible. | 1,534.4 | 2016-04-10T00:00:00.000 | [
"Environmental Science",
"Computer Science"
] |
Olistostrome and the mesozoic tectonic of the bantimala complex, South Sulawesi
The study aims to determine the presences and spreads of the olistostrome as one components of the tectonic complex of Bantimala area. The basement rocks of Bantimala area is a metamorphic rocks, which are unconformably overlain by Balangbaru Formation and radiolarian chert. Beneath in between cherts and basement rocks are presence breccia schists which give rise to various presumptions and interpretations of the environment and conditions of formation as submarine deposit, ideally cherts in the area underlain by oceanic crust.Schist breccias presence underneath of cherts in the Bantimala Complexes were suggested an olistostrome deposit. It was characterized by poorly sorting, unfoliated, shows deformed textures and composed olisthtolits which are embedded in sandy matrices, and in the cherts are presence layer sandstones and schist fragments. Olistostrome is sedimentary deposit as preserved in the trench, they will give us an interpretation that prior to formed cherts in the Bantimala area, initially tectonic subduction activity which are deforming and brecciation of the basement rocks subsequently as the constituent material of olistostrome.
Introduction
The olistostrome of Bantimala Tectonic Complexes is likely related to Western Pacific plate has subducted to Asian Continent during Mesozoic, the interaction both of the plates were caused deformation, metamorphism and olistotrome deposits. These situation is reasonably allow to occur a debris fall due to steep slope in the part of trenches as suggested by [1], [2], their fragments possibly are eroded debris materials from continent and oceanic plates, they have been mixed as a chaotic sedimentary deposits, its immediately covered by deep marine sediments of radiolarian cherts.
Our side observations that schist breccias in the study area is an olistostrome deposits characterized by present vary of olistolith schists, stratigraphically it overlain by radiolarian cherts. So far, the study of olistotrome in the area has never been studied, but previous studies it was indicated a mélange [5] thought a chaotic sedimentary processing which is more appropriately considered as olistostrome.
Methods
This study emphasizes field observations and petrographic analysis. Field observations was performed stratigraphic detail by measuring section and cross section along Patteteyang river, observing the relationship between the lithologic unit and indication deformation structures in the field and rocks sampling. Further observations were made followed by petrographic analysis to determine the texture and rock composition for all components of rocks, especially for olistolith and matrices.
Regional geology and tectonic setting
Tectonically the Bantimala Complexes is still contentious, including presence of rocks faragments and interrelationship each other such as presence of mélanges, schist basement rocks, rocks continent affinity, radiolarian cherts and Paleogene volcanics.
Geological setting
The basement rocks were exposed in the Bantimala area is part of shelf margin of East Kalimantan which was separated since Miocene coincided with the occurrence of Makassar Strait. The basement rocks is called the Bantimala Tectonic Complex where composed of: Triassic metamorphic rocks namely glaucophane schists, hornblende-mica schists, eclogites, granulites, phyllitesand metaquarzites [3]; Jurassic-Cretaceous of mélange composed of quartz schists, metacherts, metabasalts. Cretaceous sedimentary rocks which composed of fine-siliceous shales, sanstone, mudstone and radiolarian cherts.
Ophiolite blocks consisting of harzburgites and serpentinites where occur through out tectonic obduction which covered or structural contact with the Tertiary rocks in the area, whereas sedimentary type of continental shelf margin are Cretaceous of typical flysch sedimentary rocks of Balangbaru-Paremba Formations.
Tectonic setting
The Bantimala Complex is formed in two tectonics phases, the first is subduction system of oceanic plate that took place during the Mesozoic to Tertiary. The second is ophiolite obduction system that took place during Tertiary to Quaternary. The presence of high grade of metamorphic rocks associated with system during pre-Cretaceous (Triassic -Jurassic), it may cause by the West Pacific plate subducted to the East Kalimantan continent, which was produced accretion complex on the eastern shelf of Asian Continental plates. During Jurassic to Cretaceous the accretion complex has deformed to produce a mélange rocks [7][8]. During the Early Cretaceous occurred brecciation in the accretion complexes formed breccia -schist [5] or here called olisostrome [9], while continuing deposited chert and contemporaneously with deposition in the trench of terrigenic sediment of Balangbaru ( Figure 2).
More over the Cretaceous of Bantimala Complexes have been situated by some of previous study such as [10] and [11], according to them that the establishment of a high-pressure metamorphic rocks associated low-grade metamorphic rocks, mélange and ultramafic in this region are the results of system subduction of oceanic crust into the shelf margin of continental plates in the Jurassic to the Early Cretaceous (114 to 132 Ma). Based on calculations of pressure -temperature of rocks garnetglaucophaneresulted a temperature of 580 -640 °C and pressures 18-24 kbar, occurs at depth 65-85 km [5]. Furthermorethe breccia of Bantimala Complexis unconformably overlain by the Radiolarian chert which ranging from Albian to Cenomanian in age (± 100 Ma) [12].
Physical properties of Olistostrome
Lithologic characterized of olistostrome of Bantimala Complex could be described into two aspects i.e. based on components and sedimentary types.
Component Fabrics
Olistostrome component in the Bantimala Complex shows various clast of material component swhich is indicated difference origin (polymictic), it composed by deformed and reworked materials derived from subduction activity which consisting of schist, quartzite and gneiss, serpentinite, metachert, the components are vary in grain size ranging from centimeter to several meter, poorly sorting and angular to subangular. Fragments floating in reddish chert matrices suggesting retransported material within a trench, deformed textures indicated by lenses, fracturing, shows pseudofoliation with slipped surface (Figure 9 and 10).
Deposition type Olistostrome
Quartz muscovite schist in one complex with glaucophane schists, granulits, eclogites and radiolarian cherts, as an indication of sediment avalanche or slope slumping at a relatively steep slope suggested rocks that preserve from subduction between the continental plates with oceanic plates, where olistostrome interbedded in part the bottom of the radiolarian cherts [13]. To mixing materials in chert clast and fining upwards structures, cobble -granule in grain sizes, indicating that they were preserve in a deep marine or trenches on subducting tectonic conditions, based on the above determination so that the olistostrome of Bantimala Complex is categorized as subduction type.
Sediment Physiography
The appearance of interbedded between the sandstones and radiolarian chert as the top of the olistostrome, gradational structure components in the western part of the study area including fluxo deposit. Based on deposit type in the eastern part and the contact between the radiolarian chert coarse grainsof olistolith indicated proksima type deposit ( Fig. 9 and 10). Based on the spread and characteristics of layering can be interpreted the depositional direction of olistostrome from west to east. The presence of serpentinite -jadeite blocks of as exotic block in olistostrome is separate phenomenon to be determine in the future work. Body of olistostrome in Bantimala Complex suggested as a lensis in radiolarian cherts unit.
The occurrences of olistotrome
The process of subduction of the Pacific plate along continental margin of Kalimantan during Jurassic, initially the process of forming olistostrome of Bantimala. During that period waspreserved tectonic deformation, brecciation and metamorphism of the two plates subducted each other, which is accompanied by the formation of deep trenches as the depositional environment. Plate continental and oceanic plates that have undergone at low grade metamorphism and brecciated forming blocks as olistoliths that shows deformation textures. Based on experiments [14] and [15] on the slopes of the critical condition allow submarine landslide occurred in the trench, the material falling as debris flow or slumping and widespread deposited on the ocean floor which its mixture with clast of continental and oceanic components, [16] and [17], (Fig. 11).
Based on their appearance in the field, olistostrome layer at Pateteyang River (Bantimurung Village), there are at least four cycle of avalanche or debris deposition of material inter bedded with cherts, namely: -The first is preserve debris avalanches and slumping, forming as rock blocks with a thickness of ± 340 m olistolith sizes between 1-150 cm, angular and boudinage. Components consist of chlorite schist, mica schist, amphibolite schist, and quartzite gneiss. -The second, before the second deposition is preceded by the deposition of a thin layer chertsof 20 cm thick and mixed with coarse grains of schist material followa debris avalanche of gravel-blockof schist clasts (2-40 cm), relatively smoother than the first avalanche, with a thickness ± 150 cm. -The third is also inter bedded by chert layers 60 cm in thick. The third layer is composed of schist clast with 25 cm thick.
-The fourth, deposition above of layer chert (120 cm thick), a thin layer of schist clast ± 20 cm thick and (Fig. 12). The composition and structure of the sediment grain size olistostrome shows fining upwards as debris flow sediment or turbidity (Table 1).
The stratigrphy position of olistostrome by Wakita et al. (1996) is above the wedge with Radiolarian chert and intercalated with Balangbaru sandstone, and there are also schists interbedded chert on the bottom. Breccia schist by Wakita et al. (1996) and the authors include as olistostrome deposit. Stratigraphy position hardly visible in the field but by outcroping thin layer of ± 50 cm ( Figure 10) olistostrome interbedded in chert in the eastern part of the area of research, it can be interpreted that the (1996) that is deposited in the same time with the slump sedimentation system on the bottom layer of chert. The stratigraphy relationship breccia schist in this case olistostrome is unconformable overlaying cherts (Wakita, et al., 1994), conversely based on the age of the fossil in radiolarian chert, breccia or sandstone is conformable. It is also supported by the appearance of the field that showed gradational from bottom-top with fining upward structures and almost breccias composed by schist clast indicated same environment and sedimentary sources suggested that the relationship between radiolarian chert and breccia schists is conformable. | 2,277.8 | 2017-01-01T00:00:00.000 | [
"Geology"
] |
Topography generation by melting and freezing in a turbulent shear flow
Abstract We report an idealized numerical study of a melting and freezing solid adjacent to a turbulent, buoyancy-affected shear flow, in order to improve our understanding of topography generation by phase changes in the environment. We use the phase-field method to dynamically couple the heat equation for the solid with the Navier–Stokes equations for the fluid. We investigate the evolution of an initially flat and horizontal solid boundary overlying a pressure-driven turbulent flow. We assume a linear equation of state for the fluid and change the sign of the thermal expansion coefficient, such that the background density stratification is either stable, neutral or unstable. We find that channels aligned with the direction of the mean flow are generated spontaneously by phase changes at the fluid–solid interface. Streamwise vortices in the fluid, the interface topography and the temperature field in the solid influence each other and adjust until a statistical steady state is obtained. The crest-to-trough amplitude of the channels is larger than approximately 10$\delta _{\nu }$ in all cases, with $\delta _{\nu }$ the viscous length scale, but is much larger and more persistent for an unstable stratification than for a neutral or stable stratification. This happens because a stable stratification makes the cool melt fluid buoyant such that it shields the channel from further melting, whereas an unstable stratification makes the cool melt fluid sink, inducing further melting by rising hot plumes. The statistics of flow velocities and melt rates are investigated, and we find that channels and keels emerging in our simulations do not significantly change the mean drag coefficient.
Introduction
Melting and freezing processes between ice and water play an important role in the environment. For instance, the melting of ice shelves, i.e. floating glacial ice, can lead to reduced buttressing of the grounded polar ice sheets and increased sea-level rise (Pritchard et al. 2012;Rignot et al. 2013;Kennicutt 2019), while freezing of high-latitude oceans by a cold atmosphere results in sea-ice formation, increased albedo and increased ocean salinity through brine rejection (Wells, Hitchen & Parkinson 2019). Icebergs, ice shelves and sea ice are kilometre-scale objects with long lifetimes but their evolution is controlled by heat and salt fluxes across millimetre-thin ice-water boundary layers, which fluctuate rapidly (Dinniman et al. 2016). The front of a marine-terminating glacier can melt as fast as several metres per day horizontally (as recently reported for the LeConte Glacier, Sutherland et al. 2019), but an ice shelf around Antarctica typically melts at a rate of only a few centimetres per day or less (Dutrieux et al. 2014). On the other hand, ocean currents are most often turbulent and exhibit temporal variabilities on fast time scales of just a few seconds (Davis & Nicholls 2019), such that phase changes between ice and water are multi-physics phenomena with large scale separation.
An important consequence of phase changes is that topographical features can emerge at the ice-water interface when the rate of melting and freezing is spatially variable. Basal channels (Stanton et al. 2013;Gourmelen et al. 2017) and terraces (Dutrieux et al. 2014) have been observed at hundreds-of-metre to kilometre scales under ice shelves, the underside of icebergs exhibit ablation channels at the metre scale and scallops at the tens-of-centimetre scale (Hobson, Sherman & McGill 2011), and, more generally, rough features can be seen from the centimetre scale to tens-of-metre scale under sea ice (Wadhams, Wilkinson & McPhail 2006;McPhee 2008;Lucieer et al. 2016) and up to the kilometre scale under ice shelves (Nicholls 2006). The interplay between flow dynamics and phase changes leading to the generation and persistence of topographical features in the environment is of fundamental importance. The presence of topography can significantly affect the long-term flow dynamics as well as the average melting or freezing rate of the ice boundary, as suggested by, for example, the large spatial variability of melting of basal terraces (Dutrieux et al. 2014) and recent laboratory experiments on ice scallops (Bushuk et al. 2019).
Buoyancy forces play an important role in the coupling between phase changes, flow dynamics and topography generation. Buoyancy forces can be stabilizing or destabilizing depending on the relative orientation between the gravitational acceleration vector and the density gradient. In polar seas, cold and fresh melt water near the ice boundary is lighter than the surrounding water, such that buoyancy forces are restoring below a horizontal ice boundary (e.g. below an ice shelf) and drive upwellings along a vertical ice face (e.g. at the front of a marine-terminating glacier). In a cold freshwater system the thermal expansion coefficient of water is negative, i.e. it is negative for temperatures 0 • C < T < 4 • C at atmospheric pressure (Thoma et al. 2010), such that buoyancy forces are stabilizing for water under an ice cover (e.g. as in a frozen lake) but destabilizing for water above ice (e.g. for a supraglacial lake or river).
For a system dominated by destabilizing buoyancy forces, the interplay is strong between fluid dynamics and phase topography. In a thermally stratified fluid with finite depth below a solid phase, the unstable density stratification sets up a large-scale circulation known as Rayleigh-Bénard convection with alternating warm upwelling and cold downwelling regions. The warm upwellings drive stronger melting than the cold downwellings, such that a topography can emerge from spatially variable heat fluxes. The topography enhances the large-scale circulation, such that a positive feedback is obtained and the flow dynamics and solid boundary become phase locked (Rabbanipour Esfahani et al. 2018;Favier, Purseed & Duchemin 2019). Dissolution of a phase boundary, i.e. with phase changes driven by concentration gradients rather than temperature effects, in a gravitationally unstable fluid can also lead to convective motions and the generation of three-dimensional topography (even in the absence of a large-scale circulation), as shown by experiments (Kerr 1994;Sullivan, Liu & Ecke 1996) and numerical simulations (Philippi et al. 2019). Streamwise patterns also emerge in dissolution experiments when the phase boundary is not perpendicular to gravity but inclined (Allen 1971;Cohen et al. 2020).
Despite the existence of many studies on pressure-driven and shear flows (Kelly 1994;Zonta & Soldati 2018), the possibility for topography to emerge between a horizontal boundary layer flow and a solid phase, i.e. perpendicular to gravity, is not well understood, at least compared with the case of topography generation by Rayleigh-Bénard convection. Boundary layer flows strongly affected by shear, such as under ice shelves, are yet as common (if not more) as buoyancy-driven flows in the environment, such that there is significant interest in predicting their ability to generate topographical features (or roughness) at horizontal ice boundaries and the impact the sustained topography can have on overall melt rates. Using laboratory experiments, Gilpin, Hirata & Cheng (1980) demonstrated the existence of an interfacial instability and the generation of ripples at an ice boundary below a horizontal turbulent boundary layer water flow. The experiments had a modest unstable density stratification owing to the negative thermal expansion coefficient of freshwater at low temperatures (Toppaladoddi & Wettlaufer 2019). However, Bushuk et al. (2019), who conducted an experiment similar to Gilpin et al. (1980), argued that the buoyancy forcing can be neglected in the large velocity regime of the experiments, thus confirming the possibility for topography generation in the absence of vertical convection. The necessary condition for an interfacial instability to develop, regardless of the type of density stratification, is that the maximum of mass transfer from the solid to the fluid due to a heat flux or concentration gradient (resulting in ablation) at the boundary be shifted by −90 • to +90 • compared with the maximum (crest) of boundary topography (Hanratty 1981). Recently, Claudin, Durán & Andreotti (2017) demonstrated that such a shift is possible for a horizontal neutral turbulent flow and proposed a saturation mechanism for the finite amplitude of two-dimensional scallops. Three-dimensional effects and buoyancy forces are expected to play an important role in topography generation and melting rates but were not considered in the study of Claudin et al. (2017), which also relied on parameterized flow nonlinearities. Thus, additional efforts are necessary to improve our understanding of the physical mechanisms leading to the generation and saturation of phase topography by horizontal shear flows.
Here, we investigate the generation of topography at a phase boundary adjacent to a shear flow affected by buoyancy via direct numerical simulations. We focus on the case of an initially flat and horizontal solid, i.e. perpendicular to gravity, and investigate the influence of density stratification on the topography obtained and the coupled fluid-solid dynamics. Our numerical model solves the evolution of the fluid and solid phases simultaneously using the phase-field method. The phase-field method is a one-domain two-phase fixed-grid method that was originally developed by the metallurgy community for relatively smooth flows (Wang et al. 1993;Karma & Rappel 1998;Beckermann et al. 1999), but which was recently applied to the case of vigorous convective flows (Favier et al. 2019;Purseed et al. 2020). Other methods that simultaneously solve for the evolution of a fluid phase and a solid phase include the enthalpy method (Ulvrová et al. 2012), the level-set method (Gibou et al. 2007), the lattice-Boltzmann method (Rabbanipour Esfahani et al. 2018) and two-domain moving-boundary methods (Ulvrová et al. 2012). The main advantage of the phase-field method over these other methods is that it can be implemented relatively easily in any fluid solver.
Our study aims to contribute to the physical understanding of topography generation by shear flows at horizontal boundaries and the associated changes in mean melt rates, as most recently investigated theoretically by Claudin et al. (2017) and experimentally by Bushuk et al. (2019). Numerical constraints force us to consider an idealized set-up however, such that our fluid and solid phases are not exactly representative of water and ice. Notably, we assume that the fluid and solid have the same thermodynamical and transport properties, e.g. the same thermal conductivity, and we consider an anomalously warm fluid in order to minimize the time scale separation between the turbulent dynamics and generation of boundary topography. Due to computational constraints, the external flow in our simulations is also weaker than what may be expected for scallops to emerge (Claudin et al. 2017;Bushuk et al. 2019).
The main result of our paper is that topographical features spontaneously emerge at the ice-water interface due to uneven melting of the solid boundary by the shear flow. We investigate the effect of background density stratification and we demonstrate that the topography is dominated by keels and channels that are aligned with the direction of the mean flow in all cases.
We organize the manuscript as follows. In § 2 we describe the phase-field method, the dimensionless equations and the numerical method. In § 3 we present and discuss the direct numerical simulation results obtained for three different background stratifications. In § 4 we discuss the link between our results and geophysical applications and explain why we did not observe three-dimensional scallops. In § 5 we conclude. Finally, in appendices A-D, we provide additional details about the method and results.
Phase-field method
We investigate the generation of topography due to uneven melting and freezing at a fluid-solid interface. The solid is fixed and located above the fluid where a Poiseuille/channel flow develops due to an external pressure gradient (see figure 1). The initial thickness of the fluid (respectively solid) layer is H (respectively H/2), such that the channel full depth is 3H/2. The domain length (in the direction of the flow) is L x = 4πH and the transverse width is L y = 2πH. We define a Cartesian coordinate system (x, y, z) centred on the bottom of the channel with the z-axis vertically upward, i.e. opposite to gravity, and use superscripts ( f ) and (s) to denote variables in the fluid and the solid, respectively. The fluid velocity u ( f ) and pressure p ( f ) evolve according to the Navier-Stokes equations under the Boussinesq approximation. For simplicity, we assume that the solid and fluid phases have the same thermodynamical and transport properties, i.e. the same reference density ρ f , the same specific heat capacity at constant pressure c p and the same thermal conductivity k. Thus, the temperatures T ( f ) and T (s) evolve according to the same advection-diffusion (energy) equation, which turns into the heat equation where the velocity is zero. We consider a generic linear equation of state for the fluid, i.e. not specific to water, with the density related to temperature through with α the thermal expansion coefficient. For a pure component flow, the fluid-solid interface must be at the temperature of melting, denoted by T m , and the movement of the interface is governed by the Stefan condition, i.e.
where ρ s is the reference density of the solid, v n is the interface velocity in the direction normal to the interface and directed toward the solid phase (supported by unit vectorn), L is the latent heat of fusion per unit mass, q n is the heat flux in directionn, k s (respectively k f ) is the thermal conductivity of the solid (respectively fluid) and ∇ is the gradient operator (Worster 2000). We recall that we assume the same properties for the two phases, i.e. in our case k f = k s = k and ρ s = ρ f in (2.2). Note that the properties of water and ice are different under typical atmospheric pressure and near-freezing temperature conditions, i.e. ρ f ≈ 999 kg m −3 , c pf ≈ 4200 J kg −1 K −1 and k f ≈ 0.6 W m −1 K −1 , while ρ s ≈ 917 kg m −3 , c ps ≈ 2100 J kg −1 K −1 and k s ≈ 2.2 W m −1 K −1 . The relative differences are small however, i.e. within a factor of four or less, such that we do not expect fundamental differences between the physics in our model and real processes involving water and ice in nature.
Here we use a volume-penalization method (Angot, Bruneau & Fabrie 1999), which is a type of immersed boundary method, combined with the phase-field method, in order to solve for phase-change processes and the evolution of the variables in the fluid and the solid simultaneously. Specifically, we solve the Navier-Stokes equations in the Boussinesq approximation and the advection-diffusion (energy) equation for temperature combined with an equation for the fluid fraction φ, i.e. (u, v, w), p and T defined in both the fluid and the solid, i.e. assuming that the fluid and solid phases form a single domain, such that we can drop the superscripts ( f ) and (s) . In (2.3), ν is the constant kinematic viscosity, g is the standard gravity, Π is the imposed pressure-gradient force and κ = k/ρ f /c p is the constant thermal diffusivity; τ p , a, b and c are parameters related to volume penalization and the phase-field method, which we define later, andẑ andx are the unit vectors of the z-and x-axes, respectively. Note that the third term on the right-hand side of (2.3a) represents the buoyancy force. The fluid fraction, φ, also known as the phase-field variable or order parameter, satisfies a forced diffusion equation (2.3c) with parameters tuned such that φ transitions continuously from 1 in the fluid to 0 in the solid, across a diffuse interface whose thickness is artificial and must be smaller than all physical length scales in the problem, including the viscous length scale (cf. appendix A). The fluid fraction φ is introduced in the momentum (2.3a), energy (2.3b) and continuity (2.3c) equations, in order to modulate locally the importance of each physical process based on the component's phase. For instance, the last term on the right-hand side of (2.3a) is a linear (penalization) damping term, which is active in the solid but inactive in the fluid, while the second term on the right-hand side of the energy equation (2.3b) is a heat sink or source that represents the consumption or release of latent heat associated with melting or freezing. In the limit of an infinitesimally small diffuse interface thickness of the phase field, it has been shown that the dynamics of the fluid-solid interface governed by (2.3) converges to the exact Stefan conditions (2.2), and that the fluid velocity converges to 0 at the fluid-solid interface, thus mimicking a no-slip boundary. Here we multiply by φ the advective terms in (2.3a) and (2.3b), such that they are zero in the solid phase. Previous studies have used both damped and undamped advective terms and we discuss the impact of our choice on the results in appendix B.
Dimensionless equations
Equations (2.3) can be non-dimensionalized in order to identify the set of independent control parameters. Following previous studies (e.g. Favier et al. 2019), we use the thermal diffusive time scale τ κ = H 2 /κ as a normalizing time scale, i.e. the dimensionless variables, denoted by tildes, are defined as (x, y, z) = (Hx, Hỹ, Hz), t = τ κt , u = u κũ , (2.4a-f ) with u κ = κ/H the diffusion velocity scale, and ΔT = T b − T m is the temperature scale with T b the dimensional temperature on the bottom boundary. The time scale τ κ is particularly relevant for discussing the long-term dynamics of the system since temperature evolves in the solid through diffusion. We will use the shorter friction time scale to describe relatively rapid processes, such as convection in the fluid (see § 2.3). Substituting variables (2.4a-f ) into (2.3) and dropping tildes, we obtain the dimensionless equations The control parameters in (2.5) are the Prandtl number, Pr, which compares kinematic viscosity to thermal diffusivity, the centreline Reynolds number, Re, which compares the pressure-gradient force to viscous dissipation, the Rayleigh number, Ra, which compares buoyancy forces to viscous and thermal dissipation, and the Stefan number, St, which compares the available sensible heat to the latent heat. They are related to the physical parameters through The additional parameters Γ = τ p νH 2 /κ 2 , A = a/κ, B = b/(κ/H 2 ) and C are non-physical and prescribed based on numerical constraints of the volume-penalization and phase-field methods (cf. appendix A). The problem is fully specified once Pr, Re, Ra and St are known and the boundary conditions are prescribed. Here we enforce a no-slip, fixed-temperature condition at the top of the ice, i.e. u = 0 and T = T t < 0 at z = 1.5. We impose free-slip, fixed-temperature conditions on the bottom boundary, i.e. ∂ z u = ∂ z v = w = 0 and T = 1 at z = 0, such that we simulate only one half of a full channel flow (to reduce computational costs). The dimensionless melting temperature is T = 0. The initial interface position is z = 1 and we note that (l x , l y , l z ) = (4π, 2π, 1.5), the domain lengths in dimensionless space. The initial condition in the fluid is a half-channel laminar Poiseuille flow superimposed with divergence-free white noise for the velocity fluctuations. We will generally discuss our results in terms of the steady-state friction (or shear) Reynolds number, Re * , and the friction Richardson number, Ri * , i.e.
since they are more commonly used than Re and Ra in turbulent channel flow studies (García-Villalba & del Álamo 2011; Zonta & Soldati 2018). The key difference between Re and Re * is that the former is based on the velocity on the bottom free-slip boundary of the channel in the laminar regime, which is [Π H 2 (1 − z 2 /H 2 )/(2ρ f ν)]| z=0 using dimensional variables, while the latter is based on the friction velocity, which is √ −τ w = Π H/ρ f , with τ w the mean wall shear stress, again using dimensional variables. Here, we favour the friction Richardson number over the Rayleigh number as the control parameter, even when the stratification is unstable, because they are both input parameters and because the wall shear stress is an important driver of turbulence in all cases. The importance of shear forces compared with buoyancy forces can be estimated from the Monin-Obukhov length, which is in terms of dimensionless variables and which is often reported in mixed-convection experiments (Pirozzoli et al. 2017;Blass et al. 2020), with Nu the Nusselt number, which we define later (see (2.9a-c)). The Monin-Obukhov length estimates the distance from the boundary within which shear is as important or more important than buoyancy. In our simulations, we always have L MO > 0.97, such that shear is a significant source of turbulence throughout the domain.
We investigate the effect of background density stratification on the generation of topography at the fluid-solid interface by considering three distinct values of Ri * , i.e. Ri * = 40, Ri * = 0 and Ri * = −40, for which the stratification is stable, neutral and unstable, respectively. For simplicity and computational expediency, all other parameters (except T t ) are fixed such that the flow is (moderately) turbulent and phase changes are relatively rapid, i.e. we set Re * = 150 (Re = 11 250), Pr = 1 and St = 1. For each Ri * , we set T t such that the initial heat flux in the ice, −T t /2, is almost equal to the heat flux in the fluid when there is no melting. As a result, the fluid-solid interface position does not move significantly in time (at least initially) and we can maximize numerical resolution around the interface with a fixed grid. For reference, the Rayleigh number for the unstable stratification case (Ri * = −40) is Ra = 4.5 × 10 5 , which is above the instability onset for Rayleigh-Bénard convection rolls in the streamwise direction (Ra ≈ 1101) of thermally stratified plane Poiseuille flow (Chandrasekhar 1961;Gage & Reid 1968). Note that changing the sign of Ri * can be obtained by changing the sign of the thermal expansion coefficient α, which indeed can be either positive or negative depending on the fluid state. We discuss the geophysical relevance of our choice of parameters, including α, in § 4.
We solve (2.5) using the open-source pseudo-spectral direct numerical simulation (DNS) code DEDALUS . We use 256 Fourier modes in the x and y directions and a compound Chebyshev basis with a total of 288 modes in the z direction unless stated otherwise. The use of a compound Chebyshev basis allows us to have a stretched grid in the vertical direction with refined (respectively coarse) resolution near the mean fluid-solid interface (respectively in the fluid bulk). Here, the Chebyshev collocation grid has a resolution equal to approximately one-fourth of a wall unit Δz + = 1/Re * ≈ 0.0066 at and around the fluid-solid interface and equal to approximately one wall unit in the fluid bulk, whereas the Fourier collocation grids have a uniform resolution of roughly 7 and 3.5 wall units in x and y, which is within the recommended resolution for channel flow simulations (see e.g. Moin & Mahesh (1998) and appendices A and C for more details). We use a two-step implicit-explicit Runge-Kutta scheme for time integration. The Courant-Friedrichs-Lewy condition is typically set to 0.2 in the transient initial stage and 0.4 later on. At statistical steady state, the time step is typically 10 3 -10 4 times smaller than the friction time scale 1/(Re * Pr), which is equal (in terms of dimensional variables) to H divided by the steady-state friction velocity. We run each simulation for approximately 4 diffusive time scales, or 600 friction time scales, which takes roughly 2 million time steps, such that the total cost of the simulations is of the order of 1 million CPU hours. Figure 1(a) shows a snapshot of the temperature field in the fluid (red colourmap) and the solid (blue colourmap), as well as the velocity vectors (arrows) at select locations for Ri * = 0. Figure 1(c) shows the variations of the phase field along the thick solid line drawn in figure 1(b). The transition from φ = 1 in the fluid to φ = 0 in the solid occurs and Nu III b are volume-averaged and time-averaged over 50 friction time units before the end of stages I and III, respectively. Here q s is the conductive heat flux through the ice imposed as an initial condition at the beginning of stage II; ξ III , ξ − and ξ + are the mean interface position, the maximum amplitude of the keels and the maximum amplitude of the channels averaged over 50 friction time units at the end of stage III; C I D and C III D are the drag coefficients averaged over 50 friction time units at the end of stage I and stage III, respectively. Note that Re * = 150, Pr = 1 and St = 1 in all simulations. Also, Ri * = −40 corresponds to Ra = 4.5 × 10 5 . over a very thin diffuse interface of thickness ≈0.007. Simulation parameters and output variables are provided in table 1.
Variables of interest
We define the friction velocity, the bulk velocity and the Nusselt number as respectively (cf. details in appendix B), where the overbar denotes the horizontal average and · ≡ V dV/V denotes the volume average, such that φ is the mean fluid fraction and u b is the bulk velocity of the fluid phase. In (2.9a-c), τ d , τ ν and τ w are the linear damping, viscous and Reynolds shear stresses, and q = wT − ∂ z T is the heat flux. At statistical steady state, the full shear stressτ = (τ d + τ ν + τ w ) is approximately a linear function of z and q is approximately depth invariant, in agreement with channel flow simulations of a pure fluid (cf. appendix B for details on stresses and depth-independent variables using the phase-field method). Sinceτ is a linear function of z, u * can be estimated from the full shear stress as √ −τ at any depth as long as it is properly rescaled by the height at which it is estimated. Here, we use the shear stress at the top boundary z = 1.5 in (2.9a-c) for convenience but with pre-multiplying coefficient φ ≤ 1, such that u * is truly the friction velocity at the mean interface position (cf. (2.9a-c)). We denote by ξ the fluid-solid interface position, where (2.10) such thatξ = φ (note that one could alternatively define ξ such that it satisfies φ(z = ξ) = 0.5 or T(z = ξ) = 0), and we denote the melt rate byṁ = ∂ t ξ . The drag coefficient of the fluid-solid boundary is defined as the ratio of the dimensionless wall shear stress u 2 * divided by the dynamic pressure u 2 b /2, i.e. Pirozzoli et al. 2017). The temporal fluctuations of the variables of interest will be mainly reported in terms of the friction time t * = Re * Prt.
Occasionally, we will show vertical profiles of variables in terms of the distance from the interface, which we denote by χ(t, x, y) = ξ(t, x, y) − z.
Results
The key findings of our work are that (i) streamwise topographical features emerge from uneven melting and freezing at a phase boundary when the flow is driven by a pressure gradient, and that (ii) the type of density stratification affects the characteristic amplitude and spanwise wavelength of the streamwise patterns. Thus, after a discussion of the evolution of global flow variables in § 3.1, we directly present the results of the topographical features generated at the fluid-solid boundary in § 3.2. We then investigate the interplay between the turbulent flow, the topography and phase changes in § § 3.3 and 3.4, and finally discuss the evolution of the mean interface position and the statistics of melting and freezing in § 3.5.
Simulation stages and global flow variables
We show in figure 2 the friction velocity u * , the bulk velocity u b and the Nusselt number Nu for stable (top figure), neutral (middle figure) and unstable (bottom figure) stratification. Each simulation is broken down into three main stages, which are highlighted by different colours in figure 2 (note that we do not discuss the results shown in grey, which correspond to the spin-up of the fluid phase without buoyancy effects). The first stage of interest (stage I for t Ib * ≤ t ≤ t Ic * ) is shown in blue and corresponds to the spin-up of the fluid phase with buoyancy effects turned on. Importantly, stage I neglects the solid phase, which is substituted with a simple isothermal no-slip boundary, for computational expediency. The second key stage (stage II for t Ic * < t ≤ t II * ) is shown in orange and corresponds to the part of the simulation that includes the solid phase with volume penalization turned on, but neglects melting or freezing, such that the solid always occupies 1 ≤ z ≤ 1.5 and the phase field is prescribed as φ = 0.5 1 − tanh 2(z − 1)/δ , where δ is the thickness of the diffuse interface. The final third stage (stage III for t > t II * ) is shown in green and highlights results obtained when all effects are considered, i.e. buoyancy is turned on, there is both the fluid and the solid and phase changes are enabled (cf. additional details on the simulation stages in appendix C). The temperature in the solid is initialized at the beginning of stage II as where q s is the initial conductive heat flux through the solid, by imposing the fixed-temperature condition T = T t = −q s /2 at the top of the solid. The difference between the heat flux in the fluid and the conductive heat flux in the solid in stage II controls whether the fluid-solid interface melts or freezes once phase changes are turned on in stage III. Here, we set q s to be slightly smaller than the heat flux in the fluid at the end of stage I, which we denote by Nu I , such that the solid melts slowly at the beginning of stage III in all three simulations (see further discussion in § 3.5). The bulk Reynolds and Nusselt numbers at the end of stages I and III are defined as with Δ * = 50, and are reported with q s in table 1. Note that Re b Re because the flow is turbulent and, hence, experiences enhanced friction at the wall compared with the same flow in the laminar regime.
Buoyancy effects are turned off for t * ≤ t Ib * , such that the results of figure 2 are exactly the same for all three simulations until t * = t Ib * . Upon turning on buoyancy, i.e. for t * ≥ t Ib * (blue), the Nusselt number and bulk velocity deviate from the neutral case (middle figure), but with opposite behaviours: Nu decreases while u b increases with stabilizing buoyancy effects (top figure), and Nu increases while u b decreases with destabilizing buoyancy effects (bottom figure). The friction velocity, on the other hand, remains close to u * /Pr ≈ Re * in all three cases. The effect of background stratification on bulk velocity and heat fluxes are well known from channel flow studies (García-Villalba & del Álamo 2011; Pirozzoli et al. 2017), and the important point is that the heat flux is the variable that changes the most with buoyancy effects. Here, Nu I = 3.02, 4.58 and 6.74 for Ri * = 40, 0 and −40, respectively (cf. table 1). It is worth noting that while Nu remains the same between stage I and stage II (in a time-average sense), u * and u b show some variations as a result of turning on volume penalization and adding a solid phase. The large dip of u * at t * ≈ t Ic * is merely the result of a sudden deceleration of the mean flow close to the interface, due to the addition of linear damping, which is transient, as can be seen from the rapid return of u * to its statistically steady-state value of u * ≈ 150. The drop of the bulk velocity is similarly due to the added linear damping. However, unlike the dip in u * , the drop in u b persists at all times, implying that volume penalization results in anomalous drag on the mean flow. Here, the relative drop of bulk velocity is of the order of 5 % and the profiles of temperature and velocity close to the fluid-solid interface in stage II reproduce closely those obtained in stage I (see appendix A). Therefore, we consider the discrepancy to be small enough not to warrant a computationally costly increase in resolution or further tuning of the phase-field parameters. When melting is turned on, i.e. for t * > t II * (green), global flow variables show different behaviours depending on Ri * . For the stable case, u * , u b and Nu exhibit moderately large fluctuations (as in previous stages), but do not exhibit any time-mean deviation (top figure). For the neutral case, we find a small increase in u * , u b and Nu (middle figure). For the unstable case, we find that u * and u b increase slightly, while Nu increases substantially (bottom figure). The analysis presented in the next sections explains these behaviours. Eventually, all simulations reach a statistical steady state.
We show in figure 3 the temporal evolution of another global variable, namely, the drag coefficient, C D , which is of significant interest in inferring melt rates from resolved variables in coarse models (using, for instance, the three-equation model; see Holland & Jenkins 1999). The drag coefficient is of order 10 −2 and decreases (respectively increases) significantly at t * = t Ib * , i.e. when the stratification becomes stable (respectively unstable). The decrease (respectively increase) of C D results from an increase (respectively decrease) of the potential energy barrier in stirring the mean shear and bringing momentum upward with increasing stable (respectively unstable) stratification and is in agreement with previous studies (García-Villalba & del Álamo 2011; Pirozzoli et al. 2017). In stage II, C D increases because u b decreases moderately upon turning on volume penalization (cf. figure 2). In stage III, C D has similar values to stages I and II (cf. reported values in table 1), showing that it is not modified by the topographical features obtained in DNS, perhaps because they are aligned with the main flow direction (see § 3.2).
Spontaneous generation of channels and keels
The mean interface position does not vary significantly in our simulations, due to our choice of initial and boundary conditions for the solid, but uneven melting by the turbulent flow still generates large-amplitude topography, which we discuss in this section. We denote variables averaged in the x direction by a tilde ( ) and variables averaged in the x direction minus the horizontal mean by a prime ( ), such that e.g. ξ =ξ −ξ represents the spanwise variations of the streamwise-averaged topography around the horizontal mean.
We show snapshots of the two-dimensional fluid-solid interface ξ at the end of stage III in figure 4(a-c) for stable, neutral and unstable stratification, respectively (cf. movies 1-3 in the supplementary material available at https://doi.org/10.1017/jfm.2020.1064 to see the temporal evolution of the interface). In all three cases, the topography is dominated by channels (troughs in the solid; brown colour) and keels (excursions of solid into the fluid; green colour) aligned with the streamwise direction. We reach a statistical steady state relatively quickly in all cases after turning on phase changes, such that the patterns in figure 4(a-c) are representative of the interface topography throughout most of stage III (see movies 1-3 in the supplementary material). We show in figure 4(d-f ) the Hövmoller diagrams of the channels and keels by plotting ξ in the (t * , y) plane for all of stage III. It can be seen that the characteristic amplitudes of the channels and keels saturate almost immediately for stable and neutral stratification and well before the end of stage III for unstable stratification. The steady-state amplitude of the biggest channels, ξ + (maximum of ξ ), and the steady-state amplitude of the deepest keels, ξ − (minus the minimum of ξ ), increase with decreasing Ri * (i.e. from figure (d) to ( f )). The crest-to-trough amplitude is roughly 5, 10 and 45 times the viscous length scale δ ν = 1/Re * for stable, neutral and unstable stratification, respectively (note that δ ν is roughly equal to the diffuse interface thickness; cf. appendix A). Thus, the crest-to-trough amplitude is of the same order as the viscous sublayer thickness, which is approximately 5δ ν , for stable and neutral stratification, but extends beyond the buffer layer and into the log layer for the case of unstable stratification (figure 4c, f ). Figure 4(d-f ) shows that the viability of channels and keels increases with decreasing stratification: channels and keels are short lived with stable stratification but long lived with unstable stratification. For stable stratification (figure 4d), the separation of scales between the topography lifetime (about 10 friction time units) and the diffusion time scale across the solid layer (about 100 friction time units) suggests that the interface evolution is purely driven by the flow dynamics. For neutral stratification, figure 4(e) shows that channels and keels can drift, merge, split, decay and spontaneously appear over time scales of tens to hundreds of friction time units, highlighting a possible interplay between interface evolution and the fixed-temperature condition at the top solid boundary. For unstable stratification (figure 4f ), the channels and keels become time invariant and their amplitudes saturate because of the top solid boundary condition, which plays a key role in the interface evolution as discussed in the next sections.
Coupled dynamics of the fluid and solid phases
The emergence of channels and keels can be the result of either (i) a passive response of the interface to uneven melting patterns driven by the turbulent flow, or (ii) a fully coupled interplay between fluid turbulence, interface topography and temperature in the solid. Here, we investigate the relevance of regimes (i) and (ii) for each of our simulations by looking at both the flow dynamics and the temperature field in the solid. Figure 5 shows in the case of unstable stratification (figure 5c), −∂ zT fluctuations remain large near the top boundary, suggesting that there is a backreaction from the fixed-temperature top-boundary condition, T = −q s /2 at z = 1.5, on the interface evolution. The backreaction from the top boundary for unstable stratification is obtained because the position of the channels and keels becomes rapidly stationary (in contrast to the stable and neutral cases), such that the temperature field in the solid has time to adjust diffusively and balance the growth of the channels and keels (cf. appendix D for more details on the temperature field in the solid). The steadiness of the fluid dynamics and interface topography for unstable stratification can be seen to result in overlapping heat fluxes in the top figure of the bottom panel.
The emergence of streamwise channels and keels is consistent with the well-documented presence of near-wall streamwise streaks and vortices in stratified shear flows (Pirozzoli et al. 2017;Zonta & Soldati 2018), but their amplitude clearly varies with Ri * . In the case of stable stratification (figure 5a), buoyancy effects inhibit the generation of large topographical features, such that channels and keels have small amplitudes and do not feed back onto the flow (which is further discussed in § 3.4). In the case of neutral stratification, buoyancy is turned off, such that the solid boundary deforms more and can affect the flow dynamics. Figure 5(b) shows that the heat flux in the fluid close to the boundary is usually larger where there are channels (e.g. y ≈ 1.5, 2.5) than where there are keels (e.g. y = 0.9), suggesting a topographic influence on the flow. For the case of unstable stratification (figure 5c), two streamwise rolls aligned with the direction of the flow and filling the entire depth dominate the fluid dynamics. These flow features are reminiscent of Rayleigh-Bénard convection rolls as observed in channel flow simulations with unstable stratification (Pirozzoli et al. 2017), which here appear locked within the interface deformation pattern (figure 5c).
The interplay between the solid boundary and the flow dynamics for the case of unstable stratification is further highlighted in figures 6 and 7. Figure 6 shows the Hövmoller diagram of the heat flux in the middle of the fluid (figure 6a) and at 3 wall units below (figure 6b) and above the interface (figure 6c). Two mid-depth streamwise rolls whose positions are locked are evident from t * ≈ 370 onward in figure 6(a), which is about the same time as when the two channels and keels become large in figure 4( f ). Similar rolls can be inferred for t * < 370 but are weaker and meander. The two strong rolls for t * ≥ 370 are locked with the topography (cf. figure 4f ) and support large conductive heat fluxes with similar patterns right below the interface (cf. figure 6b), which further demonstrates that global modes control the interface dynamics for an unstable stratification. The conductive heat fluxes coming from the fluid are yet eventually balanced by the conductive heat fluxes through the solid (figure 6c), which adjust diffusively as the interface deforms, such that there is no net melting (figure 6d) beyond the initial transient of stage III. The decrease of conductive heat flux below the interface between t Ic * and t II * in figure 6(b) is due to the heat flux imbalance at the beginning of stage II (cf. appendix A) and has no incidence on the subsequent melting dynamics. Note that the temporal resolution is relatively coarse in figure 6(a-d) because these figures required the full three-dimensional data sets (due to interpolations in z for figure 6b,c), which we saved at a low frequency to minimize disk usage. Also, the patchiness of the melt rate in figure 6(d) is due to the interaction of the ice boundary with individual turbulent bursts, including small-scale eddies, which drive intermittent melting and freezing events but have no long-term effect on topography generation. Figure 7 shows the spanwise spectrum ofq (left axis) in the middle of the fluid (solid lines) and near the solid boundary (dashed lines) at the end of stages I (red) and III (black) for the unstable case. In all cases, the spectrum peaks at wavenumber L y k y /(2π) = 2, which is close to the critical wavenumber L y k c /(2π) = 3.11 of convection instability (λ c = 2.016), suggesting that Rayleigh-Bénard convection is already active in stage I. However, with melting turned on, the peak is significantly amplified, especially for the spectrum near the boundary (dashed lines), which also shows amplification of higher harmonics, consistent with the spectrum of the interface itself (dotted blue line; right axis). These results suggest that Rayleigh-Bénard rolls are energized more than any other fluid features once melting is turned on, because they best couple with the interface topography evolution as a result of melting and freezing. Note that the spectra of the heat flux near the ξF igure 8. Vertical profiles of the streamwise velocity U and vertical velocity r.m.s. w rms averaged in x and t * (over at least 50 friction time units) for stable (two leftmost columns), neutral (two middle columns) and unstable (two rightmost columns) stratification. The top panels show the full horizontal average of U and w rms , i.e. when they are also averaged in y, and the mean interface positionξ at the end of stage III (dotted lines). The bottom panels show the full horizontal averages of U and w rms as well (dotted lines), but also U and w rms under the largest keel (downward triangles) and channel (upward triangles), i.e. they are not averaged in y, as functions of 1 − χ, where χ is the distance from the interface. In (l), the green (respectively blue) thick and thin dashed lines show the vertical profiles shifted away from the largest keel (respectively channel) by 0.25 and 0.5 units in the +y direction. boundary and of the interface have a sawtooth-like pattern due to the non-sinusoidal shape of the interface and numerical confinement in the spanwise direction.
We next show in figure 8(a-f ) (top row) the vertical profiles of the mean streamwise velocity U and of the root-mean-square (r.m.s.) vertical velocity, w rms , where here the mean and r.m.s. are defined using a horizontal and temporal average. The results are shown for all three stages and for stable (two leftmost columns), neutral (two middle columns) and unstable (two rightmost columns) stratification. The vertical profiles are never symmetric with respect to the half-fluid depth position, which is z = 0.5 in stages I and II. This is because our velocity boundary conditions across the fluid layer are different, i.e. no-slip on the top boundary (which can be the fluid-solid interface) and free-slip on the bottom. For stable and neutral conditions, there is a strong overlap of all curves, suggesting that the topography does not influence the mean profiles, while for unstable stratification, there is a small deviation of the stage III profiles (solid lines). Note that a log-log plot of the mean velocity and temperature profiles near z = 1, which we show in figure 14 in appendix A, clearly shows that the flows follow the law of the wall on the top no-slip fluid boundary in stages I and II.
We investigate in figure 8(g-l) whether the mean profiles vary in the spanwise direction in such a way that they correlate with the x-averaged interface topography. To do so we consider the mean profiles under the largest keel (downward triangles) -i.e. at y = y k , where y k is the minimum in y ofξ at each time step -separately from the mean profiles under the largest channel (upward triangle) -i.e. at y = y c , where y c is the minimum in y ofξ at each time step -where mean now denotes streamwise and temporal averaging. For stable stratification, there is no difference between the profiles under keels and channels. However, for neutral and unstable stratification, the two profiles depart in such a way that the streamwise velocity is larger under channels than under keels, and the vertical velocity r.m.s. (defined using a temporal and (x, y) average for each χ ) is larger under keels than under channels. These results indicate a noticeable influence of the topography on the flow. For the case of neutral stratification, the separation of the profiles is maximum for χ < 0.3 and then vanishes, suggesting a local influence of the topography on the flow dynamics, while for unstable stratification, the effect of the topography is felt throughout the entire depth due to coupling with the Rayleigh-Bénard rolls. It may be noted that the profiles of w rms under the largest keels and channels are larger than the plane-average profile shown by the dotted line in figure 8(l). This is expected because Rayleigh-Bénard convection promotes both localized intense upwellings and intense downwellings under channels and keels. In fact, away from the main channel and keel the profiles decrease rapidly, as can be seen from the blue and green lines.
In order to gain further insight into the statistics of the flow interacting with the melting boundary, we show in figure 9 the probability density functions (p.d.f.s) of the streamwise velocity (a,d,g), the vertical velocity (b,e,h) and the temperature gradient (c, f,i), for stable (a-c), neutral (d-f ) and unstable (g-i) stratification. For the velocities, the p.d.f.s are shown both in the middle of the fluid, at z = 0.5, and near the boundary, at z = ξ − 0.04 (i.e. 6 wall units into the fluid). We find little difference between stages I (dashed lines) and III (solid lines) for the streamwise and vertical velocities, suggesting limited influence of the topography on the overall flow morphology, although the streamwise velocity in figure 9(g) has a negative tail with higher probability density in stage III than in stage I. The temperature gradient at the fluid-solid interface (figure 9c, f,i) does vary noticeably between stage I and stage III. However, this difference is due to the phase-field method rather than to a fundamental change in flow morphology since the p.d.f.s in stages III and II (not shown) show significant overlap.
While phase changes and the emergence of topographical features have little effect on the p.d.f.s, most p.d.f.s display flow-driven left-right asymmetries, which are worth highlighting. Most importantly, the p.d.f. of the temperature gradient at the fluid-solid interface has a rapidly decaying positive tail and a slowly decaying negative tail. This asymmetry is obtained in all stages and hence is a feature of the flow rather than a consequence of topography generation, and suggests that the phase-change dynamics should be itself asymmetric (which we show in § 3.5). The p.d.f.s of the streamwise velocity near the boundary are also asymmetric, featuring a slowly decaying positive tail and a rapidly decaying negative tail. We have further separated the p.d.f.s of the velocities in figure 9 based on the sign of the local temperature anomaly (compared with the plane and temporal mean). Blue curves denote p.d.f.s obtained for a negative temperature anomaly, i.e. representative of fluid patches influenced by the cold top boundary, while red curves denote p.d.f.s obtained for a positive temperature anomaly, i.e. representative of fluid patches influenced by the warm bottom boundary. The cold temperature p.d.f.s are shifted to the left of the warm temperature p.d.f.s for the streamwise velocity (left column), which suggests that negative streamwise velocity is more often associated with cold fluid 911 A44-20 coming from the top boundary. Also, for the vertical velocity at z = ξ − 0.04 (right panels of the middle column), the positive tail of the warm temperature p.d.f.s (red) is larger than the negative tail of the cold temperature p.d.f.s (blue), suggesting more extreme warm upwelling events than cold downwelling events just outside of the viscous sublayer. These results demonstrate that the near-wall flow dynamics have multiple asymmetries, which may be related to the asymmetry in the temperature gradient at the boundary.
3.4. Reversing the stratification In the case of stable stratification, the cool melt fluid (at freezing temperature) is more buoyant than the warmer surrounding fluid, such that it rises and accumulates in the middle of the channels. Thus, channels and keels are limited to small amplitudes in figure 5(a) because of a negative feedback between channel generation and the melt dynamics, which halts channel growth. In the case of unstable stratification the opposite is true, i.e. the cool melt fluid is denser than the surrounding fluid and, hence, is evacuated from channels. This yields a positive feedback between channel generation and melt dynamics, which results in the large-scale topographical features seen in figure 5(c) (which saturate over time due to thermal adjustment in the solid).
The hypothesis of the cool melt fluid pooling in channels and inhibiting their growth for stable stratification is difficult to verify with the results discussed previously because of the small interface deformation obtained for Ri * = 40. Therefore, we have run a fourth simulation starting from the final time of the simulation with unstable stratification (and large interface deformation), but with an increasing Richardson number such that the fluid becomes stably stratified and interacts (transiently) with the initially large topographical features. We impose the stable stratification through several intermediate steps so that the flow does not relax to a laminar state. Specifically, we use , such that Ri * starts from ≈−40 at t * = t III * and reaches ≈40 for t * > t III * + 19. We show the results of this run in figure 10, where blue/red highlight x-averaged temperature values in the solid/fluid phase, while arrows denote x-averaged velocity vectors (ṽ,w) in the ( y, z) plane. Figure 10(a) shows the results at time t * = t III * + 1 (t III * = 526), i.e. when Ri * ≈ −40. The stratification is unstable such that the flow features strong upwelling of warm fluid below the channels and strong downwelling of cool fluid along and under the keels, akin to Rayleigh-Bénard rolls locked into the deformed interface pattern. A pair of counter-rotating streamwise rolls is clearly visible below each of the two channels. These rolls persist until Ri * ≈ 0, which is in agreement with recent simulations of mixed convection that have shown that streamwise rolls extending throughout the full depth of a channel (without phase changes) are obtained for a wide range of negative Richardson numbers (Pirozzoli et al. 2017). Figure 10(b,c) show the results at times t * = t III * + 13 and t * = t III * + 21, i.e. when Ri * ≈ 0.7 and Ri * ≈ 39, respectively. At these times, the stratification is stable and the cool melt fluid produced at the keels converges toward the channels' centreline. The Rayleigh-Bénard large-scale rolls have vanished and are replaced with weaker vortices of finite vertical extent, which are most vigorous close to the interface where they are driven by the (positive) buoyancy anomaly of the melt fluid at the tip of the keels. The heat flux through the fluid goes down and freezing occurs everywhere such that the solid front advances into the fluid. Importantly, freezing is faster in the channels because of the convergence of the buoyant cool melt fluid and higher conductive heat fluxes in the solid (as the solid is thin above channels), which leads to rapid refreezing of the initial channels. While Ri * > 0 increases, the properties of the boundary-attached vortices (e.g. vertical extent and intensity) are the result of a complex interplay between the amplitude of the interface topography and of the stratification strength. The increasing stratification drives an increasing positive buoyancy anomaly of the melt fluid but also increasingly damps global modes (García-Villalba & del Álamo 2011), while the decreasing topography amplitude is expected to result in flattening and weakening vortices. Ultimately, the topography disappears and the vortices weaken significantly.
Melting
In this section we discuss the evolution of the mean interface positionξ with time and the statistics of meltingṁ = ∂ t ξ at the statistical steady state. We first show the evolution ofξ in DNS as a function of time in figure 11(a) for all three simulations (solid lines). For a solid with spatially uniform temperature equal to the melting temperature T m , we would expect a faster increase ofξ with time for an unstable than for a stable stratification, since the heat flux in the fluid is larger when buoyancy forces are destabilizing rather than restoring. However, as indicated in § 2, we have imposed a conductive heat flux in the solid (q s ) slightly less than the mean heat flux through the fluid (Nu I ) at the beginning of stage III, i.e. when we turn on melting, such that the leading-order melt rate is not controlled by the Nusselt number of the fluid-only simulations but by the difference Nu I − q s . This difference is 0.19, 0.1 and 0.05 for stable, neutral and unstable stratification (cf. table 1), such that the initial increase ofξ is faster for stable than for unstable stratification. Figure 11 shows thatξ saturates over time. This happens because the conductive heat flux in the ice increases asξ increases (in a plane-averaged sense), since the solid becomes thinner, such that it eventually balances the heat coming from the fluid. Under the assumption of small interface deformations, it is possible to predict the evolution of the mean interface position over time using a reduced model. As a first approximation, we consider that the topography has no effect on the temperature in the solid, i.e. we assume that the heat flux through the solid is simply equal to the temperature difference between the interface and the top boundary divided by the mean solid thickness h (cf. details in appendix D, which hinges on an assumption of a quasi-steady state). Then, the evolution equation forξ becomes where h 0 = 1/2 is the initial ice thickness, q f is the heat flux in the fluid and we recall thatξ = ξ 0 = 1 at t = t II * . For simplicity, we take q f to be a constant diagnosed from the simulations. The results of (3.4) for q f = Nu I , i.e. obtained when setting q f to the average heat flux before melting is turned on, are shown by the dotted lines in figure 11(a). The overlap between the reduced model and DNS results at early times is good for unstable stratification (as expected) but is poor for stable and neutral stratification. The disagreement with q f = Nu I at early times arises because the temperature in the solid is slightly above 0 because of volume penalization, such that there is some artificially large melting at the beginning of stage III (cf. appendix A). At later times, the DNS results and the model results shown by the dotted lines diverge because the heat flux Nu increases rapidly once melting is turned on, as can be seen in figure 11(b). For unstable stratification, the agreement with q f = Nu I is relatively good until t * ≈ t II * + 50 (cf. red dotted line), i.e. right until the Rayleigh-Bénard rolls are energized and a large-scale topography emerges (cf. § 3.3).
In order to account for the increase in heat flux through the fluid enabled by melting and the generation of topography, we show with dashed lines in figure 11(a) the result of (3.4) with q f = Nu III , which is the heat flux at the statistical steady state with melting turned on. For stable and neutral stratification, there is good agreement between the model results and the mean interface position at late times. For unstable stratification, however, (3.4) with q f = Nu III overestimates the final value ofξ − 1 by a factor of two (approximately), suggesting that topography plays a non-negligible role on the heat flux in the solid. We show in figure 11(a) a prediction ofξ for unstable stratification obtained using a more accurate higher-order model (red dash-dot line), which takes into account interface deformation (cf. appendix D). The higher-order prediction overlaps well with the DNS results at late times, demonstrating that melting and the generation of topography changes the heat flux through both the fluid and the solid. The topography makes the solid more efficient at evacuating heat because the anomalous (increased) heating obtained above the channels (i.e. where the solid is thin) exceeds in absolute value the anomalous (reduced) heating obtained above the keels (i.e. where the solid is thick), which is a nonlinear effect in the topography amplitude obtained for any topography with top-down symmetry (e.g. a sinusoid). The higher-order model takes into account this nonlinear effect in topography amplitude and predicts a steady-state solid layer thickness larger than that predicted by the low-order model without topography for the same forcing heat flux q f .
We finally show in figure 12 the p.d.f.s of interface deformation and melt rate at the statistical steady state, i.e. past the initial transient during which topographical features emerge and the solid melts on average. At the statistical steady state, the p.d.f.s of interface position become roughly time invariant. The mean interface position reaches a plateau (cf. figure 11a) because the mean amount of freezing balances the mean amount of melting at every time step, and the standard deviation, or topography amplitude, saturates (as can be seen in figure 4( f ) for the unstable case). For stable stratification, the steady-state p.d.f. of interface deformation is almost symmetric with respect to the mean and appears approximately Gaussian (figure 12a). This suggests that channels and keels are symmetric to each other with respect to the mean interface position for the stable case. For neutral stratification, a small asymmetry develops, i.e. the median shifts toward small positive deformations (channels) and the tail of extreme negative deformations (keels) increases slightly. The same asymmetry is amplified for unstable stratification, with a narrow peak appearing to the right of the 0 mean and the negative tail increasing further. In other words, as the stratification becomes unstable, patterns grow in size and the width-to-height ratio of channels increases (broad and flat) while the width-to-height ratio of keels decreases (narrow and deep). The asymmetry in the p.d.f.s of interface topography is consistent with the observation from figure 4 that channels are typically flatter and more widespread than keels.
The p.d.f.s for the melt rateṁ are shown in figure 12(b). The temporal and spatial average,m, which is subtracted from the p.d.f., is close to 0 in all cases, since the mean amount of melting is balanced by the mean amount of freezing at the statistical steady state. The p.d.f.s of melt rate are asymmetric, i.e. similar to the p.d.f.s of heat flux at the top of fluid-only channel simulations (see dashed lines in figures 9c, f,i). The median is shifted to the left of the mean, i.e. toward negative values representative of freezing events, and the positive tail is enhanced compared with the negative tail. In other words, the interface is typically freezing slowly (ṁ < 0), but occasionally melts rapidly (ṁ > 0). We remark that the asymmetry of the melt rate p.d.f.s is not due to the asymmetry of the interface p.d.f.s since the p.d.f.s of melt rates inside channels (upper triangle) and along keels (lower triangles) are similar, but is instead a generic feature of melting by a turbulent fluid. Indeed, while the turbulent flow can drive rapid melting independently of what happens in the solid, freezing necessarily involves slow diffusive processes in the solid. Additionally, the near-wall dynamics, which features coherent structures such as streamwise streaks and vortices, is itself asymmetric, as can be seen from the p.d.f.s of the temperature gradient in figure 9. Thus, it is not surprising that the melt rate p.d.f.s are asymmetric. While beyond our goal, it would be worthwhile in the future to try to identify flow features controlling the shape of the melt rate p.d.f.s.
Geophysical discussion
Due to the large computational costs of coupled fluid-solid simulations, all control parameters in this study were held fixed, i.e. we considered Re * = 150, Pr = 1 and St = 1, except for the friction Richardson numbers, which we varied in order to test the effect of density stratification on topography generation. Our simulation with a positive (respectively negative) Richardson number Ri * = 40 (respectively Ri * = −40), i.e. with stable (respectively unstable) stratification, assumes a negative (respectively positive) thermal expansion coefficient. Cold freshwater has a negative (respectively positive) thermal expansion coefficient at low (respectively high) pressure (Thoma et al. 2010). Thus, our stable simulation is qualitatively similar to the flow of freshwater below an ice cover at low pressure (as is the case in an ice-covered lake), whereas our unstable simulation is qualitatively similar to the flow of freshwater below an ice cover at high pressure (as is the case in a deep subglacial lake). In the case where the solid is below the fluid, the stratification is reversed, i.e. the unstable simulation results are applicable to the flow of cold freshwater above ice at low pressure (as is the case in supraglacial rivers). Our stable simulation is also qualitatively similar to the flow of salt water under ice shelves. The melt water under ice shelves is cooler but also fresher than the ambient ocean water, such that it is positively buoyant. In fact, salinity and temperature can be combined, assuming that they have the same effective diffusivities, into a single variable known as thermal driving, which has a negative expansion coefficient (Jenkins 2016). In order to minimize resolution requirements and observe large topographical changes in a relatively small amount of time, we have considered a flow that is only moderately turbulent, weakly stratified and anomalously warm. If we assume that the working fluid is water, i.e. L = 3 × 10 5 J kg −1 and c p = 4 × 10 3 J kg −1 K −1 , then St = 1 implies that the bottom boundary is held at 75 • C. If we further assume that H = 10 cm and ν = κ = 10 −6 m 2 s −1 , i.e. considering an anomalously high thermal diffusivity such that Pr = 1, then our simulation runs correspond to about 10 hours in real time, the bulk velocity is 2 cm s −1 and the thermal expansion coefficient is 5 × 10 −7 K −1 in absolute value, which is quite small for water. These calculations highlight that there is clearly a significant gap between the control parameters in our simulations and in nature. Our set of experiments is also limited to three different runs, such that we cannot offer a quantitative prediction of what would be observed in the environment. Nevertheless, Rayleigh-Bénard rolls have been observed for a wide range of Rayleigh numbers and Reynolds numbers in mixed-convection simulations with unstable stratification (Pirozzoli et al. 2017), such that large-scale channels and keels driven by these rolls may be expected in most conditions with unstable stratification, for example in supraglacial rivers or deep subglacial lakes, especially when the external flow is weak. The transverse wavelength of channels and keels maintained by global rolls is expected to be of the same order as the fluid depth, as is the case in our unstable simulation (cf. figure 10(a), in which each wavelength accommodates two counter-rotating rolls with a diameter equal to the mean fluid depth). We note that our unstable simulation has low Ra = 4.5 × 10 5 and Re * = 150 compared with state-of-the-art simulations of mixed convection (Pirozzoli et al. 2017;Blass et al. 2020), but more importantly has a relatively small bulk Richardson number Ri b = RaPr/Re 2 b ≈ 0.1 (defined as positive for an unstable stratification) and large Monin-Obukhov length L MO ≈ 1 (normalized by H). Thus, channels and keels can be expected for a broad range of bulk Richardson numbers of unstably stratified shear flows, i.e. Ri b ≥ 0.1, and to grow in size as Ri b increases. In the limit Ri b → ∞, channels and keels may be expected to eventually disappear. Indeed, as Ri b → ∞, the mean flow vanishes and buoyancy effects dominate, such that three-dimensional domes and cusps should emerge in place of streamwise features (Rabbanipour Esfahani et al. 2018). Interestingly, large channels have also been observed at the base of ice shelves. However, these channels are unlikely to originate from Rayleigh-Bénard rolls but rather from transverse perturbations of, for example, the subglacial discharge or ice thickness at the grounding line (Dallaston, Hewitt & Wells 2015), since the stratification is in this case stable.
For stable and neutral stratifications, we also observed channels and keels. Channels and keels with stable stratification (Ri * ≥ 0) are carved by boundary-attached momentum streaks rather than by global modes, however, such that their transverse wavelength is shorter than for the case of unstable stratification (although here the difference is weak given the small Re * ), and their amplitude is either equal to or smaller than the viscous sublayer thickness, i.e. small. The shape and size of the small channels and keels obtained for Ri * ≥ 0 are in stark contrast with the three-dimensional scallops, which have been observed in stably stratified polar oceans and neutral laboratory experiments. Scallops observed in the field and investigated in laboratory experiments have amplitudes of the order of a few centimetres and wavelengths of the order of a few tens of centimetres, i.e. they are tall and wide features compared with the viscous sublayer thickness, which is typically smaller than 1 mm in nature (Bushuk et al. 2019). Previous experimental and theoretical works have found that the friction Reynolds number based on the scallop wavelength λ usually satisfies Re λ * = Re * λ/H ≥ O(1000-10 000) and have always reported a scallop wavelength smaller than the fluid depth, i.e. λ < H (Blumberg & Curl 1974;Thomas 1979;Claudin et al. 2017). Considering the upper limit λ = H means that scallops are predicted to emerge for Re * ≥ O(1000-10 000) in water, which is at least one order of magnitude higher than what we selected for our study and difficult to achieve numerically. Note, however, that the minimum Re * leading to scallop formation may be different for a fluid with control parameters Pr = 1 and St = 1 (as is the case in this work) instead of Pr ≈ 10 and St ≈ 75, as is the case for water, and may also vary with the stratification strength.
Concluding remarks
We have shown that streamwise channels and keels spontaneously emerge as the dominant topographical features of a fluid-solid boundary when the flow is pressure-driven, turbulent and thermally stratified with Re * = 150, Pr = 1 and St = 1. We have investigated the effect of the background density stratification and found that the amplitude of the channels and keels increases with decreasing stratification. For unstable stratification (Ri * = −40), the channels and keels couple strongly with Rayleigh-Bénard rolls, which are energised and locked within the interface deformation pattern. For neutral stratification, a similar correlation is obtained between the flow dynamics and the interface deformation pattern. However, the full-depth rolls are replaced with smaller and weaker boundary-attached momentum streaks, which do not provide a clear locking mechanism, i.e. the topography drifts. For neutral (Ri * = 0) and stable stratification (Ri * = 40), the channels and keels saturate either because of the absence of a positive feedback between topography and momentum streaks or because stabilizing buoyancy forces inhibit channel growth. For unstable stratification, the saturation is due to the fact that we impose the temperature at the top boundary. With an imposed heat flux at the top, the entire solid would melt rapidly and entirely provided that the stratification is unstable (not shown), which means that the choice of boundary conditions at the top of the solid can be critical. Note that the growth of the fluid layer for unstable stratification is due to the positive feedback that melting has on the effective Rayleigh number of the convective fluid. As the solid melts, the effective Rayleigh number increases, leading to further melting, which is stopped only if diffusion in the solid can eventually balance the increasing heat flux in the fluid.
The analysis of the melt rate statistics indicates that there is an asymmetry in melting and freezing, which may be related to the different melting/freezing dynamics (freezing relying primarily on slow diffusive processes in the solid) but also asymmetries in the flow statistics. Specifically, melting is highly localized and intense while freezing is widespread but weak. While beyond the scope of this study, it would be useful to identify whether coherent features of the near-wall turbulent flow, such as streamwise streaks and vortices, correlate preferentially with either melting or freezing events.
The drag coefficient changes significantly depending on the type of stratification but is only weakly affected by the generation of topographical features, which is not unexpected in our case since streamwise channels and keels are smooth in the direction of the flow. Capturing three-dimensional topographical features, such as scallops, which do affect momentum and heat transfers (Bushuk et al. 2019), in coupled fluid-solid simulations would be a major achievement, which could complement fluid-only simulations at planar ice boundaries (Gayen, Griffiths & Kerr 2016;Keitzl, Mellado & Notz 2016a,b;Mondal et al. 2019;Vreugdenhil & Taylor 2019). However, as discussed in § 4, scallops may require much higher Reynolds numbers to form than Re * = 150. In fact, the minimum Re * for scallops could be too high for a phase-field method on most supercomputers. The cheapest test for evaluating the minimum Reynolds number leading to scallops would be to start the simulations with a longitudinally wavy boundary and investigate the initial evolution. The runtime would be reduced to a minimum. However, a high resolution (higher than say 1024 3 with a spectral code) would still be required. It is noteworthy that simulations of a pure fluid at a fixed wavy boundary would already be useful in helping to verify or refine the most recent theoretical predictions on scallop formation and saturation (Claudin et al. 2017) and estimate the effect of the stratification strength, which has not yet been considered. We remark that capturing scallops in a water environment would require not only higher Re * but also higher St and Pr, which would both incur significant computational overheads. Higher Pr results in thinner thermal boundary layers, which could impact the near-wall dynamics and, for example, the asymmetry between melting and freezing. Higher St results in slower melt rates, which could significantly change how interface patterns couple with transient flow features. In the case of unstable stratification, we might still expect that Rayleigh-Bénard rolls couple with the interface deformation pattern for high St, since they are relatively stationary flow features at least in the strong shear regime (Pirozzoli et al. 2017). For neutral stratification, however, the interface evolves over time scales similar to those of the flow dynamics for St = 1 (figure 4e), such that increasing St might significantly decrease the sensitivity of the interface topography to fluid anomalies. Finally, freshening effects, which are critical to ice-ocean interactions, would require adding slowly diffusing salt to the simulations, which constitutes yet another significant challenge for multi-phase DNS.
From a fundamental physics viewpoint, it would be interesting to investigate in detail how topographical features are modified when phase changes are driven by dissolution rather than by melting. The fluid-solid boundary conditions (Stefan condition) and scalar diffusivities are different between dissolution and melting experiments. However, similar longitudinal and rippled patterns have been observed in both cases (e.g. Allen (1971) for dissolution). It would be also worthwhile to explore the effect of phase changes and topographical features on the onset of global modes and the large-scale organization of mixed-convection flows, which are of interest to many fields of physics and engineering (Kelly 1994;Pabiou, Mergui & Bénard 2005;Blass et al. 2020). Finally, it would be useful to investigate potential analogies between ice patterns due to melting and freezing and the formation of sand ripples and dunes, which have been and continue to be extensively studied (e.g. Charru Figure 13. Kinetic energy in the solid divided by kinetic energy in the fluid as a function of time t * . Red, gold and blue colours denote results obtained for unstable, neutral and stable stratification, respectively. A = 6/(5St), B = (16/δ 2 ) × 6/(5St) and C = 1, with δ chosen such that it is equal to two times the local grid size at z = 1 (initial interface position). Also, Γ = (δ/2.648228) 2 and we require time steps to be always smaller than Γ /2. We first assess the effect of the phase-field method and choice of parameters on the flow variables by showing in figure 13 the ratio of the kinetic energy averaged over the solid volume, KE s , divided by the kinetic energy averaged over the fluid volume, KE f . Figure 13 shows that KE s /KE f < 10 −4 and that the fluctuations are within a factor of two of the mean, i.e. velocities in the fluid penetrate only very weakly into the solid and do not burst significantly.
We now further comment on the resolution requirements and our choice for the grid size and time step. We recall that δ is the thickness of the diffuse phase-field interface over which φ transitions from 1 in the fluid to 0 in the solid (see figure 1c). Here δ is an artificial length scale, such that it must be smaller than any physical length scale in the problem, while at the same time being larger than the grid size since it must be resolved numerically. For a boundary layer flow with Pr = 1, the smallest length scale close to the interface is the viscous sublayer thickness, which is typically equal to a few times the viscous length scale δ ν = 1/Re * . Here, we chose to have δ ≈ δ ν in all simulations, i.e. δ is equal to 1 wall unit Δz + = δ, such that the diffusive interface for the phase field is comprised within the viscous and thermal sublayers, as can be seen in figure 14. In order to resolve the diffusive interface, we selected a vertical Chebyshev basis with enough modes such that the collocation grid has a resolution dz near the mean interface position equal to or less than δ ν /2. The damping time scale Γ in (2.5) is set to Γ = (δ/2.648228) 2 in order to cancel first-order errors in the phase-field model . Figure 14(a-c) shows a semilog plot of the wall-normalized velocity U + (left axes) and wall-normalized temperature T + (right axes) as functions of wall units z + (z + ≥ 0 denote positions in the fluid while z + < 0 denote positions in the solid) in stages I and II for stable, neutral and unstable stratification, respectively. In the viscous and thermal sublayers, which extend from z + = 0 to z + ≈ 5, we expect a linear scaling for both U + and T + with z + , shown by the solid dashed lines. This linear scaling is perfectly satisfied by the DNS results in stage I (blue circles and blue crosses) as well as the DNS results in stage II (orange circles and crosses), except for |z + | < δ (shown by the vertical solid lines), i.e. within the diffuse interface, which is expected since this is where the dynamics is artificially controlled by the phase-field equation. It may be noted that T + (orange crosses) is anomalously large for z + < δ and in fact deviates from the true solution (blue crosses) slightly outside the diffuse interface. This discrepancy is due to the fact that the heat flux in the fluid is larger than the heat flux in the solid in stage II. The interface being fixed -10 0 0 10 0 10 1 10 2 -10 0 0 10 0 10 1 10 2 -10 0 0 10 0 10 1 10 2 Figure 14. Wall-normalized velocity U + = Pr u/u * (circles, left axes) and temperature T + = u * T/(PrNu) (crosses, right axes) expressed in terms of wall units as functions of the wall coordinate z + = u * (1 − z)/Pr for (a) stable, (b) neutral and (c) unstable stratification. Blue symbols denote results obtained for the simple channel flow configuration (i.e. for t * ∈ [t Ib * , t Ic * ]) whereas orange symbols denote results obtained with volume penalization (i.e. for t * ∈ [t Ic * , t II * ]). The plus symbols show the temperature profiles obtained with volume penalization but shifted to the right, i.e. into the fluid, by replacing z + with z + + 2.5, z + + 1.16 and z + + 0.78 for stable, neutral and unstable stratification. The overbars denote horizontal averaging and time averaging over 20 friction time units. The vertical dotted lines show z + = −δ, 0, δ, where δ is the diffuse interface thickness.
in stage II, the heat imbalance results in the heating of the solid, such that T + = 0 occurs at z + < 0 away from the fixed interface position z + = 0 (note that we use a symmetric logarithmic scale with a linear threshold at |z + | = 0.1). By shifting the temperature profile to the right such that T + = 0 is aligned with z + = 0 (red pluses), we recover a perfect linear scaling for the temperature within both the thermal fluid sublayer and the solid. Outside of the linear sublayer and the buffer layer, the mean vertical profiles exhibit a logarithmic behaviour.
Far from the top boundary, that is, for z + ≈ 100, U + shows a steeper scaling with z + for stable stratification than for neutral or unstable stratification. This is a consequence of buoyancy effects, which tend to decrease (respectively increase) stirring of the mean flow when the stratification is stable (respectively unstable) (García-Villalba & del Álamo 2011).
whereτ andq are the anomalous stress and heat flux due to damping of the advective terms in (2.5), i.e.
The existence of anomalous stress and heat fluxes means that the friction velocity and Nusselt numbers as defined in (2.9a-c) are based on a total stress and heat flux, which are not rigorously linearly varying or depth invariant. We show in figure 15(a-c) the Reynolds stress τ w , the viscous stress τ ν , the linear damping stress τ d and the Reynolds stress plus the anomalous stress τ w +τ for stable, neutral and unstable stratification, respectively. Importantly, τ w and τ w +τ overlap well, showing that the anomalous stress is negligible. The results of figure 15(d-f ) further confirm that the anomalous stress is negligible in all simulations: the (approximate) total stress (solid lines) decreases linearly with z in all stages and overlaps well with τ w + τ ν + τ d +τ , i.e. the total stress (thin dashed lines) that includes the anomalous stress. We show in figure 15(g-i)q (solid lines) and q +q (thin dashed lines). For stable and neutral stratification (top and middle rows),q and q +q are constants with depth and overlap perfectly, suggesting that the anomalous heat flux is negligible. For unstable stratification (bottom row), we obtain similar results for stages I and II. For stage III, however,q is not perfectly constant, but deviates from q +q and peaks at z ≈ 1.15, which is roughly the height of the channels. The relative discrepancy betweenq and q +q is of the order of 5 % and is a result of the damping of the advective terms in the momentum and heat equations (2.5). We expect that this discrepancy would decrease with increased resolution. Previous studies have alternatively considered advective terms with the same damped form as here, with a divergence damped form, i.e. (u · ∇)(φu), or without any damping. There is no proof that any of these methods is more efficient than the other two. However, we would recommend using either one of the latter two methods, i.e. not the method used in this paper, in order to simplify the analysis of the shear stress and heat flux.
Appendix C. Additional details on the simulation stages
In this section we give additional details on the simulation stages and sub-stages. For t ≤ t Ic * (including stage I), we solve (2.5a), (2.5b) and (2.5d) with φ ≡ 1 and a no-slip isothermal top-boundary condition, i.e. u = 0 and T = 0 at z = 1. We use a straightforward half-channel flow configuration, i.e. without a solid domain, with 64 Chebyshev modes in the vertical direction. In stage II we add a solid layer of thickness 0.5 on top of the fluid domain and we use a compound Chebyshev basis stitched at z = 1.2 with 256 (respectively 32) Chebyshev modes in the lower (respectively upper) region. The compound Chebyshev basis allows us to have a high vertical resolution near the interface's initial position. We solve (2.5a), (2.5b) and (2.5d) with φ prescribed, i.e. not varying in time (cf. the main text). In stage III we solve (2.5) with all variables freely evolving and we use the same spectral resolution as in stage II.
Our simulations until t = t Ic * can be broken down into three sub-stages. In stage Ia, i.e. for t * < t Ia * (cf. light grey colour in figure 2), we run a low-resolution (128 Fourier modes in the x and y directions and 32 Chebyshev modes in the z direction) spin-up simulation of an initially laminar flow superposed with three-dimensional velocity perturbations and no buoyancy effects (Ri * = 0). In stage Ib, i.e. for t Ia * ≤ t * < t Ib * , we increase the resolution 911 A44-32 -20 000 20 000 0 -20 000 20 000 0 Figure 15. (a-c) Reynolds stress τ w (dotted lines), viscous stress τ ν (dashed lines), linear damping stress τ d (solid lines) and Reynolds stress plus anomalous stress τ w +τ (for stages II and III; black dashed lines) averaged in time over 20 friction time units and horizontal planes as functions of depth z for stable, neutral and unstable stratification, respectively (see the text for more details). Blue, orange and green colours denote results obtained in stages I, II and III, respectively (same in (d-f ) and (g-i)). (d-f ) Total stress, i.e. τ w + τ ν + τ d (solid lines), and total stress plus the anomalous stress, i.e. τ w + τ ν + τ d +τ (thin black dashed lines), averaged in time over 20 friction time units and horizontal planes as functions of depth z for stable, neutral and unstable stratification, respectively. Note that the full stress is shifted to the right by 10 000 (20 000) between stage II and stage I and between stage III and stage II for the case of stable (unstable) stratification for clarity as all curves overlap otherwise. (g-i) Heat flux q (solid lines) and heat flux plus anomalous heat flux q +q (thin black dashed lines) averaged in time over 20 friction time units and horizontal planes as functions of depth z for stable, neutral and unstable stratification, respectively. Note that the heat flux is shifted to the right by 2.5 (5) between stage II and stage I and between stage III and stage II for the case of stable (unstable) stratification for clarity. and the second-order solution is Thus, the second-order-accurate formula for the mean heat flux at the top of the ice reads as which differs from the leading-order heat flux only at second order. The prediction for the evolution of the mean interface position for unstable stratification with q f = Nu III , which is shown by the red dash-dot lines in figure 11(a), is based on (D1) with q top given by (D10) and with = 0.137 and k = 2 (as obtained from best-fit of the true interface topography at steady state for unstable stratification). We note that the quasi-steady-state assumption may affect the prediction of the transient evolution of the mean interface position adversely but has no effect on the final value, which is the primary goal of the reduced model. | 20,505.8 | 2020-11-28T00:00:00.000 | [
"Physics"
] |
Nuclear scattering configurations of onia in different frames
In the scattering of a small onium off a large nucleus at high center-of-mass energies, when the parameters are set in such a way that the cross section at fixed impact parameter is small, events are triggered by rare partonic fluctuations of the onium, which are very deformed with respect to typical configurations. Using the color dipole picture of high-energy interactions in quantum chromodynamics, in which the quantum states of the onium are represented by sets of dipoles generated by a branching process, we describe the typical scattering configurations as seen from different reference frames, from the restframe of the nucleus to frames in which the rapidity is shared between the projectile onium and the nucleus. We show that taking advantage of the freedom to select a frame in the latter class makes possible to derive complete asymptotic expressions for some boost-invariant quantities, beyond the total cross section, from a procedure which leverages the limited available knowledge on the properties of the solutions to the Balitsky-Kovchegov equation that governs the rapidity-dependence of total cross sections. We obtain in this way an analytic expression for the rapidity-distribution of the first branching of the slowest parent dipole of the set of those which scatter. This distribution provides an estimator of the correlations of the interacting dipoles, and is also known to be related to the rapidity-gap distribution in diffractive dissociation, an observable measurable at a future electron-ion collider. Furthermore, our result may be formulated as a more general conjecture, that we expect to hold true for any one-dimensional branching random walk model, on the branching time of the most recent common ancestor of all the particles that end up to the right of a given position.
Introduction
Onium-nucleus scattering is an outstanding process to understand theoretically, first because it is the simplest interaction process between a (model) hadron and a nucleus, and second for its potential phenomenological applications. Indeed, if the center-of-mass energy is sufficiently large, this process can easily be factorized from deep-inelastic electron-nucleus scattering cross sections [1,2], which will be measured at a future electron-ion collider [3]. On the other hand, in proton-nucleus collisions, it turns out that an appropriate Fourier transform of the onium-nucleus total cross section is mathematically identical to the differential cross section for producing a semi-hard jet of given transverse momentum [4], at least at next-to-leading logarithmic accuracy [5]. An onium may also be a good starting point to model dilute systems, such as heavy mesons, or maybe even specific states of protons, in order to understand theoretically some of their universal properties. 1 The center-of-mass energy (or, equivalently, total rapidity) dependence of onium-nucleus forward elastic scattering amplitudes is encoded in the Balitsky-Kovchegov (BK) evolution equation [7,8] established in the framework of quantum chromodynamics (QCD). When restricted to the relevant regime for the calculation of total scattering cross sections involving a small onium, the latter belongs to the wide universality class of non-linear diffusion equations, the main representant of which is the well-known Fisher [9] and Kolmogorov-Petrovsky-Piscounov [10] (FKPP) equation. (For background on the FKPP equation, see for example the reports [11,12]; For a review on how the FKPP equation appears in QCD, see e.g. [13]). This fact can be understood quite simply from a physical point of view. On one hand, the forward scattering amplitude, the evolution of which is described by the BK equation, is tantamount to the probability that at least one gluon in the Fock state of the onium at the time of the interaction, produced by a cascade of gluon branchings, is absorbed by the nucleus. On the other hand, with an appropriate initial condition, the solution to the FKPP equation is the probability that there is at least one particle generated by a one-dimensional branching-diffusion process in space which has a position larger than some predefined number. At the level of the evolution equations, the asymptotic equivalence between the BK and the FKPP equations becomes manifest after the identification of the time variable in the latter with the rapidity variable in the former, of the space variable with (the logarithm of) the transverse dipole size, and after taking the appropriate limit of the BK equation [14]. At the level of the underlying stochastic processes, the QCD evolution towards very high energies is a gluon branching process which, when the large-number-of-color limit is taken, boils down to the iteration of independent one-to-two color dipole splittings [15]. This process results in realizations of a specific branching random walk.
Other observables, such as diffractive cross sections, can be formulated with the help of a system of BK equations, as was first shown by Kovchegov and Levin [16].
In this paper, we shall analyze the scattering cross section per impact parameter for onium-nucleus collisions in the region in which it is much smaller than unity, namely when the size of the onium is very small compared to the saturation radius, and more specifically in the so-called "scaling region", which is a well-known parametric region in which the cross section does not depend on the rapidity and on the size of the onium independently, but through a scaling variable function of the latter two [17]. In particular, we shall study the interpretation of that cross section in the framework of the parton model, in terms of fluctuations of the partonic content of the onium, in different reference frames related to each other through longitudinal boosts.
Our motivation is twofold. First, boost invariance is a fundamental symmetry of scattering amplitudes, and it is interesting to understand theoretically how it is realized microscopically in this particular regime of QCD, in which the interacting objects may be thought of as sets of independent partons generated by a branching process. Second, it is already well-known that using boost-invariance of the scattering amplitudes 2 helps to formulate the calculation of observables. For example, the sim- 1 For a review of scattering in quantum chromodynamics in the regime of high energies, see e.g. Ref. [6]. 2 Of course, generally speaking, the existence of a symmetry implies constraints on observables and on theories. Interestingly enough, in the context of (toy models for) high-energy scattering, boost invariance led to stringent constraints on the form of the elementary processes (such as parton recombination, or other nonlinear processes that slow down parton evolution in the very high-density regime) which should be taken into account at ultra-high energies [18]. plest proof of the BK equation consists in writing down the change of the partonic content of the onium in an infinitesimal boost, starting from the restframe of the onium. Here, we will take advantage of boost invariance to select a specific class of frames in which we will be able to evaluate a particular probability distribution, which is a priori very difficult to calculate, with the help of the limited available knowledge of the solution to the BK equation.
The main outcome of our work is a partonic picture of the scattering in different frames, which turns out to enable the derivation of an expression of the asymptotics of the probability distribution of the rapidity at which the slowest ancestor of all dipoles that interact with the nucleus has branched. The latter quantity characterizes the correlations of the interacting dipoles. While it is not directly an observable, it was shown to be related to the rapidity gap distribution in diffractive dissociation events [19,20]. Last but not least, it is a quantity of more general interest in the study of branching random walks.
We shall start (Sec. 2) by formulating the scattering amplitude as well as the distribution of the branching rapidity in two different ways: A formulation (in terms of evolution equations) that can be implemented numerically, and a formulation that will set the basis for an approximation scheme exposed in Sec. 3 and used to arrive at analytical asymptotic expressions. In Sec. 4, we compare our analytical predictions to numerical solutions to the complete equations, and we present our conclusions and some prospects in Sec. 5. Appendix A outines the evaluation of a useful integral, and Appendix B presents an alternative numerical model.
Two formulations for the amplitudes
In this paper, we shall address the following quantities: • The forward elastic scattering amplitude T 1 of the onium off the nucleus at a fixed impact parameter, or equivalently, the corresponding S-matrix element S ≡ 1 − T 1 ; • The probability T 2 that at least two dipoles present in the Fock state of the onium in the considered frame at the time of the interaction are involved in the scattering, as well as a particular differential G: The probability distribution of the rapidity relative to the nucleus at which the slowest common ancestor of all interacting dipoles has branched.
The most straightforward formulation of the calculation of T 1 , T 2 and G consists in writing down evolution equations with respect to the total rapidity (Sec. 2.1). We shall then introduce framedependent representations of the solutions to such equations (Sec. 2.2).
Forward elastic amplitude and the Balitsky-Kovchegov equation
The S-matrix element for the forward elastic interaction of a color dipole of transverse (two-dimensional) size r with a nucleus at relative rapidity Y obeys the Balitsky-Kovchegov (BK) equation [7,8] where we have assumed homogeneity and isotropy (S only depends on the modulus r of r, not on its orientation nor on the absolute position of the dipole in the transverse plane): In practice, this holds for nearly central collisions of small dipoles with very extended nuclei. Furthermore, this very equation is the lowest-order approximation of the QCD evolution 3 in the limit of large Y , large atomic number, large number of colors N c . The constantᾱ that controls the pace of the evolution reads α ≡ α s N c /π, α s being the QCD coupling.
The simplest way to derive this evolution equation is to start from the restframe of the onium in which the nucleus is evolved at rapidity Y , and to interpret S(Y, r) as the probability that an onium of size r, in its bare state, does not interact with the nucleus. Then, one increases the total scattering rapidity boosting the onium by dY , keeping the rapidity of the nucleus fixed. In the QCD dipole model [15], the scattering configuration of the initial onium may then either become a set of two color dipoles of sizes r and r − r (up to d 2 r ), with probabilitȳ α dY dp 1→2 (r, r ) ≡ᾱ dY d 2 r 2π or may stay a single dipole, with probability 4 1 −ᾱ dY r dp 1→2 (r, r ). Hence from which Eq. (1) easily follows. The constantᾱ always enters as a scaling factor of the rapidity: Therefore, it is convenient to absorb it into the rapidity variable, defining y ≡ᾱY . From now on, we will exclusively use this rescaled rapidity, which we will nevertheless keep calling "rapidity". With the help of these notations, the BK equation reads ∂ y S(y, r) = r dp 1→2 (r, r ) S(y, r )S(y, |r − r |) − S(y, r) .
The initial condition at rapidity y = 0 corresponds to the scattering amplitude of the dipole with an unevolved nucleus: It is usually assumed to have the McLerran-Venugopalan form [21] where the momentum Q A , called the "saturation momentum", is characteristic of the nucleus. (Its value is of the order of 1 GeV for a large nucleus). In this model, the amplitude T 1 ≡ 1 − S is steeply falling from 1 to 0 as r becomes smaller, especially since the relevant scale for the dipole sizes is logarithmic: As a matter of fact, it is a Gaussian function of this variable. The typical value of r at which the transition between S = 0 and S = 1 happens is r ∼ 1/Q A . The function S is almost tantamount to a Heaviside distribution The solution to the BK equation is known asymptotically [22,14]: 5 where up to an additive constant of order unity, which, in the limits of interest here, may always be absorbed into a rescaling of the overall constant in T 1 . This expression is correct for very large values of y. In particular, the logarithmic singularity for y → 0 would be regularized after resummation of higher orders, in such a way that ln(Q 2 s (y)/Q 2 A ) −→ y→0 0. Q s (y) is the saturation momentum of the nucleus at rapidity y, namely 1/Q s (y) is the typical value of the transverse size of a dipole that interacts with it at that rapidity above which the dipole gets absorbed with probability of order unity. Equations (7) and (8) are only valid for 1 < ln 2 [r 2 Q 2 s (y)] χ (γ 0 )y, which, up to strong inequalities, defines the scaling region. The numerical values of the parameters γ 0 , χ (γ 0 ), χ (γ 0 ) can be found e.g. in Ref. [22]. They are not relevant in our discussions: The only important point is that they are all of order 1.
Note that the BK equation is also the evolution equation for the probability that there is no dipole larger than 1/Q A in the state of the onium evolved to the rapidity y, when the initial condition is taken to be exactly the Heaviside distribution (6).
Multiple scatterings
The set of dipoles which interact with the nucleus necessarily stem from the branchings of a single dipole: their "last common ancestor". This is because we start the evolution with a single dipole (the onium), and because, as manifest in a Hamiltonian formulation of QCD, in the absence of recombination mechanism, partons evolve with rapidity through elementary 1 → 2 splitting processes. 6 We want to compute the distribution of the rapidity y 1 , with respect to the nucleus, at which this ancestor has branched.
Let us call G(y, r; y 1 )dy 1 the joint probability that the onium of initial size r interact with the nucleus and that the splitting rapidity of the last common ancestor be y 1 up to dy 1 , the total rapidity of the interaction being y. An evolution equation may be obtained, using the same method as for S. One starts with the frame in which the nucleus is boosted to the rapidity y ≥ y 1 , while the onium of size r is at rest. One then increases the total rapidity by dy, keeping y 1 fixed, through an infinitesimal boost of the onium. In this rapidity interval, the onium may split to two dipoles with probability dp 1→2 (r, r )dy, or stay a single dipole with probability 1 − dp 1→2 (r, r )dy.
For an ensemble of events restricted to those without branching, G(y + dy, r; y 1 ) is just G(y, r; y 1 ). For events in which, instead, the initial dipole branches, one and only one of the offspring dipoles may scatter. So in this case, G(y, r; y 1 ) is replaced by two terms each consisting in the product of a factor G and a factor S, the arguments of which are either r or |r − r |. Taking the sum over all possible events weighted by their probabilities, we get G(y + dy, r; y 1 ) = 1 − dy r dp 1→2 (r, r ) G(y, r; y 1 ) + dy r dp 1→2 (r, r ) G(y, r ; y 1 )S(y, |r − r |) + G(y, |r − r |; y 1 )S(y, r ) . (9) Enforcing the limit dy → 0, we obtain the evolution in the form of the following integro-differential equation: ∂ y G(y, r; y 1 ) = r dp 1→2 (r, r ) G(y, r ; y 1 )S(y, |r − r |) + G(y, |r − r |; y 1 )S(y, r ) − G(y, r; y 1 ) .
(10)
The initial condition has to be set when the total rapidity coincides with the splitting rapidity of the common ancestor: y = y 1 . In this case, there is no choice: The onium has to branch at this very rapidity y 1 , and each offspring must scatter. This translates into the following equation: G(y 1 , r; y 1 ) = r dp 1→2 (r, r ) 1 − S(y 1 , r ) 1 − S(y 1 , |r − r |) .
The equations (10), (11) for G were written for the first time in Ref. [25,26], compared to the Kovchegov-Levin equations for the rapidity-gap distribution in diffractive dissociation, and solved numerically.
Let us introduce the probability T 2 that there are at least two scatterings with the nucleus boosted to the rapidity y 0 . This is just an integral of G over y 1 : T 2 (y, r; y 0 ) = y y 0 dy 1 G(y, r; y 1 ). (12) A similar quantity has recently been identified as an estimator of the contribution of higher twists to total cross sections [27]. (Note that evidence for higher-twist effects was also found earlier in the DESY-HERA data for diffractive deep-inelastic scattering, see e.g. Ref. [28]). T 2 actually obeys an evolution equation straightforward to deduce from the evolution equation for G, which we may write as ∂ y T 2 (y, r; y 0 ) = r dp 1→2 (r, r ) T 2 (y, r ; y 0 ) + T 2 (y, |r − r |; y 0 ) − T 2 (y, r; y 0 ) − T 2 (y, r ; y 0 )T 1 (y, |r − r |) − T 1 (y, r )T 2 (y, |r − r |; y 0 ) + T 1 (y, r )T 1 (y, |r − r |) , (13) with the initial condition at y = y 0 which reads T 2 (y 0 , r; y 0 ) = 0.
The evolution equations (10) and (13) for G and T 2 respectively may be solved numerically (with the help of a solution to the BK equation (7)), but no analytical expression is known. As we will see, we can however obtain exact asymptotic expressions for these quantities (actually for the ratios G/T 1 and T 2 /T 1 ) in a model expected to capture the main features of the QCD dipole model and of more general branching random walks. The starting point will be a useful representation of T 2 in terms of dipole densities and of the dipole-nucleus scattering amplitude T 1 , that we shall expose in the next section.
Frame-dependent representations
In this section, we discuss a representation of the solutions to these evolution equations which will prove useful to set up approximation schemes, from which we shall find asymptotic expressions.
Let us choose a frame in which the nucleus is boosted at rapidity y 0 , and the onium at rapiditỹ y 0 ≡ y − y 0 in the opposite sense. We will consider frames defined by a large y 0 , butỹ 0 will not be smaller than a non-negligible fraction of the total rapidity y.
Instead of using as variables the sizes r of the dipoles or the saturation momentum at rapidity y, Q s (y), we shall express all functions with the help of the logarithms of these sizes and of the momentum, defining
S-matrix element
The following formula is an exact representation of the solution to Eq. (4): where the averaging is over all the dipole configurations of the onium at rapidityỹ 0 (with respect to the onium), represented by the set of log-sizes {x i }. The functions S that appear left and right are the same, but evaluated at two different rapidities. Because of boost invariance, S in the lefthand side must be independent of y 0 chosen in the righthand side. S defined in Eq. (15) obeys the BK equation (4). To check this statement, it is enough to see that increasing y by dy amounts to increasingỹ 0 by the same dy, and to decomposing the averaging over the dipole configurations atỹ 0 + dy as where the sets {x i } and {x i } represent the dipole configurations at rapidityỹ 0 of initial dipoles of respective sizes r and r − r . Simple manipulations and replacements lead to Eq. (4), after the limit dy → 0 has been taken. We may rewrite Eq. (15) with the help of the number density n(x) of dipoles of log-size x: where the product is now over all the bins in dipole size, of width dx . Note that n(x ) is a random density, the distribution of which depends on the size of the initial dipole and on the evolution rapiditỹ y 0 . This equation can be expressed for T 1 : Now, we assume that the dipoles that effectively contribute to the integral all have log-sizes x such that T 1 (y 0 , x ) 1. This is verified if the onium configurations which contain individual dipoles larger than the inverse saturation scale of the nucleus only bring a negligible contribution to the overall amplitude. We will check a posteriori that it is a consistent assumption. In the framework of this approximation, we can expand the logarithm in Eq. (18), and deduce an elegant formula for the amplitude T 1 = 1 − S: where we have introduced the notation for the overlap of the dipole-nucleus scattering amplitude and the dipole density in the onium.
Contribution of multiple scatterings
Let us compute the amplitude for scattering with at least two exchanges between the configuration of the onium at rapidityỹ 0 and the nucleus evolved to the rapidity y 0 . The exact formula reads In the same way as in the case of the S-matrix element, we can show that the righthand side of Eq. (21) obeys the evolution equation (13). T 2 obviously depends on y 0 . G instead, which is formally a rapidity derivative of T 2 , will be independent of the choice of frame. Assuming again that S(y 0 , x i ) 1 for all dipoles in the relevant configurations, we get We are now going to evaluate the right-hand sides of Eqs. (19) and (23). This cannot be done through a straightforward calculation, but a simple model for the realizations of branching random walks/dipole evolution can be used.
Asymptotic amplitudes from the phenomenological model for front fluctuations
In the following, we will stick to the large-rapidity limit, and pick the size of the initial onium in the so-called scaling region. This means that (A χ (γ 0 ) factor would multiply y, but it is of order 1, so it does not modify these strong inequalities.) We shall actually take a slightly stronger condition on the lower bound on x − X y : We will always assume that it be much larger than the logarithms of the rapidities y and y 0 . Setting r much smaller than 1/Q s (y), as encoded in the first strong inequality, implies that a typical realization of the dipole evolution would interact with very small probability. So we need fluctuations to create larger dipoles. We shall now introduce a model for these fluctuations and apply it to the evaluation of T 1 , T 2 , G.
Model for the dipole distribution
We present here a slightly modified formulation of the model for the evolution of branching random walks, and in particular for the QCD dipole evolution, that was initially developed in Ref. [29] and applied to particle physics in Ref. [30].
We assume that the evolution process develops essentially in a deterministic, "mean-field" way, such that the density of dipoles of log-size x at rapidity y i with respect to the nucleus, namely after evolution over the rapidity rangeỹ i , reads There may be an additive constant of order 1 inXỹ i , but to the accuracy we are considering, it can be absorbed into the overall constant C 1 . Again, if this formula is to be extrapolated to the non-asymptotic regime ofỹ i , the logarithm has to be regularized in such a way thatX 0 = 0. This formula represents the dipole density in a typical realization of the dipole evolution, in the absence of a large fluctuation, in a region of size of order √ỹ i near the typical log-size of the largest dipole, which is such that ln In practice, it is obtained from the solution of a linearized BK (or FKPP) equation with a cutoff that simulates the effect on the evolution of the discreteness of the dipoles in realizations; see Ref. [29]. On top of this deterministic particle density, we assume that one single fluctuation occurs after some random evolution rapidityỹ 1 , and that this fluctuation consists in a dipole of size larger than the largest dipole in typical configurations by a factor e δ/2 . After this fluctuation has occurred, the large produced dipole builds up into a second front in a deterministic way upon further rapidity evolution.
We need the distribution of δ. We guess that it coincides with the probability of observing the largest dipole with a log-size shifted by (−δ) with respect to the mean-field tip of the distribution. This probability solves the BK equation (see the remark at the end of Sec. 2.1.1), and thus, has the same form as Eq. (7): The rate at evolution rapidityỹ 1 reads, asymptotically for largeỹ 1 and δ, In the kinematical region we consider, the mean-field evolution of the initial onium would not alone trigger a scattering: Hence the onium always scatters exclusively through the smaller front that stems from the fluctuation. Each dipole in the state of the onium at the interaction rapidity scatters independently, with an amplitudeT 1 (y 0 , x ) that solves the BK equation (4) with S substituted with 1 −T 1 . We shall denote by the log-saturation scale of the nucleus front at rapidity y 0 , see Eq. (8) with the notation (14).
Let us now express T 1 and T 2 in this model. The overlap of the amplitudeT 1 and of the dipole number densityn that appear in Eq. (19) and (23) reads, in this model, where Ξ δ,ỹ 1 ≡ x +Xỹ 1 − δ is the log-size of the lead dipole at rapidityỹ 1 . Then and We can also obtain a formula for G itself in the framework of the phenomenological model: It is enough to use Eq. (22), from which one sees that the analytical expressions of T 2 and G only differ by the presence of the integration over y 1 in the expression of the former, a fact which is easy to understand. Indeed, the essence of the phenomenological model is to single out one dipole in the state of the onium evolved to the rapidityỹ 1 that will stand for the common ancestor of all dipoles which scatter after evolution to the rapidityỹ 0 . When the y 1 integration in Eq. (31) is left undone, then G reads Let us introduce the distance, at the scattering rapidity, between the tip of the dipole distribution and the top of the nucleus front: In other words, ∆(y 0 ; δ, y 1 ) is the logarithm of the squared ratio of the size of the smallest dipole which would scatter with probability of order unity with the nucleus in a scattering of relative rapidity y 0 , and of the size of the largest dipole in the actual state of the onium at rapidityỹ 0 . It may be rewritten as As commented above (see e.g. after Eq. (25)), the logarithmic term must be regularized in the limits y 1 → y 0 and y 1 → y. Furthermore, with the considered choice of frame and parameters, this logarithmic term is always small compared to x − X y . We shall first show that we may restrict ourselves to fluctuations such that ∆ ≥ 0. To this aim, we evaluate parametrically the contribution to T 1 (y, x) of the integration region ∆ ≤ 0, namely Starting from Eq. (30), we see that we have the following upper bound on the contribution of this region to T 1 : The Gaussian factor in p may be replaced by an effective upper cutoff on the integration over δ, set at δ 1 ≡ 2χ (γ 0 )ỹ 1 , and the integration over δ can then be performed. The condition δ 0 ≤ δ 1 for this integral not to be null implies that All in all, we get a further bound on T 1 : where we have reminded that the apparent singularity at the lower bound of the integral over y 1 needs to be regularized, in such a way that the integrand remains finite of order 1. The integral is now at most of order one. Hence, we see that T 1 is suppressed by at least a factor y 3/2 0 1 with respect to the expected result, see Eq. (7). This proves that the region ∆ ≤ 0 can be neglected. From now on, we will only consider the integration region in which ∆ ≥ 0, i.e. δ ≤ δ 0 .
Actually, a closer look would show that only the region in which ∆ 3 2γ 0 ln y 0 contributes significantly. This means physically that the scattering amplitude of all the individual dipoles in the fluctuation, at rapidityỹ 0 , is very small, consistently with the assumption that led to Eqs. (30), (31).
We see that all the functions T 1 , T 2 and G are written in terms of p and I. So let us express I in the phenomenological model. The functionn that appears in Eq. (29) is replaced by the expression given in Eq. (25). As forT 1 , since only the region x > X y 0 will be probed, we can use the solution (7) of the BK equation re-expressed in appropriate variables, namelȳ We get We shall now compute the different quantities in the framework of the phenomenological model. T 2 is explicitly frame-dependent, and although T 1 and G are boost-invariant, their evaluation will depend upon the chosen frame.
Amplitudes in a frame in which the nucleus is highly boosted
In this section, we choose the frame in which the nucleus is boosted to rapidity y 0 such that We shift the integration variable x, definingx ≡ x −Xỹ 0 −ỹ 1 − Ξ δ,ỹ 1 the log-size of the dipoles relative to the log-size of the largest dipole, at the tip of the particle distribution. The overlap integral then reads I(y 0 ; δ, y 1 ) = C 1 C 2 e −γ 0 ∆(y 0 ;δ,y 1 ) exp − ∆ 2 (y 0 ; δ, y 1 ) 2χ (γ 0 )y 1 We observe that the integral is determined by a large integration region, up tox ∼ √ y 0 . When (x −X y ) 2 y 0 , we can neglect ∆(y 0 ; δ, y 1 ) compared tox, and the integral takes a simpler form, which can be integrated exactly. Moreover, we will check a posteriori that typically,ỹ 1 y, hence y 1 ∼ y, and the Gaussian factor involving ∆(y 0 ; δ, y 1 ) can be set to unity. Then we find Replacing ∆(y 0 ; δ, y 1 ) by its expression (34), we arrive at We see that I turns out to be independent of y 0 . Let us introduce the following notation: p 1 e γ 0 δ is just the overlap of the front of the nucleus with that of an onium if the latter were evolved in a purely deterministic way and starting at a log-size x − δ.
Forward elastic scattering amplitude T 1
The amplitude T 1 is obtained by substituting Eq. (42) into Eq. (30), with the restriction on δ imposed by the condition ∆ ≥ 0: We shift δ, defining the new integration variable δ ≡ δ + 3 2γ 0 ln y y 1ỹ1 . Due to the e −γ 0 (x−Xy) factor in p 1 , the integration domain extends effectively to δ x − X y , a region much larger than the logarithm of any rapidity appearing in this problem. Hence, the lower integration bound on δ may be kept to 0, δ can just be identified to δ in the Gaussian factor, and the upper bound on δ can be released since the region δ > δ 0 x − X y gets anyway cutoff by the factors in the integrand. We get Theỹ 1 integration can be performed up to a correction of order 1/ √ y, noticing that the integral is dominated by the regionỹ 1 y, hence y 1 y. One can in particular replace the upper bound by ∞: What remains is an integration over δ : This expression can be interpreted as the scattering amplitude of an onium the state of which was determined by a front fluctuation in the terminology of Ref. [29,30], which is meant to be a fluctuation in the beginning of the evolution that essentially shifts the whole dipole distribution towards larger sizes. The size of this class of fluctuations was assigned an exponential distribution ∝ e −γ 0 δ for large δ : This is precisely the weight of δ that appears in the remaining integration.
Equation (47) may be rewritten with the help of the integral I 1 defined and evaluated in Appendix A: Using Eq. (83) and replacing p 1 by its definition in Eq. (42), the final result reads We have just recovered the scaling limit of the known solution to the BK equation, see Eq. (7). We have assumed throughout (x − X y ) 2 y, so we cannot get consistently the finite-y corrections that appear in the form of an exponential of the ratio of the first over the second scale multiplied by a negative constant factor. We have learned from this calculation that, in the considered frame, the realizations of the Fock states which trigger events look like typical realizations, as far as their shapes is concerned, but overall shifted towards larger dipole sizes by a multiplicative factor, through a fluctuation occurring at the very beginning of the evolution.
Multiple scatterings: G and T 2
As for the calculation of G, we start with Eq. (32), and substitute I by the expression obtained in Eq. (42): Again, because of the form of the integrand, δ does not exceed x − X y . This implies that the upper bound δ 0 can be replaced by +∞. Furthermore, x − X y is assumed, as a consequence of our choice of frame, to be much less than √ y 0 . Hence if one restricts oneself to values ofỹ 1 ≡ y − y 1 not smaller than y 0 , the Gaussian factor can be set to 1. Then, performing the change of variable t ≡ e γ 0 δ , G boils down to an integral computed in Appendix A, up to an overall factor: It follows that Since (x − X y ) e −γ 0 (x−Xy) = T 1 (y, x) × CC 1 C 2 π[χ (γ 0 )] 2 −1 , we arrive at the following expression for the distribution of y 1 : We see that the small-ỹ 1 region is highly singular, but the singularity has to be cut off by a factor that is subleading whenỹ 1 is taken on the order of y. This region would correspond to the production of a large dipole in the very beginning of the evolution of the onium, but this is suppressed: Indeed, the mechanism leading to a particle away from the mean position of the lead particle is diffusive, and the diffusion radius grows like √ỹ 1 . Consequently, one may expect the expression (53) to be supplemented by a multiplicative factor D(x − X y , y − y 1 ), where The scattering amplitude conditioned to having at least two scatterings between the state of the onium evolved to rapidityỹ 0 and the nucleus, T 2 , is an integral of G over y 1 . Starting with Eq. (50), its evaluation goes along the same lines as that of T 1 above. The y 1 -integration can be performed in the first place. We are then left with an integral over δ, which takes the form Using the evaluation of I 2 in Appendix A, Eq. (84), and replacing p 1 by its expression (43), we find a very simple relation between T 1 and T 2 in this frame: It turns out that we would have got the same result by integrating Eq. (53) supplemented with the diffusive factor D(x − X y ,ỹ 1 ) defined in Eq. (54) which cuts off the very small y 1 region. We note that the overall constant in front of the ratios T 2 /T 1 and G/T 1 is sensitive to the detailed form of the interaction between the dipoles and the nucleus, which was not the case for T 1 . Within the assumptions of the phenomenological model, this form is unambiguous: The number of scatterings at the time of the interaction obeys a Poisson law of parameter I. Whether the overall constant found in this model is the correct one for branching random walks and for the QCD dipole model depends on the ability of the phenomenological model to capture accurately enough the features of the latter models: We will need numerical calculations to check it (see Sec. 4 below).
Finally, for the same arguments as the ones presented in Sec. 3.2.1, the assumption that no single dipole has a significant probability to scatter also proves correct a posteriori.
What happens in a frame in which the nucleus is less boosted?
The choice of the reference frame was very important in the calculation above: We chose a frame in which the nucleus is highly boosted, such that y 0 (x−X y ) 2 . Such a choice implies that the scattering configurations are dominated by fluctuations which occur very early in the rapidity evolution. This is a perfectly valid choice, as long as we pick x in the scaling region, i.e. (x − X y ) 2 y. However, any other frame should be allowed. We shall investigate the scattering picture in frames in which the nucleus is at rest or close to rest.
Nucleus restframe
The amplitude T 1 was analyzed in the restframe of the nucleus (y 0 = 0) in Ref. [30]. In that frame, the nucleus has not developed a universal front: The scattering amplitude of a dipole becomes very small as soon as the size of this dipole gets smaller than the inverse saturation scale 1/Q A ; see Eq. (5). Therefore, in all events, the fluctuations of the partonic content of the onium must produce at least one dipole which will be completely absorbed by the nucleus, namely, which has a size larger than 1/Q A . The formulation of T 1 simply reads where T 1 (0, x ) = 1 − S(0, x ) is the McLerran-Venugopalan amplitude given in Eq. (5) that we may approximate by a Heaviside distribution with support the set of negative real numbers. The leading term in the integral over x can then be obtained quite easily. The integration over x is dominated by a region of log-size of order 1 around x = 0. The result is proportional to which is of course what is expected at the parametric level. However, the overall constant c cannot be easily related to C, C 1 and C 2 of the phenomenological model. This is because the latter are unambiguously defined for an evolved front, once a convention for the definition of the front position/saturation scale has been chosen, but the transition between the initial condition and the well-developed front is not controled analytically in the initial stages of the evolution. T 2 and G cannot be calculated in this frame. Indeed, their evaluation requires the precise understanding of the particle distribution in fluctuations happening near the boundary of the BRW, which is still an unsolved problem.
Slightly boosted nucleus
We now investigate the case of the frame in which the nucleus is boosted only slightly. We shall choose a frame in which the rapidity of the nucleus satisfies, parametrically, While for the opposite ordering between y 0 and (x−X y ) 2 one could perform a relatively straightforward calculation, this case is much trickier. We shall show how the calculations of T 1 and T 2 go, which will enable us to understand what the typical state of the onium looks like when viewed from this particular frame.
Calculation of T 1 . Anticipating that the main contribution will come from the configurations which do not overlap with the saturation region of the nucleus, we expand the exponential in Eq. (30): Substituting I, we get It is not possible to perform these nested integrals exactly, but we can extract the asymptotic expression of T 1 in a definite limit.
The Gaussian factors (I),(II),(III) set effective cutoffs: Their analysis enables us to assess which subdomains of the integration region will give the dominant contribution, and thus to judge which approximations we may afford without altering the value of the integral in the asymptotic limits of interest here.
The presence of the factor (I) implies that x − X y 0 must be at most of order √ y 0 . Since x is larger than the position of the leftmost tip of the dipole distribution, ∆(y 0 ; δ, y 1 ) defined in Eq. (33) is also at most of order √ y 0 . This in turn implies that the size of the fluctuation be δ ∼ x − X y , up to O( √ y 0 ).
The factor (II) forcesỹ 0 −ỹ 1 y 0 , i.e.ỹ 1 y up to corrections of order y 0 . So whenever a factorỹ 1 appears one can safely replace it by y. The last factor (III) is always of order unity since δ is of order x − X y , and since we have chosen to stick to the scaling region, defined by x − X y √ y ∼ √ỹ 1 . We further observe that theỹ 1 -dependence ofXỹ 0 −ỹ 1 + Ξ δ,ỹ 1 is only logarithmic, hence it can be neglected here, given that these terms do not appear in exponential factors which would have enhanced their contribution. In other words, one can afford the approximation Once this approximation is implemented, one can perform the integral over y 1 . Taking into account that the dominant region is such that y 1 is close to y 0 , we set the upper boundary to +∞, and replace y/ỹ 1 by 1. It just gives a number: Further, the integral over x boils down to the integral of an exponential. After the change of integra- Finally, the integration over δ is dominated by a region of size √ y 0 close to x − X y , namely Putting all factors together, we get Eq. (49), the overall multiplicative constants being identical: Hence we have checked explicitly that boost invariance holds. Interestingly enough, the physical pictures in both frames are very different. Indeed, from the analysis of the integration domain, we see that in the frame in which y 0 (x − X y ) 2 , the fluctuation occurs typically late in the onium evolution, at rapidityỹ 1 close to the scattering rapidityỹ 0 . It takes place in a window of size of order y 0 in such a way that the overlap with the front of the nucleus, of size √ y 0 , may be significant. This is necessary since the fluctuation needs to extend far out of the "mean field" region, and thus requires a large rapidity range to develop. The particle front that interacts with the nucleus, which results from the evolution of the fluctuation, is of size √ y 0 , that is, it has just the right size to have an optimal overlap with the front of the nucleus.
Calculation of T 2 . This quantity is significantly more difficult to compute. First, unlike in the case of T 1 , expanding the exponential in Eq. (31) is not licit, and leads to a loss of control of the constant factors multiplying the leading term. What happens physically is that requiring at least two scatterings pushes δ to take a value for which I(δ, y 1 ) ∼ 1 in each event, which limits the possible values of δ to a narrow interval. Hence the integral over δ does not bring a factor √ y 0 reflecting the size of the integration region. This is essentially the difference between the calculation of T 1 and that of T 2 in this regime. This reasoning leads to the following estimate of its parametric form: In order to get a more complete expression, we may recognize that G/T 1 is boost-invariant, and integrate its expression obtained in Eq. (53), supplemented with the Gaussian factor (54), over y 1 . In this case, the integral is dominated by the region close to y 1 ∼ y 0 , i.e.ỹ 1 y. The result has the same parametric form as the one just guessed:
Comparing the model predictions with the solutions to the exact equations
We shall now check the results we have obtained using the phenomenological model by solving numerically the exact equations. In order to compare more easily different values of y, it is useful to introduce the overlap q ≡ỹ 1 /y, representing the fraction of the total rapidity over which there is a unique common ancestor of the dipoles that eventually interact. Its distribution is just given by G/T 1 , up to the change of variable and the corresponding Jacobian: The y-dependence of this asymptotic result is very simple, consisting only in the multiplicative factor 1/ √ y: Therefore, we shall keep it implicit in the definition of the asymptotic distribution. The "∞" subscript reminds that this expression is valid for asymptotic values of y. We shall not use the QCD dipole model, but the simpler branching random walk introduced in Ref. [31] and further investigated in Ref. [32]. Indeed, it is know that the form of the asymptotics is the same for all models in the universality class of branching diffusion. Only the parameters γ 0 , χ (γ 0 ), and χ (γ 0 ), which depend on the detailed elementary processes, need to be substituted.
Definition of the implemented model
We consider a branching random walk in discrete space and time, defined by the following processes: Between the rapidities y and y + δy, a particle on site x may jump to the site on the left (i.e. at position x − δx) or on the right (x + δx) with respective probabilities 1 2 (1 − δy), or may branch into two particles on the same site x with probability δy.
The main differences with respect to the QCD dipole model is that in the latter, the diffusion and the branching actually happen at the same time through a single process, and that QCD is a theory in the continuum. But these differences should not affect the asymptotics of the observables we are considering.
The fundamental quantity for us is the probability that there is no particle to the right of the site at some position X. This is the equivalent of the S-matrix element in the QCD case. It evolves in rapidity according to the equivalent of the BK equation (4) for this model, which is the finite-difference equation with the initial condition S(y = 0, x ≤ 0) = 0 and S(y = 0, x > 0) = 1. In the numerical calculation, we shall take the following values for the parameters: The values of γ 0 , χ (γ 0 ), χ (γ 0 ) are obtained from the general solution to Eq. (70) linearized near S ∼ 1. For this model, we find [32] γ 0 = 1.4319525 · · · , χ (γ 0 ) = 1.3943622 · · · , χ (γ 0 ) = 0.96095291 · · · . (72)
Numerical calculation of the distribution of the splitting rapidity of the parent dipole
The equation to solve to get the equivalent of G in the framework of this model is the following: G(y + δy, x; y 1 ) = 1 2 (1 − δy) [G(y, x − δx; y 1 ) + G(y, x + δx; y 1 )] + 2δy G(y, x; y 1 ) S(y, x), with (Compare to Eq. (10) and (11) respectively obtained in the dipole model).
In order to satisfy in some optimal way the double constraint in Eq. (24), we set x to a value X such that X X y + √ κ y 1/4 , where κ is a constant that we shall pick in the set {1, 2, 4}. In general, we cannot achieve the equality because X is the position of a lattice site, so it is discrete, while the r.h.s. is a real number: We have picked the closest site to the left of the position in the r.h.s. Up to a numerical factor, the r.h.s. is the geometric average of the two bounds on X − X y in Eq. (24). Varying this constant enables one to go more or less deep in the scaling region.
We have collected data for y up to O(10 6 ). We plot the distribution π y (q) at finite rapidity rescaled by √ y for different y and κ in Fig. 1, together with the expected infinite-y asymptotic distribution π ∞ (q). In order to appreciate the convergence quantitatively, we pick a point of fixed q, and we compare the measured π y (q) at finite rapidity to the expected one π ∞ (q) at y = ∞. In practice, we have chosen q = 0.5, but we have also tried other values and got similar results for this ratio. We plot the complementary to one of this ratio against ln y/ √ y in Fig. 2. We expect a curve that goes through the origin: We see a quite good convergence of 1 − π y (q)/π ∞ (q) to 0 when y → ∞. Note that some non-smooth structures appear. They just reflect the discreteness of the model, which forces us to set X to discrete values according to the procedure described above. The discretization step in x is δx = 0.1 in this model, which is not that small compared to y 1/4 , even for very large y. So moving X by δx can lead to sizable differences in π y (q). Of course, these differences should get smaller and smaller as y → ∞.
In order to guide the eye, we superimpose the following function to the data: We fit the parameters a, b 11 , b 12 , to the numerical calculations. We get reasonable values for all these parameters for the considered choices of X: b 11 and b 12 are of order 1, while a is found close to zero; see Tab. 1. The constant a is of the order of a percent, when we expect it to vanish. But we can hardly aim for better, because of the structures induced by the discreteness of the model, which makes the fitting procedure by a smooth function dependent on the choice of the points. Therefore, we conclude that our analytical formula (69) is well-supported by this numerical calculation. However, these results show that the finite-y corrections are definitely very large, and the convergence slow: Although we have computed π y for values of y on the order of 10 6 , we have not managed to approach the asymptotics by better than about 3%. Figure 2 and the fitted formula (76) seem to indicate that the correction to π ∞ (q) may take the form of a multiplicative factor 1 + const × ln y/ √ y . But we have no theory that may enable us neither to understand nor to guess the form of π y (q) beyond the leading term in the limit of large y.
We have also implemented independently another branching random walk model, and we have reached the same conclusion; see Appendix B for a presentation of the model and of the obtained results.
Summary and outlook
In onium-nucleus scattering, in a frame in which the onium moves with a large rapidityỹ 0 , the onium interacts through a typically dense quantum state made of gluons, that we represented by a set of color dipoles. Only a subset of these dipoles actually exchange energy with the nucleus. Having at least one dipole in this set is necessary to have a scattering event: This requirement defines the forward elastic scattering amplitude T 1 . Calculating the joint probability T 2 to have a scattering and at least two dipoles in the set makes possible to get quantitative information on the correlations of the dipoles involved in the interaction.
In this paper, we have shown that the shape of the partonic configurations of an onium that interacts with a large nucleus depends qualitatively on the chosen reference frame. If the nucleus is highly boosted, namely if its rapidity y 0 is much larger than ln 2 [r 2 Q 2 s (y)], then the dipole distribution at the time of the interaction with the nucleus looks like a typical ("mean-field") distribution just shifted (through a "front fluctuation" occurring in the very beginning of the evolution) towards larger sizes. If instead the nucleus is less boosted, 1 y 0 ln 2 [r 2 Q 2 s (y)], then the dipoles which interact with it stem from a tip fluctuation occurring at much larger rapiditiesỹ 1 in the onium evolution, of order y such that y 1 ∼ y 0 .
Choosing a frame such that the tip fluctuation, the offspring of which scatter with the nucleus, be sufficiently developed, we were able to calculate the asymptotics of the distribution of the splitting rapidity of the slowest ancestor of the set of dipoles that effectively interact with the nucleus, including the overall constant. We have found that its ratio to the total rapidity of the scattering, a quantity that we denote by q, is distributed as This expression holds when the size r of the onium is chosen in the so-called scaling region, defined as where x and X y just correspond to 1/r and Q s (y) respectively when measured on a logarithmic scale (see Eq. (14)), and X y = χ (γ 0 )y − 3 2γ 0 ln y. Equation (69) is our main quantitative result. A particular realization of the model introduced in Sec. 4 and used to check numerically the calculations is shown in Fig. 3.
After identification of the rapidity with a time variable, we expect this expression to represent the distribution of the relative branching time of the most recent common ancestor of all particles that end up to the right of some predefined position x for any branching random walk -provided that x is chosen in the scaling region. (This q is also called "overlap" in the statistical physics literature). The constants that appear in these expressions are easily calculated from the detailed form of the elementary processes which define the branching random walk.
We observe that our result coincides with a conjecture by Derrida and Mottishaw [33] for a slightly different genealogy problem in the context of general branching random walks: They computed the distribution of the branching time of the most recent common ancestor of two particles of predefined order number counted from the tip of the particle distribution at some given large time. To make contact between our equation (69) and the formula (6) they wrote down in Ref. [33], we just need to identify our π(q) with their p(q), y with the total evolution time t, γ 0 with β, and χ (γ 0 ) with β c v (β c ). The constant χ (γ 0 ), that enters the validity condition, is just to be identified with the critical FKPP front velocity v(β c ).
While this expression was established in [33] through a calculation in the framework of the Generalized Random Energy Model [34], we have derived it in the context of the problem we were addressing from the phenomenological model for branching random walks, which just assumes that the time evolution is essentially deterministic, except for one single fluctuation. Proving our result for the genealogies rigorously, establishing a framework for the systematic calculation of corrections, crucial for applications since the approach to the asymptotics turns out to be very slow, are problems of general interest for branching processes, and exciting challenges for further investigations.
Also, our method would not apply to the specific genealogy problems Derrida-Mottishaw were addressing: While the y (or t)-dependence would turn out correct, the overall constant could not be obtained. The reason for this is that we do not have a sufficient understanding of the particle distributions and of their correlations very close to the lead particle. Trying to build a good picture of the latter [35,31,32] is a long-term program that deserves more efforts.
As for the more specialized diffraction problem, which was the initial motivation for the present work, we also intend to try and extend our calculation to the rate of rapidity gaps in high-energy onium-nucleus scattering. The main crucial difference is that the latter being a quantum mechanical observable, it has no interpretation in purely classical probabilistic terms. However, preliminary investigations seem to indicate that the technical tools developed here can be applied also to that observable, which will be measurable at a future electron-ion collider. to possess several particles to the right of the position X X y=400 + √ 2 × 400 1/4 . The grey zone is the set of non-empty lattice sites for all values of the rapidity. The black lines represent the worldlines of all the particles that end up with a position not less than X at the final rapidity y = 400. The common ancestor of these particles splits at y 1 = 192.43. The inset is a zoom on the branching region around the branching rapidity y 1 , illustrating that this common ancestor indeed stems from a large fluctuation occurring at a rapidity close to y 1 , as assumed in the phenomenological model. This rare realization was generated using the algorithm proposed in Ref. [32].
These integrals can be deduced from more general ones: where Γ is the incomplete Gamma function: We want to calculate the leading term in the small-A limit of the expansion of these integrals to order ε. To this aim, we write Note that the factor A is dimensional, while 1/A in the argument of the logarithm is just the size of the relevant integration region, which extends to ∼ 1/A. Actually, the integral could also be performed by simply restricting the integration region to [1, κ/A], where κ ∼ O(1), and expanding the exponential: The overall normalization of the leading term for A 1 would be identical to the one found from the exact calculation. The details of how the integration region is effectively cut off in the integrand do not matter.
As for I 2 and I 2 , we expand I 2,ε to order ε. We arrive at the following exact expressions, together with their expansion at lowest order for A 1: Note that I 1 and I 2 are identical for small A. However, these integrals are not logarithmic, and therefore, the overall constant factor of the leading term in the limit A 1 of interest here depends on the details of the integrand.
B Numerical calculations in an alternative model
The branching random walk model defined and studied in Sec. 4 was actually a particular discretization, in space and time, of a branching Brownian motion. The FKPP equation [9,10], that is solved e.g. by the probability that there is at least a particle to the right of some predefined X, when the diffusion constant of the Brownian process is set to D, reads Equation (70) for S = 1 − T 1 is a particular lattice discretization of (85), with D = 1 2 . In this Appendix, we shall study an alternative discretization, namely We would recover Eq. (85) with D = 1 in the joint limit δy, δx → 0. The underlying branching random walks are actually quite different in the two discretization schemes. In the present scheme, a particle on a given site has a probability to move, either left or right, of order δy, which is taken small. In the previous scheme instead, it had a probability 1 − δy to move.
In this new scheme, the evolution equation for G reads, for y > y 1 , G(y + δy, x; y 1 ) − G(y, x; y 1 ) = δy δx 2 [G(y, x + δx; y 1 ) − 2G(y, x; y 1 ) + G(y, x − δx, y; y 1 )] + δy G(y, x; y 1 )[1 − 2T 1 (y, x)], (87) and the initial condition is the same as in the previous model, see Eq. (74). We have set δy = 0.02 and δx = 0.25, which leads to the following value of the constants entering the asymptotic quantities of interest: γ 0 = 1.0120279 · · · , χ (γ 0 ) = 1.9659159 · · · , χ (γ 0 ) = 1.9065278 · · · (88) These parameters are quite different from those of the first model; Compare to Eq. (72). We have also chosen a different initial condition: T 1 (y = 0, x) = 1 − exp(e −4γ 0 x ). Our implementation of these evolution equations is independent of the one of the first model, and the very numerical methods used are different. As for the model exposed here, we evolve ln T 1 and ln G instead of T 1 and G directly, at variance with the first model, in order to make sure that we treat accurately-enough the crucial region in which these functions take very small values (see e.g. [29] for a description of such a numerical method).
The overlaps π y (q), shown in Fig. 4 for y ∈ {10 3 , 4 × 10 3 , 1.6 × 10 4 }, are compatible with a convergence to the expected asymptotics (69) -although it is more difficult to judge than for the first numerical model: The maximum value of y for which we have been able to perform the calculation is more than one order of magnitude smaller than in the case of the latter. | 14,875 | 2020-10-29T00:00:00.000 | [
"Physics"
] |
Nanoscale structural evaluation of 0-3 magnetic nanocomposites fabricated by electro-infiltration
Magnetic nanocomposites with 0-3 connectivity, whereby a 0D magnetic nanoparticle phase is embedded into a 3D magnetic metal matrix phase, have gained increased interest for use in applications ranging from integrated power inductor cores to exchange-spring magnets. The electro-infiltration process, in which a metal phase is electroplated through a nanoparticle film phase, is an inexpensive approach compatible with semiconductor fabrication methods for the formation of these nanocomposites. Past demonstrations of electro-infiltrated nanocomposites have relied on scanning electron microscopy and energy dispersive x-ray spectroscopy to evaluate the 0-3 composite structure. However, a detailed investigation of the boundary between the particle and metal matrix phases cannot be performed with these tools, and it is unknown whether the particle/matrix interfaces are dense and void-free. This detail is critical, as the presence of even nanoscale voids would affect any potential magnetic exchange coupling and hence the overall electromagnetic properties of the material. This work seeks to explore the phase boundary of 0-3 magnetic nanocomposite fabricated by electro-infiltration by using scanning transmission electron microscopy and energy-dispersive x-ray spectroscopy to analyze the nanostructure of two different composites—a nickel/iron-oxide composite and a permalloy/iron-oxide composite. High-resolution imaging indicates that the interface between the particle phase and matrix phase is dense and void-free. These results will help guide future studies on the design and implementation of these magnetic nanocomposites for end applications.Magnetic nanocomposites with 0-3 connectivity, whereby a 0D magnetic nanoparticle phase is embedded into a 3D magnetic metal matrix phase, have gained increased interest for use in applications ranging from integrated power inductor cores to exchange-spring magnets. The electro-infiltration process, in which a metal phase is electroplated through a nanoparticle film phase, is an inexpensive approach compatible with semiconductor fabrication methods for the formation of these nanocomposites. Past demonstrations of electro-infiltrated nanocomposites have relied on scanning electron microscopy and energy dispersive x-ray spectroscopy to evaluate the 0-3 composite structure. However, a detailed investigation of the boundary between the particle and metal matrix phases cannot be performed with these tools, and it is unknown whether the particle/matrix interfaces are dense and void-free. This detail is critical, as the presence of even nanoscale voids would affect any potential magnetic exchange coupling and he...
I. INTRODUCTION
The fabrication and application of magnetic nanocomposites, defined by having one or more of their phases possess a characteristic length scale of 1-100 nm, has gained interest due to the desire for new and better hard and soft magnetic materials. These new materials could decrease reliance on rare-earth magnets and improve the performance and efficiency of devices in areas such as power electronics (Sullivan, 2009;Balamurugan et al., 2012;Coey, 2012;Lewis and Jimenez-Villacorta, 2013;Sullivan et al., 2013;and Silveyra et al., 2018). Figure 1a shows the structure of a 0-3 magnetic nanocomposite, which is made from magnetic nanoparticles surrounded by a magnetic metal matrix. These composites have been made using different methods, such as cluster deposition, which requires the use of expensive and/or modified equipment (Balasubramanian et al., 2011 andLiu et al., 2011), and magnetic composite electroplating, which results in low fill ratios on the order of <5% (Guan and Nelson, 2005;2006). Over the last decade the electro-infiltration method, developed by Hayashi et al. and Wen et al., has been suggested as a relatively cheap and semiconductor fabrication compatible method for making 0-3 magnetic particle/matrix composites (Hayashi et al., 2012;Wen et al., 2014;2016;2017;and Smith et al., 2019).
The magnetic and material properties of 0-3 electro-infiltrated composites have been previously documented for various applications (Wen et al., 2016;2017;and Smith et al., 2019). These works relied on scanning electron microscopy (SEM) and energy dispersive x-ray spectroscopy (EDS) to demonstrate successful electro-infiltration, but these methods do not afford the spatial resolution to analyze the boundary between the particle and matrix phases. The phase boundary of these composites is of great importance, because the presence of voids between the particles and the metal matrix can significantly affect the properties of the material as a whole. Specifically, voids between the particle and matrix phases is of known importance for applications such as exchange-spring magnets, where voids would negatively effect the exchange-coupling interactions of the two phases (Skomski and Coey, 1993). The phase boundary is also of importance in inductor core materials, where voids would degrade the overall permeability of the core (Silveyra et al., 2018). This work explores the phase boundary between the particles and metal matrix of 0-3 nanocomposites made using electroinfiltration. To do this, high-resolution S/TEM combined with EDS, is used to probe two different composite samples-a nickel/ironoxide composite and a permalloy/iron-oxide composite. Results show that the iron-oxide nanoparticles (IONs) are surrounded by metal to form a dense, void-free boundary.
II. MATERIALS AND METHODS
Iron-oxide nanoparticles with an average diameter of 31 nm were synthesized using a two-step semi-batch thermal decomposition method adapted from Vreeland et al. and put into suspension (Vreeland et al., 2015). The electroinfiltration process flow can be seen in Fig. 1b. First, a photoresist layer is patterned on top of a silicon substrate that has a 20 nm titanium (Ti) adhesion layer and a 100 nm gold (Au) seed layer sputtered upon it, and the ironoxide nanoparticles are drop cast into the resulting molds. These molds are left to dry into a porous particle film for at least 3 hours in a fume hood. In the next step, the particle layer is placed into an electroplating bath of either Ni or NiFe, where these metals are electroplated through the film at a current density of 10 mA/cm 2 at 54 ○ C (for Ni) or 25 ○ C (for NiFe) with no agitation. A NdFeB retaining magnet, placed ∼1 cm from the backside of the substrate, is used to hold the magnetic particles in place during plating. Finally, the film is removed from the plating bath, rinsed, and the photoresist layer is stripped to leave behind a composite material. After fabrication, a FEI Helios NanoLab 600 Dual-Beam FIB/SEM was used to cut lamellae of the composite materials, and S/TEM images were taken using a FEI Tecnai F20 S/TEM instrument operating at 200 keV with an EDS detector.
III. RESULTS AND DISCUSSION
A. Identification of layers within the sample cross-sections EDS line scans taken along the thickness of each of the samples are used to reveal and identify the presence different layers formed during fabrication. Fig. 2a shows a line scan plot for the Ni/ION composite with each layer labeled, and Fig. 2b shows a line scan plot for the NiFe/ION composite with each layer labeled. In Fig. 2a, the presence of silicon and silicon oxide come from the substrate. Following these should be a 20 nm titanium adhesion layer, but this does not appear due to the low resolution of the line scan. Instead, the 100 nm gold seed layer is observed, present directly above the substrate. After this, the 2 μm Ni/ION composite layer appears, as supported by the combined presence of nickel and iron in the line scan. After the composite layer, a transition is made to a pure nickel layer. Due to the presence of an un-infiltrated layer of particles that is ∼1 μm thick above the pure nickel layer, it is suggested that stress from the electroplating step caused the nickel to stop infiltrating the particle layer and, instead, heave it upward as pure nickel continued to plate below it. The exact cause of this phenomenon requires further exploration.
In Fig. 2b, after the expected silicon, silicon oxide, and gold layers, the NiFe/ION composite layer comes into view. Unlike the Ni/ION composite, this layer does not transition into a pure NiFe layer at any point, and instead transitions to an un-infiltrated particle layer. This transition suggests the plating step had been stopped prematurely before the particle layer could be filled, or before stress caused the particle layer to be heaved up, as seen in Fig. 2a.
B. Inspection of composite layers via S/TEM
Next, inspection of the different layers of the cross-section is necessary to examine how the particles are distributed throughout the material and better understand the connectivity of the composite. Fig. 3 shows S/TEM images taken of the Ni/ION and NiFe/ION composite samples. In Fig. 3a, layer, which appears to be densely filled with nanoparticles. The presence of different shades of black and gray from one nanoparticle to the next seen in this image is a consequence of the lack of alignment of the different particles and their crystal orientations. In Fig. 3b, an image of the pure nickel layer can be seen, with a lack of the many nanoparticles that were densely packed in the Ni/ION layer. A few single IONs can be seen within this layer, but the make-up is primarily nickel metal. In Fig. 3c, the un-infiltrated layer of particles can be seen, with the various uninfiltrated nanoparticles visible and surrounded by empty space. Due to Z-contrast, elements with greater atomic mass are shown to be brighter. The presence of nanoparticles in the pure nickel layer, as seen in Fig. 3b, was previously unknown, and supports the idea that the pure nickel layer heaves up the un-infiltrated particle layer after some point during the plating step, as the particles seen in the pure nickel layer may have fell off during this lifting process and were surrounded by nickel. Fig. 3d shows the cross-section of the NiFe/ION composite layer, with nanoparticles present throughout. Fig. 3e shows the interface between the NiFe/ION composite layer and the un-infiltrated particles layer, which indicates the position where the electroplating of the permalloy was stopped. Finally, in Fig. 3f, the un-infiltrated nanoparticle layer in the sample is shown, with particles shown to be densely packed together and surrounded by empty space. These images support that there has been a successful infiltration of permalloy through the particle layer, as particles embedded in the metal matrix are clearly visible.
C. EDS confirmation of particle/matrix interface and composition
The final part of this investigation confirms that the electroplated metals do surround the particles in the composite and form a dense and void-free boundary, as evaluated with 2D EDS mapping. Fig. 4 shows two high-resolution EDS maps of an area in each of the two composites, with Fig. 4a showing the Ni/ION composite, and Fig. 4b showing the NiFe/ION composite. The color green represents the presence of nickel, and the color red represents the presence of iron. In Fig. 4a, there is a clear separation between the nickel matrix (green) and the iron-oxide nanoparticles (red) without any visible voids, indicating a dense, void-free boundary between the two phases. Likewise, Fig. 4b shows a similar behavior for the NiFe/ION composite layer, where the Ni-Fe matrix (red and green) are separate from the iron-oxide nanoparticles (red), with no visible voids present at their interfaces. These images support the conclusion that these composite materials successfully lead to a lack of voids between the particle and matrix phases.
IV. CONCLUSION
This work offers a definite confirmation that the boundary between the particle and metal matrix phase of electro-infiltrated Ni/ION and NiFe/ION nanocomposites are dense, and void-free. Previous studies of these composites confirmed the successful infiltration by examining the properties of the composites and/or through the use of SEM and EDS, but investigation of the phase boundary was limited by resolution. Through the use of S/TEM and EDS, it is now known that metals electroplated through particle film in the electro-infiltration process do fill in the spaces between particles which leads to a dense and void-free boundary. This dense, void-free interface between the particle and matrix phases in the composites is necessary for applications such as exchange-spring magnets and improved inductor cores, and further suggests the electro-infiltration fabrication process as an attractive pathway towards realizing such materials. This work provides support for past studies on electro-infiltrated 0-3 nanocomposites and encourages further investigations on the applications of these composite materials. | 2,766.4 | 2019-12-01T00:00:00.000 | [
"Materials Science",
"Physics"
] |
Decreased Sulfate Content and Zeta Potential Distinguish Glycosaminoglycans of the Extracellular Matrix of Osteoarthritis Cartilage
We aimed to determine the characteristics that distinguish glycosaminoglycans (GAGs) from osteoarthritis (OA) and normal cartilage and from men and women. Cartilage samples from 30 patients subjected to total joint arthroplasty secondary to OA or fracture (control) were evaluated, and the GAG content (μg/mg dry cartilage) after proteolysis was determined by densitometry, using agarose-gel electrophoresis. Relative percentages of carbon (C), nitrogen (N), and sulfur (S) in GAGs were determined by elemental microanalysis, as well as the zeta potential. Seventeen samples (56.6%) were from patients >70 years old, with 20 (66.6%) from women, and most [20 (66.6%)] were from the hip. The GAG content was similar regardless of patients being >/≤ 70 years old with 96.5 ± 63.5 and 78.5 ± 38.5 μg/mg (P = 0.1917), respectively. GAG content was higher in women as compared to men, with 89.5 ± 34.3 and 51.8 ± 13.3 μg/mg, respectively (P = 0.0022), as well as in OA than fracture samples, with 98.4 ± 63.5 and 63.6 ± 19.6 μg/mg, respectively (P = 0.0355). The GAG extracted from the cartilage of patients >70 years old had increase in N, and there were no gender differences regarding GAG elemental analysis. GAG from OA had a highly significant (P = 0.0005) decrease in S% (1.79% ± 0.25%), as compared to fracture samples (2.3% ± 0.19%), with an associated and significant (P = 0.0001) reduction of the zeta potential in the OA group. This is the first report of a reduced S content in GAG from OA patients, which is associated with a reduced zeta potential.
INTRODUCTION
There is a great need for accurate, reproducible biomarkers to be used in the management of osteoarthritis (OA). Without adequate parameters, it is virtually impossible to evaluate interventions to treat OA patients. Numerous attempts to quantitate soluble biomarkers or to use imaging techniques have failed to provide those most needed biomarkers (1). When subjected to OA damage, inadequate proliferation and osteophyte formation in the joint are followed by complete disruption and erosion of the cartilage leading to bare areas exposing the underlying subchondral bone. Proliferative rather than degenerative changes occurring in OA lead to osteophyte formation (2). It thus might well be that qualitative rather than quantitative changes will help distinguish OA from non-OA (normal) cartilage. Type II collagen and chondroitin sulfate (CS)-rich proteoglycans represent the major organic components of the extracellular joint cartilage matrix. Assembly of type II collagen and proteoglycans is essential for cartilage function, a highly hydrated structure that is frequently subjected to reversible deformation. Also, the number of glycosaminoglycan (GAG) chains attached to the protein backbone of proteoglycans and their aggregating pattern are relevant to cartilage homeostasis (3,4). Modifications of GAG have been found in the cartilage obtained from OA patients. The molar mass (MM) of hyaluronan, which is the main GAG in joint cartilage, was shown to be altered in areas of more severely damaged cartilage of knee OA patients (5,6). Also, CS extracted from the cartilage of areas most severely affected by OA was shown to display reduced MM (7,8).
In addition to the biological and biochemical characteristics, the electric charge of extracellular components of matrix cartilage is also relevant to homeostasis (9). The presence of both carboxylate (COO-) and sulfate (OSO3-) groups linked to the GAG attached to the protein core of proteoglycans provides a negative surface charge that is crucial to cartilage function, particularly during weight-bearing deformation (10). The zeta potential is a parameter that measures the surface charge of nanoparticles/macromolecules that can be either negative or positive, depending on the predominant electrical charge. The repulsion between surfaces of similar charge is responsible for the stability of a dispersion. When the zeta potential is reduced, compounds dispersed in a matrix have a tendency to aggregate (11). In this regard, the negatively charged groups present in GAG from joint cartilage account for its negative zeta potential. It has been shown that a zeta potential of 0 ± 10 mV compromises the stability of a polysaccharide in a solution, which then exhibits a tendency to flocculate (12). Maintenance of GAG stability may impact cell function as it was shown that in vitro chondrocyte growth is optimized in negatively charged rather than in neutral hydrogel matrices (13). We performed an elemental microanalysis of GAG extracted from the cartilage of patients subjected to arthroplasty either secondary to OA or that sustained a fracture (control) in an attempt to detect differences between those groups.
Materials
Unless otherwise stated, all reagents were purchased from Sigma-Aldrich do Brasil S.A., São Paulo, Brazil.
Collection of Human Cartilage Samples
Thirty cartilage samples from patients subjected to total joint arthroplasty in the Hospital Universitário Walter Cantídio, Fortaleza-CE, Brazil, secondary to OA or fracture (control) were collected.
Inclusion Criteria
Fifty-to 80-year-old patients subjected to total hip or shoulder arthroplasty secondary to OA or fracture, according to the clinical files, with body mass index <35 kg/m 2 ; a written informed consent was signed prior to collection of the material, as per the Brazilian rules of human experimentation.
Exclusion Criteria
refusal of the patient to participate at any time, presence of diabetes or a specific acute or chronic inflammatory arthropathy but OA. Patients who died following remote postsurgical period had the material discarded.
Prior to collecting samples, the clinical history of each patient and radiographies of the joint that would undergo arthroplasty were reevaluated by a senior rheumatologist (F.A.C.R.) both to confirm OA diagnosis and to exclude OA or other arthropathy in patients with a fracture diagnosis. GAGs were extracted within < 3 h postsurgery and analyzed as described above. The protocol was approved by our local Ethics Committee on Human Research (protocol 090.12.08) that follows the rules of the Comitê Nacional de Ética em Pesquisa, which is the Brazilian Official Committee for Ethics in Human Research. All patients signed a written informed consent prior to any procedure.
GAG Extraction
Cartilage samples were weighed after overnight drying (80 • C) and stored in acetone. Proteolysis of this material was done by incubating 1 mg with 20 µL of a 0.4% wt/vol suspension of PROLAV 750 TM (Prozyn, São Paulo, Brazil) in Tris-HCl/NaCl 50/150 mmol/L buffer (pH 8.0) for 48 h, at 56 • C. Subsequently, the NaCl concentration was corrected to 1.0 mol/L, and the mixture was kept at 37 • C during 30 min. The remaining proteins were precipitated with trichloroacetic acid to a final concentration of 10% wt/vol and centrifuged (10,000 g for 15 min at 25 • C). GAG was precipitated from the supernatant with two volumes of ethanol, followed by an overnight incubation at 4 • C and centrifugation (10,000 g for 15 min at 15 • C). The precipitated material was dissolved in 20 µL of distilled water. Protein content in the debris was considered negligible as it was undetected even using a NanoDrop apparatus.
GAG Quantification
The GAG extract was separated on a 0.6% wt/vol agarose-gel electrophoresis in diaminopropane-acetate buffer (50 mmol/L, pH 9.0). The GAG was fixed in the gel through immersion in a 0.1% wt/vol cetyl-trimethylammoniun bromide solution for 2 h. The gel was dried and stained with 0.1% wt/vol toluidine blue (in acetic acid:water:ethanol 1:49:50). For comparison, C4S, C6S, and heparan sulfate standards were subjected to the same protocol (14). Quantification was made by densitometry (525 nm). Data are expressed as µg CS/mg of dried cartilage.
Elemental Analysis of GAG
The relative percentages of carbon (C), nitrogen (N), and sulfur (S) were determined by elemental microanalysis in a Carlo Erba EA 1108 micro. The sulfate content was calculated from S% by a previously proposed equation (15). This strategy subjects the sample to combustion in pure oxygen atmosphere, and the expelled gases are detected and semiquantified using a thermal conductivity detector. Thus, the oxygen content cannot be determined.
Determination of the Zeta Potential (Pζ)
Zeta potential and conductivity were measured by a Zetasizer Nano ZS90 instrument (Malvern Instruments Ltd., Worcestershire, United Kingdom) with an λ = 633-nm laser detector with a 17 • detection angle, at 25 • C. GAG samples (50 µg/mL in deionized water) were swollen in deionized water to the equilibrium state and ground into small particles. After drying in a vacuum oven for 24 h, the particles were weighed, diluted in 1 mL deionized water, and measured.
Statistics
Results are presented as means ± SD for GAG concentration and medians [interquartile range (IQR)] for percentage of elements and evaluated using Student t-test and Kruskal-Wallis test, respectively; P < 0.05 was considered as significant.
Quantification of the GAG Extracted From the Articular Cartilage
The relative GAG content in analyzed samples was in the range of 50-85 wt%. Figure 1 illustrates that the GAG content relative to the dried cartilage weight was similar regardless of patients being >/≤70 years old (a); regarding gender, the relative GAG content was significantly higher in samples from women (b); there was also a significantly higher increase in the relative GAG content in samples from OA patients as compared to samples from patients who sustained a fracture (c).
Elemental Analysis of GAG
Using CS as reference (chemical formula C 13 H 21 NO 15 S), the relative C, N, and S percentages are roughly 34, 3, and 7 wt%, respectively. However, the observed percentage values of C (circa 23 wt%) and S (circa 2 wt%) were smaller than the predicted theoretical values, particularly regarding S content. Elemental analysis showed that GAG from the cartilage of patients >70 years old had a significant decrease in N, as compared to patients ≤70 years old (Figure 2). There were no differences in C, N, or S relative content regarding gender (Figure 3). Remarkably, the GAG extracted from the cartilage of patients with OA had a highly significant decrease in the relative S content, as compared to samples obtained from patients who sustained a fracture (Figure 4). There was also a slightly higher N relative content in GAG samples from OA patients, which reached statistical significance, probably secondary to the relative reduction of the S content in that group (Figure 4).
Analysis of the Zeta Potential (Pζ) of GAG Figure 5 illustrates the analysis of the zeta potential considering age, gender, and disease variation. As expected, all samples had a negative zeta potential, which varied within a −19 to −26 mV range. The zeta potential of GAG samples from the articular cartilage of patients >/≤ 70 years old was similar ( Figure 5A); regarding gender, there was a slight variation as cartilage samples obtained from women had a trend toward a higher modulus (−26 mV), meaning being more negative, although not reaching statistical significance, when compared to GAG samples from men ( Figure 5B); finally, there was a remarkable significant reduction of the zeta potential in GAG samples collected from the cartilage of OA patients as compared to those from patients who sustained a fracture (Figure 5C).
DISCUSSION
The present data describe a decrease in the sulfate content and a correspondent decrease in the zeta potential of GAG extracted from the cartilage of joints affected by OA. There is also an increase in the relative GAG content in samples from OA patients, as compared to those from fracture (control) patients. Both GAG content and integrity are crucial to the aggrecan role in cartilage physiology. It was reported that CS obtained from OA cartilage exhibits structural alterations, meaning different length and sulfation patterns, which may impact cartilage function (7,8). Our data reinforce those findings to suggest that qualitative changes reflect cartilage damage in OA joints. Integrity of GAG molecules is crucial to provide deformability of the cartilage particularly during weight-bearing. Additionally, GAGs are able to specifically bind to cytokines and growth factors, triggering intracellular signaling. Thus, structural modifications of the GAG structure may impact cellular responses, thus altering the function of cartilage and synovial cells (16)(17)(18).
Increased GAG content in OA cartilage has been previously shown and may illustrate the initial proliferative response of chondrocytes, as part of a repairing process. However, subjected to an OA inflammatory milieu, the "osteoarthritic chondrocytes" not only lose capacity to synthesize normal GAG but also fail to replace normal cartilage. In later stages, joint erosion occurs, with areas of denuded cartilage and exposure of the subchondral bone (4,5). In fact, our OA samples were from patients with end-stage disease, and all possible remaining cartilage was collected. Gross macroscopic evaluation does not always allow discriminating normal from damaged cartilage. Although we cannot rule out some remaining areas of normal cartilage in the OA samples, data were treated as one sample for each patient. Cartilage from dogs subjected to an experimental OA model had increased amount of proteoglycans, as compared to controls (6). Also, joint cartilage collected from humans subjected to hip arthroplasty secondary to OA was shown to have an increase in GAG content, as compared to cartilage collected from patients with fracture of that joint, used as control (19). Notwithstanding, analysis of magnetic resonance imaging of OA joints revealed an increase in matrix production in patients with recently developed OA, an aspect that was regarded as part of a repair mechanism (20,21). Although still seen as a degenerative disease, characteristic imaging findings in the OA joint reveal sclerosis of the subchondral bone and osteophyte formation, which gives an impression of de novo remodeling. In keeping with those data, the relative increase in the GAG content found in OA samples may be secondary to a proliferative, although inadequate, process happening in the affected joint.
We are not aware of previous studies showing elemental analysis of GAG isolated from the cartilage of human joints. However, at least regarding S content, a similar order of relative percentage (0.7-1.3%) was found in cartilage obtained from dogs (22). Our samples were processed just after surgical removal, aiming to avoid any possible alterations that could be due to postmortem modifications or freeze-thawing issues. There was an increase in the relative GAG content in dried cartilage seen in samples from women, which could be linked to a higher number of women in the OA group. Samples from OA cartilage had a remarkable significant decrease in the S% content as compared to control samples of patients with similar age range. Indeed, all but one of the GAG samples from the OA group had a relative S% content in the lowest value found in fracture patients.
Increased thickness of OA cartilage using delayed gadoliniumenhanced magnetic resonance was associated with increased swelling secondary to a decrease in GAG content (23). In keeping with our present data, using micro-X-ray fluorescence, it was shown that the deep zones of OA cartilage have a decrease in elemental S, which was associated with a decrease in GAG staining (24). The modulus of the zeta potential of the GAG from OA samples was significantly reduced, meaning a reduction in the negative charge of the polysaccharides probably secondary to the reduced S% relative content. We are not aware of previous studies focusing on the relevance of GAG charge to cartilage physiology. Sulfation of GAG is responsible for the negative charge of those molecules. After a compressive force applied to the cartilage, the repulsion between adjacent negatively charged GAG molecules allows the entry of water providing adequate cartilage hydration, which is crucial to a healthy joint (25). Although there is a positive association of the zeta potential and the stability of small particles, reducing its tendency to aggregate (13,26,27), there are no previous studies on the stability of GAG, let alone the relevance of the zeta potential of those molecules. However, it is reasonable to admit that a normal sulfation pattern contributes to the physiology of polysaccharides in the extracellular cartilage matrix. In this case, a reduced charge does also compromise hydration of the cartilage (28,29).
Biomarkers to be used in clinical practice are an unmet need in OA (1). Our data show that a decrease in sulfation is associated with a correspondent reduction in the zeta potential of GAG collected from OA cartilage. Current imaging studies can be designed to deliver markers able to quantitate the sulfate content or the charge of GAG in the joint cartilage. Mapping those alterations may provide semiquantitative imaging useful to evaluate interventions to modify the disease course in OA patients.
There are some limitations to our study, including time sampling. However, as mentioned previously, all material was processed within < 3 h postsurgery. Additionally, GAGs are very stable and probably would not be affected by processing. One may also argue that our samples represent solely endstage OA joint disease. Collecting enough material from living humans is very hard, and ethical rules do apply. Considering that we analyzed the whole joint, we probably saw the predominant parameter in all remaining cartilage. However, it remains to be shown if such data are reproduced in less severely affected joints. Another limitation is the low number of samples, particularly those from men, limiting gender analysis. We also restricted our samples to the hip and shoulders, and the low numbers did not allow us to compare possible differences regarding specific joints. Although knee OA is more prevalent than hip OA (30), knee fractures that lead to joint replacement are rare, making it difficult to have a suitable non-OA knee control. We also cannot completely rule out subclinical OA changes in fracture (control) samples. However, our combined clinical and imaging rheumatologic and orthopedic evaluations suggest that an OA diagnosis in the fracture group is unlikely.
In summary, we demonstrate that cartilage from OA samples displays a relative increase in the CS content. We also show that GAGs from the extracellular matrix of joints affected by OA have a decrease in sulfate content, which is associated with a decrease of the zeta potential of those polysaccharides. The possible relevance to the pathophysiology of this disease, as well as utility as a biomarker, warrants further investigation.
DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
ETHICS STATEMENT
The studies involving human participants were reviewed and approved by Conselho de Ética do HUWC-UFC (Protocol number 090.12.08). The patients/participants provided their written informed consent to participate in this study. | 4,240.6 | 2021-04-29T00:00:00.000 | [
"Medicine",
"Biology",
"Materials Science"
] |
Evaluation of the sinking processes for high-pressure-gas cylinders
High-pressure-gas cylinders are used in broad applications. Cracks on the open end would occur during the riveting stage. Such forming defects are caused by excessive hardening, although the open end has been annealed with induction heating prior to the sinking operation. Therefore, a proper design for the sinking dies is essential to the forming production of the HPG cylinders. In this paper, two die-design concepts were examined which included the conventional design for six-stage sinking with fixed die radius, and the economic design for five-stage sinking with incremental die radii. Finite element software DEFORM 2D was used to investigate the two sinking schemes. The effect of the sinking schemes on the sinking load, strain distribution, and lip thickness were analysed. The results show that the economic five-stage sinking with a large increment of die radii can provide less strain hardening as compared to other sinking schemes. Although the forming load level is acceptable and the change of lip thickness is insignificant, the production cost of the five-stage scheme is still high. A more economic measure by sinking with one-stage rotary swaging can provide an alternative scheme with advantages of simple die design and saving the lead for annealing.
Introduction
High-pressure-gas (HPG) cylinders are used in broad applications such as food dispensers, air pistols, fire extinguishers and so on.The manufacturing process of the 8gram HPG cylinder, as shown in Fig. 1, comprises blanking from steel coil and followed by deep-drawn into a slender tube in six-stage transfer stamping.The workpiece is then sunk and annealed with induction heating on its open end.The preform is further sunk with another six-stage transfer stamping, as shown in Fig. 2. The tube is subsequently milled on its open end and filled with high pressure gas and riveted to its completion.Some detailed manufacturing processes can be found in literature [1] for other examples of HPG cylinders.
Cracks on the open end would occur during the riveting stage.Such forming defects are caused by excessive hardening, although the open end has been annealed with induction heating prior to the sinking operation.Therefore, a proper design for the sinking dies is essential to the forming production of the HPG cylinders.Amirhosseini et al. [2] analyze the nosing process of empty and foam-filled circular metal tubes on semispherical die, and Lu [3] studies the preform and loading rate in the tube nosing process by spherical die.Sheu and Su [4] investigate the cold nosing process for the aluminum conical milk can.Kwan [5] investigates the cold eccentric nosing process of metal tubes with an eccentric conical die from circular tube billets.Oba et al. [6] use two-step forming for improvement of forming limit in rotary nosing with relieved die for reduction of tube tip.
In this paper, two die-design concepts were examined which included the conventional design for six-stage sinking with fixed die radius, and the economic design for five-stage sinking with incremental die radii.Finite element software DEFORM 2D was used to investigate the two sinking schemes.The effect of the sinking schemes on the sinking load, strain distribution, and lip thickness were analysed.Finally the multi-stage sinking schemes were compared with the one-stage rotary swaging process by experimental verification.
Sheet material
The sheet metal used was cold rolled steel SPCE of 0.8 mm thick.The flow stress of the sheet metal is shown in Fig. 3.The average yield strength σ y was 160 MPa.
Sinking schemes
After the completion of deep drawing, one stage sinking and annealing, the workpiece's open end was to be sunk from 14.0 to 8.5 mm in its outer diameter.For the conventional six-stage sinking scheme, a sinking ratio of 0.92 was chosen, while for the economic fivestage sinking scheme, a sinking ratio of 0.90 was used.The target outer diameters used in both six-stage and five-stage sinking schemes are shown in Tables 1 and 2, respectively.Die radii remain unchanged throughout each sinking stage in the conventional six-stage sinking scheme.However, die radii are reduced incrementally as sinking proceeds in the economic five-stage sinking schemes.Table 3 shows the die radii used in each stage of the respective sinking schemes.Increments from 0 to 9 mm were tested from the FE simulation.Fig. 4 shows the variations of die radius with outer tube half-diameter for a typical case of forming scheme "Incremental 1".Fig. 4. Variations of die radius with outer tube half-diameter for forming scheme "Incremental 1".
FE simulation
Finite element software DEFORM was used in simulating all the sinking schemes.The 2D axis-symmetric module was chosen.The planar anisotropy of the sheet material was neglected.There were 5,000 elements used in meshing the workpiece, and all the forming tools were assumed to be rigid.Fig. 5 shows the schematic of the setup for the sinking operation.The stamping speed was fixed to 300 mm/s.Because the tribological behavior of the process was not the main trust of this work, friction factors at the interfaces between the tool and workpiece were fixed to 0.05. 3 Results
Sinking load/stress
The sinking load occurring on the open end of tube should not exceed to cause distortion on the tube body which nests inside the bottom die.To ensure that the tube body is not distorted during sinking, the axial stress on the tube body should not exceed the yield stress of workpiece.Therefore, the sinking load obtained from the simulation is first divided by the cross-sectional area of tube body to obtain the axial sinking stress σ s .The sinking stress is further non-dimensionalized by dividing by the initial yield stress σ y of the sheet metal.
When the value σ s / σ y is much greater than unity, there is possibility of distortion on the tube body.
The sinking stresses for both the conventional six-stage and the economic five-stage schemes are shown in Fig. 6.The sinking stress increases as the forming stage proceeds, attributed to the increase of strain hardening in the sinking zone.The stress level is higher with the economic five-stage scheme because of greater amount of deformation was used for the respective sinking stages.Fig. 7 shows the sinking stresses for the five-stage schemes of various increments of die radius.The sinking stress increases as the increment of die radius increases.This is caused by the increasing frictional stress exerting on a larger contact interface between the die face and workpiece when the die radius becomes larger.Some values are greater than unity for the forming schemes with increments of die radius larger than 3 mm.Overall, the sinking stress shown in Fig. 7 are acceptable without causing distortion on the tube body.The tube body has experienced certain amount of strain hardening during the prior deep drawing stages.Therefore the yield stress on the tube body should be much higher than the initial yield stress 160 MPa of the sheet metal used in the calculation.The forming stress of Stage 5 is slightly less than that of Stage 4. This is because the sinking amount decreases substantially at the final stage as compared to that of Stage 4, although the sinking ratio remains constant throughout the sinking schemes.
Distribution of effective strain
Fig. 8 shows the distributions of effective strain at the final stage of selective sinking schemes.The maximum effective strain occurs on the tube inner surface and the value decreases as the die radius increases.This indicates that the strain inhomogeneity decreases with a small die inclusion angle in sinking with a large die radius [7].High strain level is unfavourable, according to the field experience, because the sinking zone would become susceptible to crack formation in the subsequent riveting operation.Fig. 9 shows the distributions of effective strain along the radial direction at Stage 5 for the economic sinking schemes.The radial position is normalized by the outer radius of the sunk tube.The distributions show a monotonic decrease of effective strain as the increment of die radius increases.The phenomenon observes that the strain level, i.e. deformation inhomogeneity [7] decreases, as the die inclusion angle increases.This trend indicates that using a large increment of die radius can obtain a less hardened sunk tube.
Lip thickness
Lip thickness increases during the sinking process.It is the primary straining direction as compared to the axial straining.Fig. 10 shows the increase of lip thickness of respective forming schemes, and the amount decreases sharply at the final stage, attributed to the accumulated strain hardening from the sinking operation.The six-stage sinking scheme exhibits a mild increase of lip thickness because of a large sinking ratio 0.92 was used for the respective stages.Fig. 11 shows the final lip thickness of respective sinking schemes.The lip thickness is the largest in sinking with an increment of die radius of 9 mm, followed by 7 mm, and so on.This is due to less strain level was generated in sinking with a large die radius, i.e. a small die inclusion angle.The analyses for the lip thickness all indicates that the variations with the change of die design are not significant.This guarantees that the amount of surface machining in the original process design remains valid.
Discussions
In Section 3, two die-design concepts are examined which include the conventional design for six-stage sinking with fixed die radius, and the economic design for five-stage sinking with incremental die radii.The results of the FE analysis show that economic fivestage with a large increment of die radii can provide less strain hardening as compared to other sinking schemes.Although the forming load level is acceptable and the change of lip thickness is insignificant, the production cost of the "economic" five-stage scheme is still high.The energy used in conduction heating, the effort in pickling the oxidation, and the maintenance of the multi-stage dies, still render the sinking scheme by transfer stamping not a lean and green manufacturing method.
A more economic measure by sinking with one-stage rotary swaging has been attempted by Lin et al. [8]. Figure 12 shows a typical hardness distribution along the axial direction for sinking with an un-annealed tube (preform).There is no significant difference between the un-annealed preform and the swaged workpiece, which indicates the feasibility of sinking with un-annealed preform.The distribution also shows that the variation of hardness number is less on the outer surface than that of inner surface.A similar trend is also observed in the FE analysis of the distribution of effective strain that the hardening level is higher on tube inner surface than that of outer surface, as shown in Fig. 9.
Conclusions
Two die-design concepts were examined in this work which included the conventional design for six-stage sinking with fixed die radius, and the economic design for five-stage sinking with incremental die radii.Finite element software DEFORM 2D was used to investigate the two sinking schemes.The results show that the economic five-stage sinking with a large increment of die radii can provide less strain hardening as compared to other sinking schemes.Although the forming load level is acceptable and the change of lip thickness is insignificant, the production cost of the five-stage scheme is still high.A more economic measure by sinking with one-stage rotary swaging can provide an alternative scheme with advantages of simple die design and saving the lead for annealing.
Fig. 5 .
Fig. 5. Setup for the sinking operation, left: before sinking; right: at the lower dead point of sinking.
Fig. 6 .
Fig. 6.Sinking stresses for both the conventional six-stage and the economic five-stage schemes.
Fig. 7 .
Fig. 7. Sinking stresses for the five-stage schemes of various increments of die radius.
5 MATEC(Fig. 8 .
Fig. 8. Distributions of effective strain at the final stage of various sinking schemes.
Fig. 9 .
Fig. 9. Distributions of effective strain at Stage 5 of economic sinking schemes.
Table 1 .
Target outer diameters used in six-stage sinking scheme, sinking ratio 0.92.
Table 2 .
Target outer diameters used in five-stage sinking scheme, sinking ratio 0.90.
Table 3 .
Die radii used in each stage of forming schemes, unit: mm. | 2,748.2 | 2018-01-01T00:00:00.000 | [
"Materials Science"
] |
Trypanosoma cruzi Sirtuin 2 as a Relevant Druggable Target: New Inhibitors Developed by Computer-Aided Drug Design
Trypanosoma cruzi, the etiological agent of Chagas disease, relies on finely coordinated epigenetic regulation during the transition between hosts. Herein we targeted the silent information regulator 2 (Sir2) enzyme, a NAD+-dependent class III histone deacetylase, to interfere with the parasites’ cell cycle. A combination of molecular modelling with on-target experimental validation was used to discover new inhibitors from commercially available compound libraries. We selected six inhibitors from the virtual screening, which were validated on the recombinant Sir2 enzyme. The most potent inhibitor (CDMS-01, IC50 = 40 μM) was chosen as a potential lead compound.
Introduction
Neglected tropical diseases (NTDs) are a diverse group of diseases that prevail in tropical and subtropical countries, affecting around 1 billion people worldwide. Among those, around 8 million cases are associated with the parasite Trypanosoma cruzi, which causes the Chagas disease [1,2]. NTDs are currently controlled through vector elimination, using insecticides, serological blood screening and social development of living conditions. In terms of chemotherapy, Chagas disease is limited to the use of nifurtimox and benznidazole, which are inefficient in the chronic phase, often present side effects and suffer from the lack of paediatric formulations [1,3,4]. Additionally, the expansion of T. cruzi resistant represent a serious public health problem. Therefore, there is an urgent need for new treatments and, therefore, the validation of new potential drug targets [5].
Trypanosomatids have a complex cell cycle in which they must respond to various environmental conditions, such as the different hosts and vectors. In this sense, transcription regulation, in particular the epigenetic machinery, is an important aspect of the parasite
It is important to compare hSir2 and TcSir2 structures to identify elements promoting selectivity and avoid pan-sirtuin inhibitors. The sirtuin enzyme family members SIRT1, SIRT2, and SIRT3 have different expressions and play different roles in trypanosomatids, depending on the context and experimental conditions. Hence, it is not clear whether a selective inhibitor of sirtuin 2, which is more expressed in the amastigote form, will be better than non-selective ones for treating Chagas disease [18].
TcSir2 has a conserved binding site for the NAD + cofactor region comprising three regions: (i) site A, where the adenine-ribose moiety of NAD + binds; (ii) site B where the nicotinamide-ribose moiety binds; and (iii) site C where the nicotinamide is supposed to bind near to the active site of the enzyme [19,20]. Some differences in the binding sites were found by superimposing the comparative model of TcSir2, LiSir2 and human Sir2 (hSir2) (Supplementary Information, Table S2). ConSurf [21][22][23] analysis ( Figure 1) reveals that the functional regions of this protein are highly conserved. For example, the residues which interact with NAD + , mainly in site C, in orange (Figure 1a), whereas hSir2 has 11 amino acids (Phe96, Leu103, Tyr104, Ile118, Phe119, Leu134, Asn160, Met168, His187 and Val266), and TcSir2 has 12 residues (Phe49, Arg50, Ile56, Pro68, Phe72, Val88, Asp92, Leu93, Asn123, Asp125, Gln217 and Val218). In this context, observing the complementarity of the cofactor with the residues for both hSir2 and TcSir2, it is noted that the trypanosomatids (TcSir2 and LiSir2) showed more polar residues (Arg50, Asp92, Asn123, Asp125, and Gln217) than hSir2 (Asn160, Met168, and His187) and the same number of hydrophobic residues. MD analysis showed that Asp125, which is found in the trypanosomatids enzyme, would be relevant for the selectivity of Sir2 inhibitors.
It is important to compare hSir2 and TcSir2 structures to identify elements promoting selectivity and avoid pan-sirtuin inhibitors. The sirtuin enzyme family members SIRT1, SIRT2, and SIRT3 have different expressions and play different roles in trypanosomatids, depending on the context and experimental conditions. Hence, it is not clear whether a selective inhibitor of sirtuin 2, which is more expressed in the amastigote form, will be better than non-selective ones for treating Chagas disease [18].
TcSir2 has a conserved binding site for the NAD + cofactor region comprising three regions: (i) site A, where the adenine-ribose moiety of NAD + binds; (ii) site B where the nicotinamide-ribose moiety binds; and (iii) site C where the nicotinamide is supposed to bind near to the active site of the enzyme [19,20]. Some differences in the binding sites were found by superimposing the comparative model of TcSir2, LiSir2 and human Sir2 (hSir2) (Supplementary Information, Table S2). ConSurf [21][22][23] analysis ( Figure 1) reveals that the functional regions of this protein are highly conserved. For example, the residues which interact with NAD + , mainly in site C, in orange (Figure 1a), whereas hSir2 has 11 amino acids (Phe96, Leu103, Tyr104, Ile118, Phe119, Leu134, Asn160, Met168, His187 and Val266), and TcSir2 has 12 residues (Phe49, Arg50, Ile56, Pro68, Phe72, Val88, Asp92, Leu93, Asn123, Asp125, Gln217 and Val218). In this context, observing the complementarity of the cofactor with the residues for both hSir2 and TcSir2, it is noted that the trypanosomatids (TcSir2 and LiSir2) showed more polar residues (Arg50, Asp92, Asn123, Asp125, and Gln217) than hSir2 (Asn160, Met168, and His187) and the same number of hydrophobic residues. MD analysis showed that Asp125, which is found in the trypanosomatids enzyme, would be relevant for the selectivity of Sir2 inhibitors. . e-An exposed residue according to the neural-network algorithm; b-a buried residue according to the neural-network algorithm; *-Insufficient data-the calculation for this site was performed on less than 10% of the sequences.
The TcSir2 3D-model, including both the Zn 2+ ion and the NAD + cofactor, was submitted to molecular dynamics simulation, aiming to verify the stability of its secondary structure elements. MD simulations improved the model quality, by reducing the number of Ramachandran outliers (from 0.5% to none, Figure S2a), and fixing the angles of Pro156 and Ser271, both in loop regions. The model was further validated and compared according to the quality indexes (Z-score −7.26 for TcSir2 before MD and −6.16 after MD) as determined by an Anolea server. Additionally, original disordered regions from the initial structure elements. MD simulations improved the model quality, by reducing the numbe of Ramachandran outliers (from 0.5% to none, Figure S2a), and fixing the angles of Pro15 and Ser271, both in loop regions. The model was further validated and compared accord ing to the quality indexes (Z-score −7.26 for TcSir2 before MD and −6.16 after MD) as de termined by an Anolea server. Additionally, original disordered regions from the initia model folded into secondary structures upon simulation with explicit solvent (Figure 2a C-terminal (residues 351-354) folded into another short strand complementing the exis ing (Tyr94-Gly96). After that, TcSir2 and NAD + analysis showed the modelled NAD + cofactor kept hy drogen interactions with the Trp94 and Asp48 (Figure 2b,c), as well as ionic interaction between the Arg50′s sidechain (Site B) and double phosphate moieties. The free amide o the nicotinamide portion had a conserved water-mediated interaction with the Asp48 an with the Ile46′s backbone, while on the other side of the molecule, transient interaction between the sugar moiety bound to the adenine and Ser240 were observed. Avalos et a (14)showed that π-cation and π-π interactions between Arg50 and adenine group an Phe49/Phe72 and the nicotinamide groups are relevant to stabilize the NAD + /Sir2 comple in the active conformation. MIF analyses corroborate this idea, presenting regions favo able for hydrophobic contacts observed at the nicotinamide binding site ( Figure S4a). H bond acceptor-favorable regions ( Figure S4b) concentrate at Site B, which is occupied b the phosphate groups of NAD + . After that, TcSir2 and NAD + analysis showed the modelled NAD + cofactor kept hydrogen interactions with the Trp94 and Asp48 (Figure 2b,c), as well as ionic interactions between the Arg50 s sidechain (Site B) and double phosphate moieties. The free amide of the nicotinamide portion had a conserved water-mediated interaction with the Asp48 and with the Ile46 s backbone, while on the other side of the molecule, transient interactions between the sugar moiety bound to the adenine and Ser240 were observed. Avalos et al. (14) showed that π-cation and π-π interactions between Arg50 and adenine group and Phe49/Phe72 and the nicotinamide groups are relevant to stabilize the NAD + /Sir2 complex in the active conformation. MIF analyses corroborate this idea, presenting regions favorable for hydrophobic contacts observed at the nicotinamide binding site ( Figure S4a). H-bond acceptor-favorable regions ( Figure S4b) concentrate at Site B, which is occupied by the phosphate groups of NAD + .
Taking together this model with information about conserved residues (Table S1), we suggest the relevant binding features cofactor NAD + (Site A) and the catalytic site (Site C, nicotinamide site, Figure 3a). Altogether, the data supported the generation of two pharmacophore models, one based on whole NAD + -binding site (pharmacophore 1), and the second based on Site C (pharmacophore 2). Thus, two virtual screenings were performed to identify compounds capable of inhibiting TcSir2 (Figure 3b,c).
Taking together this model with information about conserved residues (Table S1), we suggest the relevant binding features cofactor NAD + (Site A) and the catalytic site (Site C, nicotinamide site, Figure 3a). Altogether, the data supported the generation of two pharmacophore models, one based on whole NAD + -binding site (pharmacophore 1), and the second based on Site C (pharmacophore 2). Thus, two virtual screenings were performed to identify compounds capable of inhibiting TcSir2 (Figure 3b,c).
TcSir2 Inhibitor Virtual Screening and Experimental Validation
At this stage, purchasable and boutique subsets of the ZINC database [24] (35 million compounds) were employed for the virtual screening ( Figure 3b). This was followed by molecular dynamics simulation, as a last selection step to exclude compounds with an unwanted/unstable binding mode, which did not respect the minimal interactions identified by the pharmacophore model. Four criteria were considered for the selection of possible inhibitors of TcSir2: (i) pharmacophore model complementarity; (ii) initial ranking by the docking score values; followed by (iii) visual inspection of the compound's binding mode, analyzed according to interactions within the NAD + binding site and site C; and (iv) stability within the active site along the MD simulation. MD simulations of TcSir2-CDMS-01 complexes (Figure 4a,b) showed conserved polar interactions with the Asp125 and Ser41, together with hydrophobic interactions with Phe98 and Ala38 near 3-fluorooxadiazole ( Figure 4c). Those interactions were previously described as relevant for hSir2 binding with its substrate [25]. Additionally, the reduced RMSF values in the loop region between 122 to 144 (Figure 4d) of the TcSir2-CDMS-01 when compared with simulations of TcSir2-NAD + , suggests the loop stabilization upon ligand binding. Six compounds fitted these criteria (CDMS-05 and CDMS-06 (pharmacophore model 1- Figure 3c) and CDMS-01 to CDMS-04 (pharmacophore model 2- Figure 3d) and were selected for experimental validation.
TcSir2 Inhibitor Virtual Screening and Experimental Validation
At this stage, purchasable and boutique subsets of the ZINC database [24] (35 million compounds) were employed for the virtual screening ( Figure 3b). This was followed by molecular dynamics simulation, as a last selection step to exclude compounds with an unwanted/unstable binding mode, which did not respect the minimal interactions identified by the pharmacophore model. Four criteria were considered for the selection of possible inhibitors of TcSir2: (i) pharmacophore model complementarity; (ii) initial ranking by the docking score values; followed by (iii) visual inspection of the compound's binding mode, analyzed according to interactions within the NAD + binding site and site C; and (iv) stability within the active site along the MD simulation. MD simulations of TcSir2-CDMS-01 complexes (Figure 4a,b) showed conserved polar interactions with the Asp125 and Ser41, together with hydrophobic interactions with Phe98 and Ala38 near 3-fluoro-oxadiazole ( Figure 4c). Those interactions were previously described as relevant for hSir2 binding with its substrate [25]. Additionally, the reduced RMSF values in the loop region between 122 to 144 (Figure 4d) of the TcSir2-CDMS-01 when compared with simulations of TcSir2-NAD + , suggests the loop stabilization upon ligand binding. Six compounds fitted these criteria (CDMS-05 and CDMS-06 (pharmacophore model 1- Figure 3c) and CDMS-01 to CDMS-04 (pharmacophore model 2- Figure 3d) and were selected for experimental validation. Surprisingly, compound CDMS-04 did not show activity against TcSir2, but it was active against amastigote with the highest selectivity index (IC50 = 15.16 μM and EC50 = 62.35 μM against LLCMK2 cells and SI = 4.11) compared to the other compounds from the screening. Hence, CDMS-04 is a candidate for optimization aiming at reducing toxicity increase selectivity. As an example, these studies developed a protocol in the search for potential hSir2 inhibitors that have bulky groups reported in their structure, showed inhibition for some sirtuin isoforms on the micromolar range (Sir1, Sir2, Sir3 and Sir5), but did not show inhibition in cellular assays [27][28][29]. However, it is possible that the CDMS compounds may act as PAN inhibitors, as the activity in the recombinant enzyme is not consistent with the effect on cells. Given our high IC50 values against TcSir2, we hypothesize that we could face multi-sirtuin inhibition, and further characterization would be needed to validate off-targets.
Comparative Modelling and Conservation Analysis
The sequence of the Sir2 from Trypanosoma cruzi CL Brener strain (herein referred to as TcSir2, accession code: XP_816094) was retrieved from the NCBI database and globally Compounds were tested for TcSir2 inhibition by deacetylation, measured indirectly via fluorescent product formation (Figures S5 and S6 and Table 1). Further, the kinetic inhibitory assays with the screening hits (CDMS-01 to CDMS-06) revealed that only compound CDMS-01 (IC 50 = 40 µM, Figure 4e) efficiently inhibited TcSir2. Compounds CDMS-2 (IC 50 = 240 µM), CDMS-3 (IC 50 = 482 µM), and CDMS-6 (IC 50 = 332.5 µM) had lower potency against TcSir2 under the conditions tested. In this context, CDMS-01 K i (21.5 µM) is similar to the NAD + K m (25 µM), indicating that both would have the same affinity for the TcSir2 active site. Corroborating these results, a competitive inhibition mechanism for CDMS-01 was proposed ( Figure S7 and Table S3), due to the constant V max values and different K m values throughout the experiments. This is supported by the complementarity of CDMS-01 with catalytic pocket CDMS-01 compound. All compounds were also tested on the ability to inhibit T. cruzi intracellular replication. Compound CDMS-01 inhibited recombinant TcSir2 (Figure 4e) and T. cruzi amastigote replication (Figure 4f) in a doseresponse manner, with IC 50 values of 40 and 76.7 µM, respectively. Furthermore, CDMS-01 inhibited the mammalian cell-lines with an IC 50 value of 136 µM (Figure 4g), which indicates a slight selectivity (selectivity index = 1.77). CDMS-01 is a new and promising prototype drug that targets a previously undruggable target. It provides a starting point for the development of new inhibitors with improved activity and selectivity.
From a safety point of view, further studies should be conducted in the future with other relevant cell lines. Assuming that the hit rate of a reported virtual screening and HTS studies is not applicable as a baseline for a random screening [19], the approach used in this work was able to discover new compounds with on-target inhibition capacity, membrane permeability, and activity in amastigotes form. However, these values are similar to the approved FDA drugs amphotericin B (SI ama = 1.5) and meglumine antimoniate (SI ama = 2.46) [26].
Surprisingly, compound CDMS-04 did not show activity against TcSir2, but it was active against amastigote with the highest selectivity index (IC 50 = 15.16 µM and EC 50 = 62.35 µM against LLCMK2 cells and SI = 4.11) compared to the other compounds from the screening. Hence, CDMS-04 is a candidate for optimization aiming at reducing toxicity increase selectivity. As an example, these studies developed a protocol in the search for potential hSir2 inhibitors that have bulky groups reported in their structure, showed inhibition for some sirtuin isoforms on the micromolar range (Sir1, Sir2, Sir3 and Sir5), but did not show inhibition in cellular assays [27][28][29]. However, it is possible that the CDMS compounds may act as PAN inhibitors, as the activity in the recombinant enzyme is not consistent with the effect on cells. Given our high IC 50 values against TcSir2, we hypothesize that we could face multi-sirtuin inhibition, and further characterization would be needed to validate off-targets.
Comparative Modelling and Conservation Analysis
The sequence of the Sir2 from Trypanosoma cruzi CL Brener strain (herein referred to as TcSir2, accession code: XP_816094) was retrieved from the NCBI database and globally aligned against available protein structures using HHPred [30]. The human Sir2 crystal structure was selected as a template (PDB ID: 3ZGV, chain A, resolution 2.27 Å), due to having the highest similarity (46%) and the lowest number of outliers (0.3%). Sixty initial TcSir2 3D structure models were generated using the Modeller program [31], which were further refined by energy minimization with the steepest descent and rotamer variation. The models were ranked based on their calculated pseudo-energy values, and only the models with the lowest energy were further evaluated. The final model was chosen using Ramachandran plot analysis and Z-score validation and was tested for stability through Pharmaceuticals 2023, 16, 428 8 of 14 molecular dynamics simulation. The evolutionary conservation of the TcSir2, hSir2, and LiSir2 amino acid residues was calculated by comparing them to similar sequences using the ConSurf server [21,23]. It predicts conservation score of amino acid residues ranging from 1 to 9, which indicates the least to highest conserved, respectively (Figure 1).
Molecular Dynamics Simulation
The selected model of TcSir2 underwent molecular dynamics (MD) simulations, using a previously described protocol [32]. TcSir2 was simulated in the presence and absence of the cofactor NAD + to understand which amino acid residues contributed to the stability within the binding site ( Figure 2). MD simulations were carried out using the Desmond engine [33] with the OPLS3e force field [34]. The simulated system encompassed the protein-ligand/cofactor complex, a predefined water model (TIP3P [35]) as a solvent and counter ions (Na + or Cl − adjusted to neutralize the overall system charge). The system was treated in an orthorhombic box with periodic boundary conditions specifying the shape and the size of the box as 13 Å distance from the box edges to any atom of the protein. We used a time step of 1 fs, the short-range coulombic interactions were treated using a cut-off value of 9.0 Å using the short-range method, while the smooth particle mesh Ewald method (PME) handled long-range coulombic interactions [36].
Root mean square deviation (RMSD) values for the protein backbone were monitored along the simulation, to infer simulation equilibration and protein changes ( Figure S3a). The trajectory was clustered using the protein backbone's RMSD variance (cut-off 0.5 Å) resulting in one representative frame structure. In addition, the root mean squared fluctuation (RMSF) by residues was compared against the experimental B-factor and the observed fluctuation of residues like the reference structure that the models generated ( Figure S3b).
Molecular Interactions Fields, Pharmacophore and Virtual Screening
MD's representative conformation was used to analyze the distribution of molecular interaction fields (MIFs), performed using the GRID software (v22c, Molecular Discovery [37,38]). We evaluated three chemical probes (Dry, O 2 , and H 2 O) aiming to represent the most common intermolecular interactions (hydrophobic contacts, H-bond acceptor, and donor, respectively). Analyses were performed in a cubic box of 18 Å axis centered on the NAD + cofactor coordinates. The generated box comprises the main residues that interact with NAD + and Site C, ultimately aiming at the construction of two pharmacophoric models. MIF analyses were performed for both the human and trypanosomatid structures, and the trypanosomatid-specific regions were used to generate a pharmacophore model. Additionally, analysis of MD simulation trajectory suggests that not only hydrogen bonds, as MIFs originally proposed, but also π-cation and π-stacking interactions are involved in the NAD + stabilization.
Subsequently, we generated two pharmacophore models, one based on the NAD+ binding site and another based on the substrate site. The NAD + pharmacophore consisted of (i) three hydrogen bond acceptor points (red spheres, Figure 3d), (ii) a single hydrogen bond donor group point (blue sphere), and (iii) two hydrophobic/aromatic points (orange spheres), that can additionally represent the π-cation and π-stacking interactions. The other pharmacophore model encompassed the substrate catalytic site (Figure 3d) and considered: (i) a hydrogen bond acceptor group points (red sphere), (ii) two points of hydrogen bond donor groups (blue sphere), and (iii) three hydrophobic/aromatic points (yellow spheres) that can suggest π-cation and π-stacking interactions.
The virtual screening compound library was composed of all purchasable, all boutique subsets from the ZINC15 database [24] filtered by drug-like physicochemical properties. Namely, the number of hydrogen bond donors (HBD, with a range of 0 to 5) and acceptors (HBA, with a range of 0 to 10), the logarithm of the n-octanol/water partition coefficient (LogP, with a range of −2.5 to 5.5), and the permeability and net charge (between −2 and +2) of compounds predict their ability to enter cells. The molecular weight was limited to less than 500 Da [24,[39][40][41]. Nitro groups were limited to avoid possible toxicophoric groups [32]. The preparation of the ligands involved adjusting their protonation states and calculating their initial 3D conformations. For each ligand, 30 conformers were generated using the UNITY program within the Sybyl X 2.1 [42] package, which was used for pharmacophoric-based virtual screening using the prepared ligand library.
At the next stage, molecular docking was used to identify compounds that could bind to potential binding sites found in the pharmacophore results. Before the docking simulations, the ADP-ribose and water molecules were removed from the binding site in the model, but the zinc ion and NAD + cofactor were kept. The structure was prepared prior to the MD simulation. The docking was done using the GOLD 5.6 program [43] within a binding site defined around 8 Å from NAD + . A genetic algorithm was employed to generate 30 different poses with default parameters. Ligands were ranked by the docking score (GOLD-Score [44]) and ligand affinity (docking score divided by the molecular weight), and the best-ranked ligands were visually inspected and underwent MD simulations to ensure stable interaction pattern [45]. Simulations were conducted as previously described and further elaborated in the Supplementary Methods section. Compounds that remained stable within the binding site during the simulations were further considered for acquisition.
Chemicals and Biological Reagents
Cell culture medium and supplements were purchased from Hi-Media Laboratories (Delhi, India). Acetylated peptide was synthesized from Aminotech Co. (São Paulo, Brazil). All other reagents were purchased from Sigma Aldrich (St. Louis, MO, USA). The microplate was purchased from Corning Life Sciences (Tewksbury, MA, USA).
TcSir2 Recombinant Expression
The DNA segment encoding the full-length TcSir2 (XP_816094) cloned into the plasmid pET24a (+) was previously available in our laboratory. This vector codes for TcSir2 with a Poly-L-histidine-tag in the C-terminal end. The plasmid was transformed into chemically competent E. coli (DE3) ArticExpress cells using heat shock method. Briefly, cells and plasmid were incubated for 30 min on ice, 40 s at 42 • C, followed by another 5 min on ice. Cells containing the plasmid were grown at 30 • C and 150 rpm in Luria Broth medium (LB) containing 30 µg/mL kanamycin until reaching optical density at 590 nm of 0.6. Afterwards, the protein production was induced by adding 0.1 mM isopropyl β-thio-galactopyranoside (IPTG), and cells were further incubated at 12 • C and 200 rpm for 16 h. Cells were pelleted by centrifugation (6000× g, 4 • C, 20 min), and the pellet was then resuspended in lysis buffer (25 mM HEPES pH 7.5 with 200 mM NaCl, 5% glycerol, and 5 mM 2-mercaptoethanol). Cell lysis was done by sonication in a Branson Sonifier 250 (Branson instruments, Stanford, CT, USA) using four pulses of 12 s at 30% output potency intercalated with a cool down time (1 min) on an ice bath. Lysate and cell debris were separated by centrifugation (16,000× g for 1 h, 4 • C) and the resultant supernatant was added to a pre-equilibrated Ni-NTA resin (Qiagen, Valencia, CA, USA) for 30 min in a cold room. Weakly bound contaminants were removed by washing the resin with the lysis buffer containing 20 mM imidazole. TcSir2 and chaperone expressed by the ArticExpress (DE3) were eluted with lysis buffer containing 300 mM imidazole. The eluted was then subjected to an ion exchange chromatography with the column MonoQ 5/50 (GE HealthCare) with 50 mM Tris-HCl pH 8 buffer containing a gradient of 0 to 1 M NaCl. Protein concentration was determined by absorption at 280 nm [46] using the extinction coefficient calculated for TcSir2 [47]. Purified fractions containing proteins, as measured by 280 nm absorption, were applied to a SDS-PAGE [48] to confirm TcSir2 purity. Purified proteins were transferred to a buffer containing 25 mM Tris-HCl pH 8, 137 mM NaCl, 2.7 mM KCl and 1 mM MgCl 2 using High Trap Desalting Columns (GE Healthcare, Chicago, IL, USA).
Recombinant TcSir2 Activity Assay
The peptide substrate (Abz-Gly-Pro-AcetylLys-Ser-Gln-EDDnp, where Abz stands for ortho-aminobenzoic), based on the TcSir2 specificity [6], was acquired from AminoTech (São Paulo, Brazil). This substrate contains an acetyl-Lys, which is available for trypsin cleavage only when the TcSir2 deacetylates the lysine side chain. After peptide deacetylation and cleavage, the fluorescent moiety (Abz) and the EDDnp suppressor group (N- [2,4dinitrophenyl] ethylenediamine) are separated, and fluorescence (420 nm) can occur upon excitation at 320 nm. The TcSir2 activity assays [6] were performed at 37 • C in a 25 mM Tris-HCl buffer with pH 8 and containing 137 mM NaCl, 2.7 mM KCl, and 1 mM MgCl2. The assays used 0.1 µM purified TcSir2 enzyme, 10 µM Abz-Gly-Pro-AcetylLys-Ser-Gln-EDDnp peptide, and 25 µM NAD + cofactor. After a 15 min incubation, 12 mM nicotinamide and 0.6 mg/mL trypsin (Sigma Aldrich, Sao Paulo, Brazil, N3376) were added to stop the reaction and cleave the deacetylated peptide (30 min at 37 • C). A control was also prepared without trypsin to detect the baseline fluorescence of the uncleaved peptide. Increasing concentrations of the putative inhibitory compound prepared in DMSO were also tested under the same conditions, with controls to detect the potential effect of DMSO at a maximum concentration of 5% (v/v) on TcSir2. Fluorescence was measured using a TECAN-InfinityPro2000 plate reader.
The percentage of inhibition data was calculated using the equation: Inhibition (%) = (1 − v i /v c ) × 100, where v i represents the initial rate in the presence of the putative TcSir2 inhibitor, and v c is the initial rate of the control assay containing only DMSO. IC 50 values were calculated based on the percentage of inhibition and concentration of the inhibitor using non-linear regression with least squares on GraphPad Prism (v8.1, GraphPad software, La Jolla, CA, USA). The inhibition mechanism of the most potent compound (CDMS-01) was investigated by determining its effect on the apparent Km and Vmax for NAD + as a substrate, under the same reaction conditions. Km and Vmax in the presence and absence of the inhibitor were obtained by fitting the initial rate and [NAD + ] into the Michaelis-Menten equation on GraphPad Prism (GraphPad Prism version v8.1, for Windows, GraphPad Software, San Diego, California USA, www.graphpad.com).
Evaluation of In Vitro Trypanocidal Activity with Amastigote Forms
In 96-well plates, cells from the LLCMK2 lineage were seeded at a concentration of 5 × 10 4 cells/mL (80 µL). The trypomastigote forms of the Tulahuen LacZ strain were added at a concentration of 5 × 10 5 cells/mL (in a total of 20 µL) and, subsequently incubated for 24 h at 37 • C with a 5% CO 2 atmosphere. Next, the trypomastigote forms present in the supernatant were washed, leaving only the amastigote forms. Serial dilutions of the compounds were added (1.95 to 200 µM) and the plate was incubated for 72 h at 37 • C with a 5% CO 2 atmosphere. At the end of this period, the CPRG substrate (chlorophenol red β-D-galactopyranoside, 400 µM in 0.3% Triton X-100, pH 7.4) was added and incubated for additional 6 h. After incubation, the sample's absorbance values were measured using a spectrophotometer at 570 nm. Benznidazole and dimethyl sulfoxide (DMSO) in the same concentrations as tested compounds were used as the positive and negative controls, respectively.
Mammalian Cell Viability Assay
LLCMK2 cells were maintained in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS) (Gibco, Carlsbad, CA, USA), penicillin/streptomycin (100 µg.mL −1 and 0.1 mg.mL −1 , respectively, Sigma-Aldrich, St. Louis, Mo, USA) and 2 mM L-glutamine, and treated every second day. At LLCMK2 cells were maintained in high glucose DMEM (Dulbecco's Eagle's Medium) medium, supplemented as previously mentioned. All cells were maintained in humidified CO 2 at 37 • C. Cells were obtained at the third passage, propagated, and used in experiments between the fourth and fifth passages. Cells were routinely checked for contamination with mycoplasma by PCR.
Cytotoxic effects of the compounds were evaluated using MTT (3-(4,5-Dimethylthiazol-2-yl) 2,5-Diphenyl Tetrazolium bromide), which correlates with the cell viability upon treatment. The principle of this method, as described by Mosmann et al. [49], consists in measuring the cellular viability of the enzymatic activity of oxirubutases of living cells [50]. For the test, 1.0 × 10 5 cells from LLCMK2 lineage were seeded into 96-well microplates in the absence or presence of the compounds or benznidazole (BZN) serially diluted in base 2 (1.95 to 500 µM) and incubated at 37 • C with 5% CO 2 for 72 h. At the end of the period incubation. 50 µL MTT (Sigma-Aldrich Corp. St. Louis, MO, USA) at the concentration of 2 mg.mL −1 . After 4 h incubation, 50 µL DMSO per well was added to dissolve formazan blue crystals. The absorbance was determined at 570 nm using Biotek Synergy HT microplate spectrophotometer. The percentage of cell viability was determined from the following formula (40): Cell Viability (%) = (Treatment Absorbance)/(Control Absorbance negative) × 100.
Data Analysis
Statistical comparisons were performed with ordinary or repeated measures of oneway or two-way ANOVA or Friedman test, using respective post hoc tests for multiple comparisons against controls, as recommended by the analysis software and described in the figure legends. IC 50 values were determined by non-linear fit of dose-response using the equation for sigmoidal dose-response with variable slope.
Conclusions
The TcSir2 inhibitor CDMS-01, which is active in vitro against intracellular amastigotes of T. cruzi, is a new antitrypanosomal hit compound. For this reason, future studies may apply molecular simplification to increase activity and selectivity. CDMS-01 is a good starting point for the development of new drugs for the Chagas disease. Despite the low and medium micromolar activity against TcSir2 of our hits, they represent novel scaffold, and their derivatives may be tested to improve on-target activity.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/ 10.3390/ph16030428/s1, Figure S1. Comparison between trypanosomatids' Sirtuin 2 and humans. (A) 1D alignment between trypanosomatids' Sirtuin 2 and humans and (B) 3D alignment between trypanosomatids' Sirtuin 2 and humans. TcSir2 in cyan (Trypanosoma cruzi), LiSir2 in yellow (Leishmania infantum; PDB: 5OL0) and hSir2 in magenta (Human; PDB: 3ZGV). Table S1. Selection of the Sirtuin 2 crystal structures. Table S2. Interaction regions of different Sir2 organisms. TcSir2 (Trypanosoma cruzi); LiSir2 (Leishmania infantum); hSir2 (Homo sapiens) and TbSir2 (Trypanosoma brucei). Figure S2. Ramachandran and Z-score validation*. (a) TcSir2 before MD (0.5% outlier and Z-score −7.26) and (b) TcSir2 after MD (0.0% outlier and Z-score −6.16). * These give an assessment of the overall quality of the structure as compared with well refined structures of the same resolution and also highlight regions that may need further investigation. The PROCHECK programs are useful for assessing the quality not only of protein structures in the process of being solved but also of existing structures and of those being modelled on known structures. Figure S3. (a) TcSir2 and NAD + RMSD values along MD trajectory, showing two stages the protein explored along the simulation time, after equilibration; (b) Root mean squared fluctuation by residues (RMSF) compared against the experimental B-factor. Figure S4. Molecular Interaction Field analysis. (a) Dry probe for hydrophobic regions in yellow and (b) O 2 probe for acceptor and acceptor regions in red. Figure S5. K m calculation for formation ADP-ribose (a) and the deacetylation the peptide (b) followed by CDMS-02 (c), CDMS-03 (d), CDMS-06 (e) and Nicotinamide (f) dose-response curves for enzymatic inhibition on TcSir2. Figure S6. Chemical structures for each tested compound CDMS-02 to CDMS-06, followed by dose-response curves for trypanocidal activity on amastigote stage and cytotoxicity on mammalian cells. Figure S7. CDMS-01 mechanism inhibition. (a) Plot for determination of the Ki (Ki = 21.5 µM and r 2 = 0.989); (b) mechanism inhibition of CDMS-01. Reaction rate, in seconds, and substrate concentration were analysed in the presence of different concentrations of CDMS-01. Table S3. F-test 2 results support a competitive inhibition mechanism as observed by the differences in the fluorescent unit [51]. | 7,645.4 | 2023-03-01T00:00:00.000 | [
"Medicine",
"Computer Science",
"Chemistry",
"Biology"
] |
Monte Carlo simulation of the effect of magnetic fields on brachytherapy dose distributions in lung tissue material
The aim of this work was to use TOPAS Monte Carlo simulations to model the effect of magnetic fields on dose distributions in brachytherapy lung treatments, under ideal and clinical conditions. Idealistic studies were modeled consisting of either a monoenergetic electron source of 432 keV, or a polyenergetic electron source using the spectrum of secondary electrons produced by 192Ir gamma-ray irradiation. The electron source was positioned in the center of a homogeneous, lung tissue phantom (ρ = 0.26 g/cm3). Conversely, the clinical study was simulated using the VariSource VS2000 192Ir source in a patient with a lung tumor. Three contoured volumes were considered: the tumor, the planning tumor volume (PTV), and the lung. In all studies, dose distributions were calculated in the presence or absence of a constant magnetic field of 3T. Also, TG-43 parameters were calculated for the VariSource and compared with published data from EGS-brachy (EGSnrc) and PENELOPE. The magnetic field affected the dose distributions in the idealistic studies. For the monoenergetic and poly-energetic studies, the radial distance of the 10% iso-dose line was reduced in the presence of the magnetic field by 64.9% and 24.6%, respectively. For the clinical study, the magnetic field caused differences of 10% on average in the patient dose distributions. Nevertheless, differences in dose-volume histograms were below 2%. Finally, for TG-43 parameters, the dose-rate constant from TOPAS differed by 0.09% ± 0.33% and 0.18% ± 0.33% with respect to EGS-brachy and PENELOPE, respectively. The geometry and anisotropy functions differed within 1.2% ± 1.1%, and within 0.0% ± 0.3%, respectively. The Lorentz forces inside a 3T magnetic resonance machine during 192Ir brachytherapy treatment of the lung are not large enough to affect the tumor dose distributions significantly, as expected. Nevertheless, large local differences were found in the lung tissue. Applications of this effect are therefore limited by the fact that meaningful differences appeared only in regions containing air, which is not abundant inside the human.
Introduction
In 1895, H.A. Lorentz mathematically described the theory that describes the effect that magnetic and electric fields have on moving charged particles. This electromagnetic theory states that the trajectory of a charged particle is curved perpendicularly to the plane containing the velocity and magnetic field vectors. In medical physics, this effect is relevant because of two facts. First, the presence of strong magnetic fields in clinical environments. Clinical scanners based on magnetic resonance (MR) use fields between 1.5 T and 3 T, while higher fields up to 11.7 T or more can also be found in some specialized research environments. Second, the use of image-guided radiotherapy and brachytherapy treatments as part of the clinical workflow. The enhanced contrast/resolution provided by MR imaging is an upgrade to those of traditional use like X-ray imaging. Therefore, the industry is already promoting the production of combined MRI-Linac machines. In this context, there are valid and interesting questions to be considered. For example, what are the effects that magnetic fields induce on dose distributions produced by charged particles? Is it safe for patients and health professionals to work in this environment? Can the Lorentz force be used to limit dose distributions and confine them to smaller and more precise volumes? Is the dose distribution homogeneous inside this confined area? Computational modeling using analytical or Monte Carlo (MC) methods may assist in addressing such questions through the precise control of the theoretical variables of the physics processes and geometry models involved. For example, in-patient dose distributions in regions-of-interest with useful resolution can only be obtained with computational tools (~1 mm 3 voxels size).
The effects of magnetic fields on electron beams produced by clinical linear accelerators (LINAC) and their interaction with patients have been analyzed with MC methods in the past [1][2][3]. For LINACs, where the particles were emitted in one preferential direction, findings showed reduced scattering due to the helicoidal movement of electrons along the direction of the magnetic field (B). In some cases, it was possible to optimize the dose distributions using this effect. When the direction of the beam was transversal to the magnetic field, this effect was found to be larger [4,5]. Differences between dose distributions of up to 300% were found under the presence of magnetic fields ranging from 1.5 T to 11 T. When dose distributions were studied inside an MRI scanner, two effects became apparent to researchers. The first effect was the contribution of secondary electrons leaving and then returning to the tissue caused by the bending of their trajectories [4,6,7]. The second was the shorter dose build-up regions needed due to the bending of electron trajectories [5,6]. All these effects varied linearly with the magnetic field strength and the particle energies involved, and also depended on the tissue composition. In a more recent study, Beld et al [8] showed the effects of magnetic fields on dose distributions produced by a 192 Ir high-dose rate (HDR) brachytherapy source. Differences from previous studies arose from the fact that the brachytherapy source emitted radiation isotopically, and the energy spectrum produced by the source was lower than that used in LINACs (keV vs. MeV). Results were obtained in air and water, representing the rectum cavity and prostate tissue, respectively. They reported that the clinical effect in water was negligible, but relevant changes in dose distributions near the air pockets were found. This study covered two extreme cases where two homogeneous materials were used with density values three orders of magnitude apart.
To our knowledge, the case of intermediate, low-density tissue has not been reported. A particularly important case is the lung, as it is composed of a heterogeneous distribution of soft and dense tissues ranging from air to compressed lung [9]. Given the existence of that density distribution, physical changes in the dose distributions are expected according to the extreme cases presented in Beld et al, and call for their quantification.
The aim of this work was to use MC simulations to assess the effect on dose distributions during the course of 192 Ir brachytherapy treatments of lung inside an MR scanner. Its application specifically to lung cancer is justified on two facts. First, the prevalence of this type of cancer and its mortality which accounts for 14% of all cancers diagnosed and 25% of all annual, cancer-related deaths [10][11][12]. Second, the larger "confining" effects that are expected to be found in low-density tissue and in the air environment, due to Lorentz forces. To this end, dose distributions were calculated in a homogeneous low-density lung material using monoand poly-energetic electron spectra. This allowed evaluation of the best-and worst-case scenarios in the presence or absence of an external uniform magnetic field. Subsequently, an array of HDR brachytherapy sources of 192 Ir were modeled for a clinical lung treatment under the presence of a homogeneous magnetic field. This allowed heterogeneous tissue densities and their spatial distribution to be accounted for. All the modeling was performed with the Geant4-based TOPAS MC tool [13,14]. This allowed the implementation of detailed MC simulations considering the physics of particle interactions with matter, electromagnetic fields and patient geometries through a user-friendly interface.
Monte Carlo software
The modeling software used was TOPAS [14] version 3.2, built on top of the Geant4 MC toolkit [13] version 10.5 patch 1. The radioactive decay process of iridium was modeled using the Geant4 G4RadioactiveDecay constructor. This physics constructor contains verified physical models to simulate radioactive decay of many radionuclides, as reported in Hauf et al [15]. The electromagnetic interactions of photons, electrons, and positrons were modeled using Geant4's constructor G4EmStandard_option4. For photons, this physics constructor used Livermore models for Rayleigh scattering and the photoelectric effect, a low energy model for Compton scattering [16], and models based on the PENELOPE MC code [17,18] for gamma conversion. For electrons and positrons, the constructor used PENELOPE models for the ionization process below 1 MeV, and the Goudsmit-Saunderson model for multiple scattering. For the transport of charged particles in magnetic fields, the Geant4 parameters consisted of the classical four-order Runge-Kutta (ClassicalRK4) stepper with DeltaChord value of 0.05 mm. Additional parameters were dRoverRange and finalRange with values of 0.2 and 10 μm, respectively. Experimental comparison has been performed showing reasonable agreement for electrons produced by a clinical LINAC using a 0.1 mm DeltaChord value and the Classi-calRK4 stepper [7]. All the simulations were run in the Mexican Southeast National Laboratory of Supercomputing (LNS). A computational setup consisting of 228 nodes with 5472 CPU cores.
Comparison of TOPAS with EGS-brachy and Penelope for an 192 Ir seed scenario
TOPAS, from its initial stages, has been developed to be used in proton therapy applications [14]. Nevertheless, the physical models and geometrical tools provided by this tool allow its use for other radiotherapy modalities, including brachytherapy. Therefore, TOPAS was compared to Monte Carlo modeling standards in brachytherapy, making the results of this work more reliable. To this end, the authors compared calculated dose distributions and AAPM TG-43 [19] parameters with published data calculated with PENELOPE and EGS-brachy in similar conditions [20,21]. The TG-43 parameters compared were dose-rate value in water, air-kerma strength, and the geometry and anisotropy functions. For dose distribution calculations, a cylindrical water phantom with a 15 cm radius and 15 cm height was used. The phantom was divided into bins with 0.05 cm resolution in both the radial and longitudinal axes. This resolution was the same as the one used in PENELOPE calculations [20]. For airkerma strength calculations, the fluence spectrum was scored in-vacuo on the surface of a binned sphere of 100 cm radius. The polar angle was covered with 1-degree resolution differentials. The fluence spectrum was then weighted with the air mass energy-absorption coefficients obtained from NIST [22], according to the updated AAPM TG-43 report [19]. A low energy limit of 5 keV was used.
The effect of a magnetic field on dose distributions of 192 Ir in homogeneous lung tissue
As a first step, the authors considered two idealistic scenarios, each using an electron source interacting in a cubic phantom made of homogeneous lung tissue. For the first scenario, a punctual, isotropic and monoenergetic electron source of 432 keV was positioned in the center of a cubic lung tissue phantom of homogeneous density. A constant lung density of 0.26 g cm -3 was calculated as an average from its deflated and inflated states [9]. By using this value, the necessity of making simulations for either inflated (mean density of 0.1 g cm -3 ) and deflated (mean density of 0.4 g cm -3 ) lung tissue was eliminated. The atomic composition of the lung was obtained from the International Commission for Radiation Protection database [23]. This information was available in TOPAS through the G4_LUNG_ICRP nomenclature of Geant4 [13]. The energy value of 432 keV corresponded to the maximum energy transferred to secondary electrons set in motion by 192 Ir gamma-rays interacting in water [8]. For the second scenario, the energy of the electrons was replaced by the whole energy spectrum obtained from Fig 1A in Beld et al. [8]. Thus, the first scenario allowed the authors to evaluate the largest effect on the dose distributions. The effect was largest due to the longer range that electrons could reach given the maximum energy of the spectrum. Nevertheless, this was not the most likely scenario, and therefore, the second scenario allowed the authors to evaluate the effect on the dose distributions using the most probable range of electron energies.
For both scenarios, the simulated phantom was a 1 cm 3 , voxelized cubic box with a 0.05 mm x 0.05 mm x 0.05 mm resolution. The radioactive source was positioned in the center of the phantom, and two magnetic strengths were applied: 0 T, or no magnetic field, and 3 T magnetic field. For all scenarios, the magnetic field was pointed in the positive z-direction. A production cut off value of 0.05 mm was used for the creation of secondary electrons, which is the default value in TOPAS.
Three-dimensional dose distributions were calculated and normalized to a dose value averaged over an arbitrary point, located at coordinates (x = 1 mm, y = z = 0 mm). Results were presented as normalized dose distributions with green lines being the 100% iso-dose, followed by 50%, 25%, 10%, and 1% iso-dose lines (in blue tones). In total, 200 million histories were simulated to achieve an averaged statistical uncertainty lower than 0.5% at 100% of the normalized dose distributions.
Clinical brachytherapy treatment plan using 192 Ir
A 192 Ir brachytherapy seed was modeled for this work. The brachytherapy seed reproduced the commercially available iridium seed VariSource VS2000 HDR from Varian1. This seed is regularly used with a VariSource TM iX Afterloader from Varian1 for brachytherapy treatments [24]. The geometry details were obtained from a previous publication [20]. The external dimensions of the seed were 6.59 mm long and 0.59 mm wide, with the iridium core encapsulated inside (see Fig 1A). The core consisted of two cylinders of 0.17 mm radius and 2.5 mm length, including rounded endcaps with a 0.295 mm radius. The shielding material of the source was a titanium-nickel alloy (44:56 fraction by weight), with a density of 6.5 g cm -3 . The iridium density inside the shielding was 22.42 g cm -3 .
A clinical case of pulmonary cancer treated with 192 Ir brachytherapy sources was simulated (see Fig 1B). First, anatomical images in DICOM format of a pulmonary tumor were obtained from the image bank of the Cancer Imaging Archive [25]. The selected DICOM images corresponded to an axial set of computed tomography (CT) slices crossing a pulmonary tumor (voxel resolution: 0.69 mm x 0.69 mm x 1.00 mm, 60 slices). One advantage of using this data was that the tumor was confirmed by the research group in charge of uploading the images, and they provided information on node location and diagnosis. Hounsfield units from the DICOM files were converted to density values using the Schneider stoichiometric method [26]. This procedure is automated in TOPAS [27]. The simulation moved away from the constant lung density of 0.26 g cm -3 used in previous models, to a more realistic scenario with density values ranging from 0.01 g cm -3 to 0.6 g cm -3 .
For HDR interstitial brachytherapy of lung lesions, the use of a single brachytherapy catheter in tumors up to a 4 cm diameter, has been previously reported [28,29]. The use of a single catheter showed a 0% complication rate of pneumothorax and a 75% tumor control rate for 20 Gy of prescribed dose in a single fraction [29]. Thus, in this work, four seed positions in a single catheter were used to deliver a 20 Gy single dose to the planning tumor volume (PTV). On the other hand, automatic brachytherapy planning can be optimized using multicriteria optimization algorithms that demand high-speed computing capabilities, see Belanger et al. [30]. However, it is still a regular practice in brachytherapy planning to manually fine-tune the seed positions and dwells times to obtain a case-specific valid plan [31]. In this sense, the catheter angle and the seed positions were selected to cover the 2.8 cm diameter tumor volume. For the dwell times, each seed position was maintained for a unit of time (weight = 1) or multiple units of time (weight > 1) from a position of reference in order to obtain a tumor dose distribution that was as uniform as possible. This was accomplished by weighting the number of histories at each source position. The seed positions and weights are presented in Table 1.
The transport of radiation within the seed was handled by the parallel navigation of Geant4 and facilitated by TOPAS. The seed material was designed according to the layered mass geometry feature [32]. In this way, the simulation setup was simplified as there was no need to use voxels filled with water to accommodate the seeds as reported previously [33].
Since the millions of decays modeled in MC methods correspond to small amounts of time, the activity of the sources was assumed to be constant during the simulation period [34]. The simulations were performed with and without a uniform 3 T magnetic field pointing in the caudal-cephalic direction (positive z). In total, 3 x 10 10 histories distributed in 30 jobs were simulated for each option (0 T and 3 T fields). That number of histories was enough to obtain dose distributions with statistical uncertainties better than 0.5% and 1%, for voxels with a dose bigger than 50% or 10% of the prescribed dose, respectively. The dose distributions were normalized to the prescribed 20 Gy dose at the PTV structure. Three structures of interest were contoured: the tumor, the planning target volume (PTV), and the lung. The volumes of these structures were 3.9 cm 3 for the tumor, 12.8 cm 3 for the PTV, and 895.2 cm 3 for the lung (from a DICOM selected number of slices). Finally, dose-volume histogram parameters suggested for interstitial brachytherapy were calculated [35]. V 90 , V 100 , V 150 , V 200 , D 90, and D 100 were calculated for the tumor and PTV structures, and D 2cc for the lung. Anatomical images underneath the dose images were co-registered to the same space and presented at the same resolution.
Statistics
All results are presented as a mean ± SEM (standard error of the mean). While comparison between populations was necessary, data distributions were first assessed for normality, and depending on the result either Student's t-tests or Mann-Whitney U tests were applied. Significance of the results was considered to have been achieved when p<0.05. For all calculations, SPSS 9 software (IBM1) was used. Fig 2 shows the modeling results of a 432 keV electron source emitting isotopically in lung tissue. As depicted in the top panels of the image, in the absence of a magnetic field, the dose distribution decreased radially as the inverse of the squared distance. This decreasing in dose is The coordinates are given with respect to the global coordinate system. Each coordinate corresponds to the center of the seed located between the two cylinders that c composed the core, (point P1 in Fig 1B). represented by a change in color from red to dark blue. When a magnetic field was applied in the positive z-direction (bottom panels), electrons moving along the +x-direction were subject to a force in the -y-direction, and electrons moving in the +y-direction were subject to a force in the +x-direction, due to the Lorentz force. In this way, there was a helicoidal movement in the XY plane. This movement caused the iso-doses to be distributed in both directions (x and y) from 1.02 mm ± 0.06 mm to 0.94 mm ± 0.05 mm distance for the 100% iso-dose, and from 4.08 mm ± 0.26 mm to 1.43 mm ± 0.05 mm for the 10% iso-dose. These differences were statistically significant for both cases (p<0.001) and represented 7.8% and 64.9% reductions, respectively. The helicoidal movement in the XY plane caused the pattern in the iso-dose distribution shown in the XZ and YZ panels. The results of modeling the spectrum of secondary electrons from 192 Ir gamma-rays in pulmonary tissue can be seen in Fig 3. As seen in the top panels of Fig 3, with the absence of a magnetic field, the dose distribution was isotropic, decreasing as the inverse of the square root of distance. Because the electron energies of this spectrum were a composite of energies in the 0-432 keV range, the dose deposition was closer to the source position than in the scenario using the monoenergetic electron source. When a magnetic field was applied in the positive zdirection (bottom panels), the same effect appeared (Fig 2 bottom panels) even though it was of a smaller magnitude. In the cases without magnetic field, the iso-doses at 100% and 10% were distributed at 1.00 mm ± 0.06 mm and 2.01 mm ± 0.29 mm, respectively. In the cases with the magnetic field, the iso-doses at 100% and 10% were redistributed to 0.92 mm ± 0.02 mm and 1.48 mm ± 0.10 mm in both directions (X and Y), respectively. These differences were statistically significant for both cases (p<0.001) and represented 8.0% and 26.4% reductions, respectively. The differences in the XZ and YZ planes due to the magnetic field were of the same magnitude in either x or y direction and are not presented for clarity.
Results
In Fig 4, the TG-43 parameters shown are the geometry function (G L ), anisotropy function (F), and dose-rate constant (Ʌ). For the geometry and anisotropy functions, the point-to-point ratios of TOPAS to EGS-brachy published calculations are displayed at the bottom of panels A and B, respectively. The geometry function calculated with TOPAS and EGS-brachy differed by 1.2% ± 1.1% at 15 cm, one standard deviation. For the anisotropy function, results from
PLOS ONE
both codes were within statistical uncertainties, except for angles smaller than 6 degrees where TOPAS exceed EGS-brachy calculations by 6.4% ± 6.1% at 0 degrees. For the dose-rate constant, differences with respect to EGS-brachy and PENELOPE calculations were 0.09% ± 0.33% and 0.18% ± 0.33%, respectively. Finally, panel D shows the dose distributions on the central plane of the brachytherapy seed. The good agreement between iso-doses and the effect of the wire attached to the source in the lower region can be visually appreciated. Fig 5 shows the brachytherapy treatment case modeled with an 192 Ir source using the configurations from Table 1. Fig 5A shows the iso-dose curves superimposed on the anatomical CT image with dose values displayed in percent of the prescribed dose. Note that the 100% isodose line covered the tumor completely. The volume of lung tissue receiving 5 Gy was 19.5%, a volume consistent with the clinical use case reported in Sharma et al. [28]. As shown, the isodose curves of simulations with and without a 3 T magnetic field, overlap in denser tissues but are misaligned in softer tissues. To quantify this difference, Fig 5B shows the relative dose differences in each voxel with respect to the simulation without the magnetic field. Only those voxels with differences bigger than 2%, and one standard deviation of combined statistical uncertainty are displayed. One average, differences circa 10% are visible at the distal area of seed position P 1 (Fig 1B) with a hot spot of more than 25%. At the boundaries between lung and muscle tissue, differences of 5% on average can be found. DVH´s obtained for the tumor, PTV, and lung are shown in Fig 6. The DVH´s and point-to-point differences between no field and 3 T field results are also shown. Systematic differences outside 1 standard deviation were found for the tumor below 1% (dose > 35 Gy), the PTV below 1.5% (dose > 45 Gy), and the lung below 2% (dose > 40 Gy).
The dose-volume histogram parameter V 100 for the tumor with and without the magnetic field was 100% in both cases. For the PTV, V 100 was 97.32% ± 0.09% (without field) and 97.20% ± 0.07% (with field). Finally, for the lung, the parameter D 2cc was 96.79% ± 0.11% and 97.95% ± 0.14% (1.18% ± 0.12% difference). Differences in percentage for all the parameters are summarized in Table 2. As shown, except for D 100 , differences were below 1% (one standard deviation) for the parameters found for the tumor and PTV. For parameter D 100 , the presence of the field reduced the dose by 11.43% and 13.95% for the Tumor and PTV, respectively. These results show that clinically, inside the tumor, there were no major differences in dose distributions due to the presence of a magnetic field.
Discussion
This work presented a theoretical assessment using Monte Carlo simulations of the effect of magnetic fields on dose distributions from brachytherapy. The Monte Carlo method has proved its validity when an experimental measurement is difficult to perform, as in the assessment of theoretical predictions, development of prototyped technology, and analysis of uncertainties [6,8,36,37]. In these kinds of applications, well-controlled conditions and prior verification of the components involved (physics models and geometry models) must be ensured. In this work, such conditions were satisfied in the following way. First, the TOPAS Monte Carlo model of the brachytherapy source was verified by comparing calculated TG43 parameters with data from the literature. Second, this work relies on the accuracy of Geant4 toolkit to model transport of charged particles in magnetic fields, which has been already shown experimentally [7,38,39] and theoretically [40,41]. Third, we relied on the TOPAS tool for the implementation of magnetic field, DICOM patient, detailed geometry and source of the brachytherapy seed, and scoring. TOPAS minimizes the possible error introduced by the user by providing a robust user interface [14]. Finally, the simulated scenarios were performed in well-controlled conditions by varying only the magnetic field. Authors acknowledge that the lack of a real physical experiment supporting the findings presented here, even if properly argued and justified, will always be a limitation of any kind of study which uses modeling. Based on the conditions described before, the main findings of this work were as follows. The dose distributions from an 192 Ir source irradiating lung tissue with a homogeneous density were affected by the magnetic field. The effect occurred using either monoenergetic 432 keV electrons or the polyenergetic spectrum. This effect reduced the distance, from the source initial position, of the 10% iso-dose line by 64.9% for monoenergetic electrons and 26.4% for the whole spectrum. These differences motivated the exploration of a clinical case, where a heterogeneous density distribution existed. In this case, the difference in dose-volume histograms of the tumor with and without magnetic fields was small. However, for PTV and lung regions which received high doses of radiation (100-300% of planned), differences in dose-volume histograms of 0% to 3% were found. Finally, the TOPAS tool agreed with results from EGS-brachy and PENELOPE codes when modeling the dose depositions of a brachytherapy seed.
The magnetic field locally affected the energy deposition in the lung tissue. In Fig 5B, a hot region close to the tumor volume is shown. The intensity was accentuated in regions closer to the tumor. In lung tissue, the produced electrons had extended ranges compared to electrons of the same energy produced in denser tissue. Hence, the bending effect of Lorentz forces on the trajectory of electrons in soft tissue was more pronounced. This caused a more confined dose distribution compared to the distribution calculated in the absence of the magnetic field. At the same spatial distance far from the source, the dose was higher in the absence of a magnetic field, as shown in Fig 3. This was also present in the ratios of dose-volume histograms shown in Fig 6. The electrons produced in the denser tissue that reach the soft tissue experienced the electron return effect [8] and produced hot spots near the interfaces, as seen in Fig 5B. The tumor volume did not show major differences in dose distributions at almost any prescribed dose due to the presence or absence of the magnetic fields, as shown in Fig 6. This was an expected result caused by the tumor density. The tumor, because of its vascularity, presents larger amounts of blood (fluid) rather than air. Beld at al. [8] already demonstrated that the bending effect in water was negligible. Our results were consistent with Beld's findings.
TOPAS was suitable for brachytherapy calculations in homogeneous water. Although TOPAS was developed for proton therapy applications, its features and Geant4 physics models allowed its use in other radiotherapy modalities. In this work, we verified the feasibility of using TOPAS through the calculation and comparison of the TG-43 parameters. Comparison of TOPAS calculations was performed with results from EGS-brachy and PENELOPE codes for the VariSource VS2000 192 Ir seed. For the geometry function, systematic differences of 0.62% with respect to EGS-brachy were found at positions closer to the seed (r < 0.5 cm). These were caused by the voxel size used in this work (0.5 mm) compared to that used in EGSbrachy (0.1 mm < 1 cm). The systematic error introduced by the voxel size was investigated previously in [42]. In there, differences of 1.8% at 0.3 cm were found when going from scoring spherical shells of thickness 0.1 mm to 0.5 mm. The differences found in this work were consistent with the reported error from [42]. The geometry size effect was also responsible for the highest differences found in the anisotropy function at angles closer to 0 degrees. Nevertheless, the dose-rate calculated in this work agreed with other MC codes within statistical uncertainties of 0.33%.
Conclusions
Differences between models with and without magnetic field were minimal (absolute difference of approximately 3.0% as it can be derived from Fig 6C). Based on these results it can be argued that the concomitant use of a magnetic field in brachytherapy did not bring any advantage to this specific clinical lung case. Nevertheless, it is important to highlight the existence of a major redistribution of dose deposition in other tissues due to the magnetic fields. At the moment, because the Lorentz effect is only significant for air cavities, very high energy deliveries or strong magnetic fields, there are limits to its applicability in clinical brachytherapy. | 6,837.6 | 2020-10-09T00:00:00.000 | [
"Physics",
"Medicine"
] |
Multivariate Data Envelopment Analysis to Measure Airline Efficiency in European Airspace: A Network-Based Approach
: In this paper, data envelopment analysis (DEA) is applied to exhaustively examine the efficiency of the main airline companies in the European airspace by using novel input/output parameters: business management factors, network analysis metrics, as well as social media estimators. Furthermore, we also use network analysis to provide a better differentiation among efficiency values. Results indicate that user engagement, as well as the analysis of the position within the airspace-from an operative perspective, influence the efficiency of the airline companies, allowing a more comprehensive understanding of its functioning.
Introduction
As has occurred in many other business sectors, the airline industry has grown considerably as a consequence of the arrival of the internet, which has allowed multiple new opportunities, from online ticket purchase to real-time information updates for customers about flights or company functioning. The position of airlines in the virtual world of the internet is considered a key point to attract new customers and keep them faithful. Twitter accounts, Facebook pages, or YouTube channels, just to mention a few, are online media used for this purpose.
We can find in the literature many works assessing the efficiency of airlines [1][2][3][4][5][6][7][8][9][10]. Analysis of the efficiency of the airline industry has traditionally been carried out from an economic perspective, mainly using micro-and/or macro-economic estimates to assess the viability of a company, its business plan, its fleet optimization, or its route planning, just to mention a few [2,7,[10][11][12][13][14]. Some authors focus on specific economic variables such as the cost or passenger revenue [2,15,16], while others analyze efficiency by taking into account other airline features, such as fleet or aircraft departures [17,18], and some even include variables such as network size [1,16] or aircraft manufacturers [19].
The maximum efficiency an institution can attain is determined by the heterogeneity of its [1][2][3][4][5][6][7][8][9][10] resources and the capabilities of its management [20]. Resource-based theory [21][22][23] has been widely adopted as the reference point for research into efficient production management within institutions as it can help explain why some firms consistently outperform others. Actually, only a few approaches include some other more heterogeneous aspects of study when considering how to evaluate the efficiency of an airline [1,24]. However, the emergence of new features able to describe the functioning and the position of an airline in the market-such as its relative presence on the internet or its development with regard to other companies-have brought about new opportunities to compare its behavior in the airline business. Designing new models to utilize those non-economic parameters in order to study efficiency has become an interesting topic to address. This paper attempts to add into the equation some important features not often included into the efficiency analysis of airlines, namely social media indicators or network analysis measures.
We propose a novel approach that addresses the aforementioned issues by using data envelopment analysis (DEA) techniques along with several variables relevant to the efficiency of an airline, such as business management factors, the position of the company in terms of network analysis, and social media indicators. We aim to compare how including the two latter factors in the analysis of efficiency may produce different results from relying on traditional economic variables. To the best of our knowledge, this is the first paper on this topic and should help to build a more holistic view of the airline industry.
The paper is organized as follows. In the next section, we motivate our work by reviewing some related work in the literature. In Section 3, we present the methodology we use to study efficiency throughout the paper. Section 4 sets out the data collected related to the European airlines we study. We present the results obtained in Section 5. Finally, we discuss the results, draw some conclusions, and sketch some future avenues for research in Section 6.
Background
Air transport contributed more than €110 billion to the European GDP(gross domestic product) in 2014, and airlines have usually been effective in improving productivity [25]. The efficiency of the internal management of institutions is a function of the costs and benefits of their activities [26]. Literature has plenty of examples of the use of DEA to measure efficiency in airlines (see Table 1). This technique has often been utilized to evaluate efficiency through the use of financial resource indicators, both at corporate and at market level [27]. However, there exist two main trends in which airlines have been evaluated in the last years, in terms of performance: economic and/or operative [28][29][30][31] and economic and/or environmental [7,14,32].
There have been a considerable number of studies measuring the efficiency of European institutions through DEA [14,33,34], between others. Barros and Peypoch [35] used DEA-CCR(Charnes-Cooper-Rhodes) and two-stage regression to analyze technical efficiency and concluded that airlines should consider managerial causes of inefficiency scores when developing strategies, whilst Lee and Worthington [36] employed a double bootstrapping DEA model to analyze the period following deregulation of the European airline industry and financial turmoil in the US airline industry. Another study [14] used network slacks-based measures to analyze how including aviation in the European Union Emission Trading Scheme has affected airline efficiency since 2008. Merkert and Williams [37] used two-stage DEA to measure the efficiency of 18 European public service obligation (PSO) airlines over two fiscal years and concluded that operators with a large number of PSO contracts appear to be more efficient than those with only a few such contracts.
There are also studies of European airlines that have used DEA to explore environmental factors [34]; these studies suggest that the average productivity of airlines, which take an environmentally sensitive approach to growth, is lower than that of airlines taking a more traditional approach. More recent research [33] has used network-based DEA and shows a dynamic component, since performance differed across types of airlines during the 2000s.
However, in the above works, although the methodology applied to obtain conclusions is based on DEA, and they have brought about some important advances in gaining a better understanding of the functioning of the airline industry, they have not addressed airlines in Europe as an ecosystem, where other factors may be considered. As an example, in [38], the authors present a review specific to European airlines and social media research, which concludes that new information and communication technology has had a profound impact on airlines' delivery and consumption, increasing customer engagement, and loyalty in a technologically connected world [32,[38][39][40][41][42]. With this in mind, in this paper, we add analysis of social media indicators to a DEA-based approach measurement of the efficiency of airlines.
To understand the functioning of a complex system, it is essential to analyze its structure [43]. There is some research which has used network analysis to study the development of the airline industry. For instance, Brueckner [44] presented a simple rule for the computation of airport congestion tolls, reflecting the internalization of congestion by analyzing a complex network formed by hub airports. Bagler [45] analyzed the Indian airport network (as a weighted network), concluding that it follows a small-world model and comparing it to the worldwide airport network. To the best of our knowledge, however, network analysis has not previously been used to estimate airline efficiency. In this paper, we propose to study the network formed by the airports in which the airlines operate. By analyzing such a network, we will be able to extract certain centrality measures that may provide useful information about the airline business positioning. As we will detail, these centrality scores will be incorporated into a DEA process as input variables.
Data Envelopment Analysis (DEA)
DEA can be roughly defined as a non-parametric method of measuring the efficiency of a decision-making unit (DMU) with multiple inputs and/or multiple outputs. DEA is used to measure the relative productivity of a DMU by comparing it with other homogeneous units, transforming the same group of measurable positive inputs into the same types of measurable positive outputs. Charnes, Cooper, and Rhodes [49] introduced the DEA method to address the problem of efficiency measurement for DMUs with multiple inputs and multiple outputs in the absence of market prices.
They coined the expression 'decision making units' (DMU) to include non-market agencies such as schools, hospitals, and courts, which produce identifiable and measurable outputs from measurable inputs but generally lack market prices for outputs (and often for some inputs as well). Supposing that there are firms, each producing outputs from inputs, firm uses the input bundle to produce the output bundle. As noted above, measurement of average productivity requires the aggregation of inputs and outputs. Charnes et al. [49] proposed a minimization problem to obtain the efficiency score, assuming a constant returns scale (CRS). This represents the global technical efficiency of a DMU, known as the CCR (Charnes-Cooper-Rhodes) model. They assume that there are DMUs to be evaluated. Each DMU consumes varying amounts of different inputs to yield different outputs. More in particular, DMU consumes amount of input and produces amount of output . Equation (1) presents the multiplier CCR model in an output-oriented version displayed as a linear programming problem: where > 0 is an element defined to be a negligible positive real number, and , are the weight vectors of output variables. If the constraint ∑ = 1 holds, then the model is called BCC (Banker-Charnes-Cooper) proposed by Banker et al. [50]. This model incorporates the property of variable returns to scale (VRS).
Bootstrapping DEA Technique
Data errors may severely affect the efficacy of DEA. The DEA method uses a sample for the efficiency analysis, so the specific differences that can be found in estimations may be only motivated by sampling noise to a great extent than precise deviation in efficiency scores of the units. In this paper, we use a bootstrapping approach to address to this problem. Bootstrapping, introduced by Efron [51], is based on the idea of re-sampling from an original sample to create replica datasets which we can use to make statistical inferences. The 'smoothed bootstrap' approach of Simar and Wilson [52,53] is used in this study. The theoretical foundations of this method are set out in [53].
The key assumption underlying this approach is that the known bootstrap distribution will mimic the original, unknown distribution provided that the known data generating process (DGP) is a consistent estimator of the unknown DGP. The most common approach is to estimate the original densities of the performance scores non-parametrically using kernel smoothing methods, combined with a reflection method [54]. If this approach is followed, the bootstrap process will generate values that mimic the distributions that would be generated from unobserved and unknown DGP [52,55]. This work uses the bootstrapping technique proposed by Martínez-Núñez and Pérez-Aguiar [56] which can be summarized as follows: 1. Calculate the DEA efficiency score ; = 1,2, … , with the original data ( , ). 2. Use Kernel density estimation and the reflection method to generate a random sample * ; = 1,2, … , with replacement from the original DEA efficiency score . 3. Generate * using: 5. Generate resampled pseudo-efficiencies * using * = 2 − * * , * * < 1 * * , . . . .
Calculating Centrality Measures to Feature DEA
In this section, we intend to show how network analysis may be an important factor to consider when studying the efficiency of airline companies. We claim that using this kind of analysis allows additional relevant input properties to be included in the analysis of an organization's efficiency. In particular, we are interested in studying how centrality measures affect efficiency in the airline industry. With that in mind, we propose to develop a network model extracting relationships among airlines in terms of common operations airports, as well as incorporating centrality measures such as degree and eigenvector centrality into DEA efficiency analysis.
Network Model
Network analysis focuses on the use of different metrics with sets of entities linked to one another in some way. First, we define the sorts of networks we examine in this paper-i.e., those generated from European airline companies and their relationships in the European airspace. Let = ( , ) be a graph in which represents the set of airlines and stands for the set of links or connections between them. Let , ∈ , with , ∈ , be an edge in representing that airlines and operate in the same airport. For the purpose of this paper, we consider relationships to be bidirectional-that is, if there is , ∈ , , ∈ necessarily also exists.
Thus, the graph generated by the network is undirected.
Airline Analysis Using Centrality Network Metrics
There are many different metrics that make it possible to uncover some interesting individual behaviors and global properties in a network. We claim that in the domain we consider (airlines), centrality metrics are necessary to shed light on the importance of the position of an airline in the network generated. In other words, centrality metrics attempt to measure how well placed an airline is in relation to the other companies in the airspace-in this case, European airspace. The concept of centrality encapsulates micro measures that allow us to compare nodes and to say something about how a given node relates to the overall network [57,58]. Our hypothesis lies in finding out whether deep knowledge of the position of the airlines in the context of the overlying network generated from the relationships that emerge in the European airspace may bring about certain benefits when studying the efficiency. Let us now set out the different metrics we use to accomplish this task.
Degree centrality. The simplest way of measuring the position of a node in a network (henceforth we use the terms 'node' and 'airline' synonymously when referring to an entity in the network) is to consider the so-called degree centrality. This represents the number of links that a node has. Formally: where ( ) denotes the degree of node in the network. This metric shows how well a firm is connected in terms of direct links. We can also provide an average degree centrality value as follows: This value represents an average estimate for the number of relationships between the companies as a macro indicator of its presence in Europe.
Although degree centrality may be an important factor when analyzing an airspace network, it misses certain aspects that should be considered. While it captures the geographical distribution of an airline perfectly, it does not reflect the strength of its position within the network-i.e., the importance of the firm relative to the whole airline industry. In any case, it may be that an airline has relatively few relationships but lies in a critical location within the network. In other words, different airlines might yield different performances depending on the airports they operate in, even if the number of links they have remain the same. We use eigenvector centrality to overcome this issue.
Eigenvector centrality. Eigenvector centrality (we also use eigencentrality interchangeably throughout the paper), proposed by Bonacich [59], is able to represent the importance of a node in the network. Let ( ) be the eigenvector centrality associated with a network: the crux is that the centrality of a node is proportional to the sum of the centrality of its neighbors. Formally: in which is a scalar called eigenvalue and is an adjacency matrix representing the network. The Perron-Frobenius theorem from linear algebra guarantees that when we work with an undirected network, being a connected component-as occurs in our case-iterating over Equation (4) always converges on a fixed-point equivalent to: As our airlines' network is undirected, then = , so resulting in: As a set of separate equations (with being elements of the adjacency matrix ), we obtain: To solve this set of equations, we need to find eigenvalue-eigenvector pairs ( , ( )). A matrix can have several eigenvalues (to find the eigenvalues of matrix , we simply solve the determinant | − |, where stands for the identity matrix.), and, in turn, several associated eigenvectors. Again, following the Perron-Frobenius theorem, we choose the highest eigenvalue in order to assure positive values in the correspondent eigenvector (positive centrality scores). If we take as the highest eigenvalue in Equation (7) then, after bringing the right-side terms to the left-hand side, we have: Computing eigenvector centralities can be done in a reasonable time (it presents a complexity of ( )).
In summary, eigenvector centrality allows us to calculate how important an airline is in the airspace being analyzed by taking into account the importance of the surrounding companies.
A Network-Based Approach to Refine DEA
As mentioned, DEA proves to be a rather good method to assess the efficiency of organizations. In particular, DEA outputs an efficiency frontier that may be used to determine a reference for efficiency performance. However, this efficiency frontier usually contains more than one unique efficient organization. In this paper, we adhere to the approach by Liu and Lu [60] to better discriminate between the efficiency of the efficient organizations involved in the DEA process. This approach transforms DEA output results (organizations' efficiency) into a weighted directed network and then uses eigenvector centrality to rank the importance of each organization, so better distinguishing between efficient entities. Since we use a completely different domain of study, we have adapted Liu and Lu's approach as follows in the next description. Stage 1. Run DEA with the different proposed models to obtain the efficiency of the studied organizations.
Stage 2. For each model, benchmark the individual efficiencies of each airline with regards to the efficiency frontier (i.e., efficient airlines). This process outputs a lambda value of , if organization i is an exemplar of organization j, with , ∈ [0,1]. In other words, lambda values represent estimates to achieve target efficiency. A value of , = 0 means airline i is not a model of reference for airline j. If , > 0, then there exists an endorsement relationship between airlines i and j. Note that , does not have to be confused with algebraic eigenvalues in Section 3.3.2. Since notation for both terms is accepted worldwide, we prefer not to change either of them.
Stage 3. We aggregate the lambda values resulting from each model for which we ran DEA, so obtaining a lambda matrix as follows:
Stage 4.
We use the aggregated lambda values from the previous step to generate a network in the following way: Airlines are represented by network nodes. A link between two nodes i and j, denoted as ( , ), is created if there exists a lambda value ∑ , > 0, which is, in turn, the weight of the link between the two nodes.
Stage 5.
Calculate the eigenvector centrality for every node in the network. This step will return a ranking, from which a better discrimination of the efficiency of the airlines involved will be easily extracted.
Model Data
We have used data from 43 European airlines to demonstrate how DEA can be used to assess efficiency, and to offer additional data about European airlines' competitiveness. Firstly, we collected information concerning the balance sheet for the year 2014 from the AMADEUS database (https://amadeus.bvdinfo.com) edited by Bureau van Dijk. Secondly, data relating to the airlines' operations was obtained from Openflights.org. (http://openflights.org/) Data concerning social media networks was collected by trawling different web channels, such as YouTube, Twitter, and Facebook. It has been compulsory to normalize negative values due to the characteristics of the DEA model applied in this work.
This paper presents an incremental DEA multistage model whose main purpose is the measurement of efficiency at different layers from a strategic perspective. The comprehensive model integrating all layers is the so-called overall DEA model. Three layers are analyzed using DEA: business DEA layer, network DEA layer, and social media DEA layer. Each layer has several steps in which the variables presented in Table 2 are examined.
Input-Output Data for Bus-DEA Layer
Regarding inputs, in this study a combination of these companies' economic and human resource data was collected; these inputs were used in previous works, among other purposes, to analyze the best practice frontier performance in companies. We selected the following variables: total assets in balance sheet in 2014 [36,61]; number of employees for year 2014 as a measure of human resources [7,19,24,32,36]; and number of destinations as an indicator of geographical diversification [4]. On the other hand, the selected outputs measuring performance of business management in European airlines are sales as the profitability measure supported in the literature in previous works by Cui, Banker, and Thore [28,62,63], and millions of passengers in 2014, as a quantum of the service's users). The source of information was the AMADEUS database edited by Bureau van Dijk.
Input-Output Data for Net-DEA Layer
Besides traditional business management variables, we claim that other types of analysis may be used to include in DEA layers, in particular, when studying airlines. We adhere to the use of network analysis to obtain centrality measures, representing how well positioned each company is with regards to its competitors. With that in mind, we work on an airlines network built as detailed in Section 3. Over this network, we calculate degree and eigenvector centrality measures that we will use in the net-DEA layer.
Input-Output Data for SM-DEA (Social Media-DEA) Layer
The present boom of social media in almost all consumption areas and habits is unquestionable. Social networks contribute to strengthening relationships among and between airline organizations and individuals. To do so, interaction and engagement are the most important functions of social media regarding the management of airlines. Social media had an important effect on companies' performance through a sustainable brand [38,42]. Given the increasing importance of social media in airlines, the necessary next step has to do with the integration of related indicators to assess the efficiency and performance of airlines.
We add to previous DEA layer inputs-outputs data that represent support for and acceptance of a company in social networks for the SM-DEA layer.
As Martínez-Núñez and Pérez-Aguiar [56] assert, Web 2.0 technologies are features with two types of variables-firstly, response levels: number of fans, followers, hits, comments, and tweets received; and secondly, company activity levels: comments, tweets, followings, responses, publications/posts, applications, events, surveys. We incorporate different features to the SM-DEA model as inputs: tweets/days, publication/days, and number of released videos as measure of activity level in social networks [56,64,65]. As outputs we use Twitter likes, Facebook likes, and views in YouTube channels as the user response.
The information on practice of Web 2.0 technologies is based on indicators of social media activity in 2014; the data were obtained directly from the Web 2.0 tools themselves.
Model Summary and Statistical Analysis
The variables selected in each step analyzed in the different layers are displayed in Table 2. This first approach presents an incremental DEA multistage model whose main purpose is the measurement of efficiency at different layers from a strategic perspective. The model is created from a resources and capabilities view. The starting point of this model is that the network development and its impact on organizations modify the way airlines manage their services. Several factors linked to the analysis of the types of network defined in the previous sections have been selected as efficiency components in order to corroborate this starting point. On the one hand, operational networks as physical resources are studied; on the other, the incorporation of social media networks and their relationships with the customer have also been evaluated. Social media estimates must be understood as intangible components that are measured in different intermediate homogeneous steps, thereby analyzing their influence on the airline sector efficiency. Step 2 Step 3 Step 4 Step 5 Step 6 Step 7 Step 8 Step 9 Table 3 shows all the quantitative variables used and its descriptive statistics for the 43 European airlines analyzed.
Results
We first show the results obtained by using DEA with the different layers selected to analyze the efficiency of the airlines.
Then we use the approach adopted from Liu et Lu [60] (see Section 3.4) to rank airlines, sorted by their importance in a referential network created by benchmarking DEA outputs.
DEA Model
We propose an incremental multistage DEA model based on Martínez-Núñez and Pérez-Aguiar [56], starting with a basic layer model (the bus-DEA layer) where business management is measured according to the economic, human, and physical resources. Later, physical variables are incorporated into the operational network resource, building the net-DEA layer. Several variables from social media networks are then integrated to form the SM-DEA layer.
The effects and implications of each social media and operation network variables were analyzed and those most representative were selected to generate a holistic comprehensive model (the overall model). All layer models were created with an output orientation. This technique aims to maximize outputs with the same set of input values [66].
Bus-DEA Layer (Business-DEA)
The influence of business management data in the DEA assessment has been analyzed in steps 1 and 2. The efficiency of the various airline companies is shown in Table 4, which gives a descriptive analysis of the efficiency coefficients. Step 1 analyses only the economic and human resources of these companies, understood as businesses where the special features of the air transport sector are not considered. The average efficiency is 0.4973 and 0.6477, depending on the DEA model orientation (CCR and BCC, respectively). This suggests that companies in this sector could improve their levels of activity in terms of revenue with the same levels of input use by 51% based on constant returns, and by 36% based on variable returns. Returns may seem low in this step, but one must bear in mind that this step lacks contextual variables linked to the air transport sector, such as passengers and destinations.
These contextual variables are incorporated into Step 2 where 14 airlines are globally efficient, all of them working at an optimal scale. There are 22 efficient firms in the BCC model, which means that eight of them are technically efficient but could scale their activity to achieve global efficiency. A high value (0.93) of scale efficiency versus pure technical efficiency is also noted. The standard deviation of efficiency scores in the two steps (both in CCR and BCC orientation) exceeds 0.20, except in scale step 2 (0.1108). This means that the global inefficiencies can be assigned to managerial or operational inefficiencies. In the light of the results and the analysis of previous works, step 2 appears to be the step that best fits a basic step based on variable returns and output orientation.
As a result of the deterministic nature of the DEA method, the analysis of efficiency estimates in the presence of sample variations through the bootstrapping DEA approach is the suitable next step in an efficiency analysis [53] for the following step. For each step, we present a table showing the bootstrapping DEA results, including the average of the bias-corrected efficiency scores, bias, standard deviation, and 95% confidence intervals (lower bound and upper bound) for the biascorrected efficiency scores.
Net-DEA Layer (Network-DEA)
This kind of layer provides the efficiency score changes of all the airline companies when reassigning as input some operational resources variables that, at a network level, provide information on the European destination distribution of the airlines. Table 5 shows the original values and bootstrapped performance estimates of each step of the network-DEA layer. Therefore, the number of air routes of each airline and its strategic positioning with respect to other firms are also analyzed.
Network Model
Step 3 Step 4 Step As is pointed out in Section 4.2, two selected metrics make it possible to discover some interesting individual behaviors and global properties of a network.
Fried et al. [67] indicates that bias-corrected efficiency is preferable to original efficiency estimates because the estimated bias is larger than the standard deviation in every case. Therefore, the analysis described hereafter takes into account the average bias-corrected efficiency scores. Conclusions concerning efficiency changes in steps can be deduced by analysis of the bootstrap confidence intervals. A two-unit increase in efficiency score can be observed and a three-unit increase in both average bias-corrected efficiency score and inefficient average bias-corrected efficiency DMUs-three points in Step 5 that incorporates the two network variables with respect to Step 3. However, there are no significant differences between Step 4 and Step 5, as they have the same number of efficient DMUs and less than one percentage point difference in the bootstrap median score.
The eigenvector centrality has been selected as the critical factor for operational resources network management, given the results of the net-DEA Llyer and according to the higher correlation of Step 4 in relation to Steps 3 and 5 (0.956 and 0.997 Spearman's coefficient, respectively).
SM-DEA Layer (Social Media-DEA)
The results for the different steps of the SM-DEA layer calculated are shown in Table 6. A trend can be appreciated when associating the efficient airline firms with indicators of their involvement in social media networks. The virtual network platforms analyzed in this work are public and external resources to the firms. These resources do not have shortage and any stakeholder can isolate one network from the others. Thereby, the use of one or another virtual network would not suggest any increment of efficiency. But this is not so. The results of the SM-DEA layer show an increase in average efficiency for companies that use Twitter (Step 7) versus the other platforms studied. There is an increase of six and nine efficient companies (respectively) with respect to Step 6 (Facebook) and Step 8 (YouTube). In addition, Step 7 (Twitter) has the lowest interval that does not overlap with the other two steps. Furthermore, Step 6 (Facebook) presents the highest average efficiency score of inefficient DMUs. The potential increasing output of incorporating at least the Facebook platform is 0.0875 and 0.0142 (average efficiency increase of Step 7 and Step 8 with respect to Step 6, respectively). That is, when a sector-average firm starts to participate in social media, the performance of its outputs enhances by this ratio.
The Facebook metrics have been selected as critical factors for new technological resources, given the results of the SM-DEA layer and according to the higher correlation of Step 6 in relation to Steps 7 and 8 (0.842 and 0.713 Spearman's coefficients, respectively).
Overall Model
Specific input reallocation for each type of network (either physical or virtual) gives rise to an efficiency improvement in the overall model. The most efficient airlines combine different resources linked to networks in a way that gives rise to current, future, or potential competences that are interesting to analyze from a strategic point of view.
Analysis of the different steps indicates that the overall model (Step 9) obtains an increase in efficient enterprises of more than 18.5% and an improvement of more than 12% in the corrected average efficiency compared to Step 2. As pointed out in Table 7, the incorporation of factors linked to operational networks (Step 4)-understood as nodes with the potential for developing new business opportunities-generates an improvement in efficiency on the overall model. However, the efficiency improvement in social networks (Step 6) is much clearer, since the increase in the number of users and the social networks management generate valuable data and information with direct and positive implications for efficiency. The comparison between Step 6 (social network layer) and Step 4 (network layer) indicates a clear efficiency increase in Step 6. Three firms have been incorporated into the efficient frontier and there is an improvement in the average efficiency of the inefficient firms of more than seven percentage points, without any overlap between intervals. Therefore, these results indicate that the influence of the direct effects of the physical networks on efficiency improvement is lower that the indirect effects linked to social networks, where user enrolment generates profit but without any direct compensation.
Discriminating Efficiencies by Ranking Centralities
Many methods of discriminating between DEA results have been proposed [68][69][70]. Liu and Lu [60] suggested a method that turns DEA results into a network for the second stage analysis, and uses the tools developed in the social network analysis community to further discriminate DEA results. At this point, and using the results previously presented in this section, we run the mechanism adapted from Liu and Lu [60]. Aggregated lambda values are presented in Appendix B. As follows from our adaptation of Liu and Lu's model, we create a network based on the lambda values resulting from benchmarking the efficiencies obtained by the DEA analysis with the different proposed models. Then we calculated the eigenvector centrality of different companies in order better to discriminate efficiencies of DMUs. We present the ranking obtained in Table 8. Results show that bigger companies, in terms of incomes or engagement in social media, are not the most efficient ones when sorted after applying the mechanism. This leads us to think whether the management strategy those big companies are following has been correctly designed. More recent companies, compared to the rulers of the market in the last 50 years (Turkish Airlines, KLM, or TAP), such as AirEuropa, Air Berlin, or SAS seem to better perform in the current market. Reasons might be very diverse: from a more specific passenger target to the decision of operating in a more local area or region. Figure 1 represents the weighted network of airlines by using their lambda values. Bigger nodes denote a higher efficiency score, while edges are determined by lambda values between two different companies.
Discussion on the Results
We illustrate our approach clearly, using a bubble chart to visualize the information so that the relationship between airline centrality and efficiency can be checked. We try to show how eigenvector centrality values influence the efficiency of airlines taken into account other factors (engagement, fleet size, and income). The following bubble charts combine efficiency (X axis) and the eigenvector centrality calculated as detailed in Section 3.3.2 (Y axis). The size of the bubbles is determined by engagement, (Figure 2 The graph is divided into four quadrants. The first quadrant is the upper right-hand corner of the graph, where we see airlines with efficiency and eigencentrality values close to 1 (Air Europa, Scandinavian Air Lines, Air Berlin, Monarch Airlines, Norwegian Air)-that is, 10.6% of the sample. The second quadrant, in the lower right-hand corner, includes low values of eigencentrality and a range of efficiency between 0.5 and 1, and corresponds to 72.3% of the European airlines studied. The third quadrant, the lower left-hand corner, includes low values of both x and y (Luxair, Travel Service, Meridiana, TAROM, Turkish Airlines), and that corresponds to 17.2% of the sample. Finally, the fourth quadrant, the upper left-hand corner, includes high values of eigencentrality and low values of efficiency: we can observe that there is not one airline in this quadrant, which is consistent with benchmark's definition. From a quality perspective, we can find three groups of airlines. The first group (first quadrant) corresponds to the group leaders: the majority are efficient and are the most referenced for the rest of the DMU (despite not having a big income and their fleet not being particularly large, nor even their engagement in social media). We might suggest that this type of airline chooses a segmentation market strategy and operates on the optimal scale. In other cases, companies cannot achieve their objectives (growth, profitability, market presence, social recognition, etc.) within their current activity and opt for another type of strategy: diversification. This is the situation of the second group (second quadrant, right side): efficient airlines with low values for eigencentrality (such as Air France or Lufthansa), large fleet, their own way of operating, and singular features. In the third group, we find low values of eigencentrality, inefficient airlines, with low-medium income and fleet and with a reasonable value for engagement in social networks. These airlines should leverage their recognition and positioning in social media as an intangible value and improve the way they operate. For further details on the results, please check Appendix A and Appendix B.
Conclusions
Formulating a strategy to compete in one or more countries is complex [71]. Nowadays, the network concept is largely used, being applied in multiple scenarios-IT, logistics or business, among others. Networks represent the ability to establish links and make exchanges, whether physical or in the form of information. The use of network (social or physical) knowledge and the benefit obtained is applied by the organizations, making a resource from the opportunities. The measures of centrality provide valuable information on the positioning of airlines in the European airspace, so providing a more complete benchmarking than the one used so far in DEA.
In this work, we calculate the efficiency scores for 43 selected European airlines from 2014, using an incremental DEA multistage model [56] whose main purpose is the measurement of efficiency at different layers from a strategic perspective. Following [36,55], the bootstrap method is also used to address problems caused by DEA models' sensitivity to errors in the data. This suggests the importance of seeking mechanisms to coordinate corporate and competitive strategies-in this sense, including not only economic indicators, but also network and social media indicators, in DEA models, taking advantage of the amount of data available today. These should be considered key factors in making management decisions. This confirms similar findings in Duygun [33], which include intangible inputs related to customer satisfaction levels in a network DEA from the point of view of customer satisfaction, cost minimization, and an efficient route system. Airline organizations have identified the need to offer their customers a wide range of different social media platforms [38]. This work has found that not all airlines operate the different platforms with the same efficiency. Faber [72] identified that two channels (Facebook and Twitter) were most likely to be used by airline organizations; the results of this work show that the efficiency scores of the Twitter step surpass those of the Facebook step. These findings are similar to Zarrella [73], who suggested that the efforts involved in Twitter engagement pay off quickly and result in a great buzz.
Social media marketing initiatives should not only be measured in monetary units but should also investigate consumer intentions to engage in social media applications [74]. To identify social media marketing's acceptance amongst consumers, measures like the frequency of website visits (i.e., traffic), the number of comments or 'likes' in a social network, or the number of replies on mediasharing websites can provide information on a social media application's attractiveness [75]. DEA methodology allows the efficiency measurement of all these variables. As stated by Martínez-Núñez and Pérez-Aguiar [56], this work confirms that airlines can improve their efficiency by enhancing their social media management capabilities and incorporating Web 2.0 technologies into their business strategy. This work also finds that this new and digitalized communication is becoming increasingly influential in efficiency improvement versus analysis of the route network positioning factors.
As we have shown in the paper, we adhere (and adapt) the method used by Liu and Lu [60], using the eigenvector centrality for refining efficient DMUs. However, in that paper, the authors collect and aggregate lambda values from any possible combination of input/output parameters (in different stages), while, in our case, we only use the results from steps we consider relevant in our domain. We claim our choice is correct in most cases, since domains with many potential input/output parameters (such as the airline industry) might generate combinations not relevant or even counterproductive, in terms of the assessment of efficiency scores.
We prove that, to become and to remain successful, organizations have to accept that within the social media environment, consumer power is much stronger than their own [75]. Consequently, companies need to reconsider their strategic approach to social media marketing practices, showing willingness 'to give up control of the message' [76] as well as allowing consumers to provide criticism and constructive input. This concept becomes more understandable if this process is seen as a virtual network of relationships, where efficiency means knowing how to manage and recognize different user segments in different platforms simultaneously. Results also indicate that the influence of the direct effects of the physical networks on the efficiency improvement is lower than the indirect effects linked to social networks, where user enrolment generates profit but without any direct compensation.
As future research avenues, we plan to study the results from different perspectives, for example, a group-based analysis, focusing on the results obtained by airlines belonging to the same group (e.g., IAG is a holding company which emerged after the fusion of Iberia and British Airways, including Aer Lingus and Vueling). Results obtained from individual efficiencies compared to the cluster companies they belong to might yield interesting insights into the group development and the strategy it follows. The same study might be applied to airline alliances, such as Star Alliance, One World, or Sky Team. We also intend to study the value of social media on the efficiency calculation, but from a local perspective. That is, we are interested in determining whether local companies are relatively popular in the radius of action in which they operate, instead of comparing global estimates. For example, a company only operating in a single country does not need to attract users from foreign countries or, at least, not on the same level as multinational companies do. | 9,648.8 | 2019-12-05T00:00:00.000 | [
"Business",
"Economics",
"Computer Science"
] |
Helical damping and anomalous critical non-Hermitian skin effect
Non-Hermitian skin effect and critical skin effect are unique features of non-Hermitian systems. In this Letter, we study an open system with its dynamics of single-particle correlation function effectively dominated by a non-Hermitian damping matrix, which exhibits $\mathbb{Z}_2$ skin effect, and uncover the existence of a novel phenomenon of helical damping. When adding perturbations that break anomalous time reversal symmetry to the system, the critical skin effect occurs, which causes the disappearance of the helical damping in the thermodynamic limit although it can exist in small size systems. We also demonstrate the existence of anomalous critical skin effect when we couple two identical systems with $\mathbb{Z}_2$ skin effect. With the help of non-Bloch band theory, we unveil that the change of generalized Brillouin zone equation is the necessary condition of critical skin effect.
Non-Hermitian skin effect and critical skin effect are unique features of non-Hermitian systems.
In this Letter, we study an open system with its dynamics of single-particle correlation function effectively dominated by a non-Hermitian damping matrix, which exhibits Z2 skin effect, and uncover the existence of a novel phenomenon of helical damping. When adding perturbations that break anomalous time reversal symmetry to the system, the critical skin effect occurs, which causes the disappearance of the helical damping in the thermodynamic limit although it can exist in small size systems. We also demonstrate the existence of anomalous critical skin effect when we couple two identical systems with Z2 skin effect. With the help of non-Bloch band theory, we unveil that the change of generalized Brillouin zone equation is the necessary condition of critical skin effect.
Introduction.-Research on non-Hermitian systems is attracting growing attention as they demonstrate some novel properties without Hermitian counterparts and many physical problems in photonic systems, electrical systems and open quantum systems can be converted to non-Hermitian Hamiltonian problems [22][23][24][25][26][27][28]. In general, a Markovian open quantum system can be mapped to the problem of density matrix evolution in terms of the Lindblad equation [29,30]. If the Hamiltonian of the system is quadratic and the Lindblad operators are linear, the solution of Lindblad equation can be reduced to solving quadratic non-Hermitian Liouvillian matrix [31,32]. While topological edge states of non-Hermitian Hamiltonians have been intensively studied in recent years [33][34][35][36][37][38][39][40], it is insufficient to study the unique features of non-Hermitian matrix in open quantum systems [41][42][43][44][45][46][47][48].
One of unique features of non-Hermitian systems is the non-Hermitian skin effect [13], which is characterized by the emergence of some eigenstates which corresponding to bulk continuous eigenvalues localized at the boundaries, accompanied with the inconformity of the open and periodic boundary energy spectrum, and breakdown of conventional bulk boundary correspondence [12][13][14][49][50][51][52][53][54][55][56][57][58][59][60]. Both phenomena can be understood in the scheme of non-Bloch band theory by introducing the concept of generalized Brillouin zone (GBZ). The GBZ is composed of all possible values of z = e i(k+iκ) , where k + iκ is the complex analytical continuation of Bloch momentum k, and κ is a function of k and band index. The complex number z can be derived from the characteristic equation f (z, E) ≡ det(H(z) − E) = 0. By requiring a pair of zeros of the polynomial f (z, E) to fulfill GBZ equation |z µ | = |z ν | for the same E and certain µ, ν, the GBZ of the system can be determined [61][62][63][64]. For systems with different symmetries, we note that equations for determining the GBZ may be different. By replacing BZ with GBZ both the bulk wave functions and eigenvalues of open boundary systems can be restored. Meanwhile, the skin effect is also unveiled to be originated from intrinsic non-Hermitian topology, which can be enriched by symmetry. This leads to the discovery of Z and Z 2 non-Hermitian skin effect [62,65]. For open quantum systems related to non-Hermitian Hamiltonian with skin effect, the chiral damping has been uncovered [47].
The critical skin effect (CSE) is a rather unique phenomenon of the non-Hermitian system without Hermitian analogy. Very recently, CSE was dubbed to describe a novel critical behavior in the non-Hermitian system with the energy spectrum and wave function jumping discontinuously across a critical point [66]. It is revealed by ref [66] that CSE occurs whenever one band subsystems with different GBZs are coupled by even a vanishingly small k independent perturbation. According to Ref. [66], CSE does not occur when two one-band subsystems with the same GBZ are coupled by k independent perturbation. We construct an example with CSE by using perturbation to couple systems with same GBZs, for which we call it anomalous critical skin effect. And we also construct an example that subsystems with different GBZs are coupled by perturbations but don't support CSE. We shall explain these phenomena and demonstrate that the change of GBZ equation is the necessary condition of CSE.
In this paper, we shall work in open quantum systems described by Lindblad equation as our another important motivation is to explore new physical phenomenon associated with the Z 2 skin effect and CSE in open quantum systems. We consider a system with internal spin degree and demonstrate the existence of helical damping related to Z 2 skin effect. The helical damping is characterized by the evolution of relative particle numberñ(x, t) with exponentially decreasing intervals and power decreasing intervals distinguished by sharp wave fronts with opposite propagation directions. When the coupling perturbation breaks the anomalous time reversal symmetry, we demonstrate that the corresponding damping matrix exhibits CSE which leads to the disappearance of helical damping under the thermodynamic limit. Our research provides a framework for studying CSE and symmetry protected skin effect in open quantum systems and reveal the origin of CSE. Helical damping.-Open Markovian quantum systems satisfy the Lindblad master equation [29,30]: where ρ is the density matrix, H is the Hamiltonian and L µ are the Lindblad operators describing quantum jumps induced by the coupling to the environment. Consider a one-dimensional (1D) lattice with the unit cell composed of two orbits (sublattices) and each site can be occupied by spin up and spin down fermions. In the momentum space, the Hamiltonian is given by where σ x,y,z and τ x,y,z act on orbit and spin degree of freedom, respectively. Here we consider quantum jump processes described by the following Lindblad operators: where s =↑, ↓ and o = A, B refer to the spin and orbit index, respectively. And x is cell index. Define ∆ m,n = T r(c † m c n ρ) with m, n = (x, s, o), and ∆ = ∆ − ∆ s with ∆ s denoting the stead value of ∆.
After some derivations [67], the dynamical evolution of ∆ is governed by which gives rise to∆(t) = e Xt∆ (0)e X † t with the damping matrix in the momentum space given by where ∆ s = γg γ I with γ = γ l + γ g and takes the same form of the non-Hermitian Su-Schrieffer-Heeger (SSH) model [4,9,13]. When δ 2 = 0, X has anomalous time reversal symmetry, as it fulfills CX(−k) T = X(k)C with C = iτ y [19]. We can get the eigenvalues of X under open boundary condition (OBC) and periodic boundary condition (PBC) as shown in Fig.1 Fig.1(c), we show the distribution of A x under the OBC. All the eigenstates of X are localized on left and right boundaries, which is another sign of skin effect. If we put two identical models together and add a small symmetry-allowed perturbation, the skin effect is disappeared (SII in [67]). This is the characteristic of Z 2 skin effect.
Given n x,s,o ≡ ∆ (x,s,o),(x,s,o) denoting the particle number with spin s and orbit o at site x, we define the ) 2 and relative local particle numberñ x (t) = s,o∆ (x,s,o),(x,s,o) . In Fig.2 (a) and (b) we display log(D x (t)) as a function of t for different x. While D x (t) under PBC is always a power law function of t, D x (t) under OBC changes from a power law function to an exponential function of t during the evolution. We find that the transition time t c decreases as x increases for 0 < x < 20, and increases as x increases for 30 < x < 50. In order to see more clearly the dependencies between t c and x, we plot the relative local particle number evolution in Fig.3(a) and (b) for the periodic and open boundary system, respectively. We find that there are three main colors: dark blue, blue and purple, which are separated by two straight lines as shown in Fig.3(b). The separatrix of the dark blue area and the purple area is the transition line. Such a phenomena is dubbed as helical damping. Nevertheless, the Z 2 skin effect is not the sufficient condition of helical damping (see SIII in [67]), and we also require the Liouvillian gap of periodic lattice to be zero and the open boundary Liouvillian gap to be nonzero, where the Liouvillian gap is defined as Λ g = min[2Re(−λ n )] with λ n the eigenvalues of X. We notice that [ñx(t)] OBC [ñx(t)] P BC may exist helical behavior even if periodic boundary system is gapped [67]. When the periodic boundary system is gapped (gapless), the short-time behavior of damping fulfills exponential (power) law for both the periodic and open boundary systems, since it costs time for sites located not on the boundary to get the boundary information. On the other hand, long-time behavior of OBC's damping fulfills exponential (power) law when the open boundary system is gapped (gapless), which will be explained further below. Now we use non-Bloch band theory to explain helical damping. For open boundary system the bulk wave function and eigenvalue of X matrix can be obtained by replacing X(k) with X(k + iκ). All possible values of e i(k+iκ) constitute GBZ. In Fig.3(c) and (d), we display the GBZ of the system with different parameters, which is composed of two closed curves with one inside and one outside the Brillouin zone (BZ). The relative local particle number can be decomposed into every GBZ modes: ), where α is band index, λ k+iκ,α is the eigenvalue of X(k + iκ) corresponding to α band. For simplicity we use v to label v k+iκ in the following text. If the parameter settings are the same as in Fig.3 The decay factor cancels out, particle number damping fulfills a power law. Similarly, for v = −1, κ = 0.2 and x = x 1 + vt ≤ L − t, the decay factor also cancels out. For x = x 1 + vt<t and x = x 1 + vt>L − t, this factor cannot be canceled out, and relative particle number damping obeys an exponential law. Due to the anomalous time reversal symmetry, we have n x (t) = n L−x (t), which distributes symmetrically about x = L 2 . Dynamic Critical Skin Effect.-When the system exhibits CSE, the open boundary energy spectrum is not continuous under the small change of parameters in the thermodynamic limit. For the finite size system, the open boundary spectrum is always continuous under the small change of parameters. Therefore, if CSE occurs, the energy spectrum of the system varies greatly with the size of the system. Here we study whether the perturbation δ 1 or δ 2 will cause CSE. And we want to detect this effect in dynamic experiments. With parameter set as δ 1 = 0.02, δ 2 = 0 or δ 1 = 0, δ 2 = 0.02, we calculate the spectrum of X of the system with different sizes. The result is shown in Fig.4. In Fig.4(a)-(c), we set δ 1 = 0, δ 2 = 0.02 and display the spectrum of damping matrix for different system sizes. While the periodic spectrum is not sensitive to the system size L, the obvious change of open boundary spectrum with the increase of L indicates the existence of CSE, and the open boundary Liouvillian gap Λ g decreases as the system size increases.
To measure the Liouvillian gap from dynamic experiment, we will deduce the relationship between relative particle number evolution and Liouvillian gap. The relative particle number isñ x (t) = i,j,s,o e (λi+λ * j )t x, s, o|Ψ i LR Ψ i |Ψ j LR Ψ j |x, s, o , where subscripts R and L denote the right and left eigenvectors of X. Consider the case with large enough t. In this case, modes with −Re(λ i + λ * j ) > 2Λ g can be omitted, and it followsñ x (t) ≈ ce −2Λgt . Assume that log(ñ x (t)) = α(t)t + β, then α ≈ −2Λ g . We numerically calculate the values of 2Λ g and α for different size systems. As illustrated in Fig.4(g), the numerical results are consistent with our theoretical analysis. We also analyze the scaling of the Liouvillian gap with the system size, which indicates log(2Λ g ) ≈ −2.3log(L) + 6.8 around L = 200 and the absolute value of this slope increases as L increases. When the system is large enough that −2Λ g > α max , the helical damping is hidden. When the system is small enough that −2Λ g < α min , the helical damping is manifested. Here α max /α min is the maximum/minimum slope of log(ñ x (t)) in the power law interval. In Fig.4 the system size increases. Therefore we can detect the presence of CSE by measuring the damping spectra of systems of different sizes in this case. Remarkably, we construct an example that two irreducible subsystems with different GBZs are coupled together but no CSE occurs. Specifically, X is constructed by coupling two systems iH nSSH (k) − γ 2 and iH T nSSH (−k) − γ 2 , which have different GBZs. The perturbation term of δ 1 couples iH nSSH (k) − γ 2 and iH T nSSH (−k) − γ 2 , but there is no CSE.
Here we give an explanation in terms of GBZ. The bulk spectrum of open boundary X is composed of eigenvalues of X(k + iκ), and thus it is a continuous function of GBZ. On the other hand, the solution of a certain GBZ equation is a continuous function of parameters of X. Therefore the only origin of energy spectrum's discontinuity is the change of GBZ equation. For case 1 with δ 1 = δ 2 = 0, X is reducible, and the characteristic polynomial is f 1 (z, λ) = det(iH nSSH − γ 2 − λI) and The roots of f 1 = 0 and f 2 = 0 are z a 1 , z a 2 and z b 1 , z b 2 , respectively, where |z a 1 | ≤ |z a 2 |, z b i = 1/z a i (i = 1, 2). The GBZ equations are |z a 1 | = |z a 2 | and |z b 1 | = |z b 2 | [61,62]. If X is irreducible non-Hermitian matrix, the characteristic polynomial is f (z, λ) = det(X − λI), and the solution of f (z, λ) = 0 is z 1 , z 2 , z 3 , z 4 , where |z 1 | ≤ |z 2 | ≤ |z 3 | ≤ |z 4 |. For case 2 with δ 1 = 0 and δ 2 = 0, X does not have any symmetry, and the GBZ equation is |z 2 | = |z 3 |. For case 3 with δ 1 = 0 and δ 2 = 0, X has anomalous time reversal symmetry. The GBZ equations are |z 1 | = |z 2 | and |z 3 | = |z 4 | and the roots satisfy z 2+i = 1/z 3−i (i = 1, 2) [64]. We emphasize that the GBZ equation of case 1 and case 3 are equivalent [67]. So there is a change of GBZ equation from case 1 to case 2, which causes the discontinuity of eigenvalues and wave functions under the time-reversal-breaking perturbation. The GBZ equation will not change from case 1 to case 3, therefore CSE does not occur in this process.
Anomalous Critical Skin Effect.-Here we construct a model that couples two irreducible subsystems with same GBZs but shows CSE. It is anomalous critical skin effect. We consider the model described by The parameters are set as t 1 = t 2 = 1, γ = 0.4, δ 1 = 0.1, δ 2 = 0 and δ 3 = 0.01. It couples two identical systems of X described by Eq.(5), but exhibits CSE (See SII in [67]). When δ 3 = 0, the GBZ equations of X 1 are |z 1 | = |z 2 | and |z 3 | = |z 4 |. When δ 3 = 0, the characteristic polynomial of X 1 is g(z) = det(X 1 (z) − λI), with the solution of g(z, λ) = 0 given byz 1 ,z 2 , ...,z 8 , where |z 1 | ≤ |z 2 | ≤ ... ≤ |z 8 |. The GBZ equations are |z 3 | = |z 4 | and |z 5 | = |z 6 |. The GBZ equation also changes when CSE occurs. The reason for the discontinuity is the change in GBZ equation. Furthermore, changes in GBZ equation require changes of symmetries of the system or number of roots of characteristic equation. In this case, when δ 3 changes from zero to non-zero, the unitary symmetry σ z X 1 = X 1 σ z disappears. Therefore, changes in GBZ equation are necessary conditions for CSE, but the reverse is not correct. An example is that no matter how GBZ equation changes for a Hermitian system, there will be no CSE. The symmetry H † = H keeps the GBZ to be still a unit circle, even if the GBZ equation changes. Conclusion.
-When the open boundary spectrum of the damping matrix of system is gapped and the periodic boundary spectrum is gapless, this system has a non-trivial damping spectrum. If the system also has Z 2 skin effect, helical damping occurs. Adding perturbations which break anomalous time reversal symmetry, CSE occurs and helical damping disappears for systems in the thermodynamic limit, but still exists for small size systems. Coupling two identical models with Z 2 skin effects by perturbation, we can realize anomalous critical skin effect. And we unveil the origin of the discontinuity coming from the change of GBZ equation. These phenomena can be verified in open quantum system by measuring local particle number evolution. Next we derive the expression of evolution equation of single-particle correlation matrix ∆ with ∆ mn = T r(c † m c n ρ).
Inserting the Lindbald equation into into
Here H = j,k h jk c † j c k , L µ = L g µ or L l µ with L g µ = k D g µk c † k and L l µ = k D l µk c k , j, k, m, n, is fermion index and µ, ν is Lindblad operator's index. And we define M g jk = µ D g * µj D g µk and M l jk = µ D l * µj D l µk . The first term gives:
SII: MODEL FOR EXHIBITING ANOMALOUS CRITICAL SKIN EFFECT
In this section, we show that the skin effect and helical damping is not stable to symmetry-allowed perturbation δ 3 in Eq.(7) of the main text. The considered damping matrix is given by The open boundary and periodic boundary spectrum of X 1 are displayed in Fig.5. While the periodic boundary spectrum is not sensitive to the lattice size, the shape of open boundary spectrum changes obviously with the increase in the lattice size. Such an obvious change of open boundary spectrum is induced by the perturbation term of δ 3 . In Fig.6, we show the distribution of A x under the OBC, where A x = α |Ψ xα | 2 is sum of modular squares of the amplitude of the wave function of of X 1 in each unit cell, α is degrees of freedom in each cell and x is the cell index. It is clear that some eigenstates of X 1 spread over all the lattices. We find that there is no skin effect for this model. In this section, we show that Z 2 skin effect is not a sufficient condition for the occurrence of helical damping of particle number. When particle number damping in the periodic boundary system fulfills an exponential law, the particle number damping in the open boundary system always follows an exponential law. In this case, the Liouvillian gap is not zero. Consider a one-dimensional lattice with each cell having one orbit and spin degree of freedom. described by the following Hamiltonian in the momentum space: where τ α (α = x, y, z) act on spin space. Suppose that the Lindblad operators are given by it follows∆(t) = ∆(t) − ∆ s = e Xt∆ (0)e X † t , where X = 2it 1 cos k x − 2γ g τ 0 + (2iδτ x + 2γ g τ z ) sin k and ∆ s = I. We display the spectrum of X matrix under PBC and OBC in Fig.7, which indicates the existence of nonzero Liouvillian gap for both periodic and open boundary systems. The disappearance of skin modes, after putting two identical models together and adding a small symmetry-allowed perturbation, indicates the existence of Z 2 skin effect.
Set the initial state as the state without particle occupation, and we can get∆ = −I. We show the relative local particle number damping for different cells in Fig.8(a). And we calculate the evolution of |ñ x (t)| and display the numerical results in Fig.8(b) and (c), which indicates the particle number damping under both PBC and OBC fulfilling exponential law. In Fig.8(d), we display the evolution of (|ñx(t)|) OBC (|ñx(t)|) P BC , which exhibits helical behavior. And we call it generalized helical damping. The generalized helical damping is a more inclusive physical phenomenon than helical damping, and it don't need the periodic boundary Liouvillian gap to be gapless. The mismatch of open and periodic boundary damping spectrum is the necessary condition. | 5,269.6 | 2020-05-06T00:00:00.000 | [
"Physics"
] |
Operationalising ethics in artificial intelligence for healthcare: a framework for AI developers
Artificial intelligence (AI) offers much promise for improving healthcare. However, it runs the looming risk of causing individual and societal harms; for instance, exacerbating inequalities amongst minority groups, or enabling compromises in the confidentiality of patients’ sensitive data. As such, there is an expanding, unmet need for ensuring AI for healthcare is developed in concordance with human values and ethics. Augmenting “principle-based” guidance that highlight adherence to ethical ideals (without necessarily offering translation into actionable practices), we offer a solution-based framework for operationalising ethics in AI for healthcare. Our framework is built from a scoping review of existing solutions of ethical AI guidelines, frameworks and technical solutions to address human values such as self-direction in healthcare. Our view spans the entire length of the AI lifecycle: data management, model development, deployment and monitoring. Our focus in this paper is to collate actionable solutions (whether technical or non-technical in nature), which can be steps that enable and empower developers in their daily practice to ensuring ethical practices in the broader picture. Our framework is intended to be adopted by AI developers, with recommendations that are accessible and driven by the existing literature. We endorse the recognised need for ‘ethical AI checklists’ co-designed with health AI practitioners, which could further operationalise the technical solutions we have collated. Since the risks to health and wellbeing are so large, we believe a proactive approach is necessary for ensuring human values and ethics are appropriately respected in AI for healthcare.
Introduction
Although the exponential growth of Artificial Intelligence (AI) for healthcare is promising, its benefits are being increasingly overshadowed by its propensity to cause individual or societal harm quickly at a large scale [1,2]. AI for healthcare encompasses numerous approaches, including machine learning (ML) algorithms on structured text-or image-based data, and natural language processing (NLP) on unstructured data such as clinical notes or medical journals [3]. These approaches are being applied to the prediction (e.g. predicting the presence of type 2 diabetes from clinical risk factors, [4] or the risk of suicide from social media posts [5]), detection (e.g. detecting breast cancer tumours from mammography scans [6]), and management of diseases (e.g. using NLP-driven chatbots to deliver cognitive behavioural therapy [7]). This progress has endowed AI with significant hype [8], despite many ethical challenges threatening health and human rights remaining unaddressed [9]. A prominent example of this is how melanoma-detecting AI algorithms are presently trained largely on images of white skin, making them inaccurate at detecting melanoma in darker-skinned people (despite the fact that melanoma is more lethal in African populations) [10]. If ethical issues such as these are not promptly addressed, we risk an 'AI winter' taking place, whereby public trust and the potential benefits of AI for healthcare could be swiftly lost [6]. Regulatory policy to address these issues at the legal level [11] and governance frameworks to address these issues at the organisational level [12] are slowly emerging, but there remain few practical recommendations that developers and users of AI for healthcare can utilise throughout the AI lifecycle [13].
3
Existing ethical AI guidelines have two issues: firstly, very few are specific to healthcare [14], despite the fact that AI for healthcare involves unique ethical issues [2]; and secondly, they emphasise adherence to 'ethical principles' [15] without complementary translations into actionable practices [16,17]. As such, there remains a pressing need to operationalise ethics throughout the development pipeline of AI for healthcare [18].
To address this gap, we first provide an overview of the unique ethical issues that arise when human values are compromised in AI for healthcare. We then propose a framework for operationalising ethics, grounded in existing guidelines that provide actionable solutions. To ensure our framework is accessible and utilisable, we organise it in accordance to the AI lifecycle. In going through each stage of the development pipeline, we outline implementable recommendations that adopt, but also go beyond an abstract consideration of 'ethical principles' [19]. Our framework is of direct relevance to those developing AI for healthcare (including software developers and data scientists, who we collectively refer to as 'developers'), and those utilising AI for healthcare (including clinicians and health informaticians, who we collectively refer to as 'users'). Although we focus on healthcare, our framework may also be useful to those developing and utilising AI in other domains. Important facilitators of our framework include adherence to governance models [20] and the creation of best-practice ethics checklists [2,21], both of which are emerging as AI for healthcare continues to develop.
Methodology
To contextualise and assemble our framework, we performed two literature reviews. Firstly, we performed a scoping review on the range of ethical issues that may arise in AI for healthcare. This involved searching PubMed and Web of Science on literature at the intersection of AI, healthcare, and ethics. Titles and abstracts of the first 200 articles from each repository were assessed, and relevant articles were reviewed in full.
We then used two conceptual frameworks to organise the ethical issues identified. First, we use Schwartz' theory of basic human values, an empirically-validated framework comprised of human values that are assumed to be universal across all cultures [22]. Of the ten total human values in this theory, we reference only the four found to be most cited in software engineering literature [23]. Acknowledging the broad nature of these human values, we further subcategorise them into specific, granular ethical principles, as outlined in a recent scoping review of AI ethics publications [14]. The one-to-many mapping of human values to ethical principles is arbitrary, and is used only as a foundation to present an organised overview of ethical issues identified in the AI literature.
Second, we performed a scoping review of existing frameworks, guidelines and recommendations pertaining to ethical AI for healthcare. We chose Scopus and Google Scholar to identify relevant articles, searching for literature at the intersection of AI, healthcare, and existing guidelines for ethical AI. Noting that the publication of generic ethical AI guidelines has increased exponentially over recent years [14], we focussed on scholarship at the intersection of ethical AI and healthcare wherever possible. We assessed the first 200 articles identified by Scopus and Google Scholar, then adopted a forward and backward snowballing approach to identify papers offering actionable solutions for operationalising ethics throughout the AI lifecycle. Drawing upon existing literature to separate the AI lifecycle into distinct stages [2,24,25], we then present the actionable recommendations that can be operationalised at each stage of the pipeline by developers and/or users to ensure ethical AI for healthcare.
Key human values and ethical issues in AI for healthcare
A consideration of human values is largely lacking in software engineering [23], which is used for the majority of AI applications [26]. This is despite the fact that purposefully aligning AI with human values can produce many benefits, such as improving cancer care and patient engagement [27]. Conversely, if human values are compromised, a multitude of ethical issues can arise [8], which we outline extensively in the following subsections. Here we provide a taxonomy of human values from the social sciences and review the ethical issues corresponding to these values that arise in AI for healthcare. Schwartz' Theory of Basic Human Values describes ten human values validated by empirical research conducted across over 70 countries [22]. A literature review found four values to be the most frequently cited in 1,350 recently published software engineering publications: security, benevolence, universalism, and self-direction [23]. Independently, a scoping review conducted a thematic analysis of 84 AI ethics guidelines [14], identifying 11 key ethical principles referenced across guidelines. Table 1 presents these human values [22], together with the corresponding ethical principles associated with them [14]. The mapping between human values and ethical principles is arbitrary, and is used only to outline and conceptualise the many ethical issues arising in AI for healthcare.
Security
Security encapsulates feelings of safety and stability of oneself, one's relations, and society at large, comprising the ethical principle of non-maleficence.
Non-maleficence
Non-maleficence involves minimising foreseeable harm in terms of discrimination, violation of privacy, and bodily harm, and is cited considerably more than beneficence in current AI guidelines, suggesting it is a greater priority for AI to avoid harm than to do good [14]. Since AI for healthcare is evolving so rapidly, there is a concern that harms will only be recognised and addressed after they have occurred [28].
Safety is a key priority in AI for healthcare, especially since relatively few initiatives are backed by empirical evidence [2]. Technical failures, such as AI chatbots that cease to function properly [29], or AI initiatives that fail during a network failure [30], may result in unintended harm. In addition, AI may also lack interpersonal or cultural competency, which could hinder therapeutic relationships and cause unintended psychological distress [29].
Self-direction
Self-direction encapsulates a sense of personal independence, comprising the ethical principles of freedom and autonomy; dignity; and privacy.
Freedom and autonomy
Freedom and autonomy refer to the preservation of selfdetermination, which includes informed consent and the right to withdraw consent [14]. Informed consent involves the disclosure of relevant information, the individual being competent and fully comprehending this information, and the individual voluntarily accepting participation [31]. The process of informed consent also involves clarifying concerns or misconceptions, including the possibility of third parties accessing confidential data [32].
Obtaining consent may be difficult for black box algorithms that are too opaque to be fully understood by humans [33], and may not be practically feasible for social media or other large datasets encompassing millions of individuals [34]. In terms of AI-based mobile health apps, users may assume as high an ethical commitment to confidentiality as in professional healthcare, even though this is often not the case [35]. Although the terms and conditions of health apps may be presented, these are seldom read or comprehended by users [31].
Dignity
Dignity refers to the preservation of human decency and rights [14]. One consideration is the dignity of developers, who may have no training in healthcare and could find emotional content traumatic [34]. Other considerations relate to how individuals might form therapeutic relationships with AI. Possible ethical issues with this include mentally ill individuals wrongly believing they are interacting with a human or other force; individuals [22] and [14] respectively Human value [22] Ethical principle [14] Security (safety, harmony, and stability of society, of relationships, and of self) Non-maleficence (protection from harm, precaution, prevention, nonsubversion) Self-direction (independence in thought and action; creating, exploring, being curious) Freedom and autonomy (consent, choice, self-determination, liberty, empowerment) Dignity Privacy (protection of personal or private information) Benevolence (preserving and enhancing others' welfare, voluntary concern for others' welfare) Beneficence (benefits, well-being, peace, social good, common good) Responsibility (accountability, liability, acting with integrity) Trust Transparency (explainability, understandability, interpretability, acts of communication and disclosure) Solidarity (social security and cohesion) Universalism (understanding, appreciation, tolerance, and protection for the welfare of all people and for nature) Justice and fairness (consistency, inclusion, equality, equity, non-discrimination, respect for diversity, plurality, accessibility, redress) Sustainability (conserving environment and natural resources) 1 3 feeling upset if a humanoid robot invokes a creepy sense of repulsion (termed the "uncanny valley"); and individuals not having an easy way to safely end the therapeutic relationship [30].
Privacy
Privacy is a human right, involving individuals' information being carefully protected and securely managed [9]. Information relating to individuals' health can be highly sensitive [35], and clinicians therefore have an ethical obligation to maintain confidentiality. To what degree this obligation to maintain confidentiality extends to AI in healthcare-which often require access to large amounts of sensitive health data to be effective-remains a matter of debate [36]. The right to privacy is recognised to be in ongoing tension with the open advancement of AI for healthcare [31], with key issues relating to data collection, data management, and working with individuals' social media data.
Data collection can entail numerous ethical issues. Data may be collected by multiple sources (e.g. smartphone geolocation and online forum activity) [31], by AI robots [29], or by passive means (e.g. screen taps or voice inflection) [35]. A key issue is whether individuals feel comfortable about their data being collected in each of these cases, including when they are unaware that it is taking place [31].
Data management involves additional considerations. Privacy can be threatened through poor security practices, such as leaving a laptop with sensitive data unattended in public. Privacy can also be threatened through hacking by nonauthorised parties [30]. Regarding social media and mental health, users generally expect their privacy to be respected online [37], with most feeling their data should not be used for mental health research without their explicit consent [34]. Social media data may be problematic on numerous counts: it may be an inaccurate representation of an individuals' mental state (given one's self-portrayal on social media can differ markedly from offline behaviours); it may include other users who have been 'tagged' (who themselves have a right to privacy); and it may be difficult to anonymise [32,34]. Finally, if a user decides to 'drop out' from social media after their data has been collected, their health data may still be stored [34].
Benevolence
Benevolence encapsulates a sense of enhancing and maintaining 'good' for oneself and others, comprising the ethical principles of beneficence; responsibility; trust; transparency; and solidarity.
Beneficence
Beneficence involves contributing to individual and societal wellbeing [14]. Although AI has the potential to do much good, key ethical issues relate to the known limitations of AI, and the impact of AI on clinicians' decision-making.
The known limitations of AI should be recognised. Whilst a clinician may be able to monitor societal risks as per their professional code (e.g. the risk of domestic violence or child abuse), AI initiatives built for other purposes could miss these signs entirely [30]. Additionally, those who are highrisk, such as those with severe depression, may need more comprehensive treatment than AI initiatives alone [35].
AI-driven decisions could also have unforeseen impacts on clinicians' decision-making. For instance, if a patients' high-risk genetic mutation is known to radiologists, the number of missed breast lesions on MRI scans is known to decrease considerably [38]. AI-driven predictions could therefore sway clinicians' own assessments of risk. If only patients deemed high-risk by the AI are offered further screening or treatment (without appropriate input from clinicians), this has the potential to become a self-fulfilling prophecy [39].
Responsibility and trust
Responsibility involves attributing accountability and liability, as well as acting with transparency and integrity in a way that builds trust [14]. For developers, clearly stating the limitations of their work is a key consideration [37]; for instance, most suicide risk predictions do not predict when an individual may attempt suicide, and hence whether involuntary restraint is warranted [40].
The lines of responsibility between developers, implementers and AI are not clearly defined, particularly with 'black box' algorithms that cannot be easily understood [34]. For instance, if a suicide occurs and was not detected by the AI algorithm, is it the fault of the algorithm, the developer, the clinician, or the manufacturer? This question remains unresolved [40]. Other areas of ambiguity are how conflicts between an AI-driven decision and a clinician's impression should be resolved, such as if an AI algorithm detects an individual to be at high risk, but the clinician disagrees [40]. Although clinicians have professional standards (e.g. a duty of care) to which they are held accountable, AI have no in-built responsibility to maintain these standards [35], and have no ability to experience the moral consequences of poor decisions (e.g. emotional distress) [30].
Trust is built through a transparent culture amongst developers, implementers, and patients, ensuring that practices fulfil public expectations [14]. Trust may be lost when AI have many false positive or negative results [41], are perceived to be incompetent [30], or make use of public data in a disrespectful way [37]. A breakdown of trust is not only harmful for AI initiatives, but could also be harmful for clinicians and healthcare as a whole [30].
Transparency
Transparency involves a sense of interpretability or explainability of AI-based decisions [14]. Interpretability can be understood as 'how' a model arrives at a decision, whereas explainability can be understood as 'why' a model arrives at a decision [41]. Key issues relate to AI having low interpretability and/or explainability, and shortcomings of AI not being fully disclosed.
Although some AI algorithms are more interpretable (e.g. regression models) compared to others, empirical findings suggest that uninterpretable algorithms (e.g. deep learning algorithms developed from millions of data points) tend to perform better [31]. The issue with 'black boxes' algorithms is that humans can understand the inputs and outputs, but cannot clearly understand the process connecting the two [34]. In many cases, such algorithms also lack explainability, which poses issues; for instance, an individual may be told they are at high risk of developing an illness for reasons that remain undeterminable from the AI model. Individuals remain divided about whether they would like to be told their AI-derived high risk status of an illness if it is unexplainable, given this can be distressing [31]. Additionally, if implementers cannot understand models in the first place, they may be unable to discover sources of bias or challenge AI-driven decisions [31].
Other issues of transparency relate to disclosure of an AI intervention's shortcomings. This includes the rate of false positive or false negative results (including for specific groups) [2], changes to model performance over time [31], and the presence of bias [41]. The amount of information to be disclosed to implementers or patients about AI algorithms, and how this should best be done, remains contested [14].
Solidarity
Solidarity refers to the special consideration of vulnerable populations and those of low socioeconomic status, including the potential need to redistribute benefits of AI to these groups [14]. For instance, NLP algorithms may be developed only in the English language, such that they cannot be applied to cultural groups using other languages [32].
The implementation of AI for healthcare can cause harm to vulnerable populations. For instance, some individuals could find the notion of AI (and the tracking of their behaviour) distressing, such as those with schizophrenia fearing mass surveillance [34]. Moreover, AI could be used for ulterior motives, such as health insurance companies using AI to identify high-risk individuals whose premiums they wish to raise [42]. Finally, although AI can provide some form of healthcare in resource-poor areas, this could be used as a justification to not further develop physical (as opposed to virtual) mental health services in these areas of need [29].
Universalism
Universalism encapsulates a sense of appreciation towards people and the planet, comprising the ethical principles of justice and fairness, and sustainability.
Justice and fairness
Justice and fairness refer to representing the full diversity of society (rather than the privileged few) while safeguarding against discrimination towards vulnerable groups, and upholding individuals' right to challenge AI-based decisions [14]. If people of different gender, ethnic and other sociodemographic backgrounds are not represented in the research, design and development of AI, these interventions could implicitly ignore the needs of these groups [2]. For instance, gaps in training data could stem from the lack of non-binary gender identities in electronic health records, or from undocumented migrants with low access to healthcare [40]. Moreover, training data may also reflect systemic biases based on gender, race, and other sociodemographic characteristics-for instance, the disproportionate number of African-Americans who suffer from schizophrenia [33]. If such data is used for predictions, these algorithms could simply exacerbate existing disparities [40,41]. Economic factors (e.g. different billing rates for specific groups) could also influence what and how information is collected in healthcare [41]. Finally, social media users may be more likely to be young and Caucasian compared to the overall population, meaning the health of other groups could be ignored [34].
Since AI-driven decisions are not absolute truths, another potential issue is whether implementers can challenge questionable outputs. This is particularly important in high-risk situations; for instance, suicide prediction algorithms' false positives could result in an individual being wrongly detained by a health service [43], whilst false negatives could mean an otherwise preventable suicide is missed [40]. Due to these possibilities, delegating decision-making to "machines alone" has been criticised [44].
Sustainability
Sustainability involves considering the environment and minimising the ecological footprint of AI initiatives [14]. This is of importance for all AI initiatives, with relevant factors specific to healthcare not yet identified.
3 4 A framework for operationalising ethics in AI for healthcare
Although there has been significant work in mapping the ethical principles in AI for healthcare, this understanding alone does not translate readily into actionable practices [16,19]. To address this gap, here we present our framework for operationalising ethics in AI for healthcare across the development pipeline, as illustrated in Fig. 1. Firstly, ethical AI is a product of numerous layers of influence: the professional practices of developers and users; the governance of these individuals at an organisational level; and the regulation of individuals and organisations at a legal level. Secondly, the AI lifecycle involves three major stages, which must progressively be fulfilled before the subsequent stage is initiated. These stages are as follows [2, 24, 25]: 1. Data management involves: (A) data being collected; (B) data being appropriately protected with best-security practices; (C) data being cleaned (including pre-processing and augmentation where appropriate); and (D) data being reported. We present our operationalised ethical AI framework across these three stages, providing actionable solutions to prevent, mitigate and address ethical issues. Throughout our work, we identify guidelines that provide actionable recommendations on specific ethical needs (e.g. protecting classifiers from adversarial attacks). All of these guidelines are further explained in our framework, and have been summarised in Table 2 for the reader's convenience (see Supplemental Material for unabridged version). We end by highlighting how these solutions cannot simply be actioned in a vacuum, but must be supported by rapidly-evolving governance and regulatory frameworks.
Data collection
Data can be collected via numerous methods. People are more likely to report behavioural symptoms to digital agents than humans, but sociocultural sensitivity can only be provided by a human agent; hence, care should be taken to ensure methods of data collection are appropriate [41], and the collected data itself is more diverse and inclusive [46][47][48][49]. Moreover, the variables and collected data should be justified by having demonstrated pertinence to healthcare; unnecessary variables should not be collected [18]. This includes avoiding datasets known to be imbalanced or biased [50][51][52][53] and data types that will not be used, such as photographs of people when training text-based [32]. Although social media data may be easy to collect, the privacy of users should be acknowledged and respected [152]. There remains no consensus of the threshold of health data that requires individual consent for use, but in some cases, deidentification (including of name, date of birth, address, and/or health card number) may be deemed satisfactory for bypassing this step [153]. Since training AI on imbalanced datasets data can fuel inequalities, a special effort should be made to gain data from understudied, underserved or vulnerable populations [2,154]. Applying data pre-processing techniques [68,69] to improve the quality of data (regarding diversity and inclusion) often proves more effective than dealing with unfairness resulting from feeding biased data into models for decision making and predictions [54,55].
Technical efforts should be complemented by more holistic, humanistic initiatives. This involves raising practitioners' awareness of the contextual nature and demands of particular ethical principles such as fairness, and actively building ethically-aware and diverse teams to address potential systemic biases [45,[56][57][58][59][60][61][62][63][64]. Table 2 Guidelines addressing specific needs in AI for healthcare and the ethical principles they address
Data protection
There is always a relatively high risk of adversarial attacks on health data systems [155], and hence, data should be stored, handled and used with best security practices e.g. the use of de-identification, anonymization [51,[70][71][72] and other approaches such as Differential Privacy to to procted data from malicious attacks and maintain and data integrity [50, 73-75, 77, 78, 82]. A number of practical tools, techniques and frameworks have been developed to used for procting data used in ML [79-81, 84, 85] some of which allow a degree of transparency to the applied privacy processes [78]. One such framework is detailed in [84] for building data security into health information systems. While there is an inherent trade-off between the integrity of data and maintenance of privacy, a verification system for ensuring that both integrity and privacy are preserved in health data systems is detailed in [156]. As evidenced by the numerous attacks and security breaches commonly reported in the media, many malicious actors continue to develop a variety of adversarial attacks on sensitive data, posing ongoing threats to data security and privacy measures [51,74,76]. This requires practitioners to ensure data privacy and security measures are up to date, using state-of-the-art defence techniques such as DP [86,87] to prevent unnecessary data exposure and preserve data integrity [82].
Datasets should also be scrutinised for the possibility of 'poisoning attacks' that falsely skew the data. Practical methods for assessing this are detailed in [85] for quantitative datasets and in [87] for image-based datasets. If a deep neural network is planned, steps to protect classifiers from poisoning attacks are as detailed in [132].
'Sanitising' datasets (i.e. creating an identical dataset with identifying details purposely altered) has been shown not to guarantee individual privacy [157]. However, a novel 'DataSynthesizer' tool for creating synthetic databases ensures strong privacy features, and is further outlined in [78]. There are innovative hybrid models specifically designed to support healthcare big data protection (e.g. [84]), that support data security with techniques like masking encryption, activity monitoring, granular access control, and end point validation. Other methods involving k-anonymity and l-diversity have been shown to have numerous limitations in protecting the privacy of individuals; hence, a novel t-closeness requirement, further detailed in [72], is recommended instead. [113,114] 2. Train models using ethics-aware scenarios concerning fairness, bias/discrimination and social justice [115] 3. Decompose model bias into 'bias-variance-noise' to identify and separate sources of discrimination [64,68] 4. Add ethics-inspired regularization to penalize unfair decisions made by models under training [116,117] 11. Apply human-centered approaches and keep Explainability to (potentially impacted) humans in mind [118,119] 12. Keep humans in the loop to achieve ethical model training and desired ethical outcomes [120,121] 13. Adhere to interpretability based on regulatory and efficiency demands and users' needs [118,119,[122][123][124][125] 14. Apply hybrid approaches to enhance interpretability [126,127] 15. Apply causal reasoning to identify and understand and discrimination based on protected attributes [128] 16. Implement both mathematical and ethically grounded fairness definitions [67,129] 17. Make models and classifier safe from adversarial attacks using game theory and GANs [130][131][132][133] 18. Secure models through the use of adversary-aware learning algorithms [134][135][136] Data protection (and the protection process transparency) also entails clearly defining those allowed to access the data, with access levels layered for different agents (users, administrators, etc.) [93], and a log populated that shows who has accessed the data, when they accessed the data, and what actions they have taken [93]. In addition, data sharing practices should be transparently disclosed [18]; data should never be shared with private companies without explicit consent [153]. If data is to be shared between parties, one method to do so securely is detailed in [158]. An alternative approach is to share synthetic data that mimic original data with differential privacy; an open-source tool that be utilised to achieve this is further detailed in [73].
Data cleaning
When assessing a dataset, vulnerable groups should be explicitly identified and should each have sufficient completeness of data [159]. In some cases, simply avoiding the use of biased or imbalanced datasets (e.g. those used in Fitzpatrick Skin Type Classification System [45]) can help address potential social injustice resulting from the AI system developed. Methods to automatically identify the appropriate data cleaning activities for an incomplete or biased dataset are still emerging, but could be achieved with the use of frameworks such as MLClean (that removes data poisoning to help training accurate and fair models [100]) and tools such as like HoloClean, Activeclean [98,99] and Universal Cleanser [160]. A number of techniques that can be employed to prevent discrimination during data cleaning are outlined in [161]. If a classification algorithm is planned, specific pre-processing methods can be applied to prevent discrimination, including minimally intrusive data modifications that lead to an unbiased dataset [92,96] with additional techniques involving selective instances relabelling ('massaging'), suppression and reweighing [96,162]. If data from vulnerable groups remains inadequate despite a concerted effort, numerous techniques may be employed such as synthetically creating diverse datasets [54,55] using techniques such as convex optimization [95] to address the inadequacy of available data. missing data and training the models. Other techniques, including resampling (oversampling and under sampling [163]) and removing poisoned samples [101], can also be applied to tackle data imbalance [93][94][95].
Data reporting
All datasets should have accompanying documentation to enhance transparency and help ML engineers identify and understand issues in training data [104,154,164]. For ML, documentation should encompass the following areas: motivation, composition, collection process, pre-processing, uses, distribution, maintenance, impact and challenges [102]. For NLP, documentation should encompass the following areas: curation rationale, language variety, speaker demographic, annotator demographic, speech situation, text characteristics, and recording quality [110]. The full documentation processes are outlined in [45,164] for ML datasets, and in [110,111] for NLP data. Similarly, to engender trust in AI, systems can be accompanied by AI service FactSheets shared by service providers that declare the purpose, performance, safety, security of the AI service, i.e. something that can be examined and audited by AI service consumers and regulators [108]. One such FactSheet template is provided by [109], describing the diversity of contexts and circumstances in which AI systems are developed and deployed.
Model training
Before a model is trained, a pre-analysis plan should be written, clearly stipulating the goals of the model, technical approaches to be taken (e.g. the type of loss function), and the research question/s and outcome/s of interest [2]. Training and test datasets should be clearly separated, such that entries do not occur in both groups simultaneously [28].
Where feasible, ethical or value-based constraints should be set up as guiding criteria beyond the usual cost and efficiency optimisation focussed on during model development to ensure models are developed in a responsible manner [165]. Models can then be 'rewarded' and reinforced when outputs aligned with predefined ethical or regulatory constraints. A practical example of this process is outlined in [165]. This can also be achieved by 'raising' intelligent systems using storytelling that communicates tacit knowledge and cultural values, e.g. using (deep and inverse) reinforcement learning to 'punish' agents when their actions are misaligned with our values or otherwise undesirable to humans [113][114][115][116]. Model training can also be nudged towards ethics using training them on ethical datasets such as ETHICS and presenting the model with contrasting ethical and unethical scenarios to enhance its ability to analyse ethical expectations and behaviours [115]. Where feasible, predictor variables that are actionable should be preferentially selected over those that are not [41]. Chosen outcome variables should be clinically relevant [28], unbiased for marginalised groups [2], and could include 'soft' endpoints such as patient satisfaction [147]. Specific criteria that can be used to ensure the model is unbiased are specified in [123]. For classification algorithms that use stream-based data, a method to ensure algorithmic fairness is outlined in [166]. Although models that are less interpretable tend to have higher accuracy [123], interpretable models are generally preferred, as this aids transparency and trust [9,167].
Model training could be made more ethical by moving away from predominantly algorithm-centred approaches and towards human-centred approaches, e.g. keeping 'humans in the loop' and also by selecting models with the most probability of satisfying users' explainability goals [119]. The principle of explainability can be made more achievable by preferring inherently interpretable models to train for applications that make high-stakes decisions impacting humans instead of using black box models known to cause problems e.g. in healthcare, criminal justice systems and other domains [122]. Creating awareness about the level of explainability in ML models among engineers can go a long way to increase the chances of building explainable AI systems. ML engineers working on potentially explainable health care AI systems can benefit from a comprehensive classification of ML models, regarding their degree of explainability and other related techniques, as presented in [123]. These principles can be further operationalized to target explainable medicine by developing explainable interfaces that allowing experts to understand machine diagnosis/outcome and its influencing factors or causality [124,125]. For a particular project context or specific parameter, It may be possible to convert black box models to 'glass box' explainable models, as argued in [122]. The notion of fairness can be highly contextual, and therefore challenging to satisfy from both a group and individual perspective. For quantitative simplicity, it is the mathematical definitions of fairness that are predominantly implemented, e.g. similar percentages of false positives and/or false negatives for the different socioeconomic groups under consideration [67]. However, depending on the context, these mathematical implementations of fairness could be complemented by more ethically grounded ones [67,129] Models should also be trained to minimise the effect of adversarial attacks or of poisoned data samples. Evasion attacks can be protected against by devising systems composed of multiple classifiers, as outlined in [137], or other adversary-aware learning algorithms (e.g. one-and-a-halfclass multiple classifiers) as outlined in [135]. A 'hardness of evasion' score for measuring a model's sturdiness against adversarial attacks is detailed in [138]. Another aspect that protects against adversarial attacks is the 'resilience' of a model (i.e. perturbations in input causing minimal changes to output), which can be increased by smoothing ML decision boundaries [168]. Models may be trained only considering the subset of features that cannot be manipulated by an attacker, an approach further outlined in [136]. For training sturdy models using a Generative Adversarial Network approach, two models may be trained on similar datasets as outlined in [131]; alternatively, a 'MinMax game' technique for adversarial regularization may be taken, as outlined in [130]. After a model has been trained, poisoned samples in training data may be removed via data sanitization processes outlined in [101], or via 'machine unlearning' to remove the effects of poisoned data as outlined in [142]. To plan against an attack, the equilibrium between an attacker and the system can be identified using zero-sum game theory, as detailed in [134].
Model verification
The performance metrics of an AI algorithm should be carefully chosen depending on its purpose, including whether it is more important to have few false positives or few false negatives [28]. Importantly, the performance of an AI algorithm must not only be considered on an overall scale; the performance on marginalised groups should be specifically evaluated [2,41]. Methods to slice data and identify subsets of the data where the model has poor performance are detailed in [141].
Post-hoc classifier auditing can also be performed periodically to discover under-performing subgroups, thus iteratively improving model performance for subgroups of concern and improving fairness [143]. Practices to address fairness need to be accompanied by the team's realistic acknowledgment of the inherent challenges of fairness, including contextual trade-offs to balance fairness with accuracy. Some biases may even be correctible with posthoc techniques involving re-labelling, such as for racial bias [140]. Biases may be balanced with variance of performance estimates using out-of-sample bootstrap techniques [139]. Importantly, post-hoc model correction methods (e.g. severity scoring in clinical tasks based on randomizing predictions) must be carefully considered and should be justifiable from an ethical perspective [68]. While retraining models is possible and often necessary, other 'machine unlearning' techniques can also be applied to remove the effects of poisoned data without having to retrain the model [142].
Where possible, the AI algorithm should be tested across multiple healthcare systems, socioeconomic groups, or age ranges [28,169]; if the proposed scale of application is particularly large, proportionally greater scrutiny is warranted [140,170].
Model reporting
When reporting the final, verified model, the data source, participants, predictors and outcome variables should be clearly reported, as should the features of the model (e.g. regression coefficients) and the specific environment/s in which it has been verified [28]. Such comprehensive disclosure of model-related information significantly improves the transparency of the model and thus trust in the built technology. Clear documentation with 'model cards' is recommended for summarising a model's attributes and performance; this is further detailed in [144]. These cards clarify model's intended use cases, and disclose conditions, context and intended application domains as well as the scenarios that are suitable or unsuitable for trained ML models [144]. The stakeholders are likely to appreciate model-related documentation when it is aligned with their explainability needs. Identifying the target audience and the reasons why they need these explanations can boost their trust in AI-based systems, significantly more than handling black box models to them without any explanations of how these models come up with the decisions [123]. Besides enhancing transparency, documentation detailing benchmarked evaluation (e.g. across different cultural, demographic, or phenotypic groups) can encourage reuse for practitioners trying to implement these models in their context [144]. Transparency can also be promoted by ad-hoc and external post-hoc techniques that make models more interpretable, some of which are further explained in [123].
Stakeholder engagement and user-centered design
To ensure needs are being met, stakeholders should be consulted extensively throughout the AI lifecycle, and particularly during deployment [171]. Relevant stakeholders include knowledge experts (clinicians, ML researchers, health informaticians, implementations experts), decisionmakers (hospital administrators, institutional leadership, regulatory agencies, and government bodies), and users (clinicians, laboratory technicians, patients and their family members) [28,147]. To assist in thoughtful design, efforts should be made for the team developing and implementing AI to feature sociocultural diversity [159], including an appropriate representation of women [172], ensuring that the needs of these groups are not implicitly ignored by the AI initiative. Crucially, there should be a clearly identified clinical problem that is to be tackled [28], noting how ML outputs are not always actionable and hence may not necessarily help solve a problem [169]. A questionnaire for reviewing whether development aligns with data ethics principles is available at [173].
Moreover, there is currently a paucity of resources for clinicians to use AI [41]. Hence, a purposeful effort should be undertaken to appropriately train users, encompassing numerous dimensions as outlined in [174]: understanding how AI works; building patient trust; appraising evidence; assessing training data; and mitigating bias. The nature of AI in healthcare as one of shared responsibility should also be emphasised [175]. To increase understanding, outputs should be visualised where possible, making them easier for users to interpret [169]. To assist in transparency, users should be provided with sufficient meta-information about the model (including addressing the questions 'who', 'which', 'what', 'why', 'when' and 'where'), an approach further detailed in the explainable AI framework provided by [146].
When integrating AI into clinical practice, what the system can do (and how well it can do so) should be clearly specified to the user [51]. Information and alerts should be provided to the user in a context-specific manner (i.e. taking into account the user's current tasks and environment) and should be easy to dismiss if not relevant [51]. Clinicians should be able to disagree with AI-driven recommendations [41], and should be able to override them if they believe they have sound clinical reasons for doing so [147]. Additionally, the AI system should be built to accommodate the diverse needs of all its different users (e.g. doctors, nurses, and patients), and to integrate as smoothly as possible into the existing healthcare workflow [159,176].
The final AI model should be accessible, particularly for marginalised populations [18,147]. If model developers are charging a licensing fee to healthcare services, one method to promote accessibility is to reduce or waive fees for organisations working in disadvantaged settings [147]. AI should not be used for the purposes of allocating treatment, as most patients are opposed and believe those decisions require their collaborative input [153]. Patients should be informed about the use of AI, the risks, and expected shortcomings of its predictions [167]. Patients should be clear when they are communicating with a human and when they are communicating with an AI system [123]. The language of the AI system should not reinforce unfair stereotypes; a validated set of guidelines for human-AI interaction is available in [51]. Moreover, where possible, implementers and patients should be asked for their preferences, upholding their freedom of choice; choice architecture is further detailed in [147]. If possible, the AI system should alert a trained healthcare professional if the patient is exhibiting acute high-risk behaviours [30].
Updates and ongoing validation
The AI system should have a manual or automatic method to be updated over time [28]. Where possible, the AI should learn from users' behaviour and continually update its operation as a result of user interactions [51]. The AI system should also have a mechanism for users to provide feedback [169], and where unforeseen or unjust mistakes have occurred, a mechanism for adequate redress or reparations to take place [123]. To validate the AI system, performance metrics should be systematically and continuously evaluated after the AI has been deployed [159]. For formally evaluating real-life performance, a prospective clinical trial design may be considered [159], which should be conducted across diverse population groups [167].
Supervision and auditing
AI systems in healthcare bring new risks and amplifies existing ones, in large part due to their capacity for operating automatically and at scale [177]. Operationalising ethics in AI for healthcare necessitates an active governance model, one that implements policies, procedures, and standards to align practices with some overarching ethical principles [178]. Adherence to such governing principles ensures desired outcomes are delivered appropriately and with minimum risks [20]. However, on its own, a "principles based" approach is insufficient for ensuring the implementation of ethical principles [13,19,179,180]. As such, in ethical AI literature, an active governance approach is preferred [181]. Many AI governance approaches emphasise the necessity of regulatory influence to ensure compliance [181,182], the importance of human-centred governance to achieve shared goals [177], and even the possibility of going beyond the Westernised human rights-based approach to ethics [183].
Once a model of governance-such as that presented in [20]-is in place, the governance can be supported through mechanisms such as internal and external ethics-based auditing [184]. Internal audits can be facilitated by performing algorithmic-use risk analyses that are informed by social impact assessment of similar models. Ongoing internal audits help to avoid potential negative societal impact from the developed AI systems [148]. Organisations that design and deploy AI systems can apply ethics-based auditing through structured processes to assess adherence of their AI systems to ethical principles to take necessary corrective actions to address the identified gaps and enhance user satisfaction and trust in AI systems [150,184].
Of late, due to high-profile failures reported in the media, AI systems have not enjoyed positive public repute. To make amends and restore trust in these systems, a holistic and transparent approach is needed, one that requires necessary buy-in from an institutional perspective. The institutional approach should ensure clarity about the guiding organizational values used to develop the system, and may involve carrying out red teaming exercises to identify AI risks, conducting third party auditing to identify areas of concern, and communicating any known or potential incidents related to the system [151].
Ethical principles and actionable solutions
The ethical principles identified by [14] are preferentially addressed at different stages of the AI development pipeline. Since the distribution of ethical issues is not uniform across the AI lifecycle, we suggest that rather than considering a single ethical principle as done in some guidelines [17], developers should focus on the principles most pertinent to the specific stage of the AI lifecycle they are working on. Trade-offs between ethical principles should be carefully reasoned in this context and clearly documented [123], making use of frameworks such as that offered in [185] to evaluate ethical tensions. In addition to the AI lifecycle perspective, institutional and cultural values must also be proactive upheld for implemented changes to be meaningful and effective.
Improving organisational engagement amongst AI developers
At present, a key barrier and opportunity is that AI developers have only limited awareness of existing solutions that support ethical AI development, with organisational barriers persisting [186]. For example, developers may be unaware how to document their datasets [102] or to present their models in an explainable manner [146], despite these guidelines being available. As suggested by [186], these issues could be overcome by simple organisational mechanisms, rather than technical solutions. Such measures could include conducting ethics workshops; sending out newsletters with information on ethical AI developments; inviting external speakers to raise awareness; increasing peer networks; and following ethical AI news streams. If adequate resources were available, additional team roles such as a 'Responsible AI Champion' could also be created [186,187]. These organisational norms could promote collective responsibility and help overcome attitudes such as "it's not my job".
Another opportunity is to associate ethical values with an organisation's beliefs. Aligning AI ethics to organisational values could get more engagement from the employees and consumers, especially when those values are translatable into actionable and operational messages such as: 'we will anonymise your personal data and it is never going to be sold to any third party" [187].
Evolution rather than revolution of existing practices
Empirical research in software engineering has noted that human values can be embedded into software development by evolution, rather than revolution, of existing practices [186]. Hence, we believe that actionable solutions for addressing ethical issues are most likely to appeal to practitioners if they can be easily integrated into existing workflows. Examples of this include guidelines of reporting data [102,110] and reporting models [144] which provide an unambiguous, user-friendly structure for completing tasks. Such solutions are straightforward to implement in AI development through evolution of workflows, avoiding a need to adopt an idealistic and impractical revolutionary approach.
From a technical standpoint, since AI development is so closely related to software engineering [26], AI developers should draw from existing relevant practices in software engineering [188,189]. Exploratory processes in ML share many similarities to scientific programming, and would benefit from lessons learned in embracing uncertainty [190]. Moreover, software engineering has had decades of development in data management, which could be readily adapted to AI practices in many cases [190]. To be effective, AI for healthcare must also consider accessibility, user-centeredness and product design, aspects that can be readily informed by software engineering literature [51].
Limitations of AI governance and regulation
Governance and regulatory mechanisms may provide the necessary support to align AI development with ethical principles. However, organisational leaders may view the adoption of ethical procedures as costly, time-consuming, and counterproductive for holding a competitive edge. Another barrier is how governance and regulatory mechanisms often develop at a relatively slow pace, with regulations like the GDPR providing little support to small business to facilitate easy implementation [191]. Additionally, if auditors do not adequately understand ever-developing AI techniques, discriminatory features (such as image-based AI systems discriminating against darker skin shades) could slip by unnoticed.
Future directions
Although we have collated recommendations for addressing ethical issues at each stage of the AI lifecycle, a next step is to further expand and formalise these guidelines into categories linked to individual ethical principles such as transparency, fairness, and justice. This can enable convenient accessibility for developers to the set of practices required to embed a particular ethical principle into the system. We then plan to identify and link these guidelines to the AI development team roles that would action them, as well as to governance roles who would be responsible for monitoring and overseeing effective implementation of these guidelines. We also plan to experiment by containerizing and colorcoding guidelines and practices based on their relevance to individuals, teams, governance mechanisms, and leadership initiatives. This would enable the implementation of guidelines from both a technical and non-technical standpoint. Integration of guidelines into existing AI pipelines can take many shapes, but an approach supported by current literature is developing 'AI ethics checklists' [2,17,21,192].
We are beginning to see some attempts to utilize AI ethics checklists to address some ethical concerns, such as the checklist for fairness in AI by Madaio et al. [17]. Issues with current ethical AI checklists include a lack of detailed, technical, and actionable activities to embed ethical principles into AI development, and restricted scope dealing with only a particular ethical concern (e.g. fairness) [17]. In any case, such checklists are an expressed priority when ML practitioners are asked directly about their needs, and should therefore be co-designed with their close input [2,17,21,192].
A comprehensive ethical AI checklist that could address every aspect of AI development lifecycle, would need to be developed and comprehensively validated through active involvement of other stakeholders [193]. This approach would allow raising of necessary flags, and notifying relevant stakeholders when and where ethical concerns are raised, which could be linked to existing solutions that we have identified in this work. Although ethical come with their own limitations (such as the risk of people delegating their thinking to checklists alone) [17], they provide a promising next step for further operationalising ethics in AI for healthcare, one that should be taken in conjunction with advancements in governance and regulation.
Conclusion
The exponential development of AI in healthcare is promising, but a naïve application of AI to healthcare may lead to a wide array of ethical issues, resulting in avoidable risks and harms. As such, there is a growing need for ensuring AI for healthcare is developed and implemented in an ethical manner. In this paper, we recognise and go beyond solutions that offer principle-based guidance (e.g. adherence to 'fairness'), adopting a solution-based framework that AI developers can use to operationalise ethics in AI for healthcare across all stages of the AI lifecycle-data management, model development, and deployment and monitoring. We emphasise solutions that are actionable (whether technical or non-technical), and therefore utilizable by AI developers. Finally, we acknowledge the growing need for 'ethical AI checklists' co-designed with health AI practitioners, which could further operationalize existing solutions into the AI lifecycle.
Author contributions PS: manuscript writeup and editing; JG: manuscript review and editing; WH: methodology, framework conceptualization, literature review, supervision, funding acquisition, manuscript review and editing.
Funding Open Access funding enabled and organized by CAUL and its Member Institutions. This work is supported by a grant from HumaniSE Lab, Monash University.
Data availability Data sharing not applicable to this article as no datasets were generated or analysed during the current study.
Conflict of interest
The authors have no relevant financial or non-financial interests to disclose. On behalf of all authors, the corresponding author states that there is no conflict of interest.
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:// creat iveco mmons. org/ licen ses/ by/4. 0/. | 12,040 | 2022-07-19T00:00:00.000 | [
"Computer Science",
"Medicine",
"Philosophy"
] |
Article Analysis of Resource and Emission Impacts: An Emergy-Based Multiple Spatial Scale Framework for
Abstract: The development of the complex and multi-dimensional urban socio-economic system creates impacts on natural capital and human capital, which range from a local to a global scale. An emergy-based multiple spatial scale analysis framework and a rigorous accounting method that can quantify the values of human-made and natural capital losses were proposed in this study. With the intent of comparing the trajectory of Beijing over time, the characteristics of the interface between different scales are considered to explain the resource trade and the impacts of emissions. In addition, our improved determination of emergy analysis and acceptable management options that are in agreement with Beijing’s overall sustainability strategy were examined. The results showed that Beijing’s economy was closely correlated with the consumption of nonrenewable resources and exerted rising pressure on the environment. Of the total emergy use by the economic system, the imported nonrenewable resources from other provinces contribute the most, and the multi-scale environmental impacts of waterborne and airborne pollution continued to increase from 1999 to 2006. Given the inputs structure, Beijing was chiefly making greater profits by shifting resources from other provinces in China and transferring the emissions outside. The results of our study should enable urban policy planners to better understand the multi-scale policy planning and development design of an urban ecological economic system.
Introduction
Due to the complex and multi-dimensional urban socio-economic system, the knowledge of the organizational structure, urban energy and material inward-outward flows [1], capturing the trade-offs between natural, economic and social capital [2] is a major step towards the design of sustainable development schemes.Meanwhile, as governments are increasingly concerned about international negotiations, cooperation, and conflicts on climate change issues, the traditional environmental problems in cities such as waste water removal, sanitation, water supply, indoor and exterior air pollution, etc. have been proven to have a cross-regional impact [3].Analysis of individual urban processes is not enough for understanding the inherent functional principles and evaluating their environmental performance since such a narrow view may indistinguishably consider international and regional trading [4] or simply shift the environmental impacts to the other parts of local economic activities life cycle [5].Challenges for urban development from a sustainable perspective have been divided into two categories: (1) modes of resource supply; (2) the activity boundary that outlines the emissions emitting operations for which a city is responsible and that must be accounted for in the city's mass/energy balance.As a consequence, urban socio-economic performance metrics must be capable of linking local scales and extend further to the economy and ecosystems scales [6].With the international trading network taken into account, fruitful studies focusing on specific countries have been presented [7][8][9][10][11][12][13].There is an urgent need to develop a quantitative methodology that can evaluate both the resource supply and the adverse environmental effects of urban socio-economic systems at different scales and take into account how they affect the urban system's dynamics and sustainability.
Methods relying on input-side information have also been developed, usually based on mass [14], energy [15], exergy [16], emergy [17][18][19][20] and ecological cumulative exergy consumption [21].Emergy synthesis is a method of environmental accounting derived from energy system theory that uses the energy (in units of the same kind) required to produce a good or service as a nonmonetary measure of the value or worth of components or processes within ecosystems and the economy [22].The Emergy synthesis method transcends systems' analysis boundaries, considers resource inputs and environmental contributions, and constructs the basic emergy-based index system [23].Until now, a large number of systems have been evaluated by means of the emergy method on regional and national scales [19,20,[24][25][26][27][28][29][30][31].Most of these studies, however, did not focus on the multiscale analysis of resource supply and emissions impact, although important steps ahead have been taken in that direction.Chen and Shonnard [32] presented a hierarchical approach for environmentally conscious chemical process design based on the Analytic Hierarchy method.Brown and Ulgiati [33] applied the emergy method to suggest a system view of ecosystem integrity and also to assess the emergy investment needed to restore ecosystem health.Bakshi and his colleagues proposed a multiscale statistical framework for life cycle inventory analysis in some case studies of the U.S. economy and the CGAM cogeneration system [5,[34][35][36].Four kinds of spatial hierarchy structures were defined, including economy, life cycle, equipment and hybrid scale, yet the majority of these works have just developed a conceptual framework rather than specify a detailed list of the different emission categories, especially the emission impacts on different scales.As a smaller control unit, the city nests in a nation's economic system, which is different from the doubly-nested world economic system.A city's expanding resources consumption requires its neighbors to expend (considering the factor of state socioeconomic regulation) and purchase from abroad; meanwhile, a city should be held accountable for its "external" emissions.Here, emergy algebra was used to quantify the values of human-made and natural capital losses which were considered as indirect inputs of ecological services for airborne and waterborne pollutants dilution and damage repair or replacement to "internalize" the "externalities" with emphasis on a joint application of the emergy synthesis and LCA methods.The results obtained are potentially useful in understanding the supply networks of a city belonging to different hierarchical levels of the economy and would enable decision-makers to target emission policy measures by purchasing third-party offsets.
This paper proposes an emergy-based multiple spatial scale analysis framework and a rigorous accounting method that can quantify the values of human-made and natural capital losses.With the intent of comparing the trajectory of Beijing, the characteristics of the interface between different scales are considered for explaining the resource trade and the impacts of emissions.In addition, we examine our improved determination of emergy analysis and acceptable management options that are in agreement with Beijing's overall sustainability strategy.
As a follow-up work of our earlier effort to make an assessment evaluating the environment and economic development in Beijing's socio-economic system on a common base [29,30,37,38], this work serves as a further attempt to assess both the energy resource consumption and the adverse environmental effects in a unitary manner from a cross-regional perspective based on emergy analysis.
Emergy-Based Multiple Spatial Scale Analysis Model
Emergy is formally defined as all the available energy of one kind previously used up directly and indirectly to make a product or service [39,40].As a thermodynamic-based environmental accounting approach, the emergy synthesis converts all materials, energy sources, human labor and services required directly and indirectly into emergy unit that are summed up to yield the total emergy [22].Emergy analyses are carried out using transformities, specific emergies and other factors that are determined according to a particular planetary baseline [22,41], which is decided by the solar equivalences of the three primary energy inputs to the biogeosphere, i.e., solar radiation, residual and deep heat of the Earth, and the gravitational attraction of the sun and moon.In this study, transformities were converted from global emergy baseline of 9.44 × 10 24 to 15.83 × 10 24 seJ/yr recommended by Brown and Ulgiati [41].
A typical diagram describing an urban system is shown in Figure 1, where the standard energy system symbols are used [22].At the planetary level of organization, there are no substantial exchanges with the larger system, except for solar and gravitational energy entering the system from external sources.Within the large box, they indicate the spatial boundaries of the urban system, renewable emergy (R), i.e., the rain, wind, tides, waves, etc., Nature also does work that indirectly supports the activities of the world socioeconomic system (e.g., the photosynthesis of natural ecosystems that fixes carbon and replenishes oxygen in the atmosphere, which is necessary for all life, the movement of clean air that replaces contaminated air over cities and water flows that supply the capacity to dilute municipal wastes).The emergy provided by fuels and electricity is modeled on a separate pathway that acts on the material products arranging and ordering them.Humans extract and process the slowly renewed material products of natural work in the environmental system, i.e., fossil fuels and minerals.These inflows are considered to be nonrenewable because they are being used by socioeconomic system at a rate that is much greater than their natural renewal rate.Human work is also used to carry out economic production using these raw materials and to carry out the other processes and functions of society.The multiple spatial scales between the environment and human socioeconomic systems can most easily be understood by modeling the world system as a whole (see Figures 1 and 2).Emergy and associated systems language also provides a tool for illustrating energy and material flows between regions, as well as controlling (feedback) mechanisms.In other words, it becomes possible to visualize the flows and interactions referred to above, and simplify them through aggregation, thereby enabling the human mind to conceptualize small and large parts simultaneously, and hence see the bigger picture more clearly.By drawing on the results in this thesis, and the example above, Figure 1 therefore illustrates the urban system's position and role in both the national and world system, depicted in emergy systems language.However, even if resources or pollutions embodied in trade are fully understood and international responsibilities are reallocated on a consumption basis, it is far from enough to promote global cooperation to combat economic loss and ecological impacts.It would be great progress, though, not to mention that net pollution importers may not accept consumption-based methods.Figure 2 represents the waste released and its interaction with the urban system itself.Air and water emissions and solid waste are controlled based on additional input of fuels, goods and labor force.Thus, simply trying to seek a single global solution that is implemented by national governmental units because of global impacts is far from satisfactory.The essential role of smaller-scale effects must be recognized.In this sense, a polycentric approach might be an alternative for the problem, which means actions at various levels with active oversight of urban, regional, and national boundaries.
As we live in a highly globalized world, economies of scale and comparative advantages exist in certain areas, rendering trade and commerce highly valuable and emissions "ownership" more complex.The processes described in Figures 2 are similar in many ways, but have one major difference; on a more aggregate scale, not only is control fed back, but waste is also generated by the world system.Put together, this means that the city draws on local environmental and human resources, together with non-renewable energies from other peripheries, hence facilitating the process of accumulation in the core, only to risk exhausting the local, national and global resource base, and building up stored waste.With this current voluntary set-up and focus solely on emissions produced in each country, and in the globalized world in which we live, a perverse incentive exists for industrialized countries to transfer high emitting activities to the developing world.
Emergy Algebra
Emergy algebra comprises two parts: (1) resources and energy inputs and (2) emission impacts.The emergy embodied into an imported product is made up of two parts, one is from geobiosphere work and the other one is from services needed for its production during previous manufacturing steps [23].
Here, a monetary measure is used to account the indirect labor embodied in the production and delivery of imported goods.
In the waste side, Ulgiati et al. [42] focused on the emergy resources required in order to prevent or fix reversible damages.Moreover, they pointed out that: (1) additional emergy resources are needed to replace the lost assets or units, when irreversible damages occur, and that (2) when replacement is not possible, at least a conservative estimate of the natural or human capital loss should be attempted, based on the resources previously invested for its generation, in order to ascertain the true cost of a process product.Following Ulgiati et al. [42] and Ulgiati and Brown [43], additional emergy cost terms should be included in order to account for: (a) dilution and abatement of emissions by natural processes, (b) abatement, uptake and recycle of emissions by means of technological devices, (c) repair of damages to human-made assets by means of maintenance activities, (d) reversible and irreversible damages to natural capital (e.g., loss of biodiversity), and finally (e) reversible and irreversible damages to human health.As a consequence, the total emergy cost U (here, U = used) can be calculated as: where R and N are respectively the locally renewable and nonrenewable emergy resources, F is the emergy of imported goods and commodities (including their associated services) and where the F i terms include the environmental or human-driven emergy investments (here, F = feedback) needed to prevent or fix the damages occurred and charged to the process: In this study, a preliminary damage assessment of losses is performed according to the framework of the Eco-Indicator 99 assessment method [44] as well as the authors' own preliminary work [30,31].Such a method, like all end-point life cycle impact assessment methods, suffers from very large uncertainties intrinsically embodied in its procedure for assessment of final impacts.Damages to natural capital are expressed as the Potentially Disappeared Fraction (PDF) of species in the affected ecosystem, while damages to human health are expressed as Disability Adjusted Life Years (DALY), according to references [44][45][46].The impact of emissions on human health can be viewed as an additional indirect demand for resource investments.Human resources (considering all their complexity: life quality, education, know-how, culture, social values and structures, hierarchical roles, etc.) can be considered as a local slowly renewable storage that is irreversibly lost due to the polluting production and use processes.The emergy loss can be calculated as: where, L w,1 * is the emergy loss in support of the human resource affected, i refers to the i-th pollutant, m * is the mass of chemicals released, DALY is its E.I. 99 impact factor and τ H is the unit emergy allocated to the human resource per year, calculated as τ H = total annual emergy/population.
The effect of Potentially Disappeared Fraction of Species (PDF) can be quantified as the emergy of the loss of local ecological resources, under the same rationale discussed above for the human resource: where, L w,2 * is the emergy equivalent of the impact of a given emission on urban natural resource, PDF(%) is the fraction potentially affected, measured as PDF × m 2 × yr × kg í1 .Finally, damage associated with solid waste generation can be measured by land occupation for landfill and disposal.This may be converted to emergy via the emergy/area ratio (upper bound, average emergy density of economic activities) or even via the emergy intensity of soil formation (lower bound, average environmental intensity).Thus the related emergy loss (L w,3 ) can be obtained using the total occupied land area multiplied by the economic or environmental emergy intensity of such an area (choice depends on the area of the investigated system).
Case Study
Beijing (N115°25-117°30´, E39°26´-41°03´) lies at the eastern edge of the Eurasian continent and belongs to the Bohai sea rim economic circle, with small plains in the south and mountains in the west and north, covering an area of 16,807.8km 2 .Characterized by its long history and central political and cultural position, Beijing is amongst the most developed cities in China with a fully integrated industrial structure, including electronics, machinery, chemicals, light industry, textile and automobile manufacturing.Like other metropolis in developing countries, Beijing faces the dilemma of urban economic development versus social and ecological problems comprising the large floating population, high-yield agricultural land loss, resource shortages, high levels of pollution, ecological deterioration, and increasing risks of disaster.The evolution of the Beijing urban system can be treated as a history of resource consumption and accumulation, which has, in turn, brought about the changes in the urban structure and organization.As mentioned above, most of these intensive resources consumed in Beijing are purchased from outside with the exception of a small proportion of the fuels and minerals.Also, all the flows of resources are accompanied with human services and money flows.
The reason for choosing Beijing as the primary study site for this research is that Beijing is a major node linking China and the world and the nation and its provinces.The strong nationwide support for the 2008 Olympic Games in Beijing-hosted at the expense of investments elsewhere in the countryoffers a typical example of the Chinese desire for global recognition.The successful hosting of this globally significant event is seen as firmly demonstrating China's winning a central position on the world stage.This interpretation is possible because Beijing as the capital city represents China to the outside world.This paper focuses on how Beijing-in terms of both its physical reality and how it is imagined-mediates the interesting dynamics among various local, national, and global processes.
Main Emergy Flows in Beijing
Figure 3 shows the inputs and internal structures of Beijing that were quantified in this study.We evaluated the emergy inputs supporting economic activities of urban system and compared the emergy inflows to measures of economic activity in Beijing.The following major classes of emergy inputs supporting Beijing from 1999 to 2006 were documented: (1) renewable energy sources, (2) soil erosion, (3) energy consumption, (4) minerals consumed, (5) imported goods other than fuels and minerals, (6) imported services in goods, fuels and minerals, (7) imported services, and (8) immigrants.Out-of-scale
The Determination of Pollutants
Our study deals with the emissions harmful to human health and the ecosystem listed in Table 1.Air emission discharges from both urban production and use include SO 2 , dust, NO x and CH 4 (respiratory disorders), CO 2 , N 2 O and CH 4 (climate change).Eight waterborne pollutants (mercury, cadmium, hexavalent chromium, lead, arsenic, volatile phenol, cyanide, oil) were selected as indicated.The loss fractions of human health and ecosystem quality (DALY/kg of emission, PDF × m 2 × yr) are collected in the reference [44].The emission data related to SO 2 , dust, and NO x were collected from governmental publications, such as the Beijing Statistical Yearbook and the Chinese Environmental Statistical Yearbook [47,48].Data about CO 2 , N 2 O and CH 4 are calculated as greenhouse gases released at local and global scales, based on direct and indirect energy consumption, that in turn are evaluated according to the Embodied Energy Analysis method [49,50].The embodied energy of materials and energy flows is calculated by multiplying local inputs by appropriate Oil Equivalent Factors.
Emergy Accounting of Beijing Socio-Economic System
Examination of various aspects of the Beijing economy includes a discussion of Beijing's emergy resources, emergy consumption patterns, emergy conservation, and emergy yields.The results in 2006 are shown in the Appendix.In accordance with the system picture of Beijing (see Figure 3) and the consequent calculations shown in the Appendix, main flows introduced to the Beijing urban socio-economic system for the studied years are summarized in Table 2.
Emergy Inflows in Beijing Socio-Economic System
Since 1999, Beijing as the capital of China has adhered to the policy of reform and opening-up, and focused on economic construction.Gradually, it has stepped onto the road of establishing a market-oriented economy system.As a result, the consumption of energy, material and labors increased correspondingly.Total emergy actually used (U), as potential investment in emergy yield of the city, increases with an annual average of 19.88% with a peak in 2004 (25.11%).
As the primary impetus for the economy, environmental free renewable resources (R) involving sunlight, rain, wind, and geothermal heat remains approximately unchanged at this temporal scale (Figure 4).For the Beijing economy, the specific flow of the geothermal heat with emergy is much more than that from the sunlight, wind and rain.It is worth noticing that, of all the renewable inputs, only the largest item, rain, is taken into account though all the emergy inputs are estimated to avoid double-accounting, see Appendix.Participation of non-renewable emergy flows from urban local sources (N) fluctuated in this period so that the obvious fluctuation in constructed local input includes limestone, sand and gravel and iron ore.Construction materials are the largest individual N flows, which is much more than the natural topsoil losses for plant growing and from the degraded soil erosion.The Appendix also lists main imported inputs in terms of emergy flows for 2006 in Beijing.The total imports increased from 1.51 × 10 23 to 3.49 × 10 23 seJ/yr.Of the total imported resources, fuels grew by 1.52-fold, with emergy rising from 8.84 × 10 22 to 1.35 × 10 23 seJ/yr, while the total imported building materials (including iron ores, sand and gravel, iron and steel) increased by 3.92-fold from 4.32 × 10 22 to 1.70 × 10 23 seJ/yr.This indicates that Beijing's economic development is increasingly dependent on the infrastructure construction, which has even replaced the fuel-consuming industry for nearly a decade.In emergy to money terms, imported goods have become the most important item in this category.The export emergy flows could be highlighted petroleum derived products, minerals and mechanical and transport equipment.As shown in Figure 4, the service associated with imports in Beijing was a total of 2.04 × 10 23 seJ/yr in 2006, 4.04 times more than that in 1999.This increase in import results in decreased self-sufficiency, so the purchased component of the total economy was more important, supporting the growth of the economy.The services imported from other provinces were 7.6 times more than that from abroad, indicating that the imports of Beijing were still increasingly dependent upon the transmission of domestic market.And it's worth mentioning that, in emergy to money terms, the tourism and emergy paid for imported labor are increasing strongly, more than 2.58 and 5.52 times respectively.
The Components of the Energy Consumed in Beijing
The Appendix and Table 2 give detailed information on the consumption of energy by source for Beijing from 1999 to 2006.During this period, coal was still the most primary energy source for the region, as measured by both heat content and emergy.The coal input decrease from 2000 is evident in the figure, but this was followed by a rapid rebound over the next three years when Beijing won the bid to stage the 2008 Olympiad.Meanwhile, the consumption of petroleum from other provinces steadily decreased; however, the imported oil increases along with coal consumption in these ten years.Imported electricity is a large fraction of the total emergy use (11% to 13%) that grew fast from 1999 to 2006.During the decade, exported electricity in these years remained low.Natural gas also became the fourth largest energy source.The consumption of natural gas showed a similar trend to that of imported electricity, but with a damped response to fluctuations.
Mineral Use in Beijing
The emergy of iron and steel made the largest contribution to the emergy of minerals consumed followed by the emergy of lead up to 1999, when it was overtaken by the emergy of sand and gravel for construction.Compared with sand and gravel, hi-tech products, machinery and electrical equipment increased greatly from 1999 to 2006 and it consistently occupied the position of the 3rd largest emergy input after labor and services.
The Human Capital and Natural Capital Losses
From 1999 to 2006, the total human capital losses caused by the six air pollutants increased dramatically from 4.17 × 10 20 to 1.15 × 10 21 seJ/yr and reached a maximum peak of 1.31 × 10 21 seJ/yr in 2005 (Table 3), while losses due to the urban production sectors fluctuated with a maximum at 1.70 × 10 21 seJ/yr in 2005, as shown in Table 3.The natural capital losses showed that such losses, different from human capital losses, were assessed on the basis of acidification and ecotoxicologic emissions.The loss due to NO x shows a very large increase in the investigated period, especially after 2004.Results seem to suggest that NO 2 has overtaken SO 2 as the ever-bigger issue in Beijing's environmental pollution treatment during 1999-2006.The growth rate of damage that is caused by the emissions from urban consumption processes climbs up faster.Nitrogen dioxide and sulfur dioxide provided the largest contribution to natural capital loss while the greenhouse gases (CO 2 ) and dust play the larger role in human capital loss.Note: L w,1 * is emergy of the human life losses caused by the emissions; L w,2 * is emergy of the ecological losses due to the emissions; L w,3 is emergy of the land occupation caused by the emissions.
Analysis of the Emergy Indicators for Beijing
In this section, a series of emergy indicators based on the emergy accounting for Beijing economy are analyzed, discussed and compared with those of other Chinese cities.These indicators lend insight to the emergy support basis, the economic structure and the characters of the Beijing economy.
Emergy Intensity
Empower density or the emergy flow per unit area is a related measure that indicates the spatial concentration of economic activity or the intensity of development in a city.As shown in Table 4, the empower density of the Beijing economy developed from 1.85 × 10 13 seJ/m 2 in 1999 to 4.59 × 10 13 seJ/m 2 in 2006, revealing that Beijing maintained a rapid economic growth and scored a new high in economic aggregates during the past years.Accounting results shows that this growth was mainly caused by the input from goods and services which hold relatively high emergy transformity.Combined with the emergy use structure and the value of emergy use per person in Beijing, we find that of the total resource consumed in Beijing, most is correlated with goods and services purchased from outside, with little from free natural inputs.It also means that the development both in the living standard of local residents and in urban economy depends completely on the purchase of resources from outside.
Import/Export Structure
For an urban ecosystem, the emergy welfare enjoyed by its residents also can be revealed by comparing the resource imports and exports, which are accounted through two ratios here, one is the difference between exports and imports; the other is exports to imports.From 1999 to 2006, the exports/imports ratio of Beijing is less than 1.During this period, the rapid development of Beijing industry brought about a quick need for energy, which made the fuel consumption increase from 1.01 × 10 23 seJ in 1999 to 1.90 × 10 23 seJ in 2006.Most of these fuels are used by industry, construction and transportation.In this period, the imported emergy was much more than the exported emergy with the largest difference appearing in 2005 and 2006.
Environmental Sustainability Index
This index is an aggregate measure of the economic benefit (EYR) per unit of environmental loading (ELR).It shows that the long-term capacity of the renewable emergy sources to support life is being degraded.A quick estimate of the renewable carrying capacity of a state at the current standard of living is obtained by multiplying the fraction of use that is renewable by the present population of the state [22].As a consequence of EYR and ELR trends, the sustainability index ESI dropped significantly, thus suggesting that emissions greatly reduced the sustainability of the urban socio-economic system by pulling resources for damage repair and for replacement of lost natural and human-made capital.
Emergy-Based Multiple Spatial Scale Analysis of Beijing Socio-Economic System
We modeled the interaction of economic activities with environmental resources as a production function situated at the interface between the environment and society in cross-hierarchy perspective.Minerals, fuels, natural products as timber, water and human labor, are all required emergy inputs from different regions that support economic activities.Meanwhile, the conceptual model also showed that the environment provides emergy to support society as a whole and the emissions have the direct and indirect impacts on the economic production function on all urban, country and global scales.The multiple spatial scale analysis separates inflow and outflow, which are ordered into societal assets and structures.Indeed, when materials and emergy were placed on an equal basis using emergy, we found that materials (steel and iron, et al.) from local and other provinces in China were the dominant input to Beijing's economy during the period considered.Thus, it is plausible that the total emergy used by a prosperous city, which takes into account the materials as well as the energy used, might explain more of the variance of economic activity than energy alone.If mass alone was used as the measure of relative importance, the construction materials, sand and gravel and crushed stone would dominate the signature of inputs to the Beijing's economy by an order of magnitude.However, when specific emergies are used to transform the mass inputs to emergy, the picture of the relative importance of materials is very different.Construction materials are still the largest input at times, but other unexpected materials show their strategic importance when converted to emergy.Except for iron ore and steel, from the 1999 to the 2006, purchased/traded commodities including hi-tech products, machinery and electrical equipments from other provinces in China was the third largest contributor of emergy to the urban system.But the growth of traded rate from other countries is much slower than that from other provinces.As Figure 5 shows, emergy flows from the national system are 52.7 times higher than those from the world system in 2006.Also the emergy values are much more than that returned via monetary transactions.This way of illustrating net emergy flows is also possible to be applied to regional relations in general.From another point of view, the multi-scale environmental impacts of waterborne and airborne pollutions continue increasing from 1999 to 2006 and the values cannot be neglected.As illustrated in Figure 5, denser flows of emergy move from left to right than what is fed back via monetary transactions, hence resulting in a net accumulation in the global system.In addition, this unfair exchange is self-reinforcing due to neo-colonial controls.Therefore, the distribution of resources and wealth can never be fully understood without explicit emphasis on power relations and fairness.In this figure, control resulting from such power relations is represented by the pathway line moving from right to left.So as long as there is trade, a part of the regions will specialize in resourceintensive production.In that case, in a prosperous city, such as Beijing, the urban consumption demand will shift to other regions and increase their production, which includes the shift of environmental impacts.If resources intensity of these regions is lower than others, then global emissions will decrease.On the contrary, other regions' resource use reduction will even increase global emissions.In this sense, a more systematic consumption-based approach, which can eliminate pollution leakage and encourage reductions to occur where the costs are the lowest, will be more favorable.Thus, given the structure of input, Beijing was chiefly making greater profits by shifting resources from other provinces in China and transferring the emissions outside.
Conclusions
Increasing interest among the general public in reducing environmental impacts has fueled the aspiration of eco-cities to achieve healthy status.This study contributes to the current body of relevant literature by exploring Beijing's organizational structure of the socio-economic system.Our research focused on the characteristics of the interface in Beijing between different scales based on an emergy synthesis approach, in order to highlight the resource trade and environmental impact and separate the inwards/outwards flow, economic and ecological losses between different scales.Clarifications regarding the emissions scope and system boundaries are essential in order to develop a strategy for monitoring the intimate relationship between the resource base and economic structure.Detailed trends of the resource base and performance indicators are examined from a historical perspective for the contemporary Beijing urban system after China's Economic Reform and Opening Policies in the latest decade.
The results demonstrated that the development of economy in Beijing was closely correlated with the consumption of the nonrenewable resources and exerts rising pressure on the environment.Of the total emergy use by the economic system, the imported nonrenewable resources from other provinces contribute the most, with increasing use of imported nonrenewable resources.The multi-scale environmental impacts of waterborne and airborne pollutions continue to increase from 1999 to 2006.
Considering the structure of inputs, Beijing was chiefly making greater profits by shifting resources from other provinces in China and transferring the emissions outside.Our findings suggested several policy implications, assuming the goal is to achieve urban ecosystem health.In this regard, a regional emission trade scheme (ETS) should be enacted in the Chinese market system, via which the eastern developed regions (such as Beijing) could buy emissions directly from the west area via local governmental transactions.Via such an ETS, emissions in the eastern area can be reduced; furthermore the less developed area also benefit from such an emissions trading system.These policies may seem harsh.Curtailing development, trade or limiting access to resources may end up neo-colonial controls.Our message is not to abandon all development, stop all trade and limit resource access, but rather to highlight the limitations and consequences that might result from not taking such actions.
It is in this sense that we see this article as a step along a research path rather than its final statement.Going beyond a conceptual framework, this study specified a detailed list of emissions categories and regarded them as indirect inputs of ecological services for airborne and waterborne pollutants dilution and damage repair or replacement to "internalize" the "externalities" of different scales with emphasis on a joint application of the emergy synthesis.It is expected that such a detailed listing will evolve over time as formal standards are proposed and revised, and as the technical capacity to increase the pollution types and local impacts continues to improve.Calculations:
Figure 1 .
Figure 1.The urban socio-economic system position and role in the national and world system, described in terms of net emergy flows and trade (without considering transport of pollutants).
Figure 2 .
Figure 2. The urban socio-economic system position and role in the national and world system, described in terms of net emergy flows and trade (considering transport of pollutants).
Figure 3 .
Figure 3. Summary flow diagram for the main emergy flows in Beijing.
Figure 4 .
Figure 4. Temporal variations of emergy inflows in Beijing socio-economic system during concerned period of 1999-2006.
Figure 5 .
Figure 5. Diagrams of emergy cross-hierarchy inflows in different years.Note: R is renewables; NS is local natural system; NR L is local Non-renewables; NR C is Non-renewables from other provinces; NR W is Non-renewables from other countries; W is waste emissions.
Table 1 .
Lists of emissions and environmental impacts.
Table 2 .
Comparison of main emergy indexes and flows for time series emergy synthesis of Beijing.
Note: U = R + N | 7,987 | 2011-03-23T00:00:00.000 | [
"Economics"
] |
High-Dynamic-Range Integrated NV Magnetometers
High-dynamic-range integrated magnetometers demonstrate extensive potential applications in fields involving complex and changing magnetic fields. Among them, Diamond Nitrogen Vacancy Color Core Magnetometer has outstanding performance in wide-range and high-precision magnetic field measurement based on its inherent high spatial resolution, high sensitivity and other characteristics. Therefore, an innovative frequency-tracking scheme is proposed in this study, which continuously monitors the resonant frequency shift of the NV color center induced by a time-varying magnetic field and feeds it back to the microwave source. This scheme successfully expands the dynamic range to 6.4 mT, approximately 34 times the intrinsic dynamic range of the diamond nitrogen-vacancy (NV) center. Additionally, it achieves efficient detection of rapidly changing magnetic field signals at a rate of 0.038 T/s.
Introduction
At present, magnetic field measurements have significant applications in various fields.Therefore, rapid and accurate monitoring of magnetic field changes in different environments is particularly important.NV centers are correlated with measurements of external magnetic fields through the Zeeman splitting of degenerate electronic states within their energy structure.The measurement of magnetic field magnitude and direction is achieved based on the extent of splitting observed in the energy structure of the NV center's electronic states [1][2][3][4][5][6][7][8].Moreover, its inherent long spin coherence time [9][10][11][12][13][14][15] and high sensitivity to magnetic fields allow it to simultaneously meet the requirements of high sensitivity and high spatial resolution [16][17][18] for magnetic field detection, ensures the stability and accuracy of magnetic field change measurements.Compared to traditional magnetic detection methods, magnetometers based on NV centers can significantly improve the signal-to-noise ratio, thus achieving better magnetic measurement sensitivity.Furthermore, optical means of determining magnetic resonance enable the capture of even minute signal variations and facilitate real-time data acquisition and processing.This enhances the reliability of real-time monitoring of subtle changes in magnetic field signals [19][20][21][22][23][24][25].Building upon the inherent optical readability of diamond NV centers, a diamond NV magnetometer can be designed for rapid and precise measurement of external magnetic fields.
In recent years, there has been a significant focus on advancing the integration of NV magnetometers while maintaining their high sensitivity.This trend underscores a primary research direction in the field.Kapildeb Ambal et al. designed a differential photometry system for optically detected magnetic resonance under low photon-counting rates.The noise figure of this design reaches 4.1 µT/Hz 1/2 , allowing the continuous measurement of magnetic fields and tracking magnetic field scan rates of up to 50 µT/s [26].However, the system still relies on bulky desktop equipment, making it difficult to meet the demands of modern portable magnetometers.To achieve the miniaturization of conventional quantum sensing platforms, Wang Xuemin et al. proposed a diamond NV magnetometer integrated with a 520 nm laser diode.The probe volume of this magnetometer was 4 × 4 × 3 cm 3 , with a magnetic field standard deviation of 75 nT and a bandwidth of 25 Hz, and this achieved a magnetic measurement noise figure of 20.77 nT/Hz 1/2 .By minimizing the volume of the light source and complex optical paths, they integrated it into the NV magnetometer system [27].However, the dynamic range of this integrated magnetometer is only −214 µT to +214 µT.Subsequently, the United States Naval Academy utilized NV center vector magnetometers to calibrate geomagnetic anomalies, achieving a magnetic measurement noise figure of up to 200 pT/Hz 1/2 [28].However, the system also faced limitations in its dynamic range, effective only within a range of ±120 µT.When the magnetic field intensity exceeded the intrinsic dynamic range of NV resonance, the system would be unable to detect magnetic field changes in real-time, thus limiting its detection range in practical applications.To overcome this challenge, Wang Cao et al. employed a rapid-frequency-hopping technology to detect rapidly changing magnetic signals at 0.723 T/s [29].Their designed NV center magnetometer extended the inherent dynamic range to 4.3 mT while maintaining a noise figure of 4.2 nT/Hz 1/2 .This achievement provides important insights for the development of integrated magnetometers with high dynamic range.Therefore, to further expand the application scope of NV magnetometers, the development of magnetometers with high dynamic range, high sensitivity, and good integration has become a key research task.
This study proposes a frequency-tracking scheme for external time-varying magnetic fields.By utilizing the resonance frequency shift caused by magnetic field changes as a feedback variable, it achieves frequency adjustment of the microwave source, thus tracking and locking the NV resonance peak.This provides a continuous measurement method for time-varying magnetic fields [30][31][32].
Principle
NV centers are point defects formed by the substitution of a nitrogen atom for a carbon atom in the diamond lattice, accompanied by a nearby vacancy.Carbon atoms in diamonds are arranged in a tetrahedral structure; hence, the axial direction of NV centers is constrained to [111], 111 , 111 , 111 , exhibiting C 3v rotational symmetry.As shown in Figure 1a, the spin triplet system consists of the m s = 0 state, m s = +1 state, and m s = −1 state.NV centers exhibit two spin triplet states: the ground state 3 A 2 and the excited state 3 E.
The NV spin state in the ground state 3 A 2 transitions to the excited state 3 E under laser pumping, and then decays back to the 3 A 2 state through two pathways.Some NV centers directly return to the ground state from the excited state, emitting red fluorescence.Another portion of NV centers returns to the ground state through singlet states 1 E and 1 A 1 , which does not result in fluorescence.In the ground state, the spin triplet is degenerate for m s = ±1 states in the absence of a magnetic field, while a zero-field splitting of D = 2.87 GHz exists between the m s = ±1 and m s = 0 states.External microwave fields can induce population transfer between the m s = 0 and m s = ±1 states of NV centers.For the m s = ±1 states of the electron, applying an external bias magnetic field induces Zeeman splitting between the m s = +1 and m s = −1 energy levels, resulting in two resonance peaks [33].The external magnetic field applied to the NV axis can be determined by measuring the frequency shift of the resonance peaks in the optically detected magnetic resonance (ODMR) spectrum.The relationship between the frequency shift induced by the Zeeman splitting and the applied external magnetic field intensity is given by Equation (1) [34].
where ∆ f is the frequency shift, γ is the gyromagnetic ratio (taken as 2.8 × 10 10 Hz/T), and θ is the angle between the magnetic field direction and the NV axis.Typically, the diamond NV axis is adjusted so that one of its axes aligns with the direction of the measured magnetic field.In this case, the angle θ is 0, and cosθ equals 1.
By modulating the ODMR signal with a reference signal applied at microwave frequencies using a microwave source, and subsequently performing multiplication with the reference signal followed by filtering, the demodulated signal as illustrated in Figure 1b can be obtained.The NV spin state in the ground state 3 A2 transitions to the excited state 3 E u laser pumping, and then decays back to the 3 A2 state through two pathways.Som centers directly return to the ground state from the excited state, emitting red fluoresc Another portion of NV centers returns to the ground state through singlet states 1 E 1 A1, which does not result in fluorescence.In the ground state, the spin triplet is deg ate for ms = ±1 states in the absence of a magnetic field, while a zero-field splitting o 2.87 GHz exists between the ms = ±1 and ms = 0 states.External microwave fields ca duce population transfer between the ms = 0 and ms = ±1 states of NV centers.For th = ±1 states of the electron, applying an external bias magnetic field induces Zeeman ting between the ms = +1 and ms = −1 energy levels, resulting in two resonance peaks The external magnetic field applied to the NV axis can be determined by measurin frequency shift of the resonance peaks in the optically detected magnetic reson (ODMR) spectrum.The relationship between the frequency shift induced by the Zee splitting and the applied external magnetic field intensity is given by Equation (1) [3 where ∆ is the frequency shift, is the gyromagnetic ratio (taken as 2.8 × 10 10 H The above magnetic measurement method based on ODMR requires the calculation of the resonance peak shifts in the ODMR spectrum.However, when detecting time-varying magnetic fields, continuous scanning of the entire spectrum is required as the magnetic field changes in order to obtain the NV resonance frequency.Continuous scanning of the entire spectrum is time-consuming, as it involves monitoring non-resonant signals containing irrelevant information.Moreover, due to the lengthy scanning time, real-time detection of magnetic field changes is not feasible, which hampers the subsequent applications of NV magnetometers.Therefore, this study employs a microwave frequency modulation scheme to modulate the ODMR signal and multiply it with a reference signal, followed by filtering and demodulation.Within the linear range of the demodulated signal, the magnitude of the applied magnetic field intensity can be determined using Equation (2).
Here, V represents the voltage output of the step signal, γ is the gyromagnetic ratio, and max dv d f is the slope of the fitted line within the linear range of the demodulated signal.By pre-calibrating the relationship between the voltage values of the ODMR signal and the resonance frequencies, real-time monitoring of resonance point shifts in the spectrum can be achieved.However, this method is only applicable within the linear range of the demodulated signal, namely within the intrinsic dynamic range, as illustrated in Figure 1c.It cannot be used to measure magnetic fields beyond the approximately linear dynamic range interval.To enable the real-time measurement of rapidly changing external magnetic fields over a broader range, this study employs a frequency-tracking scheme.Frequency tracking is a process of monitoring real-time changes in the demodulated signal frequency and taking corresponding measures to maintain or adjust the signal frequency.In diamond NV magnetometers, frequency tracking typically refers to adjusting the microwave source frequency to match the resonance frequency changes induced by the magnetic field in the diamond sample.By continuously monitoring the resonance frequency variations induced by the magnetic field, the system is capable of dynamically adjusting the center frequency of the microwave source to keep it within the linear range of the demodulated signal, thereby extending the intrinsic dynamic range, as shown in Figure 1c.
Experimental Setup Design
This study presents a portable fiber-integrated NV magnetometer.A 532 nm wavelength laser is coupled into a single-mode optical fiber and focused on the diamond surface fixed on an antenna board using a lens (SBJ30F5-21C, Changsha, China), exciting the NV centers to emit fluorescence.The red fluorescence is filtered by a bandpass filter (LBTEK, NF-8-532-SP, Changsha, China) and focused again through a lens before being collected by a photodiode (LUNAINC, PDB-C609-2, Roanoke, VA, USA) and converted into electrical signals.These signals are subsequently precisely transmitted to a lock-in amplifier for signal demodulation and processing.The demodulated signals are visualized through the software interface of the lock-in amplifier's host computer, enabling real-time monitoring and analysis of the measured magnetic field.During this process, the microwave sweep frequency is provided by a microwave source through a circular copper wire antenna with a diameter of 1 mm.This antenna possesses omnidirectional radiation characteristics, capable of generating a uniform microwave field, while also exhibiting strong resistance to interference, ensuring the accuracy and stability of signal transmission.The fiber optic head, lens, filter, photodiode, and antenna board are all fixed within 3D-printed mounting rings, creating a closed structure that prevents external stray light from entering.Utilizing micro-nano fabrication techniques, an integrated module based on fiber optics for excitation and noise reduction can be achieved.Leveraging the efficient coupling of fiber optics and the focusing and collimating properties of self-focusing lenses, the use of traditional optical lenses in the optical path can be minimized.This greatly simplifies the optical path structure, increases the integration level of the magnetometer, and effectively reduces optical background noise in the environment.This structure allows the laser beam emitted from the laser to be focused at the center of the diamond sample.The overall structure of the magnetometer probe is shown in Figure 2a, with Figure 2b depicting the physical appearance of the probe, measuring 2.9 × 2.1 × 2 cm 3 .Combining the NV magnetometer probe with a lock-in amplifier (HF2LI, Zurich, Switzerland), microwave signal source (N5181B, Santa Clara, CA, USA), 532 nm laser (MGL-III-532-100 mW, Changchun, China), and current source (YP03-60, Changchun, China) forms a complete testing system, as shown in Figure 2c.The physical diagram of the experimental test system is shown in Figure 2d.
magnetometer probe with a lock-in amplifier (HF2LI, Zurich, Switzerland), microwave signal source (N5181B, Santa Clara, CA, USA), 532 nm laser (MGL-III-532-100 mW, Changchun, China), and current source (YP03-60, Changchun, China) forms a complete testing system, as shown in Figure 2c.The physical diagram of the experimental test system is shown in Figure 2d.
Experimental Details and Results
After setting up the magnetic sensing system, with the 532 nm laser controlled to output laser power of 80 mW and the microwave source set to output microwave power of 20 dBm, the obtained electron spin resonance (ESR) signal is depicted in Figure 3a.The contrast was measured to be 7.5%.By applying current from the current source to the three-axis Helmholtz coil (YP20201118-273-1, T/A:0.0016)along the X-axis, a uniform bias magnetic field is applied to the magnetometer probe.This resulted in obtaining the ODMR signal as shown in Figure 3b.The full width at half maximum (FWHM) was 13 MHz.Subsequent frequency modulation yields the demodulated signal as depicted in Figure 3c.
contrast was measured to be 7.5%.By applying current from the current source to three-axis Helmholtz coil (YP20201118-273-1, T/A:0.0016)along the X-axis, a uniform magnetic field is applied to the magnetometer probe.This resulted in obtaining the OD signal as shown in Figure 3b.The full width at half maximum (FWHM) was 13 M Subsequent frequency modulation yields the demodulated signal as depicted in Fig 3c.System noise is one of the primary fundamental parameters used to measure the formance of a magnetometer.Its value represents the ability of the magnetometer to de the smallest signal.Analyzing the amplitude spectral density is a common method rently used to calibrate the magnitude of magnetometer noise.It can reflect the ou noise values of the magnetic sensing system at different frequencies.The amplitude s tral density (ASD) of the output signal measured in a lock-in amplifier can be conve into a magnetic noise figure using Equation (3).The microwave source center frequency was fixed to the frequency correspondin the point of maximum slope of the demodulated signal.Using a three-axis Helmholtz an AC magnetic field signal with frequencies ranging from 1 Hz to 100 Hz was applie the diamond NV magnetometer.The recorded normalized peak-to-peak amplitu measured by the diamond NV magnetometer are depicted in Figure 4b.From these d the system bandwidth was determined to be 40 Hz.System noise is one of the primary fundamental parameters used to measure the performance of a magnetometer.Its value represents the ability of the magnetometer to detect the smallest signal.Analyzing the amplitude spectral density is a common method currently used to calibrate the magnitude of magnetometer noise.It can reflect the output noise values of the magnetic sensing system at different frequencies.The amplitude spectral density (ASD) of the output signal measured in a lock-in amplifier can be converted into a magnetic noise figure using Equation (3).
where ASD is obtained from the spectrum density module of the lock-in amplifier, as shown in Figure 4a.The method of real-time monitoring of varying magnetic fields using Equation ( 2) is limited by the intrinsic dynamic range of the magnetometer.As shown in Figure 5a, the region within the boxed area illustrates an approximate linear relationship between the amplitude of the demodulated signal and the microwave frequency.This linear region represents the intrinsic dynamic range of the NV magnetometer, with a slope of 1.9 mV/MHz for the fitted line.The intrinsic dynamic range of the magnetometer was approximately −384 µT to 384 µT.First, the microwave frequency was set to the frequency corresponding to the maximum slope of the linear fit to the demodulated signal.Subsequently, the current output of the power supply was gradually increased through programmatic control, causing the real-time magnetic field in the three-axis Helmholtz coil to continuously change.As the magnetic field gradually increases until the resonant frequency shifts beyond the intrinsic dynamic range, the resonant frequency point shifts from between points b and c in Figure 5a to the right of point c, exceeding the linear range.Consequently, the magnetic field conversion coefficient becomes ineffective, and during this period, the signal exhibits a decreasing trend, making it impossible to determine the magnitude of the measured magnetic field.Therefore, frequency-tracking technology is required to synchronously change the microwave frequency as the magnetic field gradually increases.This ensures that the resonant frequency point remains within the linear range.At this point, although the continuously increasing magnetic field has exceeded the magnetometer's intrinsic measurable range, the real-time calculation of the magnetic field magnitude The microwave source center frequency was fixed to the frequency corresponding to the point of maximum slope of the demodulated signal.Using a three-axis Helmholtz coil, an AC magnetic field signal with frequencies ranging from 1 Hz to 100 Hz was applied to the diamond NV magnetometer.The recorded normalized peak-to-peak amplitudes measured by the diamond NV magnetometer are depicted in Figure 4b.From these data, the system bandwidth was determined to be 40 Hz.
The method of real-time monitoring of varying magnetic fields using Equation ( 2) is limited by the intrinsic dynamic range of the magnetometer.As shown in Figure 5a, the region within the boxed area illustrates an approximate linear relationship between the amplitude of the demodulated signal and the microwave frequency.This linear region represents the intrinsic dynamic range of the NV magnetometer, with a slope of 1.9 mV/MHz for the fitted line.The intrinsic dynamic range of the magnetometer was approximately −384 µT to 384 µT.First, the microwave frequency was set to the frequency corresponding to the maximum slope of the linear fit to the demodulated signal.Subsequently, the current output of the power supply was gradually increased through programmatic control, caus-ing the real-time magnetic field in the three-axis Helmholtz coil to continuously change.As the magnetic field gradually increases until the resonant frequency shifts beyond the intrinsic dynamic range, the resonant frequency point shifts from between points b and c in Figure 5a to the right of point c, exceeding the linear range.Consequently, the magnetic field conversion coefficient becomes ineffective, and during this period, the signal exhibits a decreasing trend, making it impossible to determine the magnitude of the measured magnetic field.Therefore, frequency-tracking technology is required to synchronously change the microwave frequency as the magnetic field gradually increases.This ensures that the resonant frequency point remains within the linear range.At this point, although the continuously increasing magnetic field has exceeded the magnetometer's intrinsic measurable range, the real-time calculation of the magnetic field magnitude can still be achieved using Equation ( 2).This completes the extension of the intrinsic dynamic range.Within the range from point b to point c in Figure 5a, the dynamic range is increased by approximately 17 times through the frequency tracking scheme, as shown in Figure 5b.After extending the dynamic range through frequency tracking, the relationship between the applied current values within the range of 1 to 5 A provided by the current source and the measured magnetic field magnitude by the magnetometer can be obtained as shown in Figure 5c.In Figure 5a, the same results were obtained within the range from point a to point b through this approach.Therefore, the extension of the dynamic range obtained by the magnetometer through this frequency tracking scheme is approximately 34 times greater than the intrinsic dynamic range.
The maximum tracking rate of the changing magnetic field using this frequency tracking scheme can be represented by Equation (4).
𝑣
where represents the intrinsic dynamic range of the NV center, and represents the period for one frequency tracking cycle.During each measurement cycle, as the magnetic field changes in real-time, the upper computer program collects and calculates the frequency offset and controls the microwave source to change the center frequency, a process taking approximately 10 ms.Within one closed-loop cycle, if the change in the magnetic field falls within half of the intrinsic dynamic range 2 ⁄ , the system can track the resonant frequency using the frequency tracking scheme.However, if it exceeds half of the intrinsic dynamic range 2 ⁄ , it will deviate from feedback, and the extension of the dynamic range cannot continue.Therefore, using Equation ( 4), the tracking rate of the system for the changing magnetic field can be calculated to be 0.038 T/s.Subsequently, the method is validated within the range of 1 to 5 A provided by the current source.Initially, a bias magnetic field is generated by applying a 1 A current to the three-axis Helmholtz coil.Following the completion of the microwave source frequency adjustment, the magnetic field intensity within the coil is increased by gradually increasing the output of the current source.When the current source is increased to approximately 1.24 A, the signal, which originally exhibited a decreasing trend at point c in Figure 5a, continues to gradually rise after undergoing an instantaneous change in center frequency.This ensures that the resonant frequency point remains within the linear range, thus completing the extension of the intrinsic dynamic range.Thus, it is verified that the magnetic sensing system can perform real-time magnetic field calculations within the range of 1 to 5 A provided by the current source.
Within the range from point b to point c in Figure 5a, the dynamic range is increased by approximately 17 times through the frequency tracking scheme, as shown in Figure 5b.After extending the dynamic range through frequency tracking, the relationship between the applied current values within the range of 1 to 5 A provided by the current source and the measured magnetic field magnitude by the magnetometer can be obtained as shown in Figure 5c.In Figure 5a, the same results were obtained within the range from point a to point b through this approach.Therefore, the extension of the dynamic range obtained by the magnetometer through this frequency tracking scheme is approximately 34 times greater than the intrinsic dynamic range.
The maximum tracking rate v max of the changing magnetic field using this frequency tracking scheme can be represented by Equation (4).
where Γ represents the intrinsic dynamic range of the NV center, and t cycle represents the period for one frequency tracking cycle.During each measurement cycle, as the magnetic field changes in real-time, the upper computer program collects and calculates the frequency offset and controls the microwave source to change the center frequency, a process taking approximately 10 ms.Within one closed-loop cycle, if the change in the magnetic field falls within half of the intrinsic dynamic range Γ/2, the system can track the resonant frequency using the frequency tracking scheme.However, if it exceeds half of the intrinsic dynamic range Γ/2, it will deviate from feedback, and the extension of the dynamic range cannot continue.Therefore, using Equation ( 4), the tracking rate of the system for the changing magnetic field can be calculated to be 0.038 T/s.The slope of the linear fit to the demodulated signal within the linear range directly determines the frequency tracking system's ability to extend the intrinsic dynamic range of the NV center.Therefore, it is necessary to measure the effects of modulation frequency offset and modulation frequency on the slope of the linear fit to the demodulated signal within the linear range to obtain the optimal modulation parameters.As shown in Figure 6a, during the gradual increase in modulation frequency offset, the slope initially rises and then tends to saturate.By comparing the slope values at different modulation frequencies, it can be concluded that to obtain the optimal slope value, a reference signal with V dev = 5 MHz and V mod = 500 Hz should be chosen.Based on selecting the optimal modulation parameters, the slope of the linear fit to the demodulated signal within the linear range varies with the current in the range of 1 to 5 A from the current source, as shown in Figure 6b.It can be estimated that when the magnetic field reaches approximately 28.8 mT, the slope of the linear fit to the demodulated signal within the linear range approaches zero, and the slope magnitude can be considered negligible.Consequently, the magnetic field magnitude cannot be determined using Equation (2).From this, it can be concluded that due to the continuous increase in magnetic field affecting the slope of the linear fit to the demodulated signal within the linear range, there is a limit to enhancing the measurable magnetic field range through frequency tracking.Through analysis, it is determined that the theoretical limit of the extension of the magnetometer's dynamic range is approximately 28.8 mT.Compared to the intrinsic dynamic range of the magnetometer, this represents an increase by approximately 150 times, allowing for a significant enhancement in the measurable magnetic field range of the magnetometer and stable measurement of external time-varying magnetic fields.Based on selecting the optimal modulation parameters, the slope of the linear fit to the demodulated signal within the linear range varies with the current in the range of 1 to 5 A from the current source, as shown in Figure 6b.It can be estimated that when the magnetic field reaches approximately 28.8 mT, the slope of the linear fit to the demodulated signal within the linear range approaches zero, and the slope magnitude can be considered negligible.Consequently, the magnetic field magnitude cannot be determined using Equation (2).From this, it can be concluded that due to the continuous increase in magnetic field affecting the slope of the linear fit to the demodulated signal within the linear range, there is a limit to enhancing the measurable magnetic field range through frequency tracking.Through analysis, it is determined that the theoretical limit of the extension of the magnetometer's dynamic range is approximately 28.8 mT.Compared to the intrinsic dynamic range of the magnetometer, this represents an increase by approximately Micromachines 2024, 15, 662 9 of 11 150 times, allowing for a significant enhancement in the measurable magnetic field range of the magnetometer and stable measurement of external time-varying magnetic fields.
Conclusions
In order to realize the fast measurement of time-varying magnetic field, a frequency tracking method is proposed in this study.The time-varying magnetic field is tracked by feeding the resonant frequency shift caused by the magnetic field change back to the microwave source and continuously changing the center frequency point of the microwave source.In the experiment, the method successfully extends the dynamic range to 6.4 mT, which is equivalent to 34 times the intrinsic dynamic range of the NV center, while maintaining a tracking rate of 0.038 T/s, within the current range of 5 A that the current source can provide.Theoretical calculations show that the dynamic range of the system can be extended to 28.8 mT, which is significantly more than 150 times the intrinsic dynamic range of the NV center.In addition, by utilizing the focusing and collimation coupling characteristics of the fiber optic self-focusing lens, the fiber optic-integrated magnetometer designed in this study greatly simplifies the complex optical path structure.While maintaining portability, the system not only enhances the measurable range, but also realizes the efficient detection of continuously changing magnetic fields, which is more suitable for the complex environment of real scenarios with constantly changing magnetic fields.The research results provide an effective solution for real-time accurate monitoring of dynamic magnetic field changes, which is potentially valuable for a wide range of applications.
Figure 1 .
Figure 1.(a) Schematic diagram of diamond NV center energy level.The green line represen excitation of NV centers from their ground state to the excited state induced by a 532 nm lase red line indicates that, after a certain duration, some NV centers undergo direct transition fro excited state to the ground state, emitting red fluorescence with a wavelength ranging from 800 nm.The gray line illustrates the pathway for NV centers to return to their ground state v metastable states 1 A1 and 1 E; (b) schematic diagram of modulation-demodulation principle; (c) matic diagram of frequency-tracking principle.The black arrow indicates the intrinsic dy range, and the red arrow indicates the extended dynamic range.
Figure 1 .
Figure 1.(a) Schematic diagram of diamond NV center energy level.The green line represents the excitation of NV centers from their ground state to the excited state induced by a 532 nm laser.The red line indicates that, after a certain duration, some NV centers undergo direct transition from the excited state to the ground state, emitting red fluorescence with a wavelength ranging from 637 to 800 nm.The gray line illustrates the pathway for NV centers to return to their ground state via the metastable states 1 A 1 and 1 E; (b) schematic diagram of modulation-demodulation principle; (c) schematic diagram of frequency-tracking principle.The black arrow indicates the intrinsic dynamic range, and the red arrow indicates the extended dynamic range.
Figure 2 .
Figure 2. (a) Schematic diagram of the structure of an NV center magnetometer probe; (b) photograph of the magnetometer probe; (c) schematic diagram of the frequency tracking test system; (d) experimental test system physical diagram.
Figure 2 .
Figure 2. (a) Schematic diagram of the structure of an NV center magnetometer probe; (b) photograph of the magnetometer probe; (c) schematic diagram of the frequency tracking test system; (d) experimental test system physical diagram.
is obtained from the spectrum density module of the lock-in amplifie shown in Figure 4a. | | represents the maximum slope of the linear fit to demodulated signal in the linear range. = 2.8 × 10 10 Hz/T is the gyromagnetic ratio.noise figure of the magnetometer obtained is 3.58 nT/Hz 1/2 .
max d d f S represents the maximum slope of the linear fit to the demodulated signal in the linear range.γ = 2.8 × 10 10 Hz/T is the gyromagnetic ratio.The noise figure of the magnetometer obtained is 3.58 nT/Hz 1/2 .
Figure 4 .
Figure 4. (a) Noise amplitude spectrum; (b) results of the system bandwidth.
Figure 4 .
Figure 4. (a) Noise amplitude spectrum; (b) results of the system bandwidth.
Micromachines 2024, 15 , 662 8 of 11 Figure 5 .
Figure 5. (a) Linear region of the demodulated signal; (b) schematic diagram of the frequency tracking results; (c) the linear relationship between the output current of the current source and the measured magnetic field value.
Figure 5 .
Figure 5. (a) Linear region of the demodulated signal; (b) schematic diagram of the frequency tracking results; (c) the linear relationship between the output current of the current source and the measured magnetic field value.
Micromachines 2024, 15, 662 9 of 11 6a, during the gradual increase in modulation frequency offset, the slope initially rises and then tends to saturate.By comparing the slope values at different modulation frequencies, it can be concluded that to obtain the optimal slope value, a reference signal with Vdev = 5 MHz and Vmod = 500 Hz should be chosen.
Figure 6 .
Figure 6.(a) The relationship between modulation frequency and the slope of modulation frequency offset versus demodulated signal under different modulation frequencies; (b) the linear relationship between the output current of the current source and the slope of the demodulated signal.
Figure 6 .
Figure 6.(a) The relationship between modulation frequency and the slope of modulation frequency offset versus demodulated signal under different modulation frequencies; (b) the linear relationship between the output current of the current source and the slope of the demodulated signal. | 7,357 | 2024-05-01T00:00:00.000 | [
"Physics",
"Engineering"
] |
Neuregulin Signaling on Glucose Transport in Muscle Cells*
Neuregulin-1, a growth factor that potentiates myogenesis induces glucose transport through translocation of glucose transporters, in an additive manner to insulin, in muscle cells. In this study, we examined the signaling pathway required for a recombinant active neuregulin-1 isoform (rhHeregulin-β1, 177–244, HRG) to stimulate glucose uptake in L6E9 myotubes. The stimulatory effect of HRG required binding to ErbB3 in L6E9 myotubes. PI3K activity is required for HRG action in both muscle cells and tissue. In L6E9 myotubes, HRG stimulated PKBα, PKBγ, and PKCζ activities. TPCK, an inhibitor of PDK1, abolished both HRG- and insulin-induced glucose transport. To assess whether PKB was necessary for the effects of HRG on glucose uptake, cells were infected with adenoviruses encoding dominant negative mutants of PKBα. Dominant negative PKB reduced PKB activity and insulin-stimulated glucose transport but not HRG-induced glucose transport. In contrast, transduction of L6E9 myotubes with adenoviruses encoding a dominant negative kinase-inactive PKCζ abolished both HRG- and insulin-stimulated glucose uptake. In soleus muscle, HRG induced PKCζ, but not PKB phosphorylation. HRG also stimulated the activity of p70S6K, p38MAPK, and p42/p44MAPK and inhibition of p42/p44MAPK partially repressed HRG action on glucose uptake. HRG did not affect AMPKα1 or AMPKα2 activities. In all, HRG stimulated glucose transport in muscle cells by activation of a pathway that requires PI3K, PDK1, and PKCζ, but not PKB, and that shows cross-talk with the MAPK pathway. The PI3K, PDK1, and PKCζ pathway can be considered as an alternative mechanism, independent of insulin, to induce glucose uptake.
During adult life, NRGs are highly expressed in neurons and they act on the juxtaposed tissues contributing to the maintenance of differentiated patterns, such as the neuromuscular junction, NMJ (8 -12). They are expressed as either membrane or soluble forms, mainly as a consequence of differential splicing of the primary transcript (reviewed in 13 and 14). Soluble forms attach to the basal lamina through proteoglycans (e.g. agrin) and, by regulated proteolysis, they are released and bind to their receptors. NRGs bind to type I tyrosine kinase receptors that belong to the family of EGF receptors, ErbB3 and ErbB4. Binding generates homo-and heterodimerization between them and with ErbB2, which lacks a binding domain, whereas ErbB3 does not have kinase activity (reviewed in Ref. 15). Cultured myotubes express ErbB2 and ErbB3, but not ErbB4 (10,11), so the main NRG receptor is the heterodimer ErbB2/ErbB3. All three receptors are expressed in muscle tissue although only ErbB2 and ErbB4 accumulate at the postsynaptic site (16).
The PI3K/PKB pathway is involved in NRG anti-apoptotic * This work was supported by a grant from the Ministerio de Ciencia y Tecnología, SAF2002-01585 (to A. G.) and grants from the Instituto de Salud Carlos III/FIS, G03/212 and C03/08 (to A. Z.). The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked "advertisement" in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
‡ Both authors contributed equally to this work. § Supported by the Instituto de Salud Carlos III/FIS, G03/212. ¶ Recipient of a predoctoral fellowship from the Universitat de Barcelona, Spain. § § Supported by the Program Ramón y Cajal from the Ministerio de Ciencia y Tecnología, with FEDER/FSE funds, and the University of Barcelona.
NRGs induce glucose uptake in myocyte cultures and muscle fibers (31). This effect is fast and requires translocation of glucose transporters to the plasma membrane (31). NRGs and insulin induce glucose uptake in an additive manner and both require PI3K activity (31).
A recent study indicates that muscle contraction, caused by either acute exercise or electrical stimulation, induces release of NRGs, to the extracellular milieu that results in ErbBsinduced-activity (40). Both exercise and oxidative stress induce glucose uptake through translocation of glucose transporters, in an additive and independent manner to the action of insulin, and AMPK has been suggested to be involved in this effect (41)(42)(43) although controversy exists in this aspect.
Here we examine the signaling cascade triggering the action of HRG on glucose uptake in L6E9 myotubes. PI3Ks, PDK1, and PKC, but not PKB, are involved in HRG-stimulated glucose uptake. The additive effect of HRG and insulin on glucose uptake did not involve AMPK activation.
2-Deoxy-D-[ 3 H]Glucose Uptake-L6E9 cells were cultured on 6-well plates. 2-Deoxyglucose uptake assays were performed as previously described (31). Radioactivity was determined by scintillation counting. Protein was measured by BCA protein reagent assay (Pierce). Each condition was run in duplicate or triplicate. 2-Deoxyglucose transport was linear during the period (10 min) assayed (not shown).
2-Deoxyglucose uptake assays were performed in strips of rat soleus muscles as previously described (31). Protein was measured by Bradford method.
Preparation of Extracts from L6E9 Myotubes and from Strips of Rat Soleus Muscle-Homogenates were prepared from L6E9 myotubes. They were placed on ice and washed twice in ice-cold phosphate-buffered saline before adding 300 l (6-well plates) or 1 ml (100-mm dishes) of lysis buffer (50 mM Tris-HCl, pH 7.5, 150 mM NaCl, 1% (v/v) Nonidet P-40, 1 mM EDTA, 5 mM sodium pyrophosphate, 1 mM sodium orthovanadate, 50 mM sodium fluoride, containing freshly added protease inhibitors: 0.2 mM phenylmethylsulfonyl fluoride, 1 M leupeptin and 1 M pepstatin). Cells were then scraped, collected into an Eppendorf tube, and then homogenized by repetitive passes through a 25-gauge needle before being centrifuged at 15,000 rpm for 10 min at 4°C. Finally, supernatant was collected and it was stored at Ϫ80°C. Protein was measured by the Bradford method. Preparations of fractions enriched in plasma membranes (PM) or in intracellular membranes (LDM) were obtained from two 10-mm culture dishes as reported (31).
Muscle homogenates were obtained from strips of rat soleus. Muscles were frozen, collected into an Eppendorf, placed on ice containing 200 l of fresh lysis buffer, and then homogenized with an Eppendorf homogenizer. Homogenates were taken to a final volume of 1 ml/strip by adding lysis buffer and then centrifuged at 15,000 rpm for 10 min at 4°C. The pellet was discarded, and supernatant was collected and stored at Ϫ80°C until used. Protein was measured by the Bradford method.
Immunoprecipitation Assays-Immunoprecipitation was performed by conjugating 30 l of protein G-Sepharose beads with 2-5 g of the corresponding antibody, except when it is indicated, shaking for 1 h at 4°C, then washing twice in lysis buffer and incubating with 500 -1,000 g of lysate overnight with constant shaking at 4°C. After brief centrifugation the supernatant was discarded. The pellet was washed several times with the lysis buffer and boiled with 50 l of Laemmli sample buffer (LBS) for Western blot assays as described below. PVDF membranes were blotted with the corresponding antibodies.
For PKCimmunoprecipitation, 500 g of homogenate were incubated overnight with 1 g of an anti-nPKC-antibody (C-20, Santa Cruz Biotechnology), with constant shaking at 4°C, and 30 g of protein G-Sepharose beads were then added. Beads were then washed twice in lysis buffer, boiled at 95°C for 5 min in 50 l LSB, and the supernatant was used for Western blot as described below. PVDF membranes were blotted with anti-phospho-PKC/ (Thr 410/403 ) antibody.
PI3K Activity Assay-L6E9 myotubes were washed once in phosphate-buffered saline and lysed in a buffer containing 10 mM Tris-HCl, pH 7.4, 5 mM EDTA, 50 mM NaCl, 100 mM NaF, and 1% Triton X-100 supplemented with 2 mg/ml aprotinin, 1 mM pepstatin A, 1 ng/ml leupeptin, 2 mM phenylmethylsulfonylfluoride, and 10 mM sodium orthovanadate. 100 g of lysate protein was incubated with 1.5 l of mouse monoclonal anti-phosphotyrosine antibody (PY99, Santa Cruz Biotechnology) for 90 min at 4°C, and then immunoprecipitated with 2.5 mg of rehydrated protein A-Sepharose beads (Sigma) for 60 min. The immunoprecipitates were used for the kinase assay, which was performed with phosphatidylinositol as substrate in the presence of [␥-32 P]ATP (44). Lipids were separated by TLC, and the radioactivity corresponding to phosphatidylinositol 3-phosphate was analyzed using a Fuji FLA-2000 phosphoimager.
PKB, p70S6K, p38MAPK, and p42/p44MAPK Activity Assays-5 g of sheep antibodies raised against PKB␣ or ␥ (45), p70S6K (46), MAP-KAP-K2 (47), or p90rsk (48) were conjugated to 5 l of protein G-Sepharose beads for 30 min at 4°C. After washing twice in lysis buffer, 500 g of total lysate in the case of p70S6K or 100 g otherwise were added and the mixture was incubated for 1 h at 4°C on a shaking platform. It was then centrifuged briefly at 13,000 rpm. The immunoprecipitates were washed twice with 1 ml of lysis buffer containing 0.5 M NaCl, and twice with 1 ml of reaction buffer (50 mM Tris-HCl, pH 7.5, 0.1% (v/v) -mercaptoethanol, 0.1 mM EGTA). The standard assay, which is employed unless stated, otherwise contained (50 l of total volume): washed protein G-Sepharose immunoprecipitated, 50 mM Tris-HCl, pH 7.5, 0.1 mM EGTA, 0.1% (v/v) -mercaptoethanol, 2.5 M PKI (TTYADFIASGRTGRRNAIHD, peptide inhibitor of the cyclic AMP-dependent protein kinase), 10 mM magnesium acetate, 0.1 mM [␥-32 P]ATP (ϳ300 cpm/pmol) and the corresponding substrate: 30 M Crosstide peptide (GRPRTSSFAEG) in all cases except for measuring MAP-KAP-K2 activity (30 M MAPKAP-K2 substrate KKLNRTLSVA). The assays were carried out for 30 min at 30°C, the assay tubes being agitated continuously to keep the immunoprecipitated in suspension. Incorporation of [␥-32 P]phosphate into each peptide substrate was determined using p81 phosphocellulose paper. Papers were washed in 0.5% orthophosphoric acid, dried, and Cherenkov radiation was counted. 1 milliunit of activity is the amount of enzyme that catalyzes the phosphorylation of 1 pmol of substrate in 1 min.
AMPK Activity Assay-AMPK ␣1 and ␣2 activities were measured in L6E9 myotubes lysates as described previously (48). Briefly, cells were incubated in HEPES-buffered saline (HBS) containing 5 mM glucose for 30 min at 37°C. After washing the cells in phosphate-buffered saline, they were lysed into a minimal volume of buffer A (50 mM Tris-HCl, pH 7.5, 1 M EDTA, 1 mM dithiothreitol, 10% glycerol, 50 mM sodium fluoride, 5 mM sodium pyrophosphate, 1 mM benzamidine, 0.1 mM phenylmethylsulfonyl fluoride, and 1% Triton X-100), cell debris was spun down in a bench top centrifuge at 12,000 rpm for 5 min, and the supernatant was taken for assay. Protein concentration was determined using the Bradford reagent. 100 -200 g of protein was then sequentially immunoprecipitated using sheep antibodies raised against either ␣1 or ␣2 (49) conjugated to protein G-Sepharose beads. AMPK activity within the immune complex was measured by the phosphorylation of synthetic SAMS peptide using radiolabeled [␥-32 P]ATP.
Immunoblotting-Protein samples containing LSB were subjected to sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and transferred to PVDF membranes as previously described (31). The membranes were blocked with 5% nonfat dry milk in Tris-buffered saline solution (TBS) for 1 h at room temperature and then incubated overnight with the corresponding primary antibodies at recommended dilutions in 1% bovine serum albumin. PVDF membranes were then washed, incubated with appropriate secondary antibodies, containing horseradish peroxidase, for 1 h at room temperature and washed again. Proteins were detected by the ECL method and quantified by scanning densitometry.
Statistical Analysis-Data are presented as means ϩ S.E. Unpaired Student's t test was used to compare two groups and one-way analysis of variance, posthoc Duncan's t test, was used to compare more than two groups.
Neuregulins-induced Glucose Transport Requires an Intact
ErbB3 Binding Activity-In order to prove the necessity of the ErbB3 binding activity in the NRG action on glucose uptake in L6E9 myotubes, we used an antibody, Ab5, that binds to ErbB3 and blocks the NRG-binding domain. When Ab5 (10 g/ml) was added to the medium 30 min before HRG, it abolished the stimulation of glucose transport (Fig. 1A). Ab5 did not alter the insulin effect on glucose uptake (Fig. 1A).
HRG enhanced tyrosine phosphorylation of ErbB3 receptors but it did not induce tyrosine phosphorylation of the  subunit of insulin receptors (Fig. 1B). Moreover, chronic insulin treatment that generated insulin resistance characterized by a deficient induction of glucose uptake, did not alter HRG action (Fig. 1, C and D). In all, HRG required ErbB3 to stimulate glucose uptake in L6E9 myotubes and this effect did not involve activation of the insulin receptor and it was not modified by induction of insulin resistance.
HRG (3 nM) activated PKC in L6E9 myotubes, with a maximum at 30 min (Fig. 2C). HRG and insulin did not show additive effects (Fig 2, A-C). PDK1 Is Involved in the Effect of Neuregulins on Glucose Uptake in L6E9 Myotubes-PI3K activity is required in the HRG-induced pathway on glucose uptake in L6E9 myotubes, based in our previous studies using the specific PI3K inhibitor wortmannin (31). In the present study, we analyzed the role of PDK1 on the NRGs effects, based on the use of TPCK, a PDK1 inhibitor (52). 50 M TPCK inhibited both PKB and PKC activities induced by HRG or by insulin (Fig. 3, A and B). TPCK inhibited PKB Thr 308 phosphorylation without affecting Ser 473 phosphorylation (Fig. 3A). Interestingly, Ser 473 was consistently phosphorylated less under HRG action than under insulin action while no remarkable differences were observed at Thr 308 phosphorylation. 50 M TPCK also repressed both HRGand insulin-induced glucose transport (Fig. 3C).
PKB Is Not Involved in the Action of Neuregulins on Glucose Uptake in L6E9 Myotubes-In order to study the role of PKB in the HRG action on glucose uptake, we used dominant negative mutants of PKB. Initially, we transduced L6E9 myocytes with adenovirus containing mutated forms of PKB␣, AA-PKB, and AAA-PKB (MOI 300, 48 h), in order to inhibit endogenous PKB activity. L6E9 myocytes transduced with double mutant viruses showed a partial inhibition of insulin-or HRG-induced PKB activity (Fig. 4A). Almost complete inhibition of the PKB activity was obtained with the triple mutant AAA-PKB (Fig. 4A). The overexpression of the double mutant AA-PKB induced a significant inhibition of insulin-stimulated glucose transport, which was even greater in cells expressing the AAA-PKB mutant (Fig. 4B). In these conditions the dominant negative forms of PKB did not inhibit HRG-induced glucose transport or HRGinduced GLUT4 translocation (Fig. 4, B and C).
PKC Activity Is Necessary for Neuregulin-induced Glucose Uptake-The next step was to determine the involvement of PKC in the HRG-induced glucose uptake. Transduction of L6E9 myotubes with KI-PKC viruses (MOI 100, 48 h) blocked PKC activity (Fig. 5A) and repressed the effects of HRG and insulin on glucose uptake (Fig. 5B).
P70S6K, p38MAPK, and p42/p44MAPK Are Not Directly Involved in the Action of Neuregulin on Glucose Uptake-We analyzed HRG action on p70S6K, p38MAPK, or p42/ p44MAPKs activities in L6E9 myotubes and their possible involvement in the effects of HRG on glucose uptake using specific inhibitors.
HRG induced p70S6K activity (Fig. 6A), which was completely abolished by the specific inhibitor rapamycin. However, rapamycin did not alter the effect of HRG on glucose uptake (Fig. 6B). p38MAPK was moderately activated by HRG in comparison with the hyperosmotic effects of 0.5 M sorbitol (Fig. 6A). The specific inhibitor SB203580 (10 M) abolished HRG activation of p38MAPK activation by HRG action (Fig. 6A) but did not interfere with its action on glucose uptake (Fig. 6B).
Effects of Neuregulins on AMPK Activity-Due to the lack of a specific inhibitor of AMPK, many studies have used 5-amino-4-imidazolecarboxamide riboside, AICAR, as an AMPK activator. However, AICAR does not lead to activation in many cell types since its efficacy is dependent on its uptake by the cell and accumulation as a monophosphorylated form, ZMP, which acts as an AMP-mimetic, promoting allosteric activation of AMPK and phosphorylation by an upstream kinase (43). L6E9 myotubes are not a good cellular model for AICAR trials since it appears to have low effect on AMPK ␣ 1 and ␣ 2 (not shown), so AMPK was activated by increasing intracellular AMP with DNP, an oxidative phosphorylation uncoupler. DNP stimulated AMPK␣ 1 and AMPK␣ 2 activities in L6E9 myotubes. Under these conditions, HRG did not alter the AMPK activity at any FIG. 2. HRG stimulates PI3K, PKB␣/␥, and PKC. L6E9 myotubes were treated with 3 nM HRG or/and 1 M insulin for the indicated times, and each kinase was immunoprecipitated from the lysates and assayed. A, PI3K activity was assayed and phosphatidylinositol 3-phosphate (PIP 3 ) production was quantified densitometrically. Results are relative values that correspond to a representative experiment from two independent experiments, each one run in duplicate and assayed twice. B, PKB␣ or PKB␥ were immunoprecipitated, and activities were assayed as described under "Experimental Procedures." Results correspond to a representative experiment from two different experiments, each one run in duplicate and assayed in triplicates. C, total PKC was immunoprecipitated from lysates and phospho-Thr 410 -PKC was immunodetected and quantified by densitometric analysis. Results, expressed as relative to basal values, were obtained from three different experiments. B, basal; H, HRG, I, insulin. * indicates significant differences with the basal group, at p Ͻ 0.01. time or dose studied (Fig. 7A). Identical results were obtained in the H-2K b myocytes (48) (not shown).
DNP, HRG, and insulin showed additive effects on glucose transport (Fig. 7B) indicating that they activate at least partially different intracellular pathways to induce glucose transport in muscle cells.
Neuregulins Signaling Cascade on Glucose Transport in Rat Soleus Muscle-Strips of rat soleus muscle were incubated with HRG (5 nM, 15 min) or insulin (100 nM, 10 min) and lysates were obtained, immunoprecipitated with an anti-phosphotyrosine antibody, and the pellets were immunoblotted with antibodies against each neuregulin receptor. Results indicated that HRG, but not insulin, activated all the expressed neuregulin receptors in muscle fiber, ErbB2, ErbB3, and ErbB4 (Fig. 8A).
Treatment with wortmannin (2 M, 150 min) abolished both insulin and HRG action on glucose uptake in incubated strips of rat soleus (Fig. 8B), indicating that PI3K is also required in the neuregulins signaling cascade at the muscle fiber.
Next, strips of rat soleus muscle were treated with HRG or insulin during 30 min, and lysates were obtained to analyze phosphorylation levels of PKB and PKC. Whereas insulin induced phosphorylation of both kinases, HRG was unable to induce PKB phosphorylation (Fig. 8, C and D) at any time ranging from 10 to 60 min (not shown).
DISCUSSION
In this study we have demonstrated that NRGs stimulate PI3K, PKB, PKC, p70S6K, p38MAPK, and p42/p44MAPKs in L6E9 myotubes, that the pathway PI3K, PDK1, and PKC is essential to induce glucose transport and that p42/ p44MAPKs also contribute to a maximal glucose transport stimulation. Studies done in rat soleus muscles also support the implication of the PI3K pathway on NRG-stimulated glucose transport. This is the first report indicating that NRGs induce a PDK1/PKC pathway. In L6E9 myotubes ErbB3 is required to initiate this signaling cascade. NRGs do not transactivate the insulin receptors, so the NRG signaling pathway can be considered as an alternative mechanism to induce glucose transport in muscle cells which, like insulin, is initiated by PI3K activation (31,53,54). ErbB3 contains six consensus sequences, YXXM motifs that bind to the SH 2 domain of the PI3K p85 subunit after ligand-induced tyrosine phosphorylation (19). ErbB3 has been characterized as the main mediator of heregulin-dependent PI3K activation pathway (55) and expression of mutated forms of ErbB3 at the YXXM motifs in COS7 cells abolishes PI3K and PKB activation by heregulin binding to the heterodimer ErbB2/ErbB3 (56). Since ErbB4 is also expressed and activated by HRG in skeletal muscle, differences on the insulin and the HRG action in soleus and cultured myocytes could be consequence of a different pattern of receptors expression.
PI3K activity is essential for HRG action on glucose transport (31), although HRG induces lower maximal PI3K activities than insulin. Differences on PI3K activation could be a consequence of the expression levels of ErbB3 and insulin receptors in L6E9 myotubes, and they also suggest that PI3K activity might not be limiting for signaling on glucose uptake. This view is also supported by the lack of additivity of HRG and insulin on PI3K activity. Otherwise, HRG requires more time than insulin to reach maximal effects on glucose transport (31), so a critical amount of phosphatidylinositol 3-phosmin). Then, the 2-deoxyglucose uptake assay was performed. Results are obtained from three different experiments. Basal uptake: 0.62 Ϯ 0.03 pmol/g of protein ϫ 10 min. B, basal; I, insulin action; H, HRG action. * indicates significant differences with the control group, at p Ͻ 0.001.
FIG. 3. HRG-induced glucose uptake requires PDK1 activity.
L6E9 myotubes were treated with 3 nM HRG (15 min) or 1 M insulin (10 min), in the absence or presence of the PDK1 inhibitor, 50 M TPCK (35 min). A, total PKB activity was determined from one experiment assayed in triplicate. Cell lysates were also immunoblotted with antibodies that recognize total PKB (C-terminal) or PKB phosphorylated at Ser 473 (hydrophobic motif) or Thr 308 (activation loop). Similar results were obtained in three separate experiments. B, phospho-Thr 410 -PKC was immunodetected and densitometrically quantified. Relative phospho-Thr 410 -PKC levels are the mean of three independent experiments. C, L6E9 myotubes were treated with 3 nM HRG (90 min) or 1 M insulin (30 min), in the absence or the presence of 50 M TPCK (110 phate may be needed to trigger HRG stimulation of glucose transport. It is notable that HRG action is not impaired in muscle cells made insulin resistant by chronic exposure to insulin. Interestingly, soleus muscle insulin responsiveness was reduced in the adult rat compared with 1-month-old rats (2-3-fold increase in response to maximal insulin in adult rats and 5-9-fold increase in young rats, respectively); under these conditions, heregulininduced stimulation of glucose transport remained unaltered (1.5-2-fold increase both in adult and in young rats). 2 studies are being addressed to investigate the extent of neuregulin action in muscle in animal models of type 2 diabetes, obesity or aging.
HRG induces both PKB␣ and PKB␥ in L6E9 myotubes. Similar results were previously reported for insulin action in L6 myotubes (45). HRG induces lower PKB␣ activity than insulin, as for PI3K activity. This is in accordance with the lower ability of HRG to phosphorylate Ser 473 , which is involved in the activation of PKB. In contrast, HRG has stronger maximal effects than insulin on PKC, but HRG requires longer time than insulin to reach them (maximal insulin effect on PKC is reached within 10 min, not shown). These differences might explain the potentiated effects observed on glucose transport since insulin and neuregulins do not have additive effects on PKB or PKC activation levels. Nonetheless, an alternative pathway to PI3K activation cannot be ruled out. In fact, a second insulin pathway is involved in the stimulation of glucose transport in adipose cells (reviewed in Ref. 57). This second pathway requires phosphorylation of the adaptor protein Cbl by the insulin receptor and activation of TC10, a small GTP-binding protein located in lipid rafts, which modulates actin structure. At present we do not know whether neuregulin stimulates the Cbl pathway in muscle cells and, if so, whether it is involved in neuregulin-induced glucose uptake.
Several lines of evidence implicate PKB activity in insulininduced glucose uptake both in adipocytes and muscle (32)(33)(34)(35)(36) although some controversy exists (58,59). PKB has been implicated in insulin action on glucose transport in 3T3-L1 adipocytes (36). Our results indicate that PKB blockage alters only partially insulin action on glucose uptake, suggesting that PKB might not be essential in L6E9 myotubes. Moreover, there is no dependence on PKB activity in the NRGs action on glucose transport in muscle cells and HRG does not stimulate PKB phosphorylation in soleus. Then, PKB does not seem to be required in the pathway inducing glucose uptake in response to NRGs. One interesting possibility that should be explored is that a different subcellular localization of PI3K is responsible for the different downstream activities of insulin or neuregulin that trigger glucose transport.
The effect of HRG on glucose transport is also independent of AMPK activity. DNP induces glucose uptake in an additive manner to NRGs action, so AMPK might constitute another pathway, independent of insulin and NRGs, to activate glucose uptake. Both insulin and exercise induce the MAPK pathway, in a PI3K-dependent manner and an AMPK-dependent manner, respectively (60,61). Whereas PD98059 treatment does not affect insulin action on glucose uptake in muscle cells (62), it impairs AICAR stimulation of glucose transport in EDL muscle (61). Thus, the partial impairment of HRG action on glucose uptake under PD98058 treatment suggests that the PI3K-PDK1-PKB pathway activated by NRGs is modulated by the MAPK pathway. We conclude that NRGs require a signaling pathway to induce glucose uptake in L6E9 myotubes that involves ErbB3, PI3K, PDK1, and PKC and is not dependent on PKB activity. A, strips were incubated with insulin (100 nM, 10 min) or HRG (5 nM, 15 min). Lysates were obtained and immunoprecipitated with anti-phosphotyrosine antibody. Pellets were immunoblotted for ErbB2, ErbB3 and ErbB4. Results are shown as representative autoradiograms of three different samples. B, strips were incubated with insulin (100 nM, 60 min) or HRG (5 nM, 120 min), in the absence or presence of wort- | 5,796 | 2004-03-26T00:00:00.000 | [
"Biology",
"Computer Science"
] |
Regulation of choroid plexus development and its functions
The choroid plexus (ChP) is an extensively vascularized tissue that protrudes into the brain ventricular system of all vertebrates. This highly specialized structure, consisting of the polarized epithelial sheet and underlying stroma, serves a spectrum of functions within the central nervous system (CNS), most notably the production of cerebrospinal fluid (CSF). The epithelial cells of the ChP have the competence to tightly modulate the biomolecule composition of CSF, which acts as a milieu functionally connecting ChP with other brain structures. This review aims to eloquently summarize the current knowledge about the development of ChP. We describe the mechanisms that control its early specification from roof plate followed by the formation of proliferative regions—cortical hem and rhombic lips—feeding later development of ChP. Next, we summarized the current knowledge on the maturation of ChP and mechanisms that control its morphological and cellular diversity. Furthermore, we attempted to review the currently available battery of molecular markers and mouse strains available for the research of ChP, and identified some technological shortcomings that must be overcome to accelerate the ChP research field. Overall, the central principle of this review is to highlight ChP as an intriguing and surprisingly poorly known structure that is vital for the development and function of the whole CNS. We believe that our summary will increase the interest in further studies of ChP that aim to describe the molecular and cellular principles guiding the development and function of this tissue.
Introduction
The choroid plexus (ChP) is a conspicuous tissue within the brain ventricular system of all chordates about the lancelets [1] which is widely recognized as the predominant source of the cerebrospinal fluid (CSF) [2]. This greatly specialized structure protrudes into brain cavities from three distinct brain regions: the pair of Telencephalic ChPs (TelChP), also known as lateral ChP, extends symmetrically into lateral ventricles of the telencephalon, (2) Diencephalic ChP (DiChP), also known as 3 ChP, lays in the 3rd vesicle formed in the diencephalon, and finally, (3) Hindbrain ChP (HbChP), also known as 4th ChP, resides in the 4th cavity within the mesencephalic region. (Fig. 1) Regardless of ChP first observation by the Greek anatomist Herophilos already before our era [3], severe pathologies of this tissue reported at the beginning of the last century [4,5] and its fairly welldefined structure available from the 1960s [6,7] the growing interest in ChP among neuroscientists is evident only in the recent decade. Several facts lie at the root of this shift in the perception of ChP tissue and its significance. Besides model and methods availability or sufficiently demonstrated clinical salience, the disclosure of the ChP functional potential to actively secrete morphogens into the CSF and thus participate in the patterning of the adjacent neuronal tissue across the neural tube took the essential lead [8][9][10].
This review aspirates to compile the current knowledge about the ChP with an emphasis on its development. Within this manuscript, we attempted to bring a new synthetic view on the most crucial steps of the ChP development and to comprehensively summarize known functions of molecular regulators involved in the ChP ontogenesis. We distinguished and schematically outlined three key processes of ChP development and comprehensively described the key factors and molecules needed for the specification and formation of this tissue. In addition, our review also provides the reader with clearly summarized essential information about the ChP morphology assessment, the experimental models to study the development of ChP, the commonly used molecular markers, and the currently available specific mouse strains used for studies of ChP. We believe that this summary will provide a comprehensive insight into ChP development and morphogenesis, and will stimulate and efficiently direct further research toward our understanding of this crucial secretory tissue.
ChP morphology
ChP strikes the eyes of the neuroscientists certainly by their gross morphological shape in the lumen of ventricles. Within the central nervous system (CNS), this tissue represents rather atypical structures with huge differences in the overall appearance of individual ChPs (Fig. 1). In the murine brain, the HbChP displays the most complex threedimensional structure with numerous branches bifurcating in the hindbrain cavity whose length and a total number have become evaluation parameters of the HbChP morphology [10]. Moving more anteriorly across the neural tube, the morphology assessment of the frond-like DiChP is still quite challenging and sporadically realized as this miniature structure has less-defined borders [11]. Recently, Langford and colleagues approached this issue by the manual track of Transthyretin + epithelial cells of ChP (ChPEC) on the histological sections with the intention to determine the length of the DiChP [12]. Finally, the tarpaulin-like tissue of TelChPs Fig. 1 The localization of the ChP secretory system within the developing mouse brain. The set of three distinct ChP types (highlighted in red colour) occupies every ventricle formed within the main brain embryonic regions. This includes lateral ventricles marked by the presence of telencephalic TelChPs, 3rd ventricle characterized by the small DiChP and 4th ventricle fills with the robust HbChP. A growing body of evidence indicated differences between the individual types of ChP, not only in their diverse morphological appearance, which can be appreciated on the coronal (panel in dark green) as well as sagittal sections (panel in dark pink) of the murine brain, but also in the expression profiles of the ChP epithelia. This spatial regionalization translates into differences in the composition of the cerebrospinal fluid (CSF) which ChPs produces into the lumen of the ventricular system. ChP choroid plexus, TelChP telencephalic choroid plexus, DiChP diencephalic choroid plexus, HbChP hindbrain choroid plexus, CSF cerebrospinal fluid (CSF), E embryonic day Fig. 2 The distinct morphological shapes of TelChPs in mouse and human. The TelChPs outgrows gradually out of the medial telencephalic wall into the lateral ventricles. TelChPs fill up to 63% of the total cavity area in humans [16] while the murine TelChPs occupied only 27% of the mapped area [17]. These differences are due to a more complicated morphology of human TelChPs. TelChP telencephalic choroid plexus, GW gestation week, E embryonic day occupies nearly the entire telencephalic cavities with an exception of their frontal horns [13]. The morphology of this ChP type has been scored as an evaluation of its total area and length within the lumen of the ventricles [10,12].
Further divergences in the ChP morphology can be found across the individual species, namely between DiChPs and TelChPs of various animals [14]. The striking differences in the morphological appearance have been observed in the human TelChP compared to its murine form. TelChPs within the human brain display a more corrugated character of the tissue resembling the branched morphology of the HbChP (Fig. 2) [15]. However, the molecular mechanism which stands behind this morphological alternation has not been clarified and additional research is needed to determine the regulatory network which participates not only in this phenomenon but also in the general morphological specification of ChP across the ventricles as well as species.
Overall, quantitative analysis of ChP morphology is not trivial. It can be performed by measurements on several consecutive slides of the brain sections, which, however, still does not provide an overall view of ChP within the brain ventricular system. To overcome this obstacle, Perin et al. executed the three-dimensional reconstruction of the rat HbChP tissue using the light sheet microscope scans focusing on its total macrostructure as well as on the individual components such as vasculature [18]. Further information can provide an evaluation of the relative position of ChP within the ventricle. For example, the position of HbChP within the 4th ventricle has been used as a differential criterion in the diagnosis of the human posterior fossa malformations like Vermian hypoplasia [19] or Blake's pouch cyst [20].
ChP structure and cellular composition
The gross structure of ChP, shortly described below, is conserved across the aforementioned ChP types, as well as species. The ChP tissue is constituted out of two core components-the stromal tissue mass which is enveloped by the single monolayer of polarized cuboidal epithelial cells [21].
Mesodermally derived stroma [22] is composed of a plethora of cell types, which involve fibroblasts, glial, immune, neuronal, or endothelial cells [11]. The lastmentioned cells form an atypical vascular system, which is unique to ChP and has the diaphragmed fenestrations in capillary walls. These modulations ensure the influx of water and small molecules from the blood to epithelial cells which are elemental for the CSF secretion [23,24]. The formation and effective function of the ChP vascular network is guaranteed by the presence of the smooth muscle cells together with the pericytes that cover endothelial cells [11].
ChP vasculature is connected to the brain circulation system. LatChPs are supplied by the anterior as well as posterior choroidal arteries which are a continuance of the internal carotid and the vertebral artery, respectively. Similarly, the vertebral artery also delivers the necessary blood to the DiChP. Within the hindbrain region, the basilar and vertebral arteries branched to the anterior and posterior inferior cerebellar arteries which outgrow to the whole HbChP [25,26].
In addition, a broad spectrum of the immune cell types, including monocytes, lymphocytes, basophils, neutrophils, dendritic cells or B cells can be found in the ChP stroma. However, the most abundant immune cell population are macrophages [11] which are derived from the [37] haematopoietic stem cell-derived myeloid cells that are produced within the aorta-gonad-mesonephros region or later in the development by the fetal liver [27]. Interestingly, the apical surface of choroid plexus epithelial cells (ChPEC) is occupied by the specific subpopulation of macrophages termed as epiplexus or Kolmer cells, which scan the content of CSF [28]. The second-mentioned element of the ChP tissue-the ChP epithelium-originates in the neuroectoderm and represents the continuance of the ependymal lining of the ventricular system [29]. The population of ChPEC is distinguishable from others by three major features which are the results of the ChPEC maturation and reflect functions of this secretory tissue. First, the space between ChPEC is sealed by tight junctions which restrain the paracellular transport across the ChPEC and establish the physical barrier between blood and ventricles fulfilled by CSF [30]. Second, the composition and surface of epithelial cells are significantly adjusted to produce CSF. Namely, the epithelial sheet displays higher mitochondrial content [31] together with asymmetrical localisation of various transporters on the apical (e.g., Na + -K + -ATPase or water channels) as well as a basal (e.g., Na + -HCO 3 − exchanger or anion exchanger 2) side [32][33][34][35]. Third, the apical side of the epithelium is characterized by the presence of membrane modulations such as microvilli or cilia which significantly increase the area out of which the CSF is produced [7]. The most typical molecular markers used to distinguish individual cellular pools of developing ChP are summarized in Table 1.
Despite shared basics, ventricle-or age-dependent variations of individual cell types has been shown by recent single-cell analysis of all ChPs across the mouse lifespan. Interestingly, this study uncovered further regionalization within the ChPEC. These spatial divergences were most typical for the epithelium of the HbChP which could be further subdivided into the rostral (Penk + , Shh + , Wnt5a + among others) and the caudal (i.e., Fab3 + , Acad8 + , Rb1 + ) parts [11]. Similarly, despite the molecular details being missing, even the very early study in humans morphologically distinguished anterior and posterior regions within the epithelium of TelChP [36].
CSF fluid
The ChP is a superior secretory tissue that supplies the brain ventricular system with the CSF, producing up to 80% of its total volume [46]. The human ventricular system is filled with approximately 150 ml of CSF and this volume is daily changed 3-4x [47]. There is a general agreement about the CSF flow. Its directional movement starts from lateral ventricles to the 3rd and 4th cavity through the foramen of Monro and cerebral aqueduct, respectively. Afterward, CSF enters the subarachnoid space and spinal cord via the foramen of Magendie, where it is resorbed back to the bloodstream. On top of that, several studies suggested novel ways of CSF absorption highlighting cervical lymphatics or dural venous plexus [48]. The CSF flow is guaranteed by the synchronic cilia beating of the ependymal lining which is tightly regulated by the circadian rhythm [49] together with brainderived hormones [50]. Within the zebrafish ventricular system, the motile ependymal cilia appear to regulate the CSF flow merely within the induvial ventricles, while the CSF motion across cavities is generated by the heartbeats along with the body movements [51].
Formerly, CSF was mainly recognized as a clear body liquid serving a mechanistic function within the CNSensuring its protection, cleanliness, as well as preservation of intracranial pressure needed for the ventricle expansion and proper brain development [52]. Nowadays, CSF is perceived as a dynamic fluid with a heterogeneous morphogen composition that arises from the divergent secretion activity of each ChP epithelia, which is further described in the chapter ChP-mediated brain patterning. Interestingly, this phenomenon seems to decrease over the lifespan with its peak during embryonic development [53,54]. Thus, CSF regionalization contributes significantly to the diverse temporal developmental programs of stem cells pools that reside in the interface of individual ventricles [8,55].
Another noteworthy fact about CSF is its brain nourishing character as this body liquid contains several micronutrients, ions, peptides, or a variety of small RNAs [21,56]. Besides this, the macronutrients like glucose or lactate pass through the fenestrated capillaries and any imbalance in their concentration levels within the CSF can signify some pathological conditions. For example, decreased CSF level of glucose has been connected to bacterial meningitis [57] or leptomeningeal carcinomatosis [58]. Lower lactate parameters within the CSF have been observed in patients with cerebral hypoxia [59] or neurometabolic disorder [60]. On top of this, ChP tissue tightly coordinates the penetration of the leptin from the blood into the CSF fluid. This circulating hormone stimulates receptors of the hypothalamus which culminates in the control of satiety feeling [61].
Development of the ChP
The ChP ontogenesis is a complex and dynamic process that consists of several crucial consecutive steps leading to the formation of the fully-functioning ChP tissue. These stages will be individually discussed in the sections below and are schematized in Fig. 3 and further expanded to the cellular and molecular level in Fig. 4. Of note, ChP development is predominantly studied in the mouse. However, we would like to appreciate the fact that several noteworthy studies, not further discussed in this review, have been done in other models such as zebrafish (for primary references see [62][63][64][65]). Additional imbalance in the knowledge on ChP development can be found in the levels of published details about the formation of individual ChP types. Namely, the DiChP is repeatedly omitted from the analysis probably due to its unclear morphology. On top of this, the vast majority of the studies focused on the development of ChP were performed on the systemic knock-out models which impoverish our knowledge about the time and cell-type requirements for the specific proteins within the individual developing stages of ChP. Only more recently tissue-and/or time-specific ChP knockout models were reported. To help the reader, we hem (CH) in the telencephalon (C) and rhombic lips (RL) in the hindbrain (D). These regions, besides ChPEC, also give rise to the migratory neurons (green). They are represented by Cajal-Retzius cells (from CH) migrating into the developing cortex of the telencephalon (C, E) as well as hindbrain migratory neurons targeting the cerebellum (from upper RL) (D, F). The last step in the ChP ontogenesis is its maturation when ChP gets its typical shape (E, F). Maturation involves further differentiation of ChPEC and the formation of ChP stroma (pink) achieved by the migration of mesoderm-derived stromal cells summarize the mouse strains that are most commonly used to specifically study ChP development in Table 2.
Early ChP specifications
The first sights of ChP tissue within the murine brain have been observed at embryonic day (E) 11.5 when the HbChP start to bulge into the 4th ventricle. Shortly after, TelChPs arise from the medial wall of the telencephalon quickly followed by the DiChP formation within the 3rd cavity [7]. Nonetheless, the specification of the ChP occurs earlier in the development as indicated by grafting experiments of future chick ChP which was capable to form ChP tissue out of its original region [66]. This concept was further extended by the in vitro cultivation of the E8.5 and E9.5 mouse prospective ChP regions from relevant brain regions which also resulted in the formation of ChP like structures [29].
ChPs are adjacent to the roof plate which is a fundamental signaling hub that occupies the dorsal midline of the developing CNS [67]. Roof plate cells, which are located ventrally to the prospective neural crest cells within the neural tube [68], settle the dorsal midline region after the emigration of their neighbors [69]. Several studies have focused on this region to resolve the potential role of these apicobasal polarized cells in the ChP specification.
The lineage tracing experiment of Gdf7 + cells occupying the dorsal midline uncovered the contribution of these cells to the epithelium of HbChP as well as DiChP Endothelial cells [44] from E9.5. The examination of the TelChP showed the presence of a small Gdf7 + cellular population within the anterior part of the TelChP epithelial sheet. Contrarily, the posterior part of the epithelium lacked any progeny of Gdf7 + cells labeling [70] which suggests the existence of regionalization of the mouse TelChP epithelium as described previously in humans [36]. A further study addressing the involvement of roof plate in the formation of HbChP epithelium divided the hindbrain roof plate into three spatiotemporal regions which emerge from the neuroepithelium at distinct time points. Field 1, positive for Wnt1, colonized the dorsal midline at E8-E9.5 and did not participate in the formation of the ChP epithelium.
On the other side, fields 2 and 3 positive for Wnt1 and Gdf7 formed at latter E9.5 from rhombomeres 2-8 and rhombomeres 1, respectively, clearly contributed to the epithelium of HbChP [71]. These observations of Wnt1 significance in the development of HbChP can be further complemented by the phenotype of Wnt1 knock-out mice that lack this tissue [72].
In general, the ChPEC specification requires inhibition of the neural fate within the neuroectoderm from which these cells arise [29]. The regulatory network behind this process has been studied widely in the embryonic telencephalon. For instance, members of the BMP protein family, which are present within the dorsal parts of the telencephalon from E10.5, induce epithelial cell fate within the dorsal midline by the repression of the Lhx1/Foxg1 complex [37,[73][74][75][76]. This observation was validated experimentally in vitro by Watanabe and colleagues who showed that Bmp4 is sufficient to induce ChPEC fate in neural progenitors derived from mouse embryonic stem cells [77]. Furthermore, the misexpression of Bmp receptor Bmpr1 at E10.5 resulted in the expansion of ChPEC, whereas its deletion decreased the number of ChPEC [78]. This specification process is also regulated by the Otx2-Emx2 transcriptional network which is present in this region from E10.5 [79]. The ectopic expression of Emx2 within the dorsal midline has been observed to activate the neuronal cell fate trajectories. The absence of ChPEC in these mice that was accompanied by the decreased levels of Otx2 transcripts directed further research of pro-epithelial factors within the roof plate to Otx2 [80,81]. Indeed, the early deletion of the Otx2 gene at E9.5 caused the absence of all ChPs, while its later ablation (at E15.5) impaired only the development of the HbChP [38]. Altogether, these studies point to the crucial role of Otx2 in diverse time points of the ChP development.
Text Box: How to study ChP and itsdevelopment
Molecular markers To distinguish ChP from the surrounding tissue, several representative markers can be used to visualize the ChP epithelium. For instance, the transporter of thyroid hormone thyroxine T4 throughout the CSF-Transthyretin (Ttr)-is exclusively present in ChPEC from the early stages of ChP development, even before the ChP starts to protrude into ventricles, and its expression persists till adulthood [82]. The differentiated epithelial sheet can be also highlighted by the water channel Aqp1 [10] or other proteins essential for the ChP secretory functions like Slc4a10 [45]. The transition zone between the progenitor and fully differentiated epithelial cells can be visualized by ciliogenesis markers like Shisa8, Ccdc67 or Mcidas [11]. For summary overview see Table 1. Mouse models Several tissue-specific mouse models have been developed to study ChP development. A strong emphasis must be placed on the selection of the adequate Cre drivers, as the transcriptomic regionalization of ChP epithelium can significantly influence the outcome of the analysis. The potential disruption of the ChP morphology arising in these models has been evaluated by the manual measurements of several parameters (e.i. total area, length or number of branches [10,40] on several consecutive tissue slides. In the future, these interminable and challenging quantifications should be replaced by the automatized procedure capable to assess the ChP tissue properties in three dimensions, as lately was shown by Perin and colleagues [18]. For detailed overview of the mouse models see Table 2. ChP organoids The recent generation of the human ChP organoids has uncovered several possible directions of the ChP research, in the terms of the result validation of big-data analyses as well as exploration of ChP functions in vitro or evolutionary studies of ChP development. These TelChP-like protrusions on the cerebelar organoids derived from human pluripotent stem cells displays a high level of similarities to those in vivo regarding the secretome or ChPEC transcriptomic profile. Nonetheless, this state-of-art model lacks the vasculature network within the stroma and as such it is unsuited for the interaction studies of endothelial cells with other cell types within the ChP [41] Moving more posteriorly across the neural tube, the recent investigation of the zinc-finger Zfp423, expressed within the hindbrain roof plate from E8.5, revealed its strong pro-epithelial functions. The C-terminal mutants of this transcriptional factor display hypoplastic HbChP and the downregulation of another key player in the HbChPEC specification-Lmx1a [83,84]. Mutants of this transcriptional factor are characteristic of drastic malformations within the hindbrain region, including an underdeveloped roof plate that culminates in the absence of HbChP [84,85]. One of the transcriptional targets of Lmx1a is Mafb. The analysis of Mafb-deficient mice from E10.5 showed delayed differentiation of HbChP probably caused by increased apoptosis together with the decrease of the proliferation within this tissue. Interestingly, Mafb mutant tissue had lower levels of Lmx1a [39] which suggested the complex regulatory network required for the specification of HbChPEC within the hindbrain roof plate. Further effort is required to fully comprehend the early specification of HbChPEC.
Progenitor zones of the developing ChP epithelial cells
Later, development (E11.5-E14) of TelChPEC and HbCh-PEC is linked with two highly proliferative domains that are contiguous to the developing ChP-telencephalic cortical hem and rhombic lip residing within the hindbrain region. These structures represent not only the progenitor zones of ChPEC but also the origins of migratory neurons.
The cortical hem is a dorsal midline derivative whose development is regulated by the balance between functions of Shh signaling [86,87] and LIM-homeodomain transcription factor Lhx2 [88]. This Wnt/Bmp/Msx-rich region [87] is apart from TelChPEC, one out of three sources of Cajal-Retzius cells that migrate tangentially to the cortex [89] or hippocampal marginal zone [90]. The rhombic lip is a territory placed along the dorsal midline which is further divided into two regions-upper rhombic lip (URL) and lower rhombic lip (LRL) [91]. Apart from the epithelial cells of ChP, the Wls/Math1/Pax6-positive URL [92] gives rise also to tangentially migrating granule progenitor cells of the cerebellum [93], whereas migratory mossy fiber neurons or climbing fiber neurons have been observed to arise from the LRL [94].
A recent study connected these two progenitor domains by a shared expression profile typical of high expression of Rspo genes [11]. Rspo encodes the members of the R-spondin family which act as secreted enhancers of the Wnt/ß-catenin signaling pathways [95]. Moreover, the transcriptomic analysis of all embryonic ChPs identified Rspo2 as a common marker of progenitor cells which, based on the in silico lineage trajectory analysis, can further produce all relevant neurons and epithelia of ChPs [11]. Nonetheless, the regulators that control the balance between neuronal and epithelial differentiation of these progenitors are largely unknown, with few exceptions.
In the cortical hem, the interplay between the Hes-1,-2, and -3 transcriptional factors and Neurogenin2 (Ngn2) contributes to the equilibrium between ChPEC and Cajal-Retzius pools. The Hes mutants displayed upregulation of Ngn2 expression which acts in the favor of Cajal-Retzius cells numbers, while the population of ChPEC decreased [96]. Furthermore, the lineage tracing of Fzd10, encoding a Wnt receptor, showed the exclusive contribution of Fzd10 + cells to the pool of Cajal-Retzius cells [90], indicating that the quantitative balancing of the individual cell types may also occur on the receptor levels and subsequent downstream signaling cascades. Indeed, the most recent analysis of one of the main Wnt signaling branchescanonical Wnt/ß-catenin cascade-within the cortical hem revealed the necessity of its balanced level within this brain region as constitutive activation of ß-catenin enhanced neuronal cell fate trajectories within the cortical hem, to the detriment of the epithelial state [97]. Additionally to this, a recent study pointed out the vital role of double-sex and mab-3 related transcription factor (Dmrt) genes in the formation of hem-derived Cajal-Retzius cells, as the Dmtr1, Dmtr3, and Dmtr1/Dmtr3 double knockout out mice displayed a significant reduction of cortical hem region and lower number of these cells [98]. Unfortunately, the authors did not examine the effect of these transcriptional factors on the ChP epithelial pool.
The regulation cascade which fine-tunes the epithelial vs. nonepithelial pools within the URL is less described as scientists concentrated predominantly on the cerebellum. Still, several studies do exist. For instance, Lmx1a KO mice display several malformations within the cerebellum as well as a significant reduction of HbChP size [85]. The lineage tracing of the Gbx2 + cell progeny showed their contribution both to the granule cell layer of the cerebellum as well as HbChP epithelium. Interestingly, contribution to granule cells persisted to adulthood, whereas contributing to the outgrowing epithelium of HbChP decreased rapidly even during the early embryonic stages [99]. On top of this, a study performed by Huang and colleagues pointed to the role of Shh morphogen in the proliferation rate of LRL progenitors which can significantly affect the size of HbChP [42]. However, other LRL derivatives were not examined in this study.
Overall, the specification of the ChPEC is a complex multistage process that involved a myriad of transcriptional factors or morphogens which ensure the cell fate switch of ChPEC from the initial neuroectodermal character. Furthermore, an additional equitable network of regulators within later progenitor domains of ChP epithelia is needed to complete the differentiation of ChPEC as well as other cell types derived from these brain regions.
The maturation of ChP
Once the ChPEC cells are defined, the whole tissue undergoes the maturation process. Based on the morphological changes of human TelChPEC, this process has been divided into four stages [100]. Epithelial cells of Stage I are characterized by the central localization of nuclei and the absence of apical membrane folding. These cells form pseudostratified epithelium consisting of 2-3 layers. The transition to Stage II involves changes in the shape of epithelial cells to cuboidal and the shift of the nuclei toward the apical side. Additionally, the basal connective tissue starts to form. Stage III is mainly defined by the formation of the cilia on the apical side. Moreover, epithelial cells display a higher number of villi and basally located nuclei. Large modifications also occur within the stroma as the complexity of the vascular network increases gradually. The gradual decrease in the epithelial cell size defines the switch to the last 4th Stage [100].
It is worth mentioning that, under normal circumstances, epithelia of ChPs are quiescent cell populations whose cell division is dramatically reduced early after the maturation [101][102][103]. However, the examination of ChPs in some pathological conditions like stroke [104] or physical brain injuries [105,106] revealed an increased proliferation rate of ChPEC which led to the speculation that the ChP is a depository of neural progenitor cells in the adult mammalian brain [104]. Interestingly, grafting experiments of HbChP cells into spinal cord lesions significantly enhanced both the tissue repair and overall locomotor movements of the animals suffering from spinal cord injury [107,108]. Similarly, courses of stroke [109], hydrocephalus [110], or Huntington's disease [111] have been significantly improved after the transplantation of mature ChPEC into the rodent models of these diseases. This highlights the ChP as tissue with promising therapeutic potential in the spectrum of pathological conditions, mainly because of its capacity to produce the astonishing range of proteins and matrix compartments [112].
Overall, the regulatory programs that lead to the shaping and maturation of fully functioning ChP are not extensively described with a few exceptions. For instance, the Sox9-Co9a3 axis has been recently identified as a crucial factor of the proper epithelial-basal lamina assembly whose disruption leads to the destruction of the ChPEC polarity and hyperpermeability of the blood-CSF barrier [113].
Similar polarity defects and decreased epithelial cohesivity have been observed in the HbChP of Wnt5a mutants [12]. Within its epithelium, Meis transcriptional factors are embroiled in the transcription of the Wnt5a from the early stages, starting at E11.5 [40]. The epithelial-specific deletion of the Wnt5a gene at E11.5 culminates into the severe morphological defects of the typical gross appearance of HbChP, including the decrease of its total area, reduced number of branches and their shortening [10,40]. The severity of these malformations is further aggravated in the systemic Wnt5a knock-out [10,12,40] which can draw attention to the Wnt5a produced by stromal cells. Indeed, Dani and colleagues identified the Wnt5a + mesenchymal cluster [11], which suggests that Wnt5a can also mediate the crosstalk between the individual cell types fundamental for the formation of ChP structure.
Further essentiality of the fine-tuned epithelial-stroma interaction in the ChP maturation has been shown by the Shh knock-out mouse model which phenocopies the Wnt5a mutants. Here, the lack of Shh signals from the epithelium between E12.5 and E14.5 leads to the impaired function of pericytes and ultimately the reduced vascular surface area without the changes in the expression levels of typical proangiogenic genes like Ang-1, Tie-2, or Vegf [44]. Interestingly, the co-culture of endothelial cells with ChP epithelial cells, which have been observed to express Vegf [115], resulted in the increased numbers of ChP typical fenestrations within the endothelial cell population [116]. All these findings suggest a complex, multi-step maturation process of the ChP vasculature which involved a variety of signaling cascades needed during the different embryonic time points as the small number of fenestrated vessels have not been observed as early as E16.5 in the rat model [117].
Our understanding of the interplay between all compartments of maturating ChP can be further enhanced by recent bioinformatic analysis of the potential cell-cell interaction network across the ChP tissue. This extensive prediction, based on the ligand and receptor expression patterns within the individual cell types, highlighted mesenchymal cells (i.e., fibroblast and pericytes) as an organizing center that presumably communicates with the remaining cell types via several distinct signaling pathways including Wnt, Bmp, or Notch protein families [11].
ChP-mediated brain patterning
Besides the blood-CSF barrier formation or the CSF secretory function, the ChP has nowadays been appreciated also as the orchestrator of the long-range signaling in the developing brain. This function of ChP is ensured by the production of numerous signaling molecules into the CSF.
The initial work demonstrating the importance of ChP induced patterning within the embryonic telencephalon comes from Lehtinen et al. These authors identified the presence of Igf2 within the meninges and epithelium of the TelChP, and proved that its release into the CSF culminates in the enhanced proliferation of stem cells within the ventricular-subventricular zone (V-SVZ) [8]. Further patterning of this brain region was reported in the recent study where ChP-derived Semaphorins (Sema)/Neuropilins (Nrp) complexes were identified in the CSF. Here, Gerstmann and colleagues showed a decreased rate of murine progenitor differentiation caused by the Sema3b/Nrp2 induced enhancement of their adhesion properties. In turn, Sema3F/Nrp1 complex seems to decrease cell detachment, thus inducing neuron differentiation [118]. More investigation on this issue has been performed on adult samples. For instance, Otx2 knock-out mice display lower numbers of newborn neurons which integrate into the olfactory bulb. This is caused by the impairment of their migration properties which is connected to the alternation of extracellular matrix within the V-SVZ of animals lacking the CSF-derived Otx2 protein [9]. Furthermore, the transport of one of the most emblematic markers of ChPEC-Transthyretin-across the lateral ventricle seems to be fundamental to maintaining the balance in the differentiation of neurogenic as well as oligodendrogenic cell types within the lateroventral-SVZ progenitor niche [119]. Intriguingly, the recent study by Arnaud et al. also showed the interplay between TelChP and the hippocampus. Shortly, the elevated pool of App within the adult ChP leads to the impaired proliferation within hippocampus stem cells niche and behavioral defects in reversal learning [120]. Another brain region that is sensitive to the external stimuli sent out of ChP is the progenitor niche of the cerebellum. In detail, the Wnt5a protein [10] and Shh [55] have been observed to be emitted from the epithelial monolayer of HbChP and transported across the 4th ventricle to regulate cerebellar proliferation.
The secretome analysis of the media from TelChP and HbChP explants has revealed the presence of nearly 200 proteins that are likely to be produced by ChPs in vivo. The authors highlighted several proteins with the potential to participate in the further brain patterning (e.g., Ctsb or Ctsd in the case of TelChP or HbChP-derived Ec-sod or Penk) [54]; however, these findings need further validations in the identical model.
Aging of the ChP and its related diseases
Severe alterations have been observed within the structure of ChP during aging. One of the first described is the so-called Biondi bodies [121]. These filamentous, lipid dropletsassociated structures develop preferably in the cytoplasm of aged human ChPEC and it is speculated that these bodies have a destructive impact on the epithelial sheet [122]. Furthermore, the transcriptomic analysis of the adult and aged mouse ChP showed the expression shift related to the IL-1B signaling within the aging macrophages, endothelial, as well as mesenchymal cells. The authors ascribe this observation to enhance macrophage migration and subsequent infiltration of the CSF-blood barrier occurring in older individuals [11].
Similarly, the changes in the status or gene expression pattern of individual ChP cell types have been connected to the presence of various infectious agents in the body, for which the ChP represents the gateway to the CNS. For instance, the Zika virus disturbs the brain cortex via the ChP pericytes' infection, impairment of the ChP epithelium, and subsequent infiltration into the CSF [123]. Single-cell analysis of human ChP tissue from COVID-19-positive patients revealed inflammatory associated transcriptional changes within the ChP epithelium [124] which is in line with the previous study performed on the latest cutting-edge model in the ChP field-human ChP organoids [41]. Here, the authors showed the leakage of the CSF-blood barrier as well as the inflammatory transition of ChPEC induced by the COVID-19 infection [125]. Intriguingly, the recent study by Carloni and colleagues uncovers the defense mechanism within the ChP tissue restricting the agens entrance and further spreading across the CNS. The authors, focusing on inflammatory bowel diseases, first detected the upregulation of the Wnt/ßcatenin signaling pathway in the ChP endothelial cells from infected animals causing the increased permeability of the ChP vascular network. To prevent further infections, the recruitment of inflammatory cells occurs resulting in the shutdown of the CSF-blood barrier [126].
Beside the entering point, ChP may also act as a reservoir of infection agents as Liu et al. recently showed studying the Ebola virus. Within this study, the authors showed that even after the monoclonal antibody-based treatment, the virus particles were detected within the ChP macrophage cell pool and their subsequent activity resulted in the lethal recrudescence of Ebola diseases [127].
In addition to mental disorders, which have been linked to ChP tissue dysfunction, e.g., schizophrenia [128], bipolar disorder [129], depression [130] anxiety [131] or Alzheimer's disease [132], notorious disorders of this structure are ChP carcinoma and ChP sarcoma. These ChP-based tumors are rare diseases (accounting for 2-4% of intracranial tumors in children and 0.5% in adults [133]) that are most typical for the TelChP and HbChP. Interestingly, 80% of TelChP tumors have been observed in children, while the incidence of HbChP tumors is more uniformly distributed across age groups [134]. Characteristic clinical features of ChP tumors, which have been allied to several genetic conditions, such as Aicardi syndrome [135], Li-Fraumeni syndrome [136], or Von Hippel-Lindau disease [137], are hydrocephalus and increased intracranial pressure which manifest in vomiting or headaches [138]. On the molecular level, ChP tumors have been connected to the alternation in the TP53 [139], PDGFR [140], Notch [141], or Shh signaling cascade [142].
Apart from the aforementioned syndromes, mouse models of ciliopathies such as Bardet-Biedl syndrome [143], Joubert syndrome [144], or Meckel-Gruber syndrome [145], which have already been associated with the ChP cyst [146], show cilia loss at the epithelial monolayer of the ChP. This is accompanied by the disruption of the ChPEC integrity and hydrocephalus. Altogether, these studies underline the relevance of the ChP in the research and/or diagnosis of these ciliopathies.
Conclusion and future directions
The enormous effort has been recently put into the ChP research with the intention to fully comprehend every aspect of this secretory organ-its functions, morphogenesis, as well as mechanisms that result in the onset of ChP-related diseases. The range of models used in the ChP field is growing rapidly and includes zebrafish line with fluorescent ChPECs [63], the time-and/or tissue-specific conditional mouse knockouts [10,55], or human ChP organoids [41]. This is accompanied by the boost of transcriptomic [11,54,128] and proteomic datasets [41,54] of ChP tissue or CSF. Despite all of this, many key issues related to the ChP development, and its functions remain unresolved. First, are there any additional molecular mechanisms of ChP epithelium specification which would be shared by all ChP types? Then, is the predicted common origin of ChPs in the Rspo2 + progenitors evolutionary conserved? What is the full extent of the epithelium-stroma interactions and what is its impact on the morphology of individual ChPs? What is the extent of regionalization of CSF within the individual ventricles? How is such diversity maintained? What is the full scope of brain regions which respond to the biomolecules produced by the ChP? Overall, further studies within the basic research framework are needed to put the entire picture of developmental processes and the functioning of this unique structure together. Only once we broaden our basic knowledge of ChP, we can understand better the onset of the ChP pathologies, and ultimately develop the effective treatments of ChP-related diseases.
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:// creat iveco mmons. org/ licen ses/ by/4. 0/. | 9,272.2 | 2022-05-19T00:00:00.000 | [
"Biology"
] |
Run-up, inundation, and sediment characteristics of the 22 December 2018 Sunda Strait tsunami, Indonesia
A tsunami caused by a flank collapse of the southwest part of the Anak Krakatau volcano occurred on 22 December 2018. The tsunami affected the coastal areas located at the edge of the Sunda Strait, Indonesia. To gain an understanding of the tsunami event, field surveys were conducted a month after the incident. The surveys included measurements of run-up height, inundation distance, tsunami direction, and sediment characteristics at 20 selected sites. The survey results revealed that the run-up height reached 9.2 m in Tanjungjaya and an inundation distance of 286.8 m was found at Cagar Alam, part of Ujung Kulon National Park. The tsunami propagated radially from Anak Krakatau and reached the coastal zone with a direction between 25 and 350 from the north. Sediment samples were collected at 27 points in tsunami deposits with a sediment thickness of 1.5–12.7 cm. The average distance from the coast of the area with significant sediment deposits and the deposit limit are 45 % and 73 % of the inundation distance, respectively. Sand sheets were sporadic, highly variable, and highly influenced by topography. Grain sizes in the deposit area were finer than those at their sources. The sizes ranged from fine sand to boulders, with medium sand and coarse sand being dominant. All sediment samples had a well-sorted distribution. An assessment of the boulder movements indicates that the tsunami run-up had minimum velocities of 4.0–4.5 m s−1.
Introduction
A tsunami took place in the Sunda Strait on 22 December 2018, at 22:00 western Indonesia time (UTC+7). It shocked the local residents because it came without any warning signs, such as earthquake shocks. The source of the tsunami was the Anak Krakatau volcano, a sea mountain in the middle of the Sunda Strait. The southwestern slope of the mountain experienced a landslide that resulted in the movement of seawater, which propagated to land in the form of a tsunami wave. When the tsunami reached land, its large energy caused a lot of damage and casualties. Records obtained from the Indonesian National Disaster Management Authority (Indonesian -Badan Nasional Penanggulangan Bencana; BNPB) show 430 deaths, 1015 collapsed houses, and a lot of other damage (e.g., to seawalls, revetments, jetties, boats, and cars). The affected areas in Banten Province include the Pandeglang and Serang districts, and those in Lampung Province include the regencies of South Lampung, Tanggamus, and Pesawaran.
The Sunda Strait is home to the Krakatau (or Krakatoa) volcano. It is famous for the 1883 Krakatau eruption, which caused a 30 m tsunami that led to 36 000 fatalities and affected Earth's climate and weather for several weeks, as reported by Verbeek (1884). The 1883 eruption of Krakatau and the resulting tsunami have been widely discussed (e.g., Yokoyama, 1987;Camus et al., 1992;Maeno and Imamura, 2011;Paris et al., 2014). A young volcano called Anak Krakatau (Child of Krakatau) appeared above sea level in 1929. It grew to 338 m a.s.l. (above sea level) in Septem-ber 2018. This very active volcano was the source of the tsunami discussed in the present study.
The generation of the tsunami that occurred on 22 December 2018 in the Sunda Strait was triggered by the collapse of a flank in the southwest part of the Anak Krakatau volcano. Satellite imagery shows that the area of the body of the marine volcano that was lost was 64 ha; the collapsed volume was estimated to be 150 000 000-180 000 000 m 3 (Kasbani, 2019). As a result of the collapse of some of the volcano's body, the volcano's height decreased from 338 to 110 m a.s.l. The tsunami caused by the collapse of the Anak Krakatau flank was investigated by Giachetti et al. (2012). They used a numerical model to simulate an unstable-flank collapse in the southwest part of Anak Krakatau. A hypothetical volume of 280 000 000 m 3 produces a wave with an initial height of 43 m on the Sertung, Panjang, and Rakata islands that then spreads to the beaches on the western part of Java, including Merak, Anyer, Carita, Panimbang, and Labuhan, and Sumatra, reaching the city of Bandar Lampung. The actual area affected by the 22 December event is consistent with their model but different in magnitude due to the use of the worstcase scenario in the simulation. The numerical modeling of the December 2018 Anak Krakatau tsunami was performed by Heidarzadeh et al. (2020), while Muhari et al. (2019) conducted field surveys of this event to record tsunami run-up along the coasts of the Sunda Strait.
Tsunamis in the Sunda Strait are of great concern because the strait is important both locally and globally. It connects the two main islands of Java and Sumatra, whose populations account for 79 % of Indonesia's population (BPS -Statistics Indonesia, 2019). About 6.9 million people live in the coastal area of the strait in Banten Province and Lampung Province (BPS -Statistics of Banten Province, 2019; BPS -Statistics of Lampung Province, 2019). The strait, between Merak and Bakauheni, is the busiest interisland crossing in Indonesia, with 17 824 392 passengers and 4 218 548 vehicles in 2018 (Dirjen Perhubungan Darat, 2019). The strait is also an international route for large ships. It is the second-most crowded waterway after the Strait of Malacca, with 70 000 vessels a year passing it (Soeriaatmadja, 2019). There are three industrial regions at the edge of the strait, namely Cilegon, Serang, and Tanggamus. There is also a special economic zone in this region, namely Tanjung Lesung. The beaches in the strait are a tourist destination. There are two UNESCO world heritage sites across from each other: one on the western tip of Java (Ujung Kulon National Park) and the other on the southern tip of Sumatra (Bukit Barisan Selatan National Park). Bandar Lampung, which has a population of 1 million, is the provincial capital and faces the strait directly. Jakarta, the capital of the Republic of Indonesia, is relatively close to the strait.
Post-tsunami field surveys were conducted to obtain data for future mitigation and development activities in the region. The surveys began exactly a month after the tsunami, 22 January, and ended on 28 January 2019. Our team included people from Indonesia and Taiwan. We carried out measurements of the run-up height and inundation distance of the tsunami. In addition, we also identified flow directions and sediment deposits caused by the tsunami.
Study area
The tsunami has had a serious impact on life in the surrounding area. The affected area covers the coastal area on the western tip of Java (Banten Province) and the southern tip of Sumatra (Lampung Province). Banten Province covers two districts, namely Serang and Pandeglang. Lampung Province covers South Lampung Regency and the provincial capital of Lampung, namely Bandar Lampung. Post-tsunami field surveys were conducted in these areas. We selected 20 sites for observation and measurement (Table 1 and Fig. 1). These sites are located along 140 km of coast on Java and 80 km of coast on Sumatra. Sites 13 and 14 were reached by boat because of the difficult land route.
Method
Measurements of run-up and inundation were conducted using conservative terrestrial surveying methods with optical and laser devices (e.g., total stations, handheld GPS devices, and laser distance meters). We measured run-up and inundation based on the coastline at the time of the surveys. Run-up was corrected to calculate heights above sea level because the tide level at the time of the actual tsunami was different from the tidal level at the time of the surveys. We use WX-Tide software version 4.7 for correcting elevation. Elevation values of each survey site were corrected with the nearest tidal gauge available. We used three stations in Ciwandan, Labuhan, and Teluk Betung for corrections. The maximum run-up and inundation limits are based on remaining tsunami trail marks at measurement locations. The tracks were in the form of debris, fallen trees, plants that had changed color, and damage to buildings. The observed damage to buildings and structures was caused by the tsunami because there were no other causes, such as the earthquake sand liquefaction in the 2018 Sulawesi tsunami (Widiyanto et al., 2019). In addition, information regarding inundation limits and the highest run-up was also obtained from eyewitnesses. IOC Manuals andGuides No. 37 (1998 and and field survey reports (Maramai and Tinti, 1997;Farreras, 2000;Matsutomi et al., 2001;Fritz and Okal, 2008) were used as guidelines for the implementation of this field survey.
Sediment samples were collected from selected points at measurement locations that could be in the swash zone, nearshore, on berms, or in deposit areas (Table 2 and Fig. 2). The measurement of deposit thickness in sandy sheets was carried out by digging a number of shallow holes. The measured thickness was considered to be near the maximum thickness. This method was qualitative and subjective because tsunami deposits are discontinuous, sporadic, and scattered over a flooded area. Sand sheets deposited on land vary greatly due to the influence of sedimentary sources and to-pography. A pit was made at each selected point to observe layers suspected to have been produced by the tsunami. We took only one sample at each pit for laboratory testing and did not take vertical samples at intervals of 1 cm, as done by some researchers (Gelfenbaum and Jaffe, 2003;Hawkes et al., 2007;Srinivasalu et al., 2007;Srisutam and Wagner, 2010), because a detailed analysis was not our focus, particularly not the number of tsunami waves and the vertical variation in sediment. The grain size of the samples obtained from the field was tested using ASTM standard sieve analysis. In addition, we investigated boulder movement at four sites.
Run-up
The run-up was measured by determining the height difference between the highest point of seawater rise onto land and the coastline. Run-up is influenced by the characteristics of the ground surface and slope. The measurement results from our field surveys show that run-up ranged from 1 to 9 m (Table 1 and Fig. 1). The values in the table and figure have been corrected for tide to obtain elevation from sea level at the time of the tsunami. A run-up height of about 1 m was found in many locations at which no damage was found. The highest run-up was found at the Tanjungjaya 2, Cagar Alam, and Kunjir sites, with heights of 9.2, 7.6, and 7.8 m, respectively. Site Tanjungjaya 2 is located at Cipenyu Beach. reported a maximum run-up height of 13 m in the area around Tanjungjaya-Cipenyu Beach. This value is significantly different from our maximum run-up since we measure in a flat-valley part of Cipenyu Beach, while measured on the hilly coast of the beach area. The Tanjungjaya 2 site is a private resort with many tourists. The topography is relatively flat but suddenly rises at a distance of about 250 m from the coastline due to a long hill. A large boulder moved by the tsunami was found at this site. The Cagar Alam (sanctuary) site is part of the Ujung Kulon National Park. This site has a flat topography and a dense forest. The guard post and water police office were completely destroyed by the tsunami. The Kunjir site, which is densely inhabited, had the second-highest number of victims after the Sumur site. This site is located on Sumatra about 38 km from the tsunami source.
Our surveyed run-up heights are compared with the published tide gauge records of Heidarzadeh et al. (2020). Three tide gauges mentioned in the article (Ciwandan, Marina Jambu, and Panjang) are used. Maximum amplitudes at Ciwandan, Marina Jambu, and Panjang are 1.15, 2.8, and 1.25 m, respectively. The Ciwandan tide gauge is used to evaluate run-up heights at sites Karangsuraga, Pasauran, Sukarame, and Pejamben. The Marina Jambu tide gauge is used to evaluate run-up heights at sites Sukamaju, Karangsari, Tanjungjaya 1, Tanjung Lesung 1-3, Tanjungjaya 2, Banyuasih, Kertajaya Sumur, and Cagar Alam. Additionally, the Panjang tide gauge is used to evaluate the sites of Bumi Waras, Wayurang 1, Wayurang 2, Kotaguring, Sukaraja, and Kunjir. It is indicated that averaged run-up heights of each site associated with the tide gauge are 4, 1.15, and 3.1 times larger than the maximum amplitude at Ciwandan, Marina Jambu, and Panjang, respectively. The sites are relatively far from the tide gauge.
Inundation
The distance from the run-up point to the coastline is defined as the inundation distance (IOC Manuals and Guides No. 37, 2014). This distance can be easily obtained using a distance measurement instrument or GPS. We used a total station for this purpose. The coastline's elevation in our survey was corrected with the tide elevation measured by several tide gauges in the Sunda Strait. The results of our field measurements show that the inundation distance ranged from 10 to 290 m (Table 1 and Fig. 1). The wave with an inundation distance of 11 m and a run-up of 1.2 m at site 15 (Bumi Waras) was not felt by the population. This site was chosen to represent the area of Bandar Lampung, which is the capital of Lampung Province. This city has a population of 1 million (2018) and must thus develop tsunami mitigation strategies. The longest inundation distance was found at site 14 (286.8 m), in the sanctuary, which also had a high run-up. At this site, measurements were made near the mouth of a small river. Long inundation distances may be caused by relatively flat topography with relatively few obstacles. Tsunamis may also travel faster through a stream channel. A relatively long inundation (263.1 m) was also found at Tanjungjaya 2, the site with the highest run-up. This site is in the form of a valley plain with a small stream. The slope in the valley is relatively flat and suddenly changes steeply in hilly areas within 250-300 m of the coastline. The run-up point we recorded is located on the slope change from mild to steep. These mildly sloping and steep areas have slopes of approximately 0.025 and 0.06, respectively. Local people call this area Cipenyu Beach. This is a sandy beach flanked by cliffs or hilly beaches. Fortunately, not many people live around this site other than at a resort complex, which suffered severe damage.
Tsunami wave direction
The tsunami spread out from its source on the Anak Krakatau volcano to the beaches at the edge of the Sunda Strait. To determine the direction of the tsunami that arrived at the beach, we obtained information from eyewitnesses. The tsunami hit at night, and thus its arrival was difficult to see. Fortunately, it hit during a full-moon period, so there was some light. In addition to eyewitness accounts, we obtained evidence in the field related to the direction of the tsunami propagation. The evidence was in the form of fallen tree trunks, sloping vegetation and shrubs, damaged buildings, and building components carried away by the flow (Fig. 3).
Our survey results show that the direction of the tsunami propagation was radial from the source (Fig. 4). The tsunami traveled east along the coast between Anyer and Labuhan (sites 1-6). In the vicinity of Tanjung Lesung (sites 7-13), the tsunami was directed to the southeast, and in part of Ujung Kulon National Park (Site 14), the tsunami was directed southward. The tsunami was directed to the north and slightly to the northeast on Sumatra (sites 15-19). The westward tsunami toward the Tanggamus area was relatively small and insignificant. We did not include this area in the survey. The smaller magnitude of the tsunami to the west is likely due to obstruction by the island of Sertung and bathymetry factors. The Anak Krakatau mountain avalanche had a southwest direction, but the tsunami in this direction had no impact on human life because it leads to the open sea (the Indian Ocean), with increasing depth from the tsunami source. The precise tsunami wave direction from the north as it arrived in the coastal area is given in Table 1 for the field survey sites. The direction ranges from 25 to 350 • from the north, indicating radial propagation of the tsunami wave. 7 Sediment characteristics 7.1 Tsunami deposits Prehistoric (paleo)tsunamis have been identified from sediment deposits in several studies (Atwater, 1992;Dawson and Shi, 2000;Peters et al., 2007). Sediment deposits can be used to explain and reconstruct significant tsunami events (Dawson et al., 1995;Van Den Bergh et al., 2003). The present study attempts to describe the impact of a recent tsunami on sediment movement around the coastal area. The tsunami carried sediment from the coast inland. However, not all sites that we measured had significant sediment deposits. Only places with a sufficient source of material had clearly observable sediment deposits (Fig. 5). The survey sites used for sediment samples are shown in Fig. 2. Sediment deposits generally do not spread evenly and continuously but are separated at certain locations, which allow the deposits to settle. Topography controls sediment deposits; for instance, there are more sediment deposits at ground surface depressions.
The best location for the observation of tsunami sediment is about 50-200 m inland from the coastline (Srisutam and Wagner, 2010) or about 50-400 m inland (Moore et al., 2006), a range used for the 2004 Sumatra-Andaman tsunami. In this present study, the 12 deposit pits were 9-195 m from the shoreline. Four deposit pits were less than 50 m from the shoreline (Fig. 6). Three of them were at sites 1, 2, and 7, respectively, due to the short inundation and beach scarp. Another was at site 13 (Kertajaya Sumur), where high-density housing blocked the sediment transport and created a deposit at a short distance from the shoreline.
The interpretation of tsunami magnitude, especially runup and inundation based on tsunami deposits, is challenging (Dawson and Shi, 2000). However, the relationship between deposits and run-up or inundation is still not convincing because of the high variability in tsunami deposits in terms of thickness and location. Soulsby et al. (2007) proposed a mathematical model for reconstructing tsunami runup from sedimentary characteristics. The run-up distance for sediment is related to the run-up distance for water as where R s is the maximum distance inland for sediment deposition, R w is the run-up limit for water, α is as shown in Eq.
(2), and γ is a comparison factor between uprush time and total uprush-plus-backwash time.
where w s is the settling velocity, T is a period from the time of the first wetting to final drying of inundated ground, and H is the tsunami height. Figure 6 shows the distance of measured sediment deposition and water run-up compared to the distance of theoretical sediment deposition calculated using Eqs.
(1) and (2); the results are in good agreement. However, Sukarame, Tanjungjaya 1, and Cagar Alam do not fit well. The three locations have morphological conditions that may not be ideal for applying the theoretical approach. Sukarame has a beach scarp, and the tsunami flows across a stream around 90 m from the coastline. Tanjungjaya 1 also has a beach scarp, and there is a seawall, although not so high, that may block the sediment movement. Even though Tanjungjaya 1 has abundant material, a low-amplitude tsunami caused little sand transport. Cagar Alam has a relatively larger stream than Sukarame. In addition, Cagar Alam has dense vegetation since it is part of a national park. As can be seen in Fig. 6, the distance from the coast of the area with significant sediment deposits caused by the tsunami (deposit pit distance) varied in the range of 9-195 m (average 82 m) from the shoreline or 10 %-70 % (average 45 %) of the inundation distance. Moreover, the limits where tsunami sediments were found that can still be seen with the naked eye (deposit limit distance) are around 11-244 m (average 122 m) from the coastline or 30 %-90 % (73 % on average) of the inundation distance. Figure 6 also presents the elevation of the deposit pit and deposit limit to the run-up elevation. The average elevation of the deposit pit and deposit limit are 43 % and 72 % of the run-up elevation, respectively.
The sediment samples were tested for gradations in the laboratory. The results for each site are shown in Fig. 8. The sediment in the deposit areas on land is generally finer than at the sources on the beach (nearshore or swash zone). The characteristics of the sediment are discussed in Sect. 7.3. The sediment deposition thickness varied greatly from one point to another at the survey sites. We chose a test pit with significant thickness for sampling. The thickness was likely near the maximum sediment thickness in the area.
Boulder movement
Coastal boulder accumulation is usually associated with high-energy events (tsunamis, hurricanes, or powerful storms). A characteristic of many tsunamis is their ability Figure 6. Positions of deposit pits and deposit limit compared to inundation distance and run-up elevation. Theoretical approach for tsunami deposit only available for distance not for elevation. to deposit boulders across coastal zones (Dawson and Shi, 2000). Extreme storms also have the ability to deposit boulders (Morton et al., 2007;Richmond et al., 2010). The interpretation of boulders is difficult along coasts where both storms and tsunamis have occurred. We identified boulders moved by a tsunami wave and run-up at three survey sites based on information from eyewitnesses. Eyewitnesses said that these boulders were in new positions after the tsunami. In addition, from the physical criteria given by Morton et al. (2007) and Paris et al. (2010), it was most likely that the boulders were moved by the tsunami. One of the criteria by Morton et al. (2007) that we found in these sites is a relatively thin (average < 25 cm) bed composed of normally graded sand consisting of a single structureless bed or a bed with only a few thin layers. Sediment thickness around the boulders was very thin. Paris et al. (2010) reported regarding boulder and fine-sediment transport and deposition by the 2004 Sumatra-Andaman tsunami that most of the sediments deposited on land, from fine sands to coral boul-ders, came from offshore locations, with very high values of shear velocity (> 30 cm s −1 ). The boulders we found came from nearshore locations, and most of the boulders were submerged. We estimate that high shear velocities should occur to transport them. These velocities were most likely made possible by the 22 December 2018 Sunda Strait tsunami. At about the time of the tsunami event, a tropical cyclone called Kenanga formed in the Indian Ocean about 1400 km from the Sunda Strait. Kenanga had a speed of 75 km h −1 and was active from 15 to 18 December 2018 (Prabowo, 2019). The influence of this cyclone was weak in the coastal zone, and thus it was unlikely to have moved the boulders.
The largest boulder, measuring 2.7 m in diameter (10.4 t), was found at site 11 (Tanjungjaya 2), as shown in Fig. 7a. This boulder moved from its original point in the swash zone to 82 m inland. Other smaller stones were scattered around it. In Tanjung Lesung (sites 8- and villas was partly destroyed and moved ashore. A seawall chunk, measuring 1 m × 1 m × 4.2 m (9.5 t), made from rubble mound and mortar moved as far as 30 m from its place of origin (Fig. 7b). Other smaller chunks were also moved. The characteristics of the boulders moved by a tsunami can be used to estimate the associated flow velocities. For instance, the 2004 Sumatra-Andaman tsunami had flow velocities of 3-13 m s −1 . This tsunami drove a 7.7 t calcareous boulder 200 m and an 11 t coral boulder as far as 900 m (Paris et al., 2010). The velocity was calculated as where µ is the friction coefficient, m is the boulder mass (kg), g is the gravitational acceleration, C d is the drag coefficient, A n is the area of the boulder projected normal to the flow (m 2 ), and ρ w is the density of seawater (kg m −3 ). The velocities were calculated from Eq.
Sand size statistics
The results of sieve analysis, namely sand grain size distributions, are shown as a cumulative distribution curve of sand grain size. Figure 8 shows the cumulative distribution curve for 12 sites. From the curve, various diameter values were determined, including d 95 , d 84 , d 50 , d 16 , and d 5 . From these diameters, other statistics can be determined, namely the mean, standard deviation, skewness, and kurtosis (Table 2). The mean can be used for grain size classification. The standard deviation is a measure of range that shows the uniformity of a sand sample. A perfectly sorted sample will have sand of the same diameter, whereas poorly sorted sand will have a wide size range. Beach sand size distributions with a standard deviation of ≤ 0.5 are considered well sorted, and those with a standard deviation of ≥ 1 are assumed to be poorly sorted. Skewness occurs when the sand size distribution is not symmetrical. A negative skewness value indicates that the distribution tends toward a small ϕ value (large grain size). Kurtosis determines the peakedness of the size distribution. The normal distribution has a kurtosis value of 0.65. If the distribution is more diffuse and wider than the normal distribution, the kurtosis value will be less than 0.65 (Dean and Dalrymple, 2004). From our results, the mean values show that medium and coarse were the dominant types of sand in the sample. Very coarse sand, granular sand, and pebbles were found at the Tanjungjaya 2 and Sukarame sites. Fine and very fine sand were also identified at several sites. The range of grain sizes found in the study area depends on the available source material. Wentworth classification was used to assess the grain size. All samples had negative standard deviations, indicating that they had well-sorted distributions.
Of the 10 samples taken from the swash zone, 7 had negative skewness, which indicates a large ϕ value and an erosive tendency in the zone. The numbers of samples with positive and negative skewness were similar (7 and 6, respectively). Some deposit samples were taken at a distance of less than 50 m from the coastline, which may still be an erosive environment.
The kurtosis of the tsunami sediment indicates that grain size distributions were flat to peaked. Generally, the major sources of tsunami sediment are swash zones and berm or dune zone sands, where coarse to medium sands are dominant. A minor source of tsunami sediment is the shoreface, where fine to very fine sands are dominant. However, for a coastal area where the shoreface slope is mild, the major source of tsunami sediment is the shoreface. Table 2 provides kurtosis values in which the distribution of sediment ranges from very platykurtic to very leptokurtic.
Conclusion
We selected 20 sites on Java and Sumatra to observe the impact of the December 2018 Sunda Strait tsunami, which was caused by a mass movement of an Anak Krakatau volcano flank. The survey results revealed that the run-up height ranged from 1 to 9 m, the inundation distance was 10 to 290 m, and the direction of the tsunami was between 25 and 350 • . The highest run-up (9.2 m) was found at site Tanjungjaya 2. The longest inundation distance (286.8 m) was found at site Cagar Alam which contains a forest area, part of a national park, and a UNESCO heritage site. Sediment samples were taken from 27 points in tsunami deposits with a sediment thickness of 1.5-12 cm. The distance from the coast of the area with significant sediment deposits caused by the tsunami varied in the range of 9-195 m (average 82 m) from the shoreline or 10 %-70 % (average 45 %) of the inundation distance. Meanwhile, deposit limit distances were around 11-244 m (average 122 m) or 30 %-90 % (average 73 %) of the inundation distance. The average elevation of deposit pit and deposit limit are 43 % and 72 % of the run-up elevation, respectively. Sediment material larger than coarse sand (granular sand, pebbles, cobbles, and boulders) was found at several locations. The largest boulder had a diameter of 2.7 m and a weight of 10.4 t. From the boulder movement, the tsunami velocity at the ground surface is estimated to be more than 4.5 m s −1 . Sand size statistics are also given in this report. The sediment grain size ranged from very fine sand to boulders, with medium sand (diameter 0.25-0.5 mm) and coarse sand (diameter 0.5-1.0 mm) being dominant. All sediment samples tested in the laboratory had a well-sorted distribution, indicating that the grain sizes were relatively uniform.
Data availability. Data can be made available by the authors upon request.
Author contributions. WW and SCH designed the field survey. WW and WCL conducted the field survey in the disaster area. WW and SCH wrote the original draft. SCH managed funding acquisition. WBC corresponded with Taiwanese and Indonesian government offices. PBS and RTI provided early information and mobilized surveyors and measurement tools. All authors contributed to the discussion and interpretation of the results.
Competing interests. The authors declare that they have no conflict of interest. | 6,822.2 | 2020-04-06T00:00:00.000 | [
"Geology",
"Environmental Science"
] |
RNA Sequencing to Explore Dominant Isoform Switch During CCR6+ Memory T Cell Activation
Alternative splicing (AS) is an essential, but under-investigated component of T-cell function during immune responses. Recent developments in RNA sequencing (RNA-seq) technologies, combined with the advent of computational tools, have enabled transcriptome-wide studies of AS at an unprecedented scale and resolution. In this paper, we analysed AS in an RNA-seq dataset previously generated to investigate the expression changes during T-cell maturation and antigen stimulation. Eight genes were identied with their most dominant isoforms switched during T cell activation. Of those, seven genes either directly control cell cycle progression or are oncogenes. We selected CDKN2C, FBXO5, NT5E and NET1 for discussion of the functional importance of AS of these genes. Our case study demonstrates that combining AS and gene expression analyses derives greater biological information and deeper insights from RNA-seq datasets than gene expression analysis alone.
Introduction
Alternative splicing (AS) rapidly converts the product of a single gene from one isoform to another 1 . Thus, AS is crucial for generating immediate biological responses, and allows one gene to encode instructions for making multiple proteins with distinct functions 2 . CD45 is a good example. Its transcript undergoes AS ( Supplementary Fig. S1) to generate proteins with disparate functions 3 . For example, T and B cells express separate isoforms which change as the cells undergo activation and differentiation 4 . The large CD45 isoforms expressed by naïve T cells are a stronger break on activation than the shorter CD45 isoform, CD45RO, which is expressed by memory T cells. AS in uences the risk of several diseases, including neurodegenerative disorders, cancer, immune and infectious diseases, cardiovascular and metabolic diseases 2,5−7 . Studying AS holds promise for the development of clinical biomarkers and novel therapies in this era of precision medicine 8 .
AS isoforms were historically characterized by reverse transcription polymerase chain reaction (RT-PCR) and expressed sequence tags (ESTs). Experimental approaches graduated to the genome-wide scale with the development of AS microarrays. These successfully identi ed AS across tissues, cellular states, and species 9,10 . However, these technologies have low throughput (RT-PCR and ESTs), high noise (ESTs and AS microarrays), or only capture known splicing events (RT-PCR and splicing microarray). More recently, RNA sequencing (RNA-seq) has improved the study of AS in several ways 11 . Compared with microarraybased transcriptome pro ling, RNA-seq has a wider dynamic range and avoids some of the technical limitations such as varying probe performance and cross-hybridization 12 , and provides novel insights into AS 13 . Computational tools such as MISO 14 , MAJIQ 15 , rMATs 16 and LeafCut 17 have been developed to detect both known and novel splicing events from RNA-seq data. Other tools such as RSEM 18 , Kallisto 19 , and Salmon 20 can be applied to analyze and quantify known or annotated transcript isoforms.
Vitting-Seerup et al characterized isoform switching from > 5,500 cancer patients' RNA-seq data covering 12 solid cancer types and identi ed many isoform switches as powerful biomarkers: 31 switches were highly predictive of patient survival independent of cancer types 21 . Their study indicates that isoform switches with predicted functional consequences are common and important in dysfunctional cells, which in turn means that expression change should be analyzed at the isoform level. Our previous study also demonstrated additional mechanistic insights can be gained through interrogation of AS in addition to conventional gene-level analysis of RNA-seq data 22 Our understanding of immune response regulation by AS is scant compared to oncology 24 . AS microarray pro ling identi ed extensive novel AS changes in activated T cells suggesting a key role for AS in regulating the mammalian immune response 25 . The types of genes controlled by AS during T-cell activation are different from those governed by changes in transcript levels; AS is associated with cellcycle regulation, whereas alterations in transcript abundance dictate changes in immune defense and cytoskeletal architecture 25 . Previously, we compared the transcriptomes of activated CCR6 + memory Tcells by RNA-seq and microarray. A plethora of differentially expressed genes were identi ed by both platforms, but RNA-seq was superior in differentiating biologically critical isoforms 12 . Prior research de ned transcriptional changes that regulate protein expression during T-cell maturation and antigen stimulation. Here, we build upon this work by using the same RNA-seq dataset to globally analyze AS during T-cell stimulation.
Materials And Methods
RNA-seq dataset.
The raw RNA-seq data from this study was downloaded from the NCBI sequence read archive under the accession number SRP026389, and described in detail elsewhere 12 . Brie y, human PBMCs was puri ed from a healthy donor, and then CD4+ memory T cells were puri ed from PBMCs through negative selection using the memory CD4+ T cell isolation kit (Miltenyi) followed by positive selection with anti-CCR6/biotin conjugates and anti-biotin magnetic beads (Miltenyi). Puri ed CCR6+ T cells were stimulated with anti-CD3 and anti-CD28 coated beads (Miltenyi). RNA was prepared from resting and stimulated T cells at different time points over a time course of 3 days. There was a total of six time points, with two biological replicates per time point (Fig. 1A).
Isoform quanti cation and switch.
RNA-seq based transcriptome pro ling was performed by the Illumina HiSeq™ 2000 platform. Raw sequencing reads were mapped by Salmon 20 0.12.0 to the human genome GRCh38 and Gencode Release 29, and then a counting of matrix of 200,000 (transcripts) x 12 (samples) was generated as inputs for detecting isoform switch. All transcripts that did not express across RNA samples were ltered out rst. Then those protein-coding genes with two to ve expressed isoforms were kept for further analysis. The protocol for isoform switch is depicted in Fig. 1B and 1C. First, the relative abundance ( Fig 1C) of different isoforms was calculated from their corresponding expression levels ( Fig 1B) at each time point. Then the most dominant isoform was identi ed at each time point if there exists a one. All isoforms are sorted in a descending order by their relative abundance, and the top isoform is the dominant one if the relative abundance difference between the top two isoforms is greater than 0.3. Otherwise the top one is not a dominant isoform. In Fig. 1C, all dominant isoforms are colored in red circle, and the size of circles represents relative abundance of individual isoform at each time point. It is noted that at 6 and 24 hours, there are no dominant isoforms because the expression levels for the top two isoforms are too close. The reason for us to check the difference between the top two isoforms is to ensure the top one is indeed dominant. Isoform switch occurs if the dominant isoforms differ across time points. The list of candidate genes with potential isoform switch will be further checked using the Omicsoft genome browser, and only those isoform switches with strong evidence (i.e. raw sequence read coverage) are reported.
Protein sequence analysis and comparison.
Results
The isoform quanti cation was performed by using the computational algorithm Salmon 20 . By applying the isoform switch protocol described in the Methods Section, we identi ed eight genes with the most dominant isoform switching during T cell activation (Table 1 and Supplementary Fig. S3). We wanted to determine the functional relevance of the isoform switches in Table 1. Consistent with previously published data linking AS to cell cycle regulation 25 , seven of these eight genes either directly control cell cycle progression or are oncogenes. CDKN2C inhibits the activity of the cyclin D-CDK6 complex thus blocking the G1-to-S phase transition 28 . CENPM is a necessary component of the centromere, and conditional deletion halts cell division leading to cell death 29,30 . FBXO5 controls multiple cell cycle transitions as well as homologous DNA repair [31][32][33] . Methylation of piRNAs by HENMT1 is required for transposable element repression during germ cell division 34 . NET1 directly sequesters the phosphatase Cdc14 to allow cyclin-dependent kinase activity throughout the cell cycle [35][36][37] . SHLD1 is a component of the shieldin complex responsible for non-homologous end joining in the TP53 DNA damage repair response 38,39 . UNG is the major nuclear glycosylase responsible for removing mutagenic uracil from DNA during base excision repair and is necessary for class-switch recombination in B cells 40,41 . We selected CDKN2C, FBXO5, NT5E and NET1 for further exploration. CDKN2C inhibits T cell proliferation in response to TCR stimulation by binding to CDK6. This block is overcome through CD28 costimulation 42,43 . In resting T cells, the isoform CDKN2C-202 is dominant.
Within two hours of activation through both the CD3 and CD28 pathways, levels of this isoform rapidly diminish. Then at 72 hours post-activation expression of the CDKN2C-203 isoform increases to predominate ( Fig. 2A). Based on these data we hypothesized that the CDKN2C-202 isoform lacked the domains necessary to inhibit CDK6 thus leaving non-terminal effector T cells primed for proliferation in response to TCR stimulation.
CDKN2C encodes ve ankyrin repeats 44 . The second and third form the inhibitory interface with CDK6 45 .
Alignment of the splice variants demonstrated that the CDKN2C-203 isoform has a single amino acid truncation of the fourth ankyrin domain and completely lacks the fth ankyrin domain (Fig. 2B). Thus, both CDKN2C isoforms possess the second and third ankyrin domains, which are necessary for CDK6 binding. Another hypothesis we explored was that the CDKN2C-203 variant encodes an unstable CDKN2C isoform. Truncation mutants of CDKN2C lacking the fth ankyrin domain appear unstable as this deletion construct produces little protein 44,46 . This CDKN2C isoform switch is likely important to CDK6 regulation and the generation of adaptive immunity while preventing lymphomas and in ammatory diseases 42,43 .
FBXO5 is encoded at the minus strand, and has two expressed protein coding isoforms, i.e. FBXO5-201 and FBXO5-202 (Fig. 3A). FBXO5-202 is the dominant isoform at early time points, but FBXO5-201 becomes the dominant isoform at 72hr (Fig. 3A). The high expression level of FBXO5-201 at 72hr is evident from the sequence read coverage pro le (Fig. 3B). Note FBXO5 also has a non-coding isoform FBXO5-203 that is barely expressed. After protein translation, FBXO5-201 is 46 amino acids longer than FBXO5-202 in the N-terminus. These 46 amino acids encode three key features: a putative signal peptide (Fig. 3C) and two potential Cdk phosphorylation sites. As a member of the F-box protein family, FBXO5 has several protein-protein interactions that could be affected by this isoform switch [31][32][33] . Despite a large body of work on FBXO5, the functional consequences of these three features are unknown.
FBXO5 is involved in the osteogenic differentiation of mesenchymal stem cells (MSCs) 47 . The expression of FBXO5 was upregulated after osteogenic induction in human periodontal ligament stem cells (hPDLSCs). FBXO5 knockdown attenuated migration, inhibited alkaline phosphatase (ALP) activity and mineralization, and decreased RUNX2, OSX, and OCN expression, while the overexpression of two transcript isoforms signi cantly accelerated migration, enhanced ALP activity and mineralization, and increased RUNX2, OSX, and OCN expression in hPDLSCs. It was concluded that both FBXO5-201 and FBXO5-202 promoted the migration and osteogenic differentiation potential of hPDLSCs, which identi ed a potential target for improving periodontal tissue regeneration. However, whether the two isoforms have different biological roles, especially during T cell activation, remains unclear.
The protein encoded by this gene is a plasma membrane enzyme that catalyzes the conversion of extracellular AMP to adenosine. The encoded protein is used as a determinant of lymphocyte differentiation. Defects in this gene can lead to the calci cation of joints and arteries. The two CCDSvalidated transcripts of NT5E are NT5E-201 and NT5E-203, which differ with respect to the presence of exon 7 in NT5E-201 (Fig. 4). NT5E-201 encodes canonical CD73, denoted as CD73L, while the NT5E-203 transcript is predicted to encode a shorter protein CD73S. Human CD73S lacks amino acids 404-453, encoded by the missing exon 7. The dominant isoform is NT5E-203 at 2 and 4hr, but it is NT5E-201 at 0 and 72hr (Fig. 4). CD73S was expressed at low abundance in normal human tissues but was signi cantly up-regulated in cirrhosis and hepatocellular carcinoma (HCC) 48 . These two human isoforms exhibited functional differences, such that ectopic expression of canonical CD73L in human HepG2 cells was associated with decreased expression of the proliferation marker Ki67, whereas CD73S expression did not have an effect on Ki67 expression 48 . CD73S was glycosylated, catalytically inactive, unable to dimerize, and complexed intracellularly with the endoplasmic reticulum chaperone calnexin. Furthermore, CD73S negatively regulates CD73L activity and protein expression in a proteasome-dependent manner 48 .
It remains unclear the roles of CD73L and CD73S in T cell activation, though new data suggest that CD8 + CD73 + T cells may be especially important mediators of immunosuppression in human head and neck cancer 49 .
The gene NET1 is part of the family of Rho guanine nucleotide exchange factors. Members of this family activate Rho proteins by catalyzing the exchange of GDP for GTP. The protein encoded by this gene interacts with RhoA within the cell nucleus and may play a role in repairing DNA damage after ionizing radiation. Alternative splicing results in multiple transcript variants that encode different protein isoforms. Compared with NET1-202, the expression for NET1-201 is low at early time points but increases signi cantly and become the dominant isoform at 24 and 72hr. NET1-201 and NET1-202 display distinct exon usage in their 5' ends (Fig. 5A). We performed Clustal Omega alignment of the predicted protein sequences of NET1-201 and − 202, and found that NET1-202 completely lacks the rst nuclear localization sequences (NLS) in its N-terminus and that there is poor conservation of the second NLS (Fig. 5B). These NLSs are functionally important as oncogenic NET1 lacks the N-terminal 145 amino acids encoding the NLS, and deletion of the two NLS redistributes NET1 from the nucleus to the cytoplasm 50,51 . The dominance of the NLS-containing NET1-201 splice variant at 24-and 72-hours poststimulation suggests that nuclear sequestration of Cdc14 by NET1 is important for TCR-driven T cell proliferation and maximally effective adaptive immune responses.
The regulation of the two isoforms of NET1 by transforming growth factor-β (TGF-β) in keratinocytes has been studied, and the results emphasize the importance of NET1-202 in the short-and long-term TGF-βmediated regulation of epithelial-to-mesenchymal transition (EMT) 52 . It was found that short-term TGF-β treatment selectively induced NET1-202 (also termed as Net1A) but not NET1-201. Interestingly, long-term TGF-β treatment resulted in Net1A protein degradation by the proteasome. Silencing of Net1A resulted in disruption of E-cadherin-and zonula occludens-1 (ZO-1)-mediated junctions, as well as expression of the transcriptional repressor of E-cadherin, Slug and the mesenchymal markers N-cadherin, plasminogen activator inhibitor-1 (PAI-1) and bronectin, indicating that late TGF-β-induced downregulation of Net1A is involved in EMT. In conclusion, this study provides new evidence for the differential regulation of the two isoforms of the RhoA-speci c GEF NET1 by TGF-β. It points out differential regulatory effects of TGF-β on the NET1A isoform, depending on the duration of the signal 52 .
Discussion
Manual check to detect reliable isoform switch.
The accuracy of isoform quanti cation is in uenced by the complexity of gene structures and caution must be taken when interpreting quanti cation results for short and complex isofroms 53 . It was also discovered that both sequencing depth and the relative abundance of different isoforms affect quanti cation accuracy. Considering the inaccuracy and uncertainty in isoform quanti cation, we manually check all reported isoform switches and lter out those false positives. SFNX3 (sidero exin 3) has seven isoforms and SFXN3-201 and SFXN3-202 are the two mainly expressed forms ( Supplementary Fig. S4B). According to Supplementary Fig. S4A there is an evident isoform switch, but this switch is questionable considering the two isoforms SFXN3-201 and SFXN3-202 are nearly identical. As a matter of fact, SFXN3-202 is only 5 bp longer than SFXN3-201 at the 5' end of UTR (Exon #1), and virtually indistinguishable. Therefore, the reported isoform switch is most likely a false positive due to unreliable isoform quanti cation. Another example of a false isoform switch is shown in Supplementary Fig. S5. RAB43-201 is the dominant isoform at early time points, while RAB43-208 becomes the dominant isoform at 24 and 72 hr. Unfortunately, the high expression level of RAB43-208 at 24 and 72 hours is not supported by the read coverage pro les in Fig. S5B since no sequence reads are mapped to the unique exon #3 of the isoform RAB43-208.
Isoform switching in human T cells.
There Coordinated regulation of T cell proliferation through alternative splicing.
Proliferation is a crucial part of antigen-speci c adaptive immune responses 56,57 , however T cell proliferation must be tightly controlled to allow optimal protective immunity while preventing excessive in ammatory destruction and leukemia/lymphoma. Consistent with a prior publication 25 , seven of the eight genes that demonstrated stimulation-driven isoform switching through alternative splicing in this study regulate proliferation. Regulation of proliferation by alternative splicing and isoform switching is consistent with at least two alternative, but non-exclusive, hypotheses. First, one isoform contributes to cell cycle regulation while the other isoform is associated with normal functions of the cell. This appears to be true for NET1, where the switch from NET1-202 to NET1-201 at 24-and 72-hours post-stimulation ts with a functional refocusing of the T cell from migrating and scanning antigen-presenting cells to proliferation. Without NLS, NET1-202 is likely split evenly between the nucleus and cytoplasm where it interacts with RhoA to control cytoskeletal reorganization [58][59][60] . Combined signaling through the TCR and CD28 initiates T cell proliferation which is facilitated by Cdc14 sequestration of NET1-201 in the nucleus 35-37, 50,51 . So at least some of these alternative splicing events are likely due to the need to rapidly switch between functionally distinct protein isoforms encoded by the same gene during T cell activation.
A second hypothesis is that alternative splicing allows faster protein production than de novo transcription. This may be the case for CDKN2C. If the CDKN2C-202 isoform is unstable it may be a relatively weak Cdk6 inhibitor thus allowing earlier cell cycle progression or proliferation at lower TCR/costimulatory signaling thresholds. As activated T cells progress through the cell cycle, Cdk6 inhibition by CDKN2C must be removed hence the low levels of both isoforms from 2 to 24 hours-post stimulation. The isoform switch to the putatively more stable CDKN2C-201 at 72 hours post-stimulation may facilitate the transition from T cell proliferation to differentiation or be part of terminal effector T cell differentiation.
Rapid isoform switching through alternative splicing of CD73 is likely necessary for T cell activation and effective adaptive immune responses.
Adenosine generation by CD73 plays both autocrine and paracrine roles in T cell activation 61 . The A2a, an inhibitory adenosine receptor, is the predominant form expressed by T cells. Signaling through the A2a adenosine receptor inhibits T cell proliferation at least partly by limiting IL-2 production 62,63 . Dendritic cells (DC) also express A2A and A2B adenosine receptors, and signaling through the A2B receptor blocks DC maturation and co-stimulation of T cells 64 . Thus, the switch in alternative splicing of CD73 during T cell activation ts with the immediate need for proliferation and differentiation. Upon recognition of cognate antigen, T cells cease migrating and form an immunological synapse with the presenting DC 65,66 . Increased expression of a catalytically inactive CD73 that complexes with and promotes degradation of CD73L through isoform switching would provide a rapid method of clearing this inhibitor of proliferation and differentiation 48 . As T cell activation is a stepwise dialogue between DC and T cell, quickly preventing suppressive adenosine accumulation in the microenvironment around DC-T cell pairs is likely critical to costimulation by CD86, IL-2 production and effective adaptive immune responses. Support for this hypothesis comes from cancer, where adenosine is a key component of suppressing the anti-tumor immune response 66,67 .
Declarations Authors Contributions
SZ conceived and designed this study. SZ performed the RNA-seq data analysis and drafted the manuscript. AMSB and KD participated in biological interpretation of isoform switching genes and in writing the manuscript. All authors approved the nal manuscript. | 4,694.2 | 2021-01-12T00:00:00.000 | [
"Biology"
] |
An X-Band Radar System for Bathymetry and Wave Field Analysis in a Harbour Area
Marine X-band radar based systems are well tested to provide information about sea state and bathymetry. It is also well known that complex geometries and non-uniform bathymetries provide a much bigger challenge than offshore scenarios. In order to tackle this issue a retrieval method is proposed, based on spatial partitioning of the data and the application of the Normalized Scalar Product (NSP), which is an innovative procedure for the joint estimation of bathymetry and surface currents. The strategy is then applied to radar data acquired around a harbour entrance, and results show that the reconstructed bathymetry compares well with ground truth data obtained by an echo-sounder campaign, thus proving the reliability of the whole procedure. The spectrum thus retrieved is then analysed to show the evidence of reflected waves from the harbour jetties, as confirmed by chain of hydrodynamic models of the sea wave field. The possibility of using a land based radar to reveal sea wave reflection is entirely new and may open up new operational applications of the system.
Introduction
As it is well known, X-band radar can be used to extract valuable information about the sea state; the main mechanism is the interaction between the electromagnetic waves and the sea's short capillary waves, which in turn ride over longer gravity waves, and has been described and tested for many incidence angles and wavelengths (see for instance [1][2][3][4][5][6]). In the last two decades various approaches have been developed and validated to estimate sea state parameters such as the Significant Wave Height (SWH), the wave spectrum, as well as surface currents and bathymetry [7][8][9][10].
Most of the systems employed to these purposes operate in the short pulse mode (i.e., pulse duration of about 50 ns) and are equipped with a 9-ft (about 2.74 m) antenna. These features enable them to attain a range resolution of about 7 m and an angular resolution of approximately 0.9° thus providing good results in sea state monitoring [10][11][12][13]. Cheaper devices (i.e., 4 or 6-ft antenna radar) have also been applied but usually, due to their poor angular resolution, they cannot provide adequate results, above all in a low signal to noise ratio (SNR) regime [14].
It is also well known that the non-uniform bathymetry and current fields typical of coastal areas can complicate the estimation of the hydrodynamic parameters (i.e., the direction, the period and the wavelength of the dominant waves) with respect to offshore situations [10].
The present paper is aimed at describing an application of the Remocean processing system-implemented and tested at the Institute for Electromagnetic Sensing of the Environment (IREA) of the Italian National Research Council (CNR)-to provide reliable estimates of the surface currents, bathymetry and hydrodynamic parameters in such challenging conditions [15], by making use of an ordinary X-band navigation radar on a commercial ship during a brief stopover in a harbour.
The method proposed here involves the spatial partitioning of the radar data [16,17] and makes use of the Normalized Scalar Product (NSP) strategy to extract the bathymetry and surface current maps [7].
A comparison is provided between the reconstructed bathymetry field and the nautical chart of the area, in order to confirm the reliability of the NSP method in the nearshore zone; this implies that the filtering of the image spectrum by making use of the bathymetry is accurate. By analysing the reconstructed wave spectrum, a definite evidence was found of reflected waves: in order to confirm this result, and since no wave meters were available at or near the test site, a reliable estimate of the sea state was obtained through a chain of numerical models, transforming the offshore SWH values provided by the European Centre for Medium-Range Weather Forecast (ECMWF) into the nearshore wave field through the well-known Simulating Waves Nearshore (SWAN) software [18,19]. A refraction/reflection near-field model was then used to understand near-field effects and in particular to highlight the presence of reflected waves. This is perhaps the most important result of the work reported here, and it might potentially evolve into an important feature of X-band radar sea monitoring systems, since reflected waves may significantly complicate the harbour activities (e.g., berthing operations), as they interfere with the oncoming waves thus creating a confused sea [20,21].
The paper is organized as follows: in Section 2 the data processing approach to estimate the surface currents, the bathymetry and the sea state parameters is briefly discussed. Section 3 describes both the test area and the hardware used to acquire the images. In Section 4 we deal with the validation of the reconstructed bathymetry field by echo-sounder measurements of the depth. In Section 5, the detection of reflected waves is discussed.
Data Processing for Sea State Monitoring in a Coastal Area
The strategy employed to retrieve the sea state from nautical radar data typically involves the processing of a temporal sequence of Nt partially overlapping consecutive radar images. Starting from the 3D spectrum of the radar sequence and by going through a number of operations to take into account the physical processes and to filter out the distortions [22], it is possible to extract the wave spectrum and the hydrodynamic parameters, as well as the surface currents and the bathymetry.
The commonly used procedure is based on the assumption that the former quantities do not vary significantly in the scene; this assumption is normally only true for deep offshore areas, where the physical parameters can be reasonably assumed to be spatially homogeneous [11][12][13]. In many instances such as in the case considered here, the data do not meet this requirement due to the non-uniform surface currents and bathymetry fields; the retrieval of the sea parameters requires therefore a local estimation procedure. Accordingly, a strategy has been used based on the spatial partitioning of the radar images into partially overlapping patches [12,13,16,17].
The size of such patches has to be found as a trade-off between the minimum number of samples for the spectral analysis and the maximum extension of the area that may be considered homogeneous. The block diagram of this data processing procedure is summarized in Figure 1. Each radar image within the temporal sequence is partitioned into Ns uniform spatial patches, thus producing Ns temporal sub-sequences and-by using a FFT (Fast Fourier Transform) algorithm-Ns 3D spectra. Each spectrum is then high-pass (HP) filtered to compensate the power decay which affects the radar signal along the range direction. We denote this set of filtered signals with , ,…, , = , being the wave-vector and ω the angular frequency. The local joint estimation of the bathymetry and surface current is then carried out by applying the NSP algorithm to each spectrum of the set. The NSP algorithm-like most similar systems-exploits the dispersion relation of the gravity waves to extract the sea signal from the overall noisy data [7]. The analytic expression of the dispersion relation for the sea gravity waves is given by Equation (1): where is the gravity acceleration, = , is vector of the sea surface current, ℎ is the depth and = = + is the wave-number.
The relation above rules the propagation of the gravity waves and, moreover, it defines the ω − domain over which the energy of the sea waves is concentrated [22]. It is worth noting that any change of or ℎ turns into a shift of the spectral support of the sea signal. Therefore, these quantities play a key role in the analysis.
The joint estimation of the surface currents and the bathymetry performed by the NSP procedure is founded on the maximization of the following Normalized Scalar Product (Equation (2)): where , ω is the power spectrum in the jth patch, , ω, , ℎ is a (real) characteristic function accounting for the local support of the dispersion relation (Equation (1)), while and are the image power spectra , ω and , ω, , ℎ , respectively.
The bathymetry and the surface currents fields can thus be computed starting from the local estimates. Such information is of course extremely useful for various coastal and offshore applications, but it is also an essential tool to correctly estimate the wave field since, as it is well known, the depth and the current map are required to define a band-pass (BP) filter to separate the energy of the global sea signal from the noise background in the radar spectrum.
The required sea-wave spectrum , ω can be obtained from the filtered image spectrum , ω by resorting to the radar Modulation Transfer Function (MTF), which mitigates the distortions affecting the radar echoes and caused by both the acquisition geometry (e.g., shadowing and tilt modulation) and the electromagnetic (e.g., the Bragg) scattering mechanisms [13,[22][23][24]. For this purpose, we use the MTF described in [22]. For the scope of this work, however, the reconstruction of the wave field is only required to establish the presence of reflected waves, as it will be shown in the following section.
Instrument and Data Description
A field experiment was set up in order to assess the performances of the above described Remocean algorithm: the data were acquired by a marine radar installed on board the Caronte & Tourist ferry ship during her call in the Salerno harbour (see Some of the characteristic parameters of the acquisition system are summarized in Table 1. In particular, the nautical radar (JRC-JMA9122-6XA) was equipped with a 6-ft antenna rotating at about 25 rpm, transmitting short pulses (70 ns) at 9.5 GHz (X-band), with a pulse repetition frequency (PRF) of 2250 Hz and an output peak power of 10 kW. Both the transmitted and the received pulses were acquired in the horizontal plane (HH polarization). The radar field of view was 360°, with an azimuth resolution of 1.2°, while the range coverage was of about 2555 m, with a range resolution of 10.5 m. The radar system was connected to the Remocean system [22], which incorporates an analog-to-digital (AD) converter for the received signal. The radar images were stored using a 13-bit unsigned integer format, on a 1024 × 1024 pixels Cartesian grid. During the measurement campaign, the nautical radar acquired a total of 832 consecutive images, each with a pixel spacing of 5.0 m in both coordinates.
It is worth noting that the whole radar acquisition campaign was carried out in about half an hour, without any interference to the normal operation of the ship, thus proving the effectiveness and the flexibility of the system. A sample image of the collected dataset is shown in Figure 3. In this picture the dashed white lines define the angular sector considered for the reconstruction of the bathymetry and surface currents fields as well as for the retrieval of the sea state parameters, while the white squares identify the subareas (A and B) investigated for the detection of the sea waves reflected by the two jetties of the Salerno harbour.
Reconstruction of the Bathymetry
The reconstruction of the bathymetry field of the Salerno harbour was carried out by applying the inversion procedure described in Section 2. A sequence of 832 consecutive radar images was divided into 12 subsets, each containing 128 radar images with an overlapping factor of 64 radar images between two successive subsets. Joint current and bathymetry estimates were performed on these subsets, thus providing 12 single bathymetry maps, each with a pixel spacing of 50 m in both coordinates. A simple statistical analysis was then performed on the temporal set, so as to extract the maps of the bathymetry and of its standard deviation. The spatial resolution of each bathymetry map is dictated by the overlapping factor used in the partitioning procedure (see Section 2), which in turn derives from a trade-off between the required accuracy and the computational charge. In this work, the patch size of 500 m translated at 50 m intervals has been found to be acceptable, so the data have been analysed on a 102 × 102 grid with a spatial resolution of 50 m. A discussion of this procedure is reported in [13].
The bathymetry map is shown in Figure 4, while Figure 5 shows the local standard deviation as computed among the successive bathymetry estimates in time; its value is relatively low everywhere except in the vicinity of the seawalls and in particular at the harbour's entrance where, as expected, the low wave height and the presence of the jetties affect the accuracy of the results. Almost everywhere else the standard deviation is lower than 1 m, thus confirming the stability of the retrieval method. The retrieved mean bathymetry field has been compared with a nautical chart obtained by a survey performed by the Port Authority, on February 2011 with a single beam ElacHydrostar 4300 echo-sounder integrated with a GPS Leica System 1200. All the measurements are referred, as usual, to the lowest values of the astronomic tide (about −0.20 m with respect to the sea mean level). The nautical chart and the radar derived bathymetry have been reported to the same grid by using a Delaunay triangulation procedure ( Figure 6). In these pictures (as in Figure 4) the red and the black lines represent the iso-depth levels of the nautical chart with depth steps of 0.5 m and 2 m respectively. The good overall reconstruction capabilities provided by the Remocean algorithm can be appreciated in Figure 7, which shows the differences between the nautical chart and the reconstructed bathymetry. As stated before, close to the jetties the error can get as high as 3 m, but as the distance increases the error reduces to less than 10% of the actual values. The performance of the system can be appreciated in the scatterplot and in the histogram shown in Figure 8. The correlation coefficient (R 2 = 0.98) between the radar depth and the nautical chart is very high and the error between the radar estimates and the ground truth is very low (the standard deviation is about 0.37 m); however, a strong bias of about 1 m can be also observed. Part of this difference can be explained by the effect of tides; in order to do so, the records of the Italian Environmental Agency tide gauge, located within the harbour itself, were examined (Figure 9), and it was found that at the time of experiments the water height was at about the mean sea level (msl); since the sounding is referred to the standard datum, i.e., to the average low tide level, which is about 0.20 m under the msl, at least about one fifth of the bias can be accounted for. A further possible explanation for this systematic error can be found on the limits of the linear wave theory, upon which all the radar sea surface analyses are based. According to [16,21] the assumption of linear waves leads to an over-estimation of the water depth, and specially so in shallow water. This is indeed an aspect which needs to be clarified in further work, but the results obtained can be surely considered promising.
Sea State Retrieval and Detection of the Reflected Waves
As stated earlier, the NSP method provides [7] a joint estimation procedure for both the bathymetry and the surface currents. In the particular case considered here a very low surface current field has been estimated with a mean value equal to 0.1 m/s, a maximum of about 0.3 m/s, and a standard deviation of about 0.06 m/s. No field measurements were unfortunately available. Figure 10 depicts the bathymetry map reconstructed with a spatial resolution of 50 m and an overlapping surface current vector map on a 255 × 255 grid for better visualization spatial resolution of 255 m. The current and bathymetry values have been then employed to define the band-pass filter which is needed to compute the wave energy from the noisy radar data. Accordingly, the filtered radar spectrum , ω computed for the sea region bounded by the dotted white lines in Figure 3 has been converted to the sea-wave spectrum , ω through the Modulation Transfer Function. By integrating , ω in the ω domain, the sea directional spectrum can be obtained and in turn the sea state parameters can be computed. The dominant wave parameters are: source direction θ = 214°; wavelength λ = 95 m and peak period = 8.6 s.
No direct wave measurements were available at the experiment time, so offshore wave conditions had to be estimated by making use of European Centre for Medium-Range Weather Forecast (ECMWF, [25]), ECMWF data are produced by the "Mediterranean Model" which makes use of a Wave Action Model (WAM) coupled to the atmospheric HRES model on a 0.25° × 0.25° lat/lon grid at 6 h intervals.
The spatial resolution of such data is too coarse to provide an adequate understanding of the wave field in proximity of the harbour, so a local SWAN model has therefore been attached to the ECMWF grid (see Figure 11, illustrating a SWH field predicted by the Simulating Waves Nearshore (SWAN)) in order to transfer the SWH parameters from offshore to the entrance. SWAN is a widely used third generation wave model [26], developed at the Delft University of Technology [19] to compute the generation and transformation of wind-generated wave spectra in coastal areas. Its general theoretical and numerical approach is similar to many offshore wave models, such as WAM: both WAM and SWAN are driven by the predicted wind fields and use the same formulations for the source terms but, unlike the former, the SWAN model takes into account also depth induced effects, such as diffraction and wave breaking in shallow water. Spectral balance models such as WAM and SWAN can only provide a large scale picture of the average wave field, with a resolution that is necessarily larger than the average wavelength, since neither diffraction nor reflection effects are taken into account; when a greater detail is required, specialized "near field" algorithms have to be used.
As it will be shown in the following, one of the main results of the tests which have been carried out concerns the detection of reflected waves. In order to understand and clarify the presence of such reflection, results from an elliptical Mild Slope Equation (MSE) numerical model study were applied to the test area. MSE provides a solution for a monochromatic wave field once the appropriate boundary conditions are given; the method is well known and widely used in Coastal Engineering. Equation (3) shows a typical expression for the MSE in terms of free surface elevation η , : where is the wave period, and are appropriate functions of time and of the local depth ℎ , , η , is the wave function field, which provides the local wave height. The equation above is a classical tool of Coastal Engineering [21] and the particular implementation used for this work is the Pharos Model [27], by Deltares, (formerly Delft Hydraulic Laboratories) with geometry, bathymetry and boundary conditions that had been previously used for a harbour entrance model study.
Wave boundary conditions applied to the MSE were: SWH = 0.8 m; source direction = 209°; T = 6.3 s, i.e., roughly equivalent to the mean conditions computed by the external model chain. It is important to note that this monochromatic simulation is only meant to provide an insight of the reflection phenomena outside the external breakwater at the time of the tests, and in no way it to represent an accurate picture of the wave field. Figure 12 illustrates the computed wave field accounting for reflection and diffraction phenomena occurring just outside the harbour. As it can be seen, a whole system of reflected waves moves away from the SE facing pier and interferes with the incoming front: an effect obviously caused by reflection. A similar effect caused by the SW facing pier is also visible-less clearly since the wave propagation is nearly perpendicular to the breakwater.
It was found that such effects can be detected by carefully analysing the radar results for two subareas A and B in Figure 3. In both cases two main spectral wave components are clearly visible, related to the incident and to the reflected wave train respectively. The source direction of the dominant wave retrieved by the Remocean system is θ = 214°, thus impacting almost orthogonally to the SW jetty while, according to Snell's law, the reflected wave propagates along the same direction in the opposite way (i.e., its direction of propagation 214°). Both these direct and reflect waves can be observed in Figure 13, which shows a 2D section along the direction (corresponding to 214°) of the 3D wave spectrum (panel a) and 2D directional spectrum (panel b) relevant to the subarea A. In the picture the increasing (decreasing) dashed line depicts the dispersion relation for the incoming (reflected) wave train: the brighter signal represents the amplitude of the spectral density of the incident wave (coming from 214°), while the weaker signal accounts for the reflected wave (coming from 34°). In principle, the reflection coefficient could be evaluated by taking into account the relevant MTF and by considering the ratio between the spectral power of the incident and reflected wave system. In practise, as the mechanism of the radar detection of reflected is still not fully clarified, there are too many uncertainties so that only a rough estimate can be provided: for the SW breakwater its value is about 0.25, since the reflected wave contributes to the overall sea state with the 20% of the total power.
As for the subarea A, the dominant wave impacts on the SE jetty of Figure 3 with an incidence angle θ = 44° so that reflected wave propagates along a direction of about 126°; since the two wavefronts do not propagate along the same direction, they cannot be observed in a single ω − cut of the 3D wave spectrum, but it is nevertheless still possible to identify the sections of the 3D wave spectrum for subarea B which contain most of the energy of the incident and reflected waves. Figure 14 shows the normalized amplitude spectrum cut along direction (left panel), which contains most of the incident wave energy, and along the direction (right panel), which contains the reflected wave. In this case we found again that the energy reflection coefficient is approximately 0.25, on the basis of the ration between reflected and total sea wave energy. These results are of course purely preliminary and incomplete, since the technique is at its initial stages.
Conclusions
A 6 ft antenna nautical X-band radar with a Remocean wave analysis system [28] has been tested at the entrance of a commercial port with a complex bathymetry. It has been shown that even in a complex scenario such a harbour entrance, a bathymetric map can be extracted by making use of a procedure which involves the spatial partitioning of the radar data and the application of the already known and tested Normalized Scalar Product (NSP) strategy. A comparison between the reconstructed bathymetry field and a recent depth survey of the area has proven that a good accuracy can be achieved as long the areas of interest are not too close to the coastline or to the piers. Even though the procedure still seems to present some bias, the result are accurate enough to provide a useful operational tool.
Besides, the bathymetry can be used to produce the Low Pass filter which is necessary to process the signal in order to derive the wave field information. It has also been shown that, once the bathymetry is estimated, an appropriate X-band radar data analysis can detect the presence of waves reflection form structures. The importance of this possibility as a support to navigation and harbour management, as | 5,544 | 2015-01-01T00:00:00.000 | [
"Environmental Science",
"Mathematics"
] |
Distributed Regression Analysis Application in Large Distributed Data Networks: Analysis of Precision and Operational Performance
Background: A distributed data network approach combined with distributed regression analysis (DRA) can reduce the risk of disclosing sensitive individual and institutional information in multicenter studies. However, software that facilitates large-scale and efficient implementation of DRA is limited. Objective: This study aimed to assess the precision and operational performance of a DRA application comprising a SAS-based DRA package and a file transfer workflow developed within the open-source distributed networking software PopMedNet in a horizontally partitioned distributed data network. Methods: We executed the SAS-based DRA package to perform distributed linear, logistic
Background and Significance
Distributed regression analysis (DRA) is a suite of methods that perform multivariable regression analysis in multicenter studies without the need for pooling individual-level data [1,2].Data partners compute highly summarized intermediate statistics (eg, sums of squares and cross products matrices) of their individual-level data and share these statistics with a trusted third-party or analysis center (Figure 1).The analysis center aggregates the intermediate statistics, assesses model convergence, and computes the regression parameter estimates.DRA is mathematically equivalent to the conventional regression analysis of pooled individual-level data.It achieves the same level of statistical sophistication using only summary-level information, thereby offering better protection for individual and institutional privacy without jeopardizing the scientific rigor of the analysis.However, DRA is not widely used in practice due to the operational challenges in implementing the approach [3].The modeling process of common regression analyses (eg, logistic regression, Cox proportional hazards regression) is iterative and requires multiple exchanges of highly summarized intermediate statistics between the data partners and the analysis center.Manual execution of DRA is labor-intensive and highly susceptible to human errors (eg, transfer of incorrect files).There have been efforts to develop capabilities that coordinate and automate the iterative computation and file transfer process of DRA to make it a more practical analytical option in real-world multicenter studies [4][5][6][7][8][9][10][11].These efforts have focused primarily on the programming language R and specially designed applications (eg, Java applets) to facilitate semiautomated or fully automated file transfers between the data partners and the analysis center [7][8][9][10][11].The performance of these capabilities has largely been tested in simulated or relatively well-controlled environments [4][5][6][7][8], and no DRA application has been developed in SAS, another commonly used statistical software.
In our previous work, we enhanced PopMedNet, an open-source distributed networking software currently used by several large national distributed data networks (DDNs), to enable an automatable and iterative file transfer workflow for routine implementation of DRA [3].This workflow coordinates and automates the iterative transfer of files between the data partners and the analysis center.We also created a SAS-based DRA package to conduct distributed linear, logistic, and Cox proportional hazards regression analysis in horizontally partitioned DDN [12,13], environments where each data partner holds information about distinct individuals [14,15].We integrated the PopMedNet workflow with the SAS-based DRA package to create a DRA application.
Objectives
Despite the appealing theoretical properties of DRA, applications designed to perform the analysis can still be inoperable or produce biased results in real-world settings due to unappreciated factors (eg, human errors in execution, incompatible or different software versions, network or firewall restrictions, and network conditions).Evaluating the precision of DRA applications compared with the pooled individual-level data analysis and the feasibility of performing the analysis in reasonable execution times in real-world settings is needed to demonstrate DRA as a practical and valid analytical method.In this study, we demonstrate the feasibility of using the SAS-based DRA package and PopMedNet-driven file transfer workflow to perform DRA in a real-world horizontally partitioned DDN.Specifically, we quantify the precision of the SAS-based DRA package and the operational performance of the PopMedNet-driven file transfer workflow.
Study Setting: The Sentinel System
Funded by the US Food and Drug Administration, the Sentinel System is an active surveillance system designed to monitor the safety of approved medical products using longitudinal, regularly updated electronic health data from a network of 18 health plans and health care delivery systems [16,17].Sentinel data partners transform their data into a common data model [18], which enables analytical programs and tools to be centrally developed and executed across data partners with minimal modifications.Over the years, the system has developed a suite of version-controlled, customizable, and freely available modular programs to rapidly query the transformed data across the DDN [19].Among the tools is the Cohort Identification and Descriptive Analysis (CIDA) tool, a SAS program that assembles cohorts of individuals according to user-specified study parameters (eg, exposures, outcomes, inclusion and exclusion criteria) using established coding systems (eg, International Classification of Diseases, Ninth or Tenth Revision, Clinical Modification; National Drug Codes).The CIDA tool can generate a harmonized (ie, with the same covariates and covariate names) individual-level dataset at each data partner.Users can employ other tools (eg, Propensity Score Analysis Tool) or develop ad hoc analytical programs to query these datasets behind the data partner's firewall for complex inferential analyses.
Sentinel uses PopMedNet to facilitate file transfers between the data partners and the Sentinel Operations Center [20].The Sentinel Operations Center, which serves as the analysis center for all Sentinel queries, uses a Web-based portal to create and securely distribute queries to data partners via PopMedNet.The data partners use a locally installed Microsoft Windows application, known as the DataMart Client, to retrieve the query and return the requested dataset, usually in aggregate-level format, to the Sentinel Operations Center.All file transfers between data partners and the Sentinel Operations Center are accomplished through secure HTTPS, secure sockets layer, or transport layer security connections.PopMedNet security and authentication requirements ensure that only approved queries are submitted to and responses returned by prespecified and approved data partners.In addition, the PopMedNet workflow is agnostic to query types, file formats (RData, sas, .docx,etc) and can transfer individual file sizes up to 2 GB.
SAS-Based Distributed Regression Analysis Application
There are numerous algorithms (eg, secure data integration, secure summation) for DRA in horizontally partitioned DDNs, environments where each data partner holds information about distinct patient cohorts [21,22].In our previous work, we created a SAS-based DRA package comprising 2 interlinked SAS packages (one executed at the data partners and the other at the analysis center) using 2 algorithms: (1) distributed iteratively reweighted least squares to perform distributed linear and logistic regression analysis [12], and (2) distributed Newton-Raphson algorithm to perform distributed Cox proportional hazards regression analysis using the Efron or Breslow approximation for tied event times [13].Both algorithms utilize a semitrusted third-party as the analysis center to aggregate the highly summarized intermediate statistics (eg, sums of squares and cross products matrices) and compute regression parameter estimates and SEs.We define a semitrusted third-party as a party that data partners trust with their summary-level information but not with their individual-level data.This party does not share data from any data partner with other data partners without consent, does not attempt to derive the individual-level data from the intermediate statistics, does not collude with data partners to derive any information about other data partners' individual-level data, and follows the DRA algorithms [23].
We provide a brief overview of the distributed iteratively reweighted least squares and the Newton-Raphson algorithms used to implement the SAS-based DRA package for distributed linear, logistic, and Cox proportional hazards regression analysis using the Sentinel Operations Center as the analysis center in Multimedia Appendix 1.A detailed description of these algorithms is available elsewhere [12,13].
PopMedNet Enhancements to Enable Automatable Distributed Regression Analysis
Both the distributed iteratively reweighted least squares and Newton-Raphson algorithms in the SAS-based DRA package utilize a master-worker process, where the analysis center directs the iterative DRA computations and the data partners execute these computations on their individual-level data with input (eg, updated regression parameter estimates) from the analysis center.Thus, an iterative file transfer workflow is required to transfer the highly summarized intermediate statistics and the updated regression parameter estimates between the data partners and the analysis center until the model converges or the analysis reaches a prespecified maximum number of iterations.
We previously enhanced PopMedNet to create an iterative and automatable file transfer workflow to facilitate routine DRA [3].In brief, we built a back-end component, referred to as the DRA-adapter, into PopMedNet to allow the DataMart Client to upload files automatically and iteratively from and download files to prespecified folders at the data partners and the analysis center.We also developed functionalities for folder monitoring and trigger file creation and deletion in the DataMart Client to integrate the PopMedNet workflow with the two interlinked SAS packages of our SAS-based DRA package.A full description of the PopMedNet workflow and its integration with the SAS-based DRA package is available elsewhere [12,13].We collectively refer to the integration of the SAS-based DRA package and the PopMedNet-driven file transfer workflow as the DRA application hereafter.
Distributed Regression Analysis: A 3-Step Process
A typical DRA includes 3 major steps [3].Step 1 involves the assembly of a harmonized individual-level analytical dataset at each data partner.In step 2, the analysis center and each data partner execute a DRA algorithm locally.Step 3 involves the iterative transfer of the DRA algorithm outputs between the data partners and the analysis center until the regression model converges or the process reaches a prespecified maximum number of iterations.We used this 3-step process to guide our execution and evaluation of the DRA application with 3 Sentinel data partners, with the Sentinel Operations Center serving as the analysis center (Figure 2).
Step 1: Assemble a Harmonized Individual-Level Analytical Dataset at Each Data Partner
We used the CIDA tool (version 3.3.6)to assemble a harmonized individual-level analytical dataset of adult patients aged 18-79 years who received sleeve gastrectomy or Roux-en-Y gastric bypass in any care setting between January 1, 2010 and September 30, 2015 at 3 Sentinel data partners.To be eligible for cohort inclusion, patients must be continuously enrolled in a health plan with medical and drug coverage for at least 1 year before the index bariatric surgery, have at least one weight and height measurement that corresponded to a BMI ≥35 kg/m 2 in the year before surgery, and have at least one height and weight measurement in the year after surgery.We excluded patients with any bariatric procedure during the 1-year period before the index bariatric surgery.We also excluded patients with gastrointestinal cancer or a revised bariatric surgery procedure on the day of surgery.For each regression analysis, follow-up started on the day of the index bariatric surgery and continued until the occurrence of the outcome of interest (see below), death, end of health plan enrollment, or end of the study period.For distributed linear regression analysis, the outcome was a change in BMI within 1-year postsurgery, defined by subtracting the BMI measurement closest to the end of the XSL • FO RenderX 1-year postsurgery date from the last BMI measurement before surgery.For logistic regression, we created a binary outcome variable indicating 1 if the patient had weight loss ≥20% within 1-year postsurgery, and 0 if otherwise.For Cox regression analysis, we computed the time to weight loss ≥20% within the 1-year post-surgery period (Table 1).
Step 2: Locally Execute the Distributed Regression Analysis Application at Each Data Partner and the Analysis Center
We assembled 3 separate SAS-based DRA packages to perform distributed linear, logistic, or Cox regression analyses and assessed the association between bariatric procedure (sleeve gastrectomy vs Roux-en-Y gastric bypass) and weight loss within 1-year postsurgery, adjusting for prespecified confounders (Table 1).For Cox regression analysis, we used the Efron approximation to handle tied event times.To be consistent with the standard SAS regression procedures, we prespecified a convergence criterion of <0.01 and a maximum of 25 iterations for distributed logistic and Cox regression analyses.
We distributed each SAS-based DRA package to the 3 data partners through PopMedNet (version 6.7).We instructed the data partners to (1) initiate the automated PopMedNet workflow, allowing the DataMart Client (version 6.7) to automatically download and unzip the SAS-based DRA package to a prespecified local directory, (2) manually place the individual-level analytical dataset created in step 1 in a prespecified local folder, (3) specify the file path to the SAS-based DRA package, and (4) execute the SAS-based DRA package in batch mode.Similarly, we instructed the Sentinel Operations Center to (1) initiate the automated PopMedNet workflow, (2) manually place the SAS-based DRA package for the analysis center in a prespecified local directory, (3) specify the file path to the SAS-based DRA package, and (4) execute the SAS-based DRA package in batch mode.Full details of these packages and examples of their execution have been previously described [12,13].
Step 3: Iteratively Transfer Distributed Regression Analysis Files Between the Data Partners and the Analysis Center
Once the data partners and the analysis center executed their SAS-based DRA package, the package ran continuously, awaiting input files (eg, updated regression parameter estimates or intermediate statistics) and DRA computation directions (eg, compute intermediate statistics, residuals, and SEs) from the Sentinel Operations Center.We used the PopMedNet workflow to transfer input files and computation directions iteratively and automatically between the data partners and the Sentinel Operations Center.
Evaluation of Precision and Operational Performance
We requested all data partners to securely transfer their deidentified individual-level analytical datasets to the Sentinel Operations Center.We assessed the precision of the SAS-based DRA package by comparing the DRA parameter estimates and SEs to those obtained from the pooled individual-level data analyses using standard SAS procedures.For distributed linear regression, we compared the model fit statistics R 2 , Akaike information criterion (AIC), Sawa's Bayesian information criterion (BIC), and Schwarz BIC to the statistics obtained from a PROC REG run with the pooled individual-level data.For distributed logistic regression, we compared the model fit statistics log-likelihood, AIC, and Sawa's BIC to the statistics obtained from a PROC LOGISTIC run with the pooled individual-level data.For distributed Cox proportional hazards regression, we compared the model fit statistics log-likelihood, AIC, and Schwarz BIC to the statistics obtained from a PROC PHREG run with the pooled individual-level data.We considered the integration successful if the DRA parameter estimates and SEs and model fit statistics were precise to the results from the corresponding pooled individual-level data analyses (10 −6 ).
For distributed logistic regression, we also compared the receiver operating characteristic (ROC) curve and the area under the ROC curve with the corresponding curve and area obtained from a PROC LOGISTIC run with the pooled individual-level data.We considered the integration successful if the ROC curves were similar in likeliness and if the areas under the curves were comparable.To offer better privacy protection, we summarized individual-level predicted values for the distributed logistic regression analysis in bins of 6. Full details of this approximation method can be found elsewhere [12].For distributed Cox proportional hazards analysis, we also compared the survival function curve with the curve obtained from a PROC PHREG run with the pooled individual-level data.We considered the integration successful if the survival function curves were similar in likeliness and if the median times to weight loss ≥20% were equivalent.
RenderX
We extracted time stamps of status changes from PopMedNet and computed the average download, upload, SAS execution, and transfer time at the data partners and the analysis center to evaluate the operational performance of the DRA application.We also reported the average iteration time for each regression model type, and the time required to perform an end-to-end DRA in our test case.
We executed all SAS-based DRA packages in SAS versions 9.3 or 9.4, on a Windows desktop or server routinely used to perform Sentinel queries.All machines used to execute the SAS-based DRA packages and DataMart Client instance operated on a Windows 7 platform, with multiple Intel core processors ranging from 2.3 to 3.4 GHz, and 8 to 16 GB of RAM (Multimedia Appendix 2).
Overview
We identified 5452 eligible patients among the 3 participating data partners (n 1 =1706, n 2 =2728, and n 3 =1018).Of these, 981 patients received sleeve gastrectomy, whereas 4471 patients received Roux-en-Y gastric bypass during the study period.Within 1-year postsurgery, the BMI decreased on average by 9.8 kg/m 2 in sleeve gastrectomy patients and 18.7 kg/m 2 in Roux-en-Y gastric bypass patients.Five-hundred eighty-two of the 981 (59.3%) patients who had undergone sleeve gastrectomy and 3617 of the 4471 (80.10%) patients who had undergone Roux-en-Y gastric bypass had a weight loss ≥20% within the 1-year postsurgery period.The median time to a weight loss ≥20% was 223.9 days for patients who had undergone sleeve gastrectomy and 196.2 days for patients who had undergone Roux-en-Y gastric bypass.
Precision
Tables 2-4 summarize the precision of distributed linear, logistic, and Cox proportional hazards regression analyses.Table 5 shows the model fit statistics of the 3 regression models.All DRA parameter estimates, SEs, and model fit statistics were highly comparable to the estimates obtained from the pooled individual-level analyses that used standard SAS regression procedures.The ROC curve in distributed logistic regression (Figure 3) and the survival function in distributed Cox regression (Figure 4) were similar to those obtained from the pooled individual-level data analyses.The DRA application reported an area under the curve (AUC) of 0.6591 for logistic regression (vs 0.6592 from the pooled individual-level data analysis) and 184 days for Cox proportional hazards analysis (vs 184 days from the pooled individual-level data analysis) as the median time to weight loss ≥20%.
Operational Performance
As expected, the closed-form solution of distributed linear regression analysis required only two iterations, one for computing the regression parameter estimates and SEs and the other for computing the model fit statistics.Both logistic and Cox proportional hazards regression analyses required 6 iterations for model convergence in our test case.Each file transfer process transferred between 3 and 10 files with sizes of 1 to 800 KB.
We extracted 111, 271, and 271 time stamps of status changes from PopMedNet for distributed linear, logistic, and Cox analysis, respectively.Table 6 summarizes the operational performance of the DRA application.It took an average of 102.4 s to complete one DRA iteration across all regression model types.The file transfer workflow (file upload, download, and transfer to the reciprocal party) accounted for 89% of the iteration time.Downloading and uploading the DRA files at the Sentinel Operations Center required an average of 28.6 and 9.8 s, respectively.File transfer from the Sentinel Operations Center to the data partners took on average 9.4 s.Downloading and uploading the DRA files at the data partners required an average of 10.1 and 15.5 s, respectively.File transfer from the data partners to the Sentinel Operations Center took an average 22.1 s.Computing the intermediate statistics at the data partners required an average of 8.0 s, whereas computing the updated regression parameters took an average of 3.8 s at the Sentinel Operations Center.
The distributed Cox regression required the greatest amount of iteration time (113.5 s), followed by logistic regression (95.0 s) and linear regression (91.5 s).Overall, distributed linear regression analysis with our bariatric surgery test case required 440.7 s to complete, whereas logistic and Cox proportional hazards regression analysis required 925.5 and 1016.0 s, respectively.
Principal Findings
We have successfully integrated a SAS-based DRA package with PopMedNet, an open-source distributed networking software, and performed DRA in select data partners within a real-world DDN.Our application was able to compute regression parameters, SEs, model fit statistics, and model fit graphics of 3 regression model types (linear, logistic, and Cox proportional hazards) that were within machine precision or similar in likeliness to those produced using standard SAS regression procedures, without the need to share any individual-level data, in under 20 min.The study demonstrated the feasibility and validity of performing multivariable regression analysis in a multicenter setting while limiting the risk of disclosing sensitive individual or institutional information.
Previous Studies
Previous studies have used simulated or relatively well-controlled distributed environments to demonstrate the ability to perform DRA with only summary-level information [4][5][6][7][8].These studies have consistently reported that DRA produced precise (generally <10 −12 ) results compared with the results from the pooled individual-level data analysis.However, information on the operational performance (computation and file transfer time) of DRA algorithms or workflows is scarce.The closest experience to our DRA application is a Web-based DRA software developed by the SCAlable National Network for Effectiveness Research (SCANNER) [11].This software is composed of a network portal with a set of Web services and virtual machines that host data from data-contributing sites and several libraries of analytical programs.At the time of our analysis, 3 method libraries were available in the SCANNER software: a cohort discovery tool, an algorithm to perform meta-analyses with distributed data, and an algorithm to perform distributed logistic regression analysis (Grid Binary LOgistic Regression, GLORE) [6].The authors reported that GLORE produced results equivalent to those from the pooled individual-level data analysis, and software response times of 0.015 s with a dataset of 580 records (with a binary outcome variable, a treatment indicator variable, and 24 covariates) and 27.02 s with a dataset of 10,000 records (with a binary outcome variable and 5 covariates) partitioned among 3 different institutions.
Our DRA application required significantly more time for model convergence than the SCANNER software.However, this additional time for model convergence may be considered marginal in practice, where other aspects of a multicenter study are typically more time-consuming.For example, developing a study protocol and analysis plan or assembling an analytical dataset at each participating data partner for DRA may require considerably more time than the time required to perform DRA.There are also several key differences between our application and the SCANNER software that may explain the difference in operational performance.Specifically, the SCANNER software requires users to install a virtual machine and open ports to the master node hosting the SCANNER hub.This design may have shorter file upload, transfer, and download times between the execution nodes, as files are only transferred between homogeneous virtual machines on the server and not subject to impediments such as firewall security protocols, additional workload, and upload, transfer, and download speeds.
The operational performance of the SCANNER software makes it a desirable option for DRA in networks that are amenable to installing the required software and applications.We previously found that most Sentinel data partners were unwilling to install new software or make modifications to their existing hardware configurations to perform DRA [3].We chose to develop the DRA application using SAS and PopMedNet because all Sentinel data partners have both software in their systems.In addition, several other large DDNs, including the National Patient-Centered Clinical Research Network [24] and the National Institutes of Health's Health Care Systems Research Collaboratory [25], use PopMedNet as their file transfer software.In other words, our DRA application requires no new software installation or modifications to existing hardware configurations in DDNs that employ SAS as their statistical software and PopMedNet as their file transfer software.The 3 data partners that participated in this project are also members of numerous PopMedNet-based DDNs.Therefore, the successful integration of our SAS-based DRA package with PopMedNet and execution of DRA with these data partners have the potential to extend DRA beyond the Sentinel System.
Limitations
Our study is not without limitations.First, DRA requires infrastructure and processes beyond the algorithms and technology described in this paper.For example, DRA with our application requires harmonized individual-level datasets.Since its inception, Sentinel has continuously enhanced its common data model, routine analytical tools, and data quality assurance processes.Thus, Sentinel data partners can rapidly create harmonized analytical datasets for DRA.Research networks and investigators without the same infrastructure may not be able to perform DRA with our application as easily, even if data partners are willing to use PopMedNet as their data-sharing software.
Second, we tested the DRA application with only 3 Sentinel data partners, and all tests were completed in a Windows version of SAS (desktop or server).It is possible that different hardware configurations not found at these data partners or different versions of SAS (Linux or Unix) could change the precision and operational performance or even inhibit the execution of our DRA application.However, we previously found only 3 different configurations of the required hardware components (DataMart Client, SAS software, and the common folder structure) among Sentinel data partners [3].All 3 hardware configurations were represented among the 3 data partners in this study.We also found the reconfiguration of these components to be relatively straightforward.Therefore, it may be possible to have data partners with other configurations make minor changes to implement our DRA application.During the development of the DRA application, we were able to successfully execute our application on a Linux server with a fourth data partner, by placing the application on a Linux server directory accessible to the DataMart Client as a mapped XSL • FO RenderX Windows network drive.This allowed the DataMart Client to access the same file system as the DRA application.Overall, additional testing with more data partners with different hardware configurations and different versions of SAS is needed to fully ensure that our DRA application is operable across different DDNs, research networks, operation systems, and environments.
Third, our precision and operational performances were based on a small sample of successful end-to-end executions of our DRA application.These executions were limited to regression models with 23 variables and analytical datasets ranging from 1000 to 3000 patients distributed among 3 data partners.Future work should include more end-to-end executions, regression models with more variables, datasets of larger sample sizes, and more data partners.However, we found that 89% of the iteration time was attributed to file transfer time, which was largely driven by the number of files, size of the files transferred, and network conditions (upload, download, and transfer speeds, firewall security protocols, and workload).Because the files contain highly summarized information, increasing the number of variables or patients will not increase the number of files or substantially increase the size of the files to be transferred.In this study, each file transfer process transferred files that were less than 1 MB.Our internal testing of analyses with more variables, patients, and data partners did not result in file sizes larger than a few MBs or increased the iteration time.Thus, we do not anticipate DRA with more variables, patients, and data partners in a real-world setting to have a considerable impact on the operational performance of our DRA application.In addition, network conditions at each data partner can vary depending on the workload.We could not vary network conditions at each data partner to formally analyze its impact on the operational performance.However, we did complete our experiments with 3 Sentinel data partners, with machines that are routinely used to fulfill Sentinel query requests.Thus, our results on precision and operational performance likely represent what potential users of DRA will experience in practice.
Fourth, our bariatric surgery test case was relatively simplistic and not as sophisticated as an actual clinical or epidemiologic study.For example, we did not include all the potential confounders.Therefore, the results of our analysis did not have any causal interpretation.
Finally, although DRA uses the intermediate statistics at each data partner to perform multivariable regression analysis, the risk of reidentifying specific individuals is not 0.Under certain conditions (eg, uncommon individual attributes coded with indicator variables), there could be leakage of personal information that could be used to infer or identify specific individuals [26].To further protect privacy, DRA can be performed using more secure algorithms, such as encrypting or perturbing the intermediate statistics.Future work should explore the integration of these more secure DRA algorithms into our DRA application.
Conclusions
We have successfully developed and integrated a SAS-based DRA package with an iterative and automatable PopMedNet-driven file transfer workflow to create a DRA application and conduct DRA in select data partners within a real-world DDN.The application produced results that were within machine precision to the results from the pooled individual-level data analyses using standard SAS regression procedures.The end-to-end execution times were reasonable, demonstrating that DRA can be a practical and valid analytical method in real-world settings.
Figure 1 .
Figure 1.Distributed regression analysis with horizontally partitioned data.
Figure 2 .
Figure 2. Three-step process to conduct distributed regression analysis with PopMedNet.CIDA: Cohort Identification and Descriptive Analysis Tool; DRA: Distributed Regression Analysis; SOC: Sentinel Operations Center.
event (days) a ROC: receiver operating characteristic.
Figure 3 .
Figure 3.Comparison of receiver operating characteristic curves between distributed logistic regression (left) and pooled individual-level logistic regression (right).To offer better privacy-protecting, individual-level predicted values were summarized in bins of 6 and transferred to the analysis center for aggregation in the distributed logistic regression analysis.The size of the bin is user-specified.ROC: receiver operating characteristic.
Figure 4 .
Figure 4. Comparison of survival functions between distributed cox proportional hazards regression (left) and pooled individual-level cox proportional hazards regression (right).The survival curves were evaluated at the mean value of covariates for patients with events.
Table 1 .
Analytical datasets and variables.Bariatric surgery exposure, age at surgery, sex, race and ethnicity, combined Charlson-Elixhauser comorbidity score, number of ambulatory visits, number of other ambulatory visits, number of inpatient stays, number of nonacute institutional stays, number of emergency department visits, BMI before bariatric surgery, number of days between last weight and height measurement and bariatric surgery, and data partner
Table 2 .
Distributed linear regression vs pooled individual-level linear regression.
Table 3 .
Distributed logistic regression vs pooled individual-level logistic regression.
Table 4 .
Distributed Cox proportional hazards regression vs pooled individual-level Cox proportional hazards regression.
Table 5 .
Comparison of model fit statistics between distributed regression and pooled individual-level data analysis.
Table 6 .
Operational performance of the distributed regression analysis application.
a N/A: not applicable. | 6,788.6 | 2020-06-04T00:00:00.000 | [
"Computer Science"
] |
A Lightweight Dense Connected Approach with Attention on Single Image Super-Resolution
In recent years, neural networks for single image super-resolution (SISR) have applied more profound and deeper network structures to extract extra image details, which brings difficulties in model training. To deal with deep model training problems, researchers utilize dense skip connections to promote the model’s feature representation ability by reusing deep features of different receptive fields. Benefiting from the dense connection block, SRDensenet has achieved excellent performance in SISR. Despite the fact that the dense connected structure can provide rich information, it will also introduce redundant and useless information. To tackle this problem, in this paper, we propose a Lightweight Dense Connected Approach with Attention for Single Image SuperResolution (LDCASR), which employs the attention mechanism to extract useful information in channel dimension. Particularly, we propose the recursive dense group (RDG), consisting of Dense Attention Blocks (DABs), which can obtain more significant representations by extracting deep features with the aid of both dense connections and the attention module, making our whole network attach importance to learning more advanced feature information. Additionally, we introduce the group convolution in DABs, which can reduce the number of parameters to 0.6 M. Extensive experiments on benchmark datasets demonstrate the superiority of our proposed method over five chosen SISR methods.
Introduction
The single image super-resolution (SISR) image processing technique is essential in computer vision and aims to recover a high-resolution image from a single low-resolution (LR) counterpart. It has been used in a large number of computer vision applications, such as medical imaging [1], surveillance imaging [2,3], object recognition [4], remote sensing imaging [5] and image registration and fusion [6,7]. For example, image registration requires high-resolution (HR) images to provide richer details to transform different sets of data into one coordinate system. However, when the upscaling factor increases, the complexity of image registration increases. As a result, it is vital to design an appropriate architecture for the SISR technique.
Traditional SISR methods can be divided into three categories: reconstruction-based methods [8,9], interpolation methods [10][11][12], and learning-based methods [13][14][15][16][17]. Recently, with the rapid development of neural networks, Convolutional Neural Network-based (CNN-based) SR methods have achieved remarkable performances [18][19][20][21][22][23][24]. In 2015, Dong et al. [18] proposed the first Super-Resolution Convolutional Neural Network (SRCNN), introducing a three-layer convolution neural network for single image super-resolution. Afterwards, a Fast Super-Resolution Convolutional Neural Network (FSRCNN) [19] was proposed to accelerate SRCNN, speeding it up by more than 40 times with better restoration quality via a modified network structure and smaller filter sizes. To deal with multi-scale SR problems, Lai et al. proposed a progressive upsampling framework named the Laplacian pyramid SR network (LapSRN) [20], which can progressively generate intermediate SR predictions. Since He et al. [25] presented a Residual Network (ResNet) to show that the network's depth can be of great significance for various computer vision tasks, some researchers began to increase the network depth to enhance the SR effect. Kim et al. [21] proposed the very deep convolutional network for SR (VDSR). They extended the depth of the network to 20 and achieved higher performance compared with SRCNN and FSRCNN. Soon after, some works [22,23] also successfully demonstrated that deepening CNN networks could further boost SR performance. To alleviate the vanishing-gradient problem brought by the deep network structure, T. Tong et al. [24] proposed a Densely Connected Convolutional Network for SR (SRDensenet) based on dense skip connections, which achieved significant improvement in the image SR task.
SRDensenet has exhibited a good performance at ICCV 2017 due to its contribution to the effective integration of low-frequency and high-frequency features. However, as the network layer deepens, the number of model parameters grows exponentially, dramatically increasing computational complexity. In particular, SRDensenet could waste unnecessary calculations for low-frequency features. The network lacks the ability to identify and learn across feature channels in terms of staking dense blocks, ignoring the inherent feature relationship. Thus, SRDensenet will produce redundant and conflicting information in rich features, which is useless for reconstruction. To resolve the above-discussed issues, in this paper, we propose a lightweight dense connected approach with attention to single image super-resolution tasks, namely, LDCASR, which uses the attention mechanism to learn more effective channel-wise features and the group convolution structure to decrease model parameters. Experiments show that the proposed LDCASR can achieve a better reconstruction performance over the state-of-the-art SISR methods on public SR benchmarks while significantly reducing the model parameters. (About one-ninth of the parameters of SRDensenet.) In summary, the main contributions of this paper can be summarized as follows.
1.
We introduce the attention mechanism to the dense connection structure as well as the reconstruction layer, which helps to suppress the less beneficial information during model training. Extensive experiments verify the effectiveness of this attention-based structure. 2.
Our model can extract the important features by using a lightweight approach. By introducing the group convolution, we reduce the number of parameters to 0.6 M, which is around 1/9 of the original SRDensenet.
The remainder of this paper is organized as follows. In Section 2, we introduce the related work of super-resolution tasks and the attention mechanism. In Section 3, we describe the proposed network LDCASR architecture, including the details of its compositions: Recursive Dense Group (RDG), Dense Attention Block (DAB), and Channel Attention Unit (CAU). The experimental results and analysis on the comparison with other methods are provided in Section 4. Finally, we draw our conclusions in Section 5.
CNN-Based SISR
Dong et al. unprecedentedly introduced a three-layer CNN framework into the SISR and proposed a super-resolution convolutional neural network (SRCNN) [18], which exhibited a remarkable performance compared to the traditional works [8][9][10][11][12][13][14][15][16][17] and opened the way for neural network-based SR research. After that, plenty of approaches based on convolution neural networks were proposed. A fast super-resolution convolutional neural network (FSRCNN) [19] introduced the deconvolution operation to the CNN model. It not only accelerated the speed but also enhanced the performance of the SRCNN. To further speed up the SR, Shi et al. [26] presented an effective method called efficient subpixel convolutional neural network (ESPCN), which completed the upscaling progress with the aid of subpixel convolution. Lai et al. [20] proposed a progressive upsampling framework called the Laplacian pyramid network (LapSRN) to increase image size gradually. By further deepening the network structure, a very deep super-resolution, VDSR [21], achieved a better result using a deeper network, including almost 20 convolution layers. Afterwards, DRCN [22] and DRRN [23] also realized deep networks by employing recursive learning and sharing parameters. However, with the deepening of the network, the issue of the gradient vanishing appeared. Researchers found the skip connection [25] is a handy way to address the gradient vanishing. Regarding this problem,Ledig et al. [27] proposed residual neural network for SR (SRResNet) with more than 100 layers. They adopted the generator part of the SRGAN as the model structure and employed the residual connections between layers. After that, Lim et al. proposed two even deeper and wider networks: an enhanced deep SR network (EDSR) [28] and a multi-scale deep SR network (MDSR) [28], which both consisted of 1000 convolution layers. These deep SISR networks improve performance by simply stacking the different blocks. However, they ignore the channel-wise feature information. In fact, in addition to the dimension of length and width, the channel is another crucial dimension of an image. Channel attention assigns different weights to each channel to help the network pay attention to important features and suppress unimportant features. With channel attention, the model performance can be improved with a small amount of computation.
Dense Skip Connections in SISR
To deal with deep model training problems, researchers utilized dense skip connections to promote the model's feature representation ability by reusing deep features of different receptive fields. The dense skip connections were first proposed in DenseNet [29], which was the subject of best paper in CVPR 2017. Afterward, SRDenseNet [24] exhibited a good performance in SISR by introducing the dense skip connections. Then, many networks employed the dense connected structure in the SR task and exhibited remarkable performances. However, the dense connected structure simultaneously introduces redundant and useless information, which is harmful to the image's super-resolution. Different from these methods, we combined the dense connected blocks with the attention mechanism to focus on learning important information.
Attention Mechanism
The attention mechanism was derived from a study of human vision, which was first proposed in the field of visual images. Google Mind team [30] proposed Recurrent Models of Visual Attention in 2014, which used the attention mechanism for image classification on the RNN model. In recent years, attention-based methods have yielded attractive results in various tasks, for instance, image recognition [31] and natural language processing [32]. Researchers found that attention mechanism can not only reduce useless information by discriminating the effective feature information but emphasize the importance in various dimensions. Wang et al. [33] designed a stackable network structure with a trunk-andmask attention mechanism for image classification tasks. Hu et al. presented a novel block called squeeze and excitation (SE) [34], which employs channel-wise associations using average-pooled features to increase the representational power of a CNN network and ensure accurate image classification. Specifically, the SE block was introduced to deep convolutional neural networks to enhance performance more [35]. Recently, Dai et al. [36] proposed a novel module called second-order channel attention (SOCA) to gain more useful feature expression as well as feature correlation learning. All the methods mentioned above undoubtedly obtained significant results. Inspired by these works, we introduce the attention mechanism to reinforce our network and improve the effect.
Our Model
In this section, we first describe the entire architecture of the proposed LDCASR. Then, we introduce the details of different components of the proposed network, including the Recursive Dense Group (RDG) and Dense Attention Block (DAB) with a Channel Attention Unit (CAU).
Network Architecture
As illustrated in Figure 1, our LDCASR mainly includes three modules: feature extraction module, upscale module, and reconstruction module. We define the original LR input as I LR and the output as I SR .
Feature Extraction Module
The feature extraction module includes a 3 × 3 convolutional layer for shallow feature extraction and several Recursive Dense Groups (RDGs) for deep feature extraction. First, the original low-resolution images were fed into the 3 × 3 convolutional layer straightway to extract the shallow feature. This part aims to transfer the inputs from color space to feature space. The process can be expressed as: where H SF (·) represents the function of the shallow feature extraction process, including only one 3 × 3 convolution layer. Then, the extracted shallow feature F 0 passed through the stacked Recursive Dense Groups, as well as a 1 × 1 convolution layer to change the number of channels. This process produces the deep image feature, denoted as: where H RDG (·) represents the feature extraction operation by RDGs, and H DF (·) represents the 1 × 1 convolution operation. Each RDG applies multiple Dense Attention Blocks (DABs) with Channel Attention to suppress redundant information (see Figures 2 and 3), which will be discussed in the following subsection.
Upscale Module
The upscale module was placed after the feature extraction module to upscale the feature maps from small size to the size of ground truth. Instead of merely performing deconvolution or subpixel convolution [26] for upscaling, as in the existing methods, we used channel attention units (shown in Figure 3) and deconvolution operations cross-wise to better capture high-frequency information. Specifically, we used one deconvolution operation in the ×2 experiment, two deconvolution operations in the ×4 experiment, and three in the ×8 experiment.
The low-level feature information includes much original image information, a lot of which will be lost during the process of forward propagation. Thus, it is important to combine the low-level features gained by bicubic-interpolation with the high-level information to obtain the final results. Considering fusing extra original image information, the upscaled features were added to the bicubic-interpolated LR image to acquire the final output of the upscale layer F up .
The process is expressed as: where the upscale operation is marked as H up (·).
Reconstruction Module
The final SR image I SR was obtained by the reconstruction module, which contains only one convolution layer of size 3 × 3. This layer aims to recover the images from the feature space to the color space. The reconstruction progress can be expressed as: where H R (·) denotes the reconstruction layer and H LDCASR (·) denotes the whole process of LDCASR. I SR was optimized by the absolute difference between the I LR and I HR .
Dense Attention Block (DAB) with Channel Attention Unit (CAU)
As mentioned before, the recursive dense groups (RDGs) are an essential component of our model. Each RDG consists of staked Dense Attention Blocks (DABs) connected by dense connections. It is verified that a large number of dense blocks are beneficial to form a deep CNN in [24]. However, the stacked dense blocks will introduce redundant and conflicting information, causing a longer training time and unsatisfied reconstruction results. Inspired by the methods based on attention, we employed the channel-dimension attention mechanism to learn the high-frequency features and propose the Dense Attention Block (DAB) (see Figure 2), which contains two 3 × 3 convolution operations and a Channel Attention Unit (CAU). As a result, with the aids of DABs, our model is able to focus on acquiring more important and useful information. The progress of DAB can be expressed as: where F h n , F h−1 n denote output and input of the h − th DAB in the n − th RDG separately, the operation of convolution, Relu, concatenation and CAU as f conv (·), f relu (·), f cat (·), f cau (·), respectively.
We also denote the input of CAU as F in and the output as F out . The specific formula is as follows: where f avgpool (·) represents the operation of average pooling, f sigmoid (·) represents the function of sigmoid and ⊗ denotes the element-wise product.
Group Convolution
Group convolution first appeared in AlexNet [37] architecture, which is shown in Figure 4. Unlike the standard convolution operation, group convolution can divide all the R channel inputs into G groups, so each group responds to R/G channels. Then, the output is cascaded to the final output of the entire set of convolutional layers. In order to reduce the parameters of the network, we introduced group convolution in each dense attention block (DAB) to achieve a lightweight network.
Loss Functions
In SISR tasks, loss functions are used to measure reconstruction error and lead the model optimization direction.
In the previous stages, many methods [18][19][20]22,23,25] employ the L2 loss, which is also named the mean square error (MSE) loss. However, researchers found that L2 loss cannot measure the reconstruction quality precisely. Moreover, the result of the reconstruction is not satisfying using L2 loss. Afterwards, more and more researchers tend to use the L1 loss (mean absolute error), which can achieve a better effect of reconstruction. Furthermore, some researchers use the L1 Charbonnier loss function to train the models, which was first proposed in LapSRN. We mark the original high-resolution image as I HR and the loss functions can be expressed as: where h, w and c represent the height, width and number of channels of the feature maps, respectively. N = h × w × c . is a constant for numerical stability.
To compare different loss functions, we employed the three loss functions mentioned above to train the LDCASR. The comparison results are displayed in Section 4.
Training and Testing Datasets
Similar to previous works, our training dataset uses the DIV2K [38]. It consists of 800 training RGB images, 100 validation RGB images, and 100 test images. Following the previous works, we employed 800 training images as our training set, which was then augmented with a 90°rotation and horizontal flip. Furthermore, we used the bicubic kernel function to down-sample the ground truth images to generate LR image counterparts. The training data were generated with Matlab Bicubic Interpolation; training files were created with https://github.com/wxywhu/SRDenseNet-pytorch/tree/master/data, accessed on 11 April 2021.
We followed the previous works by converting the test images from RGB color space to YCbCr color space for evaluation. After that, we employed the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) to evaluate the performance of our structure only on the Y channel.
Implement Details
The number of RDB and DAB was set to 8, respectively, identical to the dense blocks and single blocks in SRDensenet. Our model was trained by ADAM optimizer with β 1 = 0.9, β 2 = 0.999, the batch size was set to 32, the initial learning rate was 0.0001, and we set the learning rate to the initial LR decayed by momentum every 30 epochs. We conducted our experiment with the scaling factors of ×2, ×4, ×8 between the HR and the LR images. We used PyTorch to implement our models with one Titan Xp GPU.
Different Loss Function Analysis
We show the training results of 60 epochs to compare the training convergence curves of L1 loss, L2 loss, and L1 Charbonnier loss. During the process of gradient descent, the L1 loss was more robust than the L2 loss. Moreover, L1 loss was not easily affected by extreme feature points. This is because the L2 loss function is based on the squared error, which increases twice as much as the L1 loss. As a result, L1 loss is more stable than L2 loss. Additionally, L1 Charbonnier loss adds a variable 2 compared with L1 loss, making L1 Charbonnier loss more robust than L1 loss. Thus, the learning curve of L1 Charbonnier loss yields the best result. We chose the L1 Charbonnier loss to conduct our following experiments. As shown in Figures 5 and 6, we found that the performance of the training using L1 Charbonnier loss is better than others as its curve is more stable and the average PSNR or SSIM is higher. Therefore, we chose the L1 Charbonnier loss to train our models. The equations of loss functions are shown in Equations (8)-(10).
In Table 1, we compare our model with the state-of-the-art methods, including SRCNN, VDSR, LapSRN, and SRDensenet; all the mentioned methods were trained under the same condition. Bold indicates the best performance. As is shown in Table 1, our model achieved was more effective than the other methods under the scale factors of ×2, ×4, ×8 on the four benchmark datasets. Obviously, under the ×2 scale factor, the value of PSNR on the Urban100 dataset increased by almost 0.47 compared to SRDensenet. In Table 2, we compare the average computational time of different methods. As can be seen, our LDCASR achieved the second fastest performance for ×2, ×4, and ×8 SR experiments, preceded only by SRCNN, which is a simple model with only three convolution layers. Furthermore, we present the performance curve on Set5 for ×4 SR in Figure 7, which intuitively reflects that our model LDCASR is superior to SRDensent. By using the bicubic interpolation, our model exhibited a good performance from the beginning. Subsequently, as the training process continued, the PSNR value of LDCASR stayed at around 32, while the PSNR of SRDensenet stayed at approximately 31.54. In addition, we provide the comparisons of evaluations in terms of visual quality, which are displayed in Figures 8 and 9.
Model Size Comparison
The comparisons of model size and performance are illustrated in Figure 10. It shows the model size and performance of different state-of-the-art methods. The abscissa refers to the model parameter sizes, while the ordinate denotes the image average PSNR obtained by these models, and each point represents a model. The red point represents our proposed network. As can be seen, SRCNN and VDSR exhibited lightweight but slightly lower performances. LapSRN exhibited a better performance than SRDensenet with fewer parameters. Our LDCASR had fewer parameters and exhibited a relatively better performance, which indicates our model achieves a better trade-off between parameter scale and performance.
Conclusions
In this paper, in order to solve the deficiency of dense network in the SR task, a lightweight dense connected approach with attention is proposed for SISR, where dense attention blocks (DABs) capture important information in the channel dimension by a channel attention unit (CAU). The design of DABs makes the whole network focus on high-frequency details and successfully suppresses useless information in smooth areas.
In addition, our model achieved a lightweight effect with fewer parameters and had a relatively superior performance. The extensive experiments on four benchmark datasets illustrate that our LDCASR achieves better results than other state-of-the-art SR methods in terms of objective evaluation and subjective visual effects. In our future work, we will explore more advanced model structures, such as the model based on improving the spatial attention mechanism or multi-attention mechanism. Moreover, we will study the application of our model in other fields.
Conflicts of Interest:
The authors declare no conflict of interest. | 4,895.8 | 2021-05-22T00:00:00.000 | [
"Computer Science"
] |
Streamlined identification of strain engineering targets for bioprocess improvement using metabolic pathway enrichment analysis
Metabolomics is a powerful tool for the identification of genetic targets for bioprocess optimisation. However, in most cases, only the biosynthetic pathway directed to product formation is analysed, limiting the identification of these targets. Some studies have used untargeted metabolomics, allowing a more unbiased approach, but data interpretation using multivariate analysis is usually not straightforward and requires time and effort. Here we show, for the first time, the application of metabolic pathway enrichment analysis using untargeted and targeted metabolomics data to identify genetic targets for bioprocess improvement in a more streamlined way. The analysis of an Escherichia coli succinate production bioprocess with this methodology revealed three significantly modulated pathways during the product formation phase: the pentose phosphate pathway, pantothenate and CoA biosynthesis and ascorbate and aldarate metabolism. From these, the two former pathways are consistent with previous efforts to improve succinate production in Escherichia coli. Furthermore, to the best of our knowledge, ascorbate and aldarate metabolism is a newly identified target that has so far never been explored for improving succinate production in this microorganism. This methodology therefore represents a powerful tool for the streamlined identification of strain engineering targets that can accelerate bioprocess optimisation.
www.nature.com/scientificreports/their conclusions based on 29 of these metabolites.Their analysis led to the overexpression of pyk and deletion of the araR repressor, achieving the simultaneous consumption of both substrates at the same rate.
Other authors looked at a smaller set of targeted metabolites for strain optimisation.For example, George et al. used targeted metabolomics to look at eight compounds of the isoprenoid pathway in E. coli to improve the heterologous production of three C5 alcohols with potential use as biofuels, particularly by engineering the Shine-Dalgarno sequence of the nudB gene, significantly alleviating the bottleneck of isopentenyl diphosphate 12 .Barton et al. performed a targeted 13 C labelling analysis using LC-MS to look at the intracellular levels of two compounds of the 1,4-butanediol (BDO) formation pathway and five up-stream by-products 13 .This way, the authors found the last two steps to be the bottleneck of BDO formation.Subsequent genetic engineering of these two steps led to improved BDO formation.Some studies have also looked at the availability of different target nutrients during the bioprocess in order to identify and prevent specific limitations to improve process performance.For example, Korneli et al. looked at 22 metabolites-19 of which were amino acids-and identified amino acid limitation during the production of GFP in Bacillus megaterium 14 , and Ho et al. looked at the composition of six fatty acids-and at least 11 other intracellular metabolites-produced by the marine microalga Chlamydomonas sp.JSC4 under different conditions of salinity and nitrogen starvation, allowing an increase in lipid accumulation by adjusting the cultivation conditions 15 .
In the examples mentioned above-and others [16][17][18][19] -metabolomics was used as a targeted tool to look at specific metabolites and pathways, in most cases using unit mass resolution mass spectrometers-such as single-and triple-quadrupoles-thus limiting the potential of finding engineering targets for bioprocess improvement to those based on prior knowledge of the biological system.However, untargeted metabolomics offers the possibility to find targets for bioprocess optimisation in a wider metabolic context in a more unbiased fashion, particularly when used with high-resolution accurate mass (HRAM) mass spectrometry.For example, Xu et al. performed an untargeted metabolomics analysis to compare a pho13Δ mutant of S. cerevisiae-with a higher capacity to catabolise xylose-with its parental strain.A total of 134 intracellular metabolites were identified, and sedoheptulose-7P-a metabolite from the pentose phosphate pathway (PPP) outside carbohydrate metabolism-showed the most significant difference between both strains.Mutants overexpressing the PPP gene TAL1 at different degrees were constructed, and their TAL1 expression levels were positively correlated to their respective xylose consumption rates 20 .
Another study used metabolite profiling to determine which nutrients were limiting in a bioprocess for recombinant IgG4 antibody production using Chinese hamster ovary (CHO) cells, a discovery experiment better suited to untargeted metabolomics due to the unknown composition of the medium 21 , and Xia et al. used untargeted and targeted metabolomics to compare the production by Streptomyces tsukubaensis of FK506-a polyketide used as immunosuppressant-in a high and low productivity media 22 .Multivariate analysis of the metabolomics results with partial least squares (PLS) allowed the authors to identify and supplement limiting nutrients in the medium, increasing FK506 production.The 13 nutrients most significantly correlated with FK506 biosynthesis included coenzyme A (CoA) esters, shikimate, amino acids, pyruvate, lactate and PPP intermediates.Once again, these would have been difficult to identify by targeting the product biosynthetic pathway.
As the examples above show, untargeted metabolomics offers the possibility to identify key metabolites for bioprocess improvement that are outside the product biosynthetic pathway, which are commonly missed when targeted methods are used focusing on prior biological knowledge about the system.However, untargeted metabolomics produces very large datasets from which it is challenging to prioritise metabolic reactions and pathways for modification.A useful tool for dealing with and obtaining useful information from untargeted metabolomics data is pathway enrichment analysis, which analyses groups of compounds that work together to carry out a biological process-like a metabolic pathway 23,24 , and ranks them in terms of their statistical significance and importance.Although pathway enrichment analysis has historically been applied more frequently to genomics, transcriptomics and proteomics data [25][26][27] , it has also been applied in the field of metabolomics.However, most of these publications are clinical studies [28][29][30][31][32] .
There have been some recent studies where metabolic pathway enrichment analysis (MPEA) has been used with metabolomics data for bioprocess improvements.For example, Morris et al. compared four different fedbatch cultivation conditions resulting in differences in monoclonal antibody titres in a CHO bioprocess 33 .The authors then manually looked in the KEGG database 34 at the pathways of the 10 most significant metabolites, as identified by PLS-discriminant analysis (PLS-DA), to find titre inhibitors and promoters that might be modulated by changing the bioprocess feeding strategy.With this approach, however, constraining pathway analysis to the 10 most significant metabolites significantly limits the vast analytical power of metabolomics.In another example, Alden et al. used untargeted metabolomics to find metabolites that accumulated in the culture medium of CHO fed-batch processes.Using MPEA with a modified Fisher's exact test, the authors found three pathways that were enriched in the cell line with the lowest cell density profile of their study, including aminoacyl-tRNA biosynthesis, tryptophan and histidine metabolism.Further investigation into 11 putatively annotated metabolites that accumulated in the culture medium led the researchers to hypothesise that products of tryptophan metabolism could behave as inhibitors of cell growth.The authors finished by suggesting targeting the tryptophan pathway with genetic engineering, or lowering the concentration of tryptophan in the cultivation medium 35 .These examples show how MPEA can be used for bioprocess improvement, however, this is still a line of research that is underexplored, particularly for the identification and prioritising of genetic targets for bioprocess optimisation.
MPEA can be performed in different ways, similar to the examples described above for targeted and untargeted metabolomics.Mainly, the analysis can focus on comparing a case and a control group, such as high-and low-productivity conditions or strains 9,10,15,20,22 , or focus on the dynamic changes of different metabolites throughout the course of the fermentation 11,12,14,21 .Although not mutually exclusive, the first type of analysis can be more tailored to identifying media deficiencies, performance biomarkers and genetic differences between differently www.nature.com/scientificreports/performing strains, whereas the second type can be used to improve an established working fermentation process further by, for example, identifying the accumulation of by-products, the presence of inhibitors, the (in)activation of specific metabolic pathways or the depletion of substrates.
In this study, a commercial E. coli succinate production process was analysed with a combined targeted and untargeted metabolomics method using HRAM mass spectrometry, and the results were used to perform MPEA throughout the time course of the fermentation to identify potential targets for bioprocess optimisation in an unbiased fashion.
Results
Escherichia coli succinate fermentation process.Three E. coli dual-phase succinate fermentation replicates were performed and samples were taken throughout the course of the fermentation for metabolomics analysis with LC-MS.Furthermore, the extracellular concentration of glucose and the main fermentative products was also determined by HPLC-UV/Vis-RI analysis (Fig. 1).The fermentation process is split into two distinctive parts.The first phase is an aerobic batch process, where all the glucose from the medium is consumed for biomass growth, with little production of fermentative products.Under aerobic conditions with glucose excess, acetate is formed in E. coli due to overflow metabolism, and part of this acetate is excreted out of the cell to prevent osmotic stress due to the accumulation of negative charges in the cytoplasm.After glucose depletion, this acetate can be reimported inside the cell and be consumed as a source of carbon and energy [36][37][38] .A small amount of lactate, malate, pyruvate and succinate was measured in one of the three replicates in the middle of the biomass exponential growth phase, potentially due to a temporary limited supply of oxygen.However, the concentration of these organic acids went back down before the end of the biomass growth phase.The second phase of the bioprocess is an anaerobic production phase (biotransformation), where additional glucose is added to the medium and this is converted into succinate and other fermentative products, being pyruvate the main byproduct.The E. coli strain used in this process cannot sustain cell growth on glucose under anaerobic conditions, which can be appreciated by the plateau and slight decrease in the biomass wet cell weight (WCW) measurements during the anaerobic phase of the process.
LC-MS analysis of fermentation samples.The three fermentation experiments were analysed by LC-MS to try to identify targets for improving succinate production.A total of 13 samples from each replicate fermentation were analysed using both intracellular and extracellular fractions of the samples (see Supplementary file "Metabolomics data.xlsx").The Principal Component Analysis (PCA) plot of the samples for both intra-and extracellular analysis shows good clustering of the different time points of the three fermentation replicates, www.nature.com/scientificreports/demonstrating reproducible data (Fig. S1 in the Supplementary materials), despite the lower succinate production for one of the three fermentation runs.The combined untargeted and targeted analysis resulted in a total of 6341 annotated and 92 identified metabolites (Table S1 in the Supplementary materials and see also the Supplementary file "Metabolomics data.xlsx").Metabolite identification was performed by matching the retention time and accurate mass to reference standards following the Metabolomics Standards Initiative (MSI) classification system 39 .
Metabolic pathway enrichment analysis.The metabolomics results were used to perform pathway enrichment analysis in order to find potential genetic targets for bioprocess optimisation without bias from preconceived process-specific biological knowledge.MPEA was performed using the Pathway Activity Level Scoring (PALS) application in the Polyomics integrated Metabolomics Pipeline (PiMP) online platform 40 (date of use: 08 Jul 2020).PALS ranks significantly changing metabolite groups over different sample sets using the pathway level analysis of gene expression for metabolomics (mPLAGE) algorithm, which converts m/z features into formulae and matches them to pathways 41 and the KEGG library 34 (see Supplementary file "Pathway enrichment analysis data.xlsx").The output of the mPLAGE algorithm is a p-value that has taken into account the multiple comparisons over all the pathways in the KEGG library.The different time points of the fermentation data were grouped by triplicate measurements and used for pathway enrichment analysis.For every triplicate time point of the fermentation, the algorithm evaluates pathways that are significantly different when compared to the first triplicate time point.This way, every metabolic pathway assessed is assigned a significance level (p-value) for every fermentation sample (except for the first time point).Each p-value indicates the probability that there is no difference between the time points for the metabolites on the pathway.The first sample analysed is a time point at the early stages of the aerobic phase of the fermentation (right after inoculation for extracellular samples and 3 h after inoculation for intracellular samples.See the "Methods" section for more details).By doing a timecourse comparison to an early time point in the fermentation where the cells had abundant access to glucose and oxygen, it is possible to identify changes in metabolic regulation at the different stages of the fermentation (e.g.end of biomass growth, beginning or end of succinate production phase, etc.).This information can be used to put the results in the context of the experiment and aid in the identification of engineering targets.
To find genetic targets for improved succinate production, the mean p-value of each pathway (as compared to the first time point) was calculated across all time points in the succinate production phase.This mean p-value was used as ranking criteria (rather than focusing on extremely small p-values at individual time points) in order to find pathways with consistent statistical significance throughout the production phase and avoid the impact of potential outliers.Intracellular and extracellular metabolite fractions were analysed separately.The ten pathways with the highest level of significance (lowest mean p-value) resulting from analysing intracellular metabolites are represented as a heatmap in Fig. 2.These include several pathways from amino acid, carbohydrate, nucleotide and vitamin metabolism.More significant changes were observed in the extracellular fraction samples (Fig. S2 in the Supplementary materials).This is caused by a significant increase in the number of metabolites found in the extracellular fraction as the fermentation progresses-potentially due to cell lysis, particularly under oxygen limitation or absence, because the E. coli strain used cannot sustain biomass growth on glucose anaerobically.The large number of significant pathways makes the selection of specific ones with significant changes in metabolic levels based on the extracellular fraction more challenging.For this reason, the selection of potential genetic targets was done purely based on the MPEA performed with the intracellular metabolites.
Selection of potential genetic targets.
Pathways showing high metabolic changes during the succinate production phase could be potential targets for metabolic reshuffling.Either due to the removal of by-products or the elimination of metabolic bottlenecks, the modulation of these pathways could potentially lead to higher titres or yields.From the 10 pathways in Fig. 2, three pathways were identified to have particularly low p-values during the succinate production phase: the pentose phosphate pathway, ascorbate and aldarate metabolism and pantothenate and CoA biosynthesis (marked with red boxes in Fig. 2).The criteria to select these three particularly was that they all had at least three time points in the succinate production phase with a p-value ≤ 0.01.
The metabolomics results for these pathways are shown in Fig. 3. Several metabolites from the PPP show a sudden increase in intracellular levels at the beginning of the succinate production phase and in many cases the levels remain high, with the exception of sedoheptulose-7P.The sudden increase at the beginning of the production phase is also observed for the precursors of pantothenate.However, pantothenate itself shows a decrease in intracellular intensity at the transition from the aerobic growth phase to the anaerobic production phase.Similarly, CoA levels drop significantly between 2 and 18 h of succinate production, indicating that CoA availability can be a potential bottleneck for succinate production, as suggested by Lin et al. 42 .Finally, 5-dehydro-4-deoxy-d-glucarate and 2-oxoglutarate-both involved in ascorbate and aldarate metabolism-show a sudden increase in intracellular levels at the beginning of the production phase, followed by a sudden decrease.However, l-gulono-1,4-lactone shows a progressive increase in intracellular intensity throughout the succinate production phase, indicating intracellular accumulation.Trying to minimise this accumulation can, therefore, be another potential genetic target to improve succinate production.
PLS-DA for identifying targets of succinate improvement.
Other authors have used PLS regression to identify targets for bioprocess engineering 9,22 .Therefore, PLS-DA was chosen as a benchmark analytical method to compare to the findings of the metabolic pathway enrichment analysis.For this, the three fermentation replicates described in this work were split into two groups, the first group containing the two experiments with higher succinate titres and the second group containing the experiment with a lower titre (10.05 ± 0.64 SD and 4.23 g/L final succinate concentration, respectively).Then, the samples from the succinate production phase of the bioprocess were analysed by PLS-DA to identify which metabolites contribute the most to the differences in succinate titre (Fig. 4).The samples from the aerobic growth phase were not included in the analysis to avoid discrimination based on the large metabolic differences between the aerobic biomass growth phase and the anaerobic succinate production phase.Succinate was also removed from the data set before performing PLS-DA in order to prevent biasing group discrimination.Good separation of the two sample groups was observed (Fig. 4, top), and the four first latent variables were selected to build the model, based on maximising the crossvalidated coefficient of determination (Q 2 = 0.7123, R 2 0.9738, Fig. 4, bottom).
From a total of 3989 mass spectrometry signals analysed in PLS-DA, 495 were found significant with a variable importance in projection (VIP) score > 1 43 in the first latent variable (Supplementary file PLSDA summary.xlsx).The VIP score is a common measure in PLS-DA of the relative importance of each metabolite with respect to the total variation of the latent variable.From these significant signals, 270 had unknown annotations and many of the 225 remaining had multiple possible metabolic annotations.Those metabolites with the highest level of annotation confidence and a VIP score above 1 in the four first latent variables were investigated.Metabolites with a high level of annotation confidence were defined as either having retention time and accurate mass matched to a reference standard, or accurate mass and fragmentation pattern matched to metabolomics library data.The 38 metabolites that met the criteria above are gathered in Table S2 (see Supplementary file PLSDA summary.xlsx), and most of them belong to amino acid metabolism (18 out of the 38 metabolites) and nucleotide metabolism (10 out of 38), which is a much more reduced breadth of metabolic pathways than the findings from Fig. 2. Nevertheless, coinciding with the results from MPEA, three metabolites involved in pantothenate and CoA biosynthesis were found important for discrimination of a successful E. coli succinate fermentation run according to the PLS-DA model, namely d-pantothenic acid, l-alanine and l-valine (Fig. 5), strengthening the hypothesis that this pathway could be a potential target of interest for bioprocess optimisation.While l-alanine and l-valine show clearer differences between both sample groups (higher levels for the higher succinate titre group), the levels of d-pantothenic acid are more similar in both groups, which is also reflected in the latter having a lower VIP score in the first two latent variables (see Supplementary file PLSDA summary.xlsx).However, it is worth mentioning that PLS-DA looks at individual metabolites and does not reveal any biological information of the relationship between them, whereas enrichment analysis groups features (metabolites in this case) under common biological themes, making it more robust to reveal real biological differences between sample groups 44 .That is, for analysing the PLS-DA results, the three abovementioned metabolites were manually linked to the same metabolic pathway assisted by the findings from MPEA from Fig. 3.In other words, MPEA makes it easier to put metabolomics results in a biological context.
Discussion
Metabolomics is an expanding field that has been previously applied to identify potential targets for genetic engineering to improve bioprocesses [9][10][11][12][13]20 . Howver, in most cases, this is still performed with targeted metabolomics using unit mass resolution single-or triple-quadrupole instruments looking at pre-selected intermediates from the biosynthetic pathway of the product of interest, thus limiting the vast detection capacity of mass spectrometry.Some examples are available using untargeted metabolomics [20][21][22] , however, these tend to identify potential genetic targets using multivariate statistical tools-often PCA or PLS-applied directly to the metabolomics data without using information from metabolic pathways.As this study shows, without the context of biological pathways it can be very hard to interpret and extract useful information from untargeted metabolomics data using multivariate statistical analysis.This was demonstrated by developing a PLS-DA model to discriminate between higher and lower succinate titre runs in an E. coli bioprocess, based on targeted and untargeted intracellular metabolomics data.Overall, although the PLS-DA model was able to achieve good separation of the two sample groups, the large number of significant signals and the high level of uncertainty in metabolite annotation makes it difficult to identify genetic targets for succinate production improvement.Focusing on those metabolites with a higher confidence in annotation it was possible to identify metabolites from amino acid, nucleotide and pantothenate Figure 3. Dynamic evolution of intracellular metabolites of the three pathways identified as potential targets for genetic engineering: the pentose phosphate pathway (purple), pantothenate and CoA biosynthesis (green) and ascorbate and aldarate metabolism (grey).A simplified version of the pathways is shown for ease of reference.The results show both phases of the bioprocess: the aerobic bacterial growth phase (blue) and the succinate production phase (red).Orange error bars represent the standard error of the mean (n = 3).and CoA metabolism as being important for group discrimination.However, this is a time-and effort-intensive exercise that left behind 448 of the 495 (90.51%) significant features according to the PLS-DA model.
Conversely, the use of MPEA allows to exploit metabolomics data further.By considering correlated changes in different compounds of specific pathways, certain enrichment analysis tools such as PALS can be used to analyse targeted and untargeted metabolomics to identify significantly modulated pathways during the bioprocess in a more streamlined way.Importantly, as MPEA highlights sets of metabolites that are consistently regulated, it is more robust to individually misannotated metabolites than techniques that look at individual metabolites without considering any biological context (such as PLS-DA).Furthermore, the PALS tool used in this work uses formulae rather than identifications to safeguard even more against the impact of potential misannotations.
The E. coli succinate production bioprocess was analysed with MPEA and the results pointed at three different pathways with significantly different levels of metabolites during the bioprocess: the pentose phosphate pathway, pantothenate and CoA biosynthesis and ascorbate and aldarate metabolism.
Interestingly, the PPP has previously been reported to play an important role during succinate production in E. coli.For example, Lu et al. 45 calculated the carbon flux down the PPP during succinate production using LC-MS and HPLC measurements and using a stoichiometric model in an E. coli AFP111 dual phase bioprocess using 13C-labelled glucose as a carbon source.Using a higher percentage of CO 2 during the succinate production phase increased both succinate production and the percentage of carbon flux channelled towards the PPP.In another study, Zhu et al. found that increased activity of the enzyme transketolase in the PPP led to an increase in succinate titre and yield in E. coli ATCC 8739 derived strains.The succinate titre and yield improved even further increasing the activity of the enzyme transhydrogenase, which can convert NADPH generated from the PPP into NADH-which can then be used for succinate production 46 .The same research group performed a further study where the authors generated a library of PPP genes under the regulation of a constitutive M1-93 promoter and different ribosome binding sites to generate different E. coli ATCC 8739 derived variants with different expression levels of the PPP genes.Their results indicate that increasing the expression levels of PPP genes can lead to higher succinate yields and titres, but that a fine tuning of the expression levels gives the best results 47 .These examples are in agreement with the identification, in this work, of the PPP as a potential target for succinate improvement using MPEA, validating this methodology for the identification of pathways for bioprocess improvement.
There is less prior literature about manipulating the pantothenate and CoA biosynthetic pathways to increase succinate production.However, Lin et al. showed increased succinate production in E. coli GJT001 derived strains by overexpressing the enzyme pantothenate kinase, which catalyses the first step in the conversion of pantothenate to CoA 42 .Succinate production was increased even further with the co-overexpression of either phosphoenolpyruvate carboxylase or pyruvate carboxylase, which catalyse the conversion of phosphoenolpyruvate and pyruvate into oxaloacetate.The authors attributed the increase in succinate production to a higher intracellular availability of acetyl-CoA and CoA.In all experiments, the fermentation medium was supplemented with 5 mM pantothenate, therefore, genetic engineering to increase the levels of intracellular pantothenate would also be required on top of the changes indicated by the authors to increase succinate production.This study suggests that genetic manipulation of the pantothenate and CoA pathway can, indeed, lead to an increase in succinate production.
Finally, as far as ascorbate and aldarate metabolism is concerned, no examples were found in the literature showing the manipulation of this pathway in E. coli for increased succinate production.Therefore, this study might be the first to identify this pathway as a potential target for genetic engineering to achieve improved succinate production in E. coli.
This work is the first example of the application of MPEA using both targeted and untargeted metabolomics to identify potential strain engineering targets for bioprocess improvement.The examples above show that the targets identified with this methodology are relevant and coincide with previous successful attempts to improve www.nature.com/scientificreports/succinate production in E. coli.The three pathways identified in this work are outside the succinate biosynthetic pathway, showing how untargeted metabolomics can identify important pathways for product formation, even if these are initially not known to have an impact on the bioprocess.Untargeted metabolomics has the potential to accelerate bioprocess optimisation, and pathway enrichment analysis is a useful tool to help analyse and interpret the results obtained with this technology putting them in a biological context.Looking forward, it is important to mention that, although MPEA is a powerful tool for the identification of significantly regulated pathways, it does not indicate how these need to be modified to achieve the desired strain phenotype.Indeed, biological regulation is complex, with many factors affecting metabolic and protein levels.Therefore, predicting and modulating phenotypes is still difficult even when knowing which metabolic pathways are highly regulated in a bioprocess.Nevertheless, MPEA can be a powerful tool to streamline and speed up the Design-Build-Test-Learn cycle of strain engineering, i.e. the identification of engineering targets.
Methods
Bacterial strain.All experiments described in this article were carried out using a proprietary industrial E. coli strain (Ingenza Ltd., UK) previously described 48 , based on the E. coli NZN111 strain with deletions of the pyruvate-formate lyase (pflB) and lactate dehydrogenase (ldhA) genes 49 .
Growth media.All 5 L scale fermentation experiments were carried out with a batch phase for biomass formation using a defined minimal medium containing 11.90 g/L glucose as the sole carbon source, 2.00 mM MgSO 4 , a mix of salts solution (2. -2000).Shake flask overnight cultures were prepared using the same medium but with 10.00 g/L glucose and no antifoam.
Fermentation process conditions.All fermentation experiments were carried out in a 5 L Applikon stirred tank fermenter (ADI 1030 Bio Controller, 1035 Bio Console), and the process consisted of an initial aerobic batch phase where the minimal medium was primarily used for biomass formation, followed by a 24 h anaerobic succinate production phase as previously described 48 .
Inoculum.Fermentation inocula were prepared by inoculating 50 µL of cell bank into 100 mL of growth medium in a 500 mL baffled shake flask and incubated at 37 °C and 165 rpm for 17-17.5 h.
Aerobic batch phase for biomass growth.The fermentation was started by inoculating 100 mL of overnight culture into 3 L of growth medium in the 5 L fermenter for a starting OD 600 of 0.21 ± 0.025.During biomass growth the conditions were maintained at 37 °C temperature, 500-900 rpm agitation (controlled to keep the dissolved oxygen (DO) > 30%), 4.00 L/min air (1.33 vvm) and pH 7.0 ± 0.1, controlled with 2.00 M H 2 SO 4 and 28% w/v NH 4 OH.
Anaerobic succinate production phase.At the beginning of the production phase, glucose from a 500 g/L solution and sodium bicarbonate from a 100 g/L solution were added to the fermenter as a single bolus addition to a final concentration of 20 g/L and 5 g/L respectively in the vessel, as previously described 50 .The sodium bicarbonate provides soluble CO 2 , which is required for the conversion of PEP to oxaloacetate 51 .Once the glucose and sodium bicarbonate were added to the fermenter, the sparged air was replaced by pure (99.8%)CO 2 at 0.50 L/ min (0.17 vvm), agitation was set to 300 rpm, temperature at 37 °C and pH at 7.0 ± 0.1 controlled with 2.00 M H 2 SO 4 and 28% w/v NH 4 OH.
Biomass measurement.Biomass levels were reported as OD 600 and wet cell weight (WCW).The former was the measured optical density at 600 nm wavelength.The latter was determined by spinning down 1 mL of sample for 5 min at 14,462 g twice in a pre-weighed Eppendorf tube, removing the supernatant and weighing the resulting pellet.The weight of the pellet in g/L was calculated from gravimetric difference.
Metabolites extraction for LC-MS analysis.
Samples for off-line liquid chromatography-mass spectrometry (LC-MS) analysis were removed from the bioreactor and collected into dry-ice-cold universal vials, placed briefly on an ethanol dry ice bath for a fast cold sample quenching and immediately spun down twice at 4 °C and 13,000 g for 10 min.The supernatant and cell pellet were collected as extracellular and intracellular fractions respectively and stored at − 80 °C until further extraction for LC-MS analysis.
Extracellular fractions.Extracellular fraction extractions were prepared by diluting 10 µL of sample into 400 µL of 1:3:1 chloroform:methanol:water (C:M:W).The samples were then mixed vigorously in a chilled microtube mixer for 5 min and then centrifuged for 3 min at 13,000g and 4 °C.At this point, 360 µL of supernatant were transferred into a new microtube and stored at − 80 °C until LC-MS analysis.25 µL of supernatant of each extracted sample were combined into one single vial to generate a pooled sample.During handling, the 1:3:1 C:M:W extraction solvent and the samples were kept on an ethanol dry ice bath.LC-MS method.Metabolite separation was performed using a zwitterionic hydrophilic interaction liquid chromatography (ZIC ® -pHILIC) column (Merck SeQuant ® ) (150 mm × 4.6 mm, 5 µm particle size) equipped with the corresponding guard column (20 mm × 2.1 mm, 5 µm particle size) (Merck SeQuant ® ).A linear gradient was applied to the column, running from 80 to 20% solvent B over 15 min, followed by a 2 min wash with 5% solvent B, and 9 min re-equilibration with 80% solvent B, where solvent B was acetonitrile and solvent A (the remaining percentage) was 20 mM ammonium carbonate in water.The total flow rate was 300 µL/min, column temperature was maintained at 25 °C, sample injection volume was 10 µL, samples were maintained at 4 °C for the duration of the analysis and a HESI probe was used on the ion source.Metabolite detection was done in a high-resolution Thermo Scientific™ Q Exactive™ Orbitrap mass spectrometer at 70,000 resolution, mass range 70-1050 m/z in polarity switching mode with a spray voltage of ± 3.8 kV.Capillary temperature was set to 320 °C, sheath gas 40 a.u., AGC target 1 × 10 6 a.u. and the lock masses in positive and negative mode were 144.9822 m/z and 100.9856 m/z, respectively.
Fragmentation was performed on pooled samples by isolating ions in a 1.2 m/z window and fragmentation with stepped HCD collision energy of 24.8, 60.0 and 94.8% for both polarities with 17,500 resolution and AGC target 1 × 10 5 a.u.Top 10 ions (intensity threshold 1.3 × 10 5 ) were selected for fragmentation and then added to a dynamic exclusion window for 15 s.
Metabolomics data processing and analysis.
Raw mass spectrometry files were converted into .mzXMLfiles in profile mode with the open-source software ProteoWizard 52 (Version 3.0).Further data processing and analysis was performed using the PiMP online platform 40 (date of use: 08 Jul 2020).Peak detection and filtering were set to 3 ppm of the theoretical monoisotopic mass, minimum intensity to 5000, noise to 0.8, retention window to 0.05 and minimum number of detections to 3. Peak retention time was corrected using the Obiwarp algorithm 53 from the xcms package (Version 1.48.0).
Metabolite annotation and identification.Metabolite annotation was performed using a local copy of the KEGG database 34 (Metabolomics Standards Initiative (MSI) class 2/3 annotation 39 ), the fragmentation library (MSI class 2 annotation) and by matching the retention time and accurate mass to reference standards, which are listed in the Supplementary file Std_list_Metabolites.xlsx.HPLC-UV/Vis-RI analysis.HPLC coupled to UV/Vis and refractive index detectors (HPLC-UV/Vis-RI) analysis was carried out using a Rezex™ ROA Organic Acid H + ion-exclusion column (Phenomenex ® ) (300 mm × 7.8 mm) equipped with a Carbo-H4 guard column (SecurityGuard™) (3.0 mm i.d.).An isocratic method was applied to the column, running a 5 mM H 2 SO 4 mobile phase solution for 30 min.The total flow rate was 800 µL/min, column temperature was maintained at 65 °C, sample injection volume was 10 µL and samples were maintained at 4 °C for the duration of the analysis.The HPLC-UV/Vis-RI data was extracted as a .csvfile and was further analysed using the ggplot2 package 54 (Version 3.3.3) in the statistical software environment R (Version 3.6.1).Metabolite levels below the limit of quantification were replaced by 1/5 of the minimum metabolite concentration quantified.
Metabolic pathway enrichment analysis.MPEA was performed using the PALS application in PiMP (date of use: 08 Jul 2020) using the mPLAGE algorithm and a local copy of the KEGG database 34 .Pathways with p-values of 0.05 or below were considered statistically significant across experimental groups compared.The experimental groups compared were the different time points, and each pairwise comparison was done comparing any given time point with the first sample of the fermentation.In order to reduce the dimensionality of the analysis, for each pathway, the p-values of all the samples in the succinate production phase were averaged together and this mean p-value was used to find the 10 "most significant" pathways (lowest mean p-value).These 10 "most significant" selected pathways were displayed as a heatmap showing the p-value of each pathway for each sample.Heatmaps were created using the gplots package 55 (Version 3.1.3)in the statistical software environment R (Version 4.0.4).MPEA for intracellular and extracellular samples was performed separately.www.nature.com/scientificreports/ditions were filtered using the interquartile range, and the data was log 10 transformed and normalised by the median to adjust for systematic differences among samples (Figs.S3 and S4 in the Supplementary materials).The number of latent variables for the PLS-DA model was chosen by maximising Q 2 using tenfold cross validation.Those features with a VIP score above 1 in the first latent variable were considered as significant for group discrimination when analysing the results.
Figure 1 .
Figure 1.Profile of the E. coli succinate fermentation process showing the main parameters measured by HPLC-UV/Vis-RI.Time is indicated with respect to the beginning of the succinate production phase.Error bars represent the standard error of the mean of the three replicates (n = 3).
Figure 2 .
Figure 2. Top 10 pathways with the lowest mean p-value in the succinate production phase based on MPEA using the mPLAGE algorithm for intracellular metabolites found with a combined targeted and untargeted metabolomics method.Pathway names are shown on the y-axis (in increasing average p-value in the production phase going down) and time points of the fermentation process on the x-axis.The values of the cells are the p-values (n = 3) comparing each time point to the first time point of the fermentation analysed (3 h after the beginning of the aerobic batch phase).The resulting colour corresponds to the level of significance: green (p-values ≤ 0.05) and beige (p-values > 0.05).The vertical black dashed line indicates the transition from the aerobic growth phase to the anaerobic succinate production phase.The horizontal red dashed lines indicate the three pathways highlighted as potential targets for strain engineering.
Figure 4 .
Figure 4. (A) Scores plot between the latent variables (components) 1 and 2 for the PLS-DA of the three fermentation replicates analysed, which were split into two experiments with a high succinate titre (n = 12) and one with low succinate titre (n = 6).The explained variances of each latent variable are shown in brackets.(B) PLS-DA classification using different number of latent variables (components).The red star indicates the best classifier based on maximising Q 2 .
Figure 5 .
Figure 5. Three metabolites involved in pantothenate and CoA biosynthesis found important for discrimination of a successful E. coli succinate fermentation run according to the PLS-DA model (VIP score > 1 in at least one of the four latent variables) and with a high level of annotation confidence.The boxplots show the peak intensity of the three metabolites in the two sample groups: high (red) and low (green) succinate titre (n = 12 and 6, respectively).The peak intensity values are shown as black dots, the black horizontal line represents the population median and the yellow diamond represents the population mean.For each metabolite, the subplot on the left shows the results with the original peak intensity data, and the subplot on the right the results after log 10 transformation and normalisation of the data.
-DA.PLS-DA was performed in MetaboAnalyst (date of use: 08 Sep 2022) with R (Version 4.1.3).Constant features across all samples were deleted and missing values were replaced by 1/5 of the minimum positive values of their corresponding variables.Variables with near-constant values throughout the experiment conhttps://doi.org/10.1038/s41598-023-39661-x 00 g/L (NH 4 ) 2 SO 4 , 14.60 g/L K 2 HPO 4 , 3.60 g/L NaH 2 PO 4 •2H 2 O, 0.50 g/L (NH 4 ) 2 H-citrate), a mix of trace elements (1.0 mg/L CaCl 2 •2H 2 O, 20.06 mg/L FeCl 3 , 0.36 mg/L ZnSO 4 •7H 2 O, 0.32 mg/L CuSO 4 •5H 2 O, 0.30 mg/L MnSO 4 •H 2 O, 0.36 mg/L CoCl 2 •6H 2 O, 44.60 mg/L Na 2 EDTA•2H 2 O), antibiotics (100 mg/L kanamycin, 34 mg/L chloramphenicol) and antifoam (33.33 µL/L polypropylene glycol P Intracellular fractions.Prior to extraction, intracellular fractions were washed by resuspending the cell pellets in 1 mL of sterile phosphate buffer solution.The phosphate buffer was removed by spinning down the samples twice for 10 min at 13,000 g and 4 °C.For metabolite extraction, 200 µL of 1:3:1 C:M:W were added for every 5 mg of WCW pellet, thus normalising all samples by their biomass concentration.Cell pellets were resuspended by pipetting, and then the samples were mixed vigorously in a chilled microtube mixer for 1 h, before being centrifuged for 3 min at 13,000g and 4 °C.At this point, 200 µL of supernatant were transferred into a new microtube.The samples were further diluted by adding 200 µL 1:3:1 C:M:W and stored at − 80 °C until LC-MS analysis.25 µL of supernatant of each extracted sample were combined into one single vial to generate a pooled sample.During handling, the 1:3:1 C:M:W extraction solvent and the samples were kept on an ethanol dry ice bath.The first two fermentation samples had a very small biomass pellet, which made the extraction process impractical, particularly the resuspension of the pellet in the amount of extraction solvent required for sample normalisation.For this reason, these first two time points from the intracellular fraction were not included in the analysis. Vol:.(1234567890) Scientific Reports | (2023) 13:12990 | https://doi.org/10.1038/s41598-023-39661-x | 9,051 | 2023-08-10T00:00:00.000 | [
"Biology",
"Engineering"
] |
Substantiation Of The Operating Mode Of The Pendulum Feeder
Today, auto-feeding is the most effective way of feeding and it is less time consuming because feeding fish is a mandatory process in all enterprises, engaged in fish farming.
INTRODUCTION
In the climatic conditions of the Republic of Uzbekistan, in the absence of large water resources and with a shortage of live food organisms, such as zooplankton, the use of high-calorie compound feeds will be very effective. Currently, for some cultivated fish species, feeds are being developed containing hydrolysis products and a large amount of low molecular weight protein.
Automated control of all processes in pond fish farming starting from pond preparation for stocking of fish to placement and production of commercial fish will significantly increase the efficiency of fish farming by 70-80% thus increasing the overall profit by 20-30% [1]. In this regard, in the industrial cultivation of fish, the use of biologically active substances that promote the absorption of plant proteins is of great importance. The metabolic rate in fish is The American Journal of Applied Sciences (ISSN -2689-0992) determined by the temperature of the water. The temperature spectrum of vital activity occurs both more intense and less intense. In cold waters, metabolism proceeds more slowly than in warm waters, because, in warm waters, oxidative processes are activated in its tissues, and oxygen shortage increases. With an increasing of water temperature, the content of dissolved oxygen decreases in fish and the intensity of respiration increases. The best digestion of feed occurs at a water temperature of 10-15 ° C, and the maximum growth rate with the least use of food energy is observed at a temperature of 16-18°C. At present, the most economical way of feeding is dry food because wet food have some disadvantages during transportation, storage, as these products contain up to 80% moisture and cannot be stored for a long time, as they quickly deteriorate [2].
THE MAIN FINDINGS AND RESULTS
Physiologically adequate to this food are protein raw material hydrolysates with a certain degree of hydrolysis, characterized by a low content of free amino acids and an increased level of easily assimilated oligopolypeptides. Therefore, research is currently needed to search for food components containing a significant amount of low molecular weight protein. The nutritional needs of fish change with age, depending on the formation of digestive functions and the stage of development of fish. One of the sources of easily digestible fodder protein is industrial protein hydrolysates, autolysates, and fermentolysates, since the products of protein hydrolysis are well absorbed by early young fish. The objective is to have a good effect on the growth rate and allow a decrease in juvenile mortality, increasing in rearing efficiency and lower food costs. Insufficient amount of food at the developmental stage increases death, leads to growth retardation, and also prevents the development of the digestive system. For normal growth and development, fish need a certain amount and ratio of basic nutrients. Protein with a set of essential amino acids, fat, carbohydrates, minerals, vitamins and other biological active substances should be in the feed composition in accordance with the needs of fish. Moreover, the need for fish varies depending on age, size, water temperature and other environmental factors. It is believed that food containing at least 15 g of fishmeal are fully provided with minerals, fish also need vitamins and other biologically active substances. To date, the need for fish in 15 vitamins and vitamin-like substances has been established. Thus, in industrial production, the basis for the nutrition of cultivated fish is compound food, composed on the basis of dry flour-like components according to special recipes [3]. Its effectiveness depends on the level of protein, fat, carbohydrates, minerals and vitamins, as well as the balance of amino acids, fatty acids and vitamins. The main nutrients of the food are protein with essential amino acids, fat with essential fatty acids, simple and complex carbohydrates, minerals and vitamin-enzyme complexes. The latter, like vitamins, do not carry energy, but the growth and development of the body is impossible without them.
If the diet for fish has the required amount of fat and carbohydrates, then protein is used in protein metabolism for the growth of the body. With a lack of fats and carbohydrates in the food, proteins can be used as an energy source in functional metabolism. This is not economical because protein is the most expensive part of the food.
One of the most important aspects of feeding juvenile fish is to establish the correct daily feeding rate, which would satisfy the need for growing juveniles for food.
The correct food has a high physiological calorie content of 17-18 thousand kJ/kg, its usage requires strict rationing, taking into account the whole complex of factors affecting the establishment of daily diets. For The American Journal of Applied Sciences (ISSN -2689-0992) The analysis show that the optimal feeding scheme is granular fooding, depending on the body weight of the fish, which contributes to increased survival. Compound food come in the form of pellets of various sizes, and their composition can be combined in different ways, taking into account the size, type and age of the fish. Fodders in the form of granules have a porous structure, a specific gravity of less than one, and water does not sink for a considerable time.
In conditions of industrial fish farming, the intensity of water exchange is of great importance for feeding fish, which should ensure not only oxygen delivery, but also removal of metabolic products.
The use of automatic feeders is very effective in growing fish. To increase the efficiency of cultivation, it is necessary to observe feeding technology, which also includes the frequency of feeding and the ratio of the size of granules and body weight of fish. The problem of feeding is one of the leading places in the technological scheme of fish cultivation. Since feed should not only meet the body's needs for energy and basic nutrients, but also consist of components available for assimilation, including protein. This can be solved by balancing the fractional composition of food components and, in particular, protein. The use of new types of high-tech compound food makes it possible to grow viable young with a deficiency or complete absence of living food organisms in the diet. However, their use is difficult due to the lack of effective technologies for feeding fishThe use of automatic feeders is very effective in growing fish and reduces feed costs. Because, the efficiency of food consumption, as well as the average daily weight gain of juveniles is slightly higher, at low feed costs, in comparison with manual feeding. The main task of commercial fish farming is to grow fish in the shortest possible time and at minimal cost. One of the main factors affecting the rapid growth of fish is the maintenance of optimal growing conditions and the usefulness of feeding [5].
Feeding fish can be done until complete eating and according to certain standards. Normalization of feeding is preferable to feeding until complete eating, since it seems possible to more reliably take into account the influence of water temperature and the mass of fish grown.
Pendulum feeders are devices that deliver feed at the request of a fish, which for this purpose deflects the pendulum or pulls the ball. Pendulum feeders are the simplest and most utilitarian in their industry. These feeders are used for fish farming in pond farms, in private reservoirs. The difference between pendulum feeders is the lack of settings for the electronics and power source. The food gets enough spill out after touching the fish to the end of the rod of the feeder. Thus, the fish quickly develops a reflex, and it eats on its own, the likelihood of the appearance of not eaten food and starvation of the fish is reduced. The mechanical operation of the feeder allows you to install devices in places where the operation of electrical devices is not possible: in remote households, without connecting to power networks, in open water bodies, in inaccessible places with rare service feeders. 1-cap; 2 -hopper; 3 -traction for opening the cap; 4 -bracket. 5 -a supporting glass; 6-crossbeam; 7-screw; 8-ball support; 9-pendulum; 10 -a nut; 11 -a table; 12 -loop-shaped dropper of granul; 13 -a protective pin; 14 -moisture protective casing.
The pendulum feeder can be equipped with a hopper of any size. The main parts of the feeder are made of special plastic, which allows you to operate the device in water for a longer time than products that undergo corrosion.
Auto feeders favorably differ from devices of this type in the ability to adjust the feed. In this model, it is possible to reduce or increase the distance between the hole in the hopper and the food platform. The amplitude of the pendulum is limited by a special plastic cone.
Daily rate of feeding fish with dry granular food,% to body weight. Obtaining the maximum and economically profitable increase in fish production in pond conditions is possible using pendulum feeders that ensure the distribution of food in accordance with the regime chosen by the fish itself. The use of pendulum feeders can increase fish productivity, reduce food losses and costs, and significantly increase planting density.
Temperature
The depth of installation of the feeders can vary from 10-1.5 to 2-2.5 m. It is important to ensure that the distance from the feeding table to the surface of the water is at least 40-50 sm to avoid getting water on it when feeding fish and wind excitement. Immersion of pendulums to a depth, as a rule, has no strict restrictions. However, care must be taken that they do not touch the bottom of the pond, plants and other objects. The desired length of the pendulum, depending on the location of the feeder, can be considered 1.5-2 m, and the distance to the bottom is 10-15 sm. In the case of very dense fish landing, the distance from the pendulum to the bottom of the pond does not matter.
The process of teaching fish to eat from selffeeders for each pond, depends on the specific conditions where it can proceed in different ways. Usually, active consumption of animal feed coincides with the spring rise of water temperature. Therefore, in order to accustom the fish to the places where the feeders are based, food should be distributed first directly under the pendulums. A sign of the active response of fish to the self-feeding trough is the formation of a "rash funnel" inside them.
CONCLUSION
Currently, the development of economically profitable fish farms in different regions of the country requires fish production using a scientifically based technological system, the basis of which is the effective feeding of fish. Operation of pendulum feeders will help to fully realization of the benefits of self-feeding. | 2,527 | 2020-11-30T00:00:00.000 | [
"Physics"
] |
Video Liveness Verification
The ubiquitous and connected nature of camera loaded mobile devices has greatly estimated the value and importance of visual information they capture. Today, sending videos from camera phones uploaded by unknown users is relevant on news networks, and banking customers expect to be able to deposit checks using mobile devices. In this paper we represent Movee, a system that addresses the fundamental question of whether the visual stream exchange by a user has been captured live on a mobile device, and has not been tampered with by an adversary. Movee leverages the mobile device motion sensors and the inherent user movements during the shooting of the video. Movee exploits the observation that the movement of the scene recorded on the video stream should be related to the movement of the device simultaneously captured by the accelerometer. the last decade e-lecturing has become more and more popular. We model the distribution of correlation of temporal noise residue in a forged video as a Gaussian mixture model (GMM). We propose a twostep scheme to estimate the model parameters. Consequently, a Bayesian classifier is used to find the optimal threshold value based on the estimated parameters.
The ubiquitous and connected nature of camera loaded mobile devices has greatly estimated the value and importance of visual information they capture. Today, sending videos from camera phones uploaded by unknown users is relevant on news networks, and ing customers expect to be able to deposit checks using mobile devices. In this paper we represent Movee, a system that addresses the fundamental question of whether the visual stream exchange by a user has been captured live on a mobile device, and t been tampered with by an adversary. Movee leverages the mobile device motion sensors and the inherent user movements during the shooting of the video. Movee exploits the observation that the movement of the scene recorded on the video stream related to the movement of the device simultaneously captured by the accelerometer. the lecturing has become more and more popular. We model the distribution of correlation of temporal noise residue in a forged video as a Gaussian del (GMM). We propose a twostep scheme to estimate the model parameters. Consequently, a Bayesian classifier is used to find the optimal threshold value based on the estimated Lecture videos, automatic video indexing, lecture video archives video has become a popular storage and exchange medium due to the rapid development in recording technology, improved video compression speed networks in the last few years. Therefore audio visual recordings are used lecturing systems. A number of universities and research institutions are taking the chance to record their lectures and represent them online for students to access independent of time and location. As a result, there has been a large growth in the amount of multimedia data on the Web. Therefore, for a user it is nearly impossible to find desired videos without a finding function within a video archive. Even when the user has get related video data, it is still difficult most of the time for him to judge whether a video is useful by only flash at the title and other international which are often brief and high level. Moreover, the requested information may be covered in only a few minutes, the user might thus want to find the piece of information he requires without viewing the complete video. In response to the ubiquitous and connected nature of mobile and wearable devices, industries such as utilities, insurance, banking, retail, and broadcast news have started information gleaned from or created using devices. Mobile apps utilize mobile and wearable device cameras for purposes varying from authentication to location verification, tracking, witnessing, and remote assistance. deposit a check using a mobile phone, from mobile phones uploaded by unknown users shown on broadcast news to a national audience. correlation measurement between the reference pattern noise image and pattern noise image is used here. The sensor pattern noise scanner model identification and tampering detection of scanned images [8]. In [8], in addition to camera source identification, sensor pattern noise was first utilized for image forgery detection. This method proposed an accurate pattern noise extraction scheme.
Scientific (IJTSRD)
International Open Access Journal number of universities and research institutions are to record their lectures and line for students to access independent of time and location. As a result, there n the amount of multimedia data on the Web. Therefore, for a user it is nearly impossible to find desired videos without a finding function within a video archive. Even when the user related video data, it is still difficult most of im to judge whether a video is useful by flash at the title and other international metadata which are often brief and high level. Moreover, the requested information may be covered in only a few minutes, the user might thus want to find the piece of information he requires without viewing the complete In response to the ubiquitous and connected mobile and wearable devices, industries insurance, banking, retail, and broadcast news have started to trust visual tion gleaned from or created using mobile devices. Mobile apps utilize mobile and wearable device cameras for purposes varying from location verification, tracking, witnessing, and remote assistance. Today, one can mobile phone, and videos from mobile phones uploaded by unknown users are shown on broadcast news to a national audience. The correlation measurement between the reference pattern noise image and pattern noise image is used here. The sensor pattern noise has also been used for scanner model identification and tampering detection of scanned images [8]. In [8], in addition to camera source identification, sensor pattern noise was first utilized for image forgery detection. This method ttern noise extraction scheme. The above methods [ Page: 2450 number of images captured from specific video cameras to extract the sensor pattern noise of the cameras. Besides, it is difficult to extract sensor pattern noise from a video without a extensive variety of video contents.
II.EXISTING SYSTEM
Information retrieval in the multimedia-based learning domain is an active and integrative research area. Video texts, spoken language, community tagging, manual annotations, video actions, or gestures of speakers can act as the source to open up the content of lectures.
A. Slide Video Segmentation F. Chang, C [13].Video browsing can be achieve by segmenting video into representative key frames. The chosen key frames can provide a visual guideline for navigation in the lecture video portal. Moreover, video segmentation and key-frame selection is also often armed as a preprocessing for other analysis tasks such as video OCR, visual concept revelation, etc. Choosing a sufficient segmentation method is based on the definition of "video segment" and usually depends on the brand of the video. In the lecture video domain, the video sequence of an individual lecture topic or subtopic is often considered as a video segment. This can be roughly resolve by analyzing the materialistic scope of lecture slides. Many methods (as, e.g., [11]) make use of global pixellevel-differencing metrics for capturing slide transitions. A disadvantage of this kind of method is that the salt and pepper noise of video signal can affect the segmentation accuracy. After analyzing the content of lecture slides, we realize that the major content as, e.g., text lines, figures, tables, etc., can be considered as Connected Components (CCs). We therefore initiate to use CC instead of pixel as the basis element for the differencing analysis. We call it component-level-differencing metric. We solve the fundamental question of whether the visual stream uploaded by a user has been captured live on a mobile device, and has not been tampered with by a malicious user attempting to game the system. We refer to this problem as video "liveness" verification. The practical attacks we consider are feeding a previously recorded video through man in the middle software ("Copy-Paste" attack) and pointing the camera to a projection of a video ("Projection" attack). This problem is a cornerstone in a variety of practical applications that use the mobile device camera as a trusted witness. Examples applications include citizen journalism, where people record witnessed events (e.g., public protests, natural or man-made disasters) and share their records with the community at large. Other applications add video based proofs of physical possession of products and prototypes (e.g., for sites like Kick starter [5], Amazon [1] and eBay [3]), and of deposited checks [11], [3] In this paper we represent Movee, a motion sensor based video liveness verification system. Movee edge the pervasive mobile device accelerometers and the fundamental movements of the user's hand and body during the shooting of the video. Movee exploits the intuition that video frames and accelerometer data captured simultaneously will bear certain relations. Specifically, the movement of the scene recorded in the video stream should be related to the movement of the device registered by the accelerometer. We conjecture that such relations are hard to fabricate and emulate. If the data from the accelerometer corroborates the data from the camera, Movee proved that the video stream was genuine, and has been taken by the user pointing the camera to a real scene. 2 In essence, Movee provides CAPTCHA like verifications, by adding the user, through her mobile device, into the visual verification process. However, instead of using the cognitive strength of humans to read visual information, we rely on their innately flawed ability to hold a camera still. Movee can also belook as a visual public notary that stamps an untrusted video stream, with data simultaneously captured from a trusted sensor. This data can later be used to verify the liveness of the video. Previous work has proposed to use audio streams in captured videos to protect against spoofing attacks in biometric authentication. The current verifications use static and International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470 @ IJTSRD | Available Online @ www.ijtsrd.com | Volume -2 | Issue -3 | Mar-Apr 2018
III. PROPOSED SYSTEM
Page: 2451 dynamic relations between the recorded voice and the motion of the user's face. In this paper we use the previously unexplored combination of video and accelerometer data to solve a different problem: verify the liveness of the video capture process. Movee consists of four modules, explained in Figure 1. The Video Motion Analysis (VMA) module processes the video stream captured by the camera. It uses video processing techniques to infer the motion of the camera, generating a time-dependent motion vector. VMA is inspired by the process used in image stabilization capable cameras.
III THE MODEL: SYSTEM AND ADVERSARY
We now describe the system and adversary models that we assume in this work.
A. System Model We consider a system that consists of a service provider, e.g. video sharing services such as Vine YouTube or check deposit services The provider offers an interface for subscribers to upload or stream videos they shot on their mobile devices. We assume subscribers own mobile devices adapted with a camera and inertial sensors (i.e., accelerometers). Devices have Internet connectivity, which, for the purpose of this work may be infrequent. Donor need to install an application on their mobile devices, which we henceforth denote as the "client". A subscriber needs to use this client to capture videos. In addition to video, the client simultaneously captures the inertial sensor (accelerometer) stream from the device. The client uploads both the video and the accelerometer streams to the provider. The provider verifies the authenticity of the video by checking the consistency of the two streams. The verification is performed using restricted information: the two streams are from independent sources, but have been captured at the same time on the same device. We assume a system where the problems of constructing trust in the mobile device, operating system and identify drivers, and the mobile client are already addressed. This added for instance a system where a a chain of trust has been developed The chain of trust ensures that the operating system, adding the camera and sensor device drivers, and the installed apps, are trusted and have not been tampered with by an attacker, see e.g., [4], [6]. A discussion of limitations is included in Section VII. In the remainder of the paper we use the terms accelerometer and inertial sensor interchangeably.
B. Adversary Model
We assume that the service provider is honest. Users however can be malicious. An adversarial user can tamper with or copy video streams and inertial sensor data. The goal is to fraudulently claim ownership of videos they upload to the provider. Let V be such a video. The adversary can use a trusted device to launch the following attacks, that generate fraudulent videos or fraudulent video and acceleration data: Fig. 2. The Video Motion Analysis module processes each consecutive video frame and finds the motion vector by computing the amount of displacement that common image components have shifted between two frames.
Algorithm Video Segmentation
Consider the frames which are converted from the video by using this code Let F(k) be the kth frames in the given video, where k will takes the values from k=1,2,3,….n, the shot boundary detection algorithm for the above frames can be explained as follows Step 1: Split the given frames into block with m rows & n Columns B(i,j,k) stands for the block at (i ,j) in the given frame.
Step 2: Computing the histogram matching difference between the neighboring blocks in consecutive frames for a video sequence H(i ,j,k) and H(i ,j,k+1) stands for the histogram of blocks at (i,j) in the kth and (k+1) th frames respectively the block difference is measured by using the flowing equation Where DB = block difference Step 3: Evaluate the histogram difference between the two consecutive frames by Where Wij is the weight of the block at (i ,j) Step 4: calculating the threshold by the use of mean and standard variance of histogram which are differ over the whole video sequence and is different for different kind of information extracted. Mean and standard variance can be calculating by using the following equations.
Step 5: Calculating the total number of frames Copy-Paste attack. Copy V and output it.
Projection attack. Point the camera of the device over a projection of the target video. Output the result. Random movement attack. Move the device in a random direction, and capture the resulting acceleration data. Output the video V and the captured acceleration stream.
Direction sync attack. Use the video to infer the dominant motion direction of V . Use the device to capture an acceleration sample that encodes the same motion direction. Output V and the acceleration sample.
Cluster attack. Capture a dataset of videos and associated sensor streams. Use a clustering algorithm (e.g., K-means [12]) to group the videos based on their movement (i.e., the values of video features extracted by the VMA module, see Section III-A and Section III-D). Assign V to the cluster containing videos whose movement is closest to V . Randomly choose one of the videos and associated sensor streams in the cluster, (V ′,A′). Output (V , A′).
Replay attack. Study the target video V . Then, holding a mobile device that captures acceleration data, emulate the movements observed in V . Let A′ be the acceleration data captured by the device during this process. Output (V,A′).
CONCLUSIONS
In this paper we have introduced the concept of "liveness" analysis, of verifying that a video has been shot live on a mobile device. We have proposed Movee, a system that relies on the accelerometer sensors ubiquitously deployed on most recent mobile devices to verify the liveness of a simultaneously captured video stream. We have implemented Movee, and, through extensive experiments, we have shown that (i) it is efficient in differentiating fraudulent and genuine videos and (ii) imposes reasonable overheads on the server. In future work we intend to integrate more sensors (e.g., gyroscope), as well as the use of MonoSLAM [12] as an alternative VMA implementation to improve accuracy. | 3,719.8 | 2018-04-26T00:00:00.000 | [
"Computer Science"
] |
Structural Biology and Crystallization Communications the Structure of Nmb1585, a Marr-family Regulator from Neisseria Meningitidis
The structure of the MarR-family transcription factor NMB1585 from Neisseria meningitidis has been solved using data extending to a resolution of 2.1 A ˚. Overall, the dimeric structure resembles those of other MarR proteins, with each subunit comprising a winged helix–turn–helix (wHtH) domain connected to an-helical dimerization domain. The spacing of the recognition helices of the wHtH domain indicates that NMB1585 is pre-configured for DNA binding, with a putative inducer pocket that is largely occluded by the side chains of two aromatic residues (Tyr29 and Trp53). NMB1585 was shown to bind to its own promoter region in a gel-shift assay, indicating that the protein acts as an auto-repressor.
Introduction
In Escherichia coli, the multiple antibiotic resistance (mar) locus regulates the expression of proteins that confer resistance to numerous exogenous factors such as antibiotics, organic solvents, oxidative stress and disinfectants (Alekshun et al., 2001;Ellison & Miller, 2006;Sulavik et al., 1995). Resistance to these antimicrobial agents or environments is believed to be determined primarily through the control of efflux pumps with a range of specificities, the expression of which is controlled locally by the binding of MarR to its cognate DNA, preventing initiation of gene transcription and thereby acting as a repressor (Alekshun et al., 2001).
E. coli MarR was the first of the MarR-family regulators to be described and forms the archetype for a family of homologous transcriptional regulators which are widely distributed amongst both archaea and prokaryotes. Some of these homologues are also known to control mar-type efflux pump operons, e.g. FarR in Neisseria gonorrhoeae, which controls the expression of the FarAB efflux pump mediating resistance to long-chain fatty acids (Lee et al., 2003), and MgrA in Staphylococcus aureus, which controls the expression of NorA, a multidrug transporter responsible for resistance to fluoroquinolones (Truong-Bolduc et al., 2005). In other cases, homologues have been recruited to different systems and regulate tissue-specific activities such as the adhesive properties of cells, haemolytic properties and regulation of protease expression (Ludwig et al., 1995;Marklund et al., 1992;Perego & Hoch, 1988;Saridakis et al., 2008).
Knowledge of the three-dimensional structures of the MarR-family regulators has contributed to understanding their mechanism of action. To date, the structures of more than 20 MarR-family regulators have been solved and deposited in the Protein Data Bank, including those of E. coli MarR (Alekshun et al., 2001), Bacillus subtilis OhrR in the unliganded state and bound to its cognate DNA (Hong et al., 2005), Deinococcus radiodurans HucR (Bordelon et al., 2006), Enterococcus faecalis SlyA (Wu et al., 2003), Methanobacterium thermoautotrophicum MarR in the unliganded state and with salicylate bound (Saridakis et al., 2008), Pseudomonas aeruginosa MexR (Lim et al., 2002), Sulfolobus tokodaii EmrR (Miyazono et al., 2007) and Xanthomonas campestris MarR (Chin et al., 2006). Many of these homologues share less than 20% sequence identity, but they all possess the same core fold. The proteins are homodimers comprising a largely helical dimerization domain linked to a DNAbinding domain that contains a winged helix-turn-helix motif. MarR proteins repress the activity of their target genes by binding as dimers to pseudopalindromic sequences in the À10 region of the regulated promoters. The repressor activity of MarR proteins is modulated by co-inducer binding, or in the case of OhrR oxidation of cysteines disrupting disulfide-bridge formation, both of which lead to a major rearrangement in the dimerization region such that the spacing of the DNA-binding domains is significantly altered, preventing DNA recognition.
Neisseria meningitidis encodes two MarR-family repressors, NMB1853, a homologue of FarR found in N. gonorrhoeae which presumably also controls expression of the FarRAB efflux pump, and a second MarR, NMB1585, of unknown function. In N. gonorrhoeae the close homologue of NMB1585 encoded by NGO1244 has been shown to be part of the RpoH regulon and is upregulated in response to temperature stress (Gunesekere et al., 2006). Given the potential importance of NMB1585 in the pathophysiology of N. meningitidis, we have targeted this protein for structural studies and in this report we describe the structure at 2.1 Å resolution.
Protein production and crystallization
The NMB1585 expression construct was generated by means of ligation-independent cloning using Gateway technology (Invitrogen). NMB1585 was amplified from genomic DNA (N. meningitidis strain MC58) with KOD HiFi polymerase (Novagen) using the forward primer 5 0 -GGGGACAAGTTTGTACAAAAAAGCAGGCTTCCT-GGAAGTTCTGTTCCAGGGCCCGATGAACCAACTCGACCA-ACTTGGC-3 0 and the reverse primer 5 0 -GGGGACCACTTTGTA-CAAGAAAGCTGGGTCTCACTATTTTTTATTTTCCGAGATT-GTTTTTTC-3 0 . The PCR product was purified using QIAquick 96 plates (Qiagen) and cloned into the expression vector pDEST17 in two steps according to the manufacturer's protocol (Invitrogen), resulting in a construct with an N-terminal His tag and 3C protease cleavage site. BP and LR reactions were carried out according to the manufacturer's instructions. Recombinant LR clones were identified by PCR using a gene-specific forward primer and a T7 reverse primer and verified by DNA sequencing. Protein was produced in E. coli strain B834 (DE3). The cells were grown at 310 K in GS96 media (QBiogene) to an A 600 of 0.6, induced by the addition of 0.5 mM IPTG and then incubated for a further 20 h at 293 K. The cells were harvested by centrifugation at 6000g for 15 min and lysed using a Basic Z Cell Disruptor (Constant Systems Ltd) at 207 MPa in 50 mM Tris pH 7.5, 500 mM NaCl, 0.2%(v/v) Tween-20. The protein was purified by nickel-affinity chromatography followed by size-exclusion chromatography using the standard His Affinity-Gel filtration program on the Ä KTA 3D (GE Healthcare). After centrifugation at 30 000g for 30 min, the lysate was loaded onto a 1 ml pre-charged HiTrap Chelating Sepharose FF column (GE Healthcare). The column was washed with 50 mM Tris pH 7.5, 500 mM NaCl, 20 mM imidazole. The protein was then eluted in 50 mM Tris pH 7.5, 500 mM NaCl, 500 mM imidazole and injected onto a 16/60 HiLoad Superdex 200 column (GE Healthcare) equilibrated in 20 mM Tris pH 7.5, 200 mM NaCl. Protein-containing fractions were analyzed on SDS-PAGE gels (Biorad). The N-terminal tag was removed by overnight incubation at 277 K with His-tagged 3C protease (prepared from pET-24/His-3C kindly provided by A. Geerlof, EMBL, Heidelberg). The 3C protease and any uncleaved protein were removed by nickelaffinity chromatography and the protein was concentrated to 9.7 mg ml À1 using a Vivaspin 15 concentrator with 5 kDa molecularweight cutoff (Vivascience) in 20 mM Tris pH 7.5, 200 mM NaCl, 1 mM tris(2-carboxyethyl)phosphine (TCEP). The protein was crys-tallized using a nanodrop crystallization procedure (Walter et al., 2005). Crystals were initially obtained in 0.1 M HEPES buffer pH 7.5 containing 25%(w/v) PEG 3350, 0.2 M ammonium chloride and growth was optimized by varying the pH of the precipitant by addition of acid/base as described by Walter et al. (2005). The crystals of NMB1585 used for data collection were partially dehydrated/cryoprotected by a three-stage transfer to 40%(w/v) polyethylene glycol 3350, 15%(v/v) ethylene glycol and were flash-frozen in a 100 K nitrogen cold stream prior to data collection. Owing to their superior diffraction properties compared with methionine-containing native crystals, data from selenomethionine-labelled crystals were used for both structure determination and refinement.
Crystallography methods
Multiwavelength X-ray diffraction data were collected from selenomethionine-labelled NMB1585 crystals on beamline BM14 at the ESRF, Grenoble. Data were indexed, integrated and scaled using DENZO and SCALEPACK (Otwinowski & Minor, 1997; Table 1). The protein was crystallized in space group P2 1 , with two subunits per asymmetric unit. The selenomethionine substructure of the NMB1585 crystal was solved by multiwavelength anomalous dispersion methods using SHELXD (Sheldrick, 2008). Three of four possible selenium sites were thus located and were used to obtain an initial phase set (SHELXE; phase extension to 2.1 Å , contrast 0.499, connectivity 0.930, pseudo-free CC 70.64%, mean FOM 0.62). Further density modification and initial model building were then performed with RESOLVE (Terwilliger, 2004), yielding a dimeric starting model with 75% of the expected number of residues and 50% of the sequence threaded. This model was then refined with CNS (Brü nger et al., 1998), iterated with several rounds of rebuilding in O (Jones et al., 1991). Final statistics are given in Table 1.
Electrophoretic mobility shift assay
A 378 bp probe corresponding to the intergenic sequence between NMB1584 and NMB1585 was amplified from genomic DNA using the following pair of PCR primers: forward primer 5 0 -CGAACAGG-ACGTTTCCGGCG-3 0 and reverse primer 5 0 -CATTGCAAATCA-GGTTGATACGG-3 0 . A fluorescence-detection method was used for the electrophoretic mobility shift assay (EMSA), as described by the manufacturer (Electrophoretic Mobility Shift Assay Kit, Invitrogen). Briefly, DNA (20 nM) was incubated at room temperature with increasing amounts of purified NMB1585 protein (up to 640 nM for the dimeric form) in a total volume of 10 ml containing 1Â EMSA binding buffer (10 mM Tris pH 7.4, 0.1 mM dithiothreitol, 0.1 mM EDTA, 150 mM KCl). Following the addition of 2 ml 6Â EMSA gelloading buffer, the samples were loaded onto a pre-cast tris-borate-EDTA (89 mM Tris base, 89 mM boric acid, 1 mM EDTA pH 8.0) gel (10%; Invitrogen) that had been pre-equilibrated for 15 min at 120 V in ice-cold 0.5Â TBE running buffer (Invitrogen). The gels were run at 120 V for 15 min followed by 160 V for a further 70 min. The gels were stained for 20 min in the dark with SYBR Green EMSA nucleic acid gel stain in 1Â TBE buffer. After washing twice in distilled water for 10 s each time, the gels were visualized on a UV-light transilluminator using the Gene Genius Bio imaging system (Syngene).
Overall structure
The structure of NMB1585 was solved to a resolution of 2.1 Å by the multiple-wavelength anomalous dispersion method using selenomethionine-substituted protein. Like other members of this family of regulators, the meningococcal MarR monomer structure is pre- Structure-based alignment of NMB1585 with MarR sequences of known structure. The sequences of NMB1585 (N. meningitidis), MarR (E. coli; PDB code 1jgs), OhrR (B. subtilis; PDB code 1z91) and MTH313 (M. thermoautotrophicum; PDB code 3bpv) were aligned using ClustalW and displayed with secondary structures using ESPript2.2. The residues proposed to be involved in DNA binding (Gln58, Thr59, Ser61, Arg83) and ligand binding (Tyr29, Trp53) are indicated by arrows. ) 26/27 † The data are essentially complete to 2.3 Å resolution. ‡ R work and R free are defined by R = P hkl jF obs j À jF calc j = P hkl jF obs j, where hkl are the indices of the reflections (used in refinement for R work ; 5% not used in refinement for R free ) and F obs and F calc are the structure factors deduced from measured intensities and calculated from the model, respectively.
dominantly -helical, with an elongated 'arm' domain (1, 5 and 6) linked to a more compact 'wing' domain that contains a winged helixturn-helix motif (wHtH; topology 2-t-3-t-4-1-W1-2; Fig. 1). Consistent with other MarR-family regulators, NMB1585 is dimeric, with the two wHtH motif-containing 'wing' domains distal to the dimer interface (Figs. 1a and 1b). The dimer interface is thus formed by a symmetrical interaction of the arm domains of the two monomers and involves contacts between both N-terminal and C-terminal regions (approximately residues 1-25 and 110-142). The orientation of the two monomers in the dimer and their ability to pivot relative to one another has previously been shown to be an essential factor affecting the DNA-binding mode of MarR-type repressors such as MexR (Lim et al., 2002), HucR (Bordelon et al., 2006;Wilkinson & Grove, 2005) and MobR (Hiromoto et al., 2006). More recently, a comparison of unliganded and DNA-bound forms of OhrR identified clear conformational changes accompanying DNA binding (Hong et al., 2005), whilst analysis of M. thermoautotrophicum MarR (MTH313) crystal structures revealed a large reorientation of the 'arm' and 'wing' domains accompanying salicylate binding (Saridakis et al., 2008).
Comparison with other MarR structures
The structure of NMB1585 was superimposed onto that of the B. subitilis OhrR-DNA complex and shown to match it closely, with an r.m.s.d. of 2.4 Å for 213 equivalent C atoms of 274 (Fig. 1c). Thus, NMB1585 appears to be pre-configured for DNA binding, similar to the structures reported for the HucR regulator of D. radiodurans (Bordelon et al., 2006) and a MarR-family protein from S. tokodaii (Kumarevel et al., 2008). However, this does not appear to be typical amongst other MarR structures (Hong et al., 2005). By reference to the OhrR structure, the recognition site of NMB1585 is likely to be approximately 20 bp, with each monomer binding into consecutive major grooves of the DNA double helix. The residues in the recognition helix (4) that are most likely to be involved in binding in the major groove of DNA are Gln58, Thr59 and Ser61 (Fig. 2). In the OhrR complex, a highly conserved arginine (Arg94) residue makes the key contact between the wing of the wHtH motif and the minor groove of the DNA target and it has been proposed that this represents a generic interaction common to all MarR-family proteins (Hong et al., 2005;Kumarevel et al., 2008). The conformation of the wing loop of NMB1585 closely resembles that of OhrR and Arg83 would make a similar minor-groove contact (Figs. 1c and 2). It follows that discrimination between different DNA-binding sites must largely depend on the sequence and orientation of the recognition helix (4) of the HtH motif which contacts the major groove of the DNA.
A feature of the MarR family is their capacity to bind to a variety of effector molecules, generally phenolic compounds such as salicylate; in most cases, this results in a reduction of DNA binding . The results of cocrystallization experiments have shown that salicylate can bind at two different locations in MarR proteins. In the structure of E. coli MarR, a salicylate molecule was observed bound in two surface pockets (termed SalA and SalB) on each subunit that were located either side of the DNA-recognition helix (Alekshun et al. (2001). These positions are indicated in the structure of NMB1585, clearly showing how occupancy is likely to directly interfere with DNA binding (Fig. 3). However, the residues in E. coli MarR that form these surface pockets and interact with the salicylate molecules are not conserved in NMB1585, indicating that binding to this part of the protein is highly unlikely. In contrast, the salicylate-binding pocket identified in MTH313, a MarR protein from M. thermoautotrophicum, is conserved (Saridakis et al., 2008). In the MTH313 structure one salicylate molecule was observed bound to each subunit of the dimer at the interface of the DNA-binding domain and the helical dimerization domain. The two salicylates were observed to interact at different sites within the binding pocket (Fig. 3). The positions of the bound salicylates in MTH313 have been mapped onto the NMB1585 structure and are shown in Fig. 3(a); detailed views of the two binding sites are shown in the superimposition of the two structures (Figs. 3b and 3c). Overlaying of NMB1585 and MTH313 confirms the presence of a potential ligandbinding pocket in NMB1585 at a similar location to those in MTH313 and other MarR proteins (Saridakis et al., 2008). However, it is clear that the side chains of the residues that line the pocket, notably Tyr29, Tyr36 and Trp53, occupy much of the internal volume of the binding pocket, which would prevent a ligand such as salicylate from binding to this conformation of the protein (Figs. 3b and 3c). Therefore, in order for a ligand to bind into the binding pocket of NMB1585 a conformational change would have to occur in the protein so that the dimerization domain moved away from the DNA-binding domain. This would increase the separation of the 2-helix with respect to 3helix in each subunit, thus opening up the hydrophobic pocket.
DNA binding
The overall structure of NMB1585 confirms that the protein is a member of the MarR family of transcription repressors. DNA binding was verified experimentally in an EMSA experiment. Purified NMB1585 protein showed concentration-dependent binding to a double-stranded DNA probe corresponding to the region between the end of the upstream gene (NMB1584) and the start of the coding sequence for NMB1585 (Fig. 4). This region contains the promoter for NMB1585 and suggests that, in common with other MarR regulators [e.g. FarR (Lee et al., 2003) and MexR (Evans et al., 2001)], NMB1585 is an auto-regulator. Two potential DNA-protein complexes were observed in the EMSA experiments: a faster migrating species at low protein:DNA ratios and a complex of slower mobility at higher protein:DNA ratios. This suggests that there is more than one binding site for NMB1585 in the region between the NMB1564 and NMB1565 genes. Typically, MarR regulators bind to relatively short (pseudo)palindromic sequences consistent with the dimeric structure of the proteins, although the lengths of the inverted repeats and the spacing between half-sites is variable . Further experiments would be required, for example DNA footprinting, to identify the cognate DNA-binding sites of NMB1585. Interestingly, the addition of salicylate, a prototypical MarR ligand, did not affect the formation of the protein-DNA complexes (Fig. 4), suggesting that the protein does not interact with salicylate, in contrast to E. coli MarR (Alekshun et al., 2001) and MTH313 (Saridakis et al., 2008). This may be explained by the occluded nature of the putative binding site observed in the crystal structure of NMB1585 and suggests that the protein may adopt this conformation in solution.
The physiological role of NMB1585 has not been characterized and therefore the identity of any natural ligand(s) that may modulate its activity is unknown. Intriguingly, the gene immediately downstream of NMB1585 is annotated as a potential integral membrane protein (NMB1586) classified as a component of an ABC-type multidrug transport system, ATPase and permease. Transcription analysis of a NMB1585-knockout strain shows an increase in NMB1586 transcript expression in the absence of NMB1585 (N. J. Saunders, unpublished data). Given the role of MarR-family proteins in the regulation of the expression of efflux systems, it is tempting to speculate that NMB1585 may be a repressor of a transport protein involved in the export of xenobiotic compounds from Neisseria.
In conclusion, we describe the crystal structure of meningococcal MarR, which represents a highly adaptable fold widely used in transcriptional regulation in many bacteria with particular significance in controlling responses to changes in their chemical environment. | 4,244.8 | 0001-01-01T00:00:00.000 | [
"Chemistry"
] |
Transient defects of mitotic spindle geometry and chromosome segregation errors
Assembly of a bipolar mitotic spindle is essential to ensure accurate chromosome segregation and prevent aneuploidy, and severe mitotic spindle defects are typically associated with cell death. Recent studies have shown that mitotic spindles with initial geometric defects can undergo specific rearrangements so the cell can complete mitosis with a bipolar spindle and undergo bipolar chromosome segregation, thus preventing the risk of cell death associated with abnormal spindle structure. Although this may appear as an advantageous strategy, transient defects in spindle geometry may be even more threatening to a cell population or organism than permanent spindle defects. Indeed, transient spindle geometry defects cause high rates of chromosome mis-segregation and aneuploidy. In this review, we summarize our current knowledge on two specific types of transient spindle geometry defects (transient multipolarity and incomplete spindle pole separation) and describe how these mechanisms cause chromosome mis-segregation and aneuploidy. Finally, we discuss how these transient spindle defects may specifically contribute to the chromosomal instability observed in cancer cells.
Mitosis, the process by which a single eukaryotic cell partitions its genetic material, has fascinated scientists for over a century, and already in the late 1800s Walther Flemming described the processes of chromosome segregation and cell division with exquisite detail [1,2]. Four main structures, consisting of centrosomes, a microtubule-based mitotic spindle, kinetochores, and chromosomes [3,4], cooperate to form the mitotic apparatus ( Figure 1A) in vertebrate somatic cells. The centrosomes are specialized organelles, each consisting of a pair of centrioles and pericentriolar material, and they play a key role in mitotic spindle assembly by serving as the primary sites of microtubule nucleation [5,6]. Molecular motors act to move the replicated centrosomes to diametrically opposing positions around the nucleus [7][8][9] (Figure 1A, Prophase), thus ensuring assembly of a fusiform and symmetric microtubule-based mitotic spindle once the nuclear envelope breaks down. At the same time, within the nucleus, the chromosomes undergo significant condensation ( Figure 1A, Prophase) while kinetochores assemble on the centromeric region of each sister chromatid of the replicated chromosomes (reviewed in [10]). Upon nuclear envelope breakdown, which marks the beginning of prometaphase, the kinetochores become available for capture by dynamically searching microtubules ( Figure 1A, Prometaphase). The kinetochore is a large protein complex that constitutes the attachment site for microtubules of the mitotic spindle on each chromatid [11]. In addition to acting as attachment sites for microtubules, kinetochores are also part of a signaling pathway, termed the spindle assembly checkpoint (SAC), that facilitates the coordinated and accurate segregation of chromosomes by preventing anaphase onset until all kinetochores are bound to microtubules (reviewed in [12]). As mitosis progresses, chromosomes establish attachments with microtubules and undergo poleward and anti-poleward movements, which are generated by minus end and plus end directed motors located at the kinetochore as well as along the chromosome arms [13][14][15][16][17][18][19]. Eventually, all chromosomes align between the two centrosomes, at the equator of the mitotic spindle, forming the metaphase plate ( Figure 1A, Metaphase). Upon chromosome alignment, the SAC becomes satisfied [20], and the sister chromosomes segregate Figure 1 Diagrammatic representation of mitosis, the mitotic apparatus, and different types of kinetochore attachments. A. During the first stage of mitosis (prophase), the replicated chromosomes, still enclosed by the nuclear envelope, undergo condensation, while the replicated centrosomes move to diametrically opposing positions around the nucleus. Nuclear envelope breakdown marks the beginning of prometaphase, when kinetochores establish attachment with microtubules of the mitotic spindle. At the end of prometaphase, the chromosomes become aligned at the spindle equator forming the metaphase plate, and the cell is said to be in metaphase. The sister chromatids separate from each other and move to opposite poles of the mitotic spindle in anaphase. During the last stage of mitosis, telophase, the nuclear envelope begins to reassemble around the recently segregated chromosomes. Mitotic chromosome segregation is followed by cytokinesis, in which an actin-myosin contractile ring cleaves the cytoplasm to generate two individual daughter cells. B. Kinetochores and chromosomes can establish different types of attachments with microtubules during the early stages of mitosis. Monotelic attachment occurs when one sister kinetochore is attached to microtubules and the other sister is unattached. This is a typical first step in establishment of attachment during prometaphase. When the unattached sister kinetochore binds microtubules from the opposite spindle pole, the chromosome establishes amphitelic attachment. Amphitelic attachment is the only type of attachment that ensures correct chromosome segregation. Due to the stochastic nature of kinetochoremicrotubule interactions, chromosomes can occasionally establish erroneous attachments. These include syntelic attachment, in which the two sister kinetochores bind microtubules from the same spindle pole, and merotelic attachment, in which a single kinetochore binds microtubules from both spindle poles instead of just one. Persistence of merotelic attachment into anaphase causes a chromosome segregation defect in the form of a lagging chromosome (see text and Figure 2 for details). and move towards opposite spindle poles (Figure 1, Anaphase). All of these events must occur in a highly coordinated manner for accurate chromosome segregation into the two daughter cells. If any aspect of this process goes awry, cells may end up with an incorrect number of chromosomes, a state referred to as aneuploidy, which is the leading cause of mis-carriage and birth defects in humans and is a hallmark of cancer (reviewed in [21,22]). Thus, fidelity of the mitotic process is important for development and growth, as well as for homeostasis, repair, and renewal of adult tissues. In this review, we will focus on how defects in mitotic spindle geometry affect the fidelity of mitosis and how transient spindle geometry defects contribute to chromosomal instability in cancer cells.
Abnormal mitotic spindle geometry: permanent vs. transient The bipolar geometry of the mitotic spindle is essential for accurate chromosome segregation, and already a century ago Theodor Boveri postulated that supernumerary centrosomes could lead to the production of aneuploid cells [23,24]. Observations of mitosis in both transformed and non-transformed cells reveal that multipolar mitotic spindles do occasionally form [25][26][27][28][29][30][31]. Typically, chromosomes within multipolar spindles form metaphase plates that display branched, Y-, V-, or T-shaped configurations as a consequence of chromosome alignment between multiple spindle poles [25,26]. Although chromosomes can align between supernumerary spindle poles, multipolar cell division has been shown to cause cell death in the progeny [32], most likely due to the fact that it causes massive chromosome mis-segregation, thus producing daughter cells that have far fewer chromosomes than is needed for survival. Moreover, anaphase lagging chromosomes (chromosomes that lag behind while all the other chromosomes segregate to the spindle poles during anaphase) are also frequently observed during multipolar cell division [25,27], thus adding to the burden of chromosome missegregation in multipolar cell division. Given the risk, it is not surprising that multipolar spindle assembly is a rare event in non-transformed cells growing under optimal conditions. Conversely, multipolar spindle assembly is very common in cancer cells [28][29][30][33][34][35][36][37][38], yet cancer cells avoid multipolar cell division and subsequent cell death by exploiting a number of mechanisms that allow centrosome clustering prior to anaphase onset [32,[39][40][41]. Microtubule-associated proteins (e.g., NuMA, TPX2, ch-TOG, and ARL2), motor proteins (e.g., dynein and HSET), central spindle components, CLASPs, components of the Augmin complex, anillin, proteins involved in sister chromatid cohesion, kinetochore components, and chromosomal passenger proteins have all been implicated in the clustering of centrosomes prior to chromosome segregation [40][41][42][43][44][45]. The molecular and structural requirements that cause some cells to undergo multipolar anaphase and others to cluster their centrosomes are not yet know. However, variations in expression levels of minus end directed motors that oppose motors which serve to separate the centrosomes are likely involved. Alternatively, the orientation of the centrosomes as well as their distance from one another may serve to facilitate supernumerary centrosome clustering. These possibilities have yet to be explored, but studies investigating these issues would undoubtedly lead to a more comprehensive understanding of the specific mechanism(s) responsible for centrosome clustering. Regardless of the specific mechanisms, however, it is clear that cancer cells employ strategies to avoid permanent multipolarity, thus experiencing this mitotic spindle defect only transiently.
Another type of abnormal spindle geometry is observed in prometaphase cells with unseparated or incompletely separated spindle poles. In these cells, the centrosomes fail to migrate to opposing positions around the nucleus (a process driven by dynein and kinesin-5; reviewed in [46]) before the cell enters prometaphase (i.e., before nuclear envelope breakdown, NEB). If the centrosomes persisted in such unseparated/partially separated configuration, the cell would arrest in mitosis with a monopolar spindle, which would then lead to either mitotic slippage or cell death [47][48][49]. However, all cell types studied to date appear capable of completing centrosome separation after NEB thanks to a number of mechanisms, including Eg5 motor activity [46,50], myosin activity at the cell cortex [50,51], and kinetochore/kinetochore-microtubule-generated forces [52,53]. Once again, cells seem to have developed ways to avoid a permanent spindle defect and to limit monopolarity to a transient stage. The phenomenon of incomplete centrosome separation at NEB has been observed in a variety of cell types [52,[54][55][56] and, while it was initially described in the mid-1970s [55,56], only sporadic reports described this centrosome behavior over a period of several decades [51,54]. Recently, this phenomenon has received renewed attention [53,57] in part due to its correlation with increased rates of chromosome missegregation [57,58]. Why in some cells the centrosomes can reach diametrically opposing positions around the nucleus prior to NEB, whereas in other cells they cannot, is not clear. However, it has been proposed that the cause may be a lack of coordination between the timing of NEB and centrosome separation [55][56][57]. In any case, these cells achieve complete centrosome separation after NEB with no noticeable defects in subsequent mitotic spindle assembly or in the timing of mitosis as defined by the time between NEB and anaphase onset [52,55,[57][58][59]. Nevertheless, cells that complete centrosome separation after NEB will experience monopolarity or near-monopolarity (spindle geometry defect) over a time window in early prometaphase, when initial kinetochoremicrotubule attachments are being established.
Studies in various types of cancer cells showed that multipolar prometaphase cells display higher numbers of merotelic kinetochore attachments (one kinetochore bound to microtubules from two spindle poles instead of just one) compared to bipolar prometaphases ( [32,39]; Figure 2A). Moreover, both experimental and computational analysis showed that the number of merotelic kinetochores increases with increasing numbers of spindle poles [60]. To explain this correlation, it was proposed that the reduced distance between each pair of spindle poles would increase the likelihood of each kinetochore to be reached by and bind to microtubules from two spindle poles instead of just one [32,39].
Experimental and computational studies were also employed to investigate the process of establishment of kinetochore attachment in cells with incomplete centrosome separation at NEB. These studies showed that when the two spindle poles are very close to each other upon NEB, chromosomes are extremely likely to establish syntelic attachments (both sister kinetochores bound to microtubules from the same spindle pole) ( [57]; Figure 2B, top). As the spindle poles separate, such syntelic attachments can be partially corrected and converted into merotelic attachments ( [57]; Figure 2B, top). When, upon NEB, the spindle poles are farther apart, but not diametrically opposed, kinetochores can form merotelic attachments without transitioning through a syntelic intermediate ( [57]; Figure 2B, bottom). What is important to emphasize here is that in both cases of spindle geometry defects (multipolarity and near-monopolarity), large numbers of merotelic kinetochores are formed before spindle bipolarization ( Figure 2). Because the SAC cannot detect merotelic kinetochore attachment [61][62][63][64][65], cells can progress through mitosis in the presence of large numbers of A. Transient Multipolarity B. Incomplete Spindle Pole Separation Figure 2 Transient defects of mitotic spindle geometry and chromosome mis-segregation. A. Transient multipolarity mechanism, in which initial assembly of a multipolar spindle (first panel) favors the formation of merotelic kinetochore attachment (second panel). Subsequently, centrosome clustering/coalescence leads to mitotic spindle bipolarization (third panel). However, merotelic kinetochore attachment can persist into anaphase and produce a chromosome segregation defect in the form of an anaphase lagging chromosome (fourth panel). B. Incomplete spindle pole separation at NEB results in a transient spindle geometry defect that promotes formation of kinetochore mis-attachments. If the centrosomes are very close to one another (top row), chromosomes are extremely likely to form syntelic attachments (first panel), which can be converted into merotelic attachments upon spindle bipolarization (second panel). If the centrosomes are not completely separated, but more than a few microns apart (bottom row), merotelic attachments can form directly without transitioning through a syntelic intermediate (first and second panel). In both cases, merotelic attachments can persist through mitosis (third panels) and induce chromosome mis-segregation in the form of anaphase lagging chromosomes (fourth panels). Adapted from [57]. merotelic kinetochores. Although merotelic kinetochores can be corrected by an Aurora B-dependent mechanism [66,67], mitosis will not halt to allow for correction, and therefore high rates of merotelic kinetochore formation invariably result in high rates of anaphase lagging chromosomes (reviewed in [21]). Indeed, both transient multipolarity and incomplete spindle pole separation have been shown to result in high rates of anaphase lagging chromosomes ( [32,39,58]; Figure 2). Thus, abnormal spindle geometry, albeit transient, can have detrimental effects on the fidelity of chromosome segregation, and represents a potential source of aneuploidy. Interestingly, over 70% of cancer cells from various sites are aneuploid [21,68], and many of them also display high rates of chromosome missegregation, a phenotype that leads to continuous changes in chromosome number, or chromosomal instability (CIN) [22,[69][70][71]. Recent studies have shown that merotelically attached anaphase lagging chromosomes represent the most common chromosome segregation defect in CIN cancer cells [32,39,72]. One cause of such high rates of anaphase lagging chromosomes appears to be the inefficiency of the correction mechanisms for kinetochore misattachments in cancer cells [73]. However, the high rates at which kinetochore mis-attachments (particularly merotelic) form are perhaps the major cause of chromosome mis-segregation in cancer cells, and the transient spindle geometry defects described above represent the most likely mechanisms of kinetochore mis-attachment formation in cancer cells ( [32,39]; Silkworth, Nardi, and Cimini, unpublished; see below for further discussion).
Transient spindle geometry defects in development, adult tissues, and cancer
The transient spindle geometry defects described here and their effects on the fidelity of chromosome segregation have been mainly characterized in tissue culture cells, with transient multipolarity exclusively observed in CIN cancer cells to date. Indeed, the frequencies of multipolar spindles in non-transformed or non-CIN cancer cells are typically very low (Silkworth, Nardi, and Cimini, unpublished). Given the causal relation between transient multipolarity and chromosome mis-segregation and that transient multipolarity is very common in CIN cancer cells, this mechanism is widely recognized as a major player in CIN. Whether this mechanism is also acting in tumors in situ has not been investigated. However, multipolar spindles have been observed in tumor tissues [74][75][76] and shortterm tumor cell cultures [77,78].
Although incomplete spindle pole separation at NEB has also been characterized mainly in tissue culture cells, the information available to date reveals that this mechanism is not exclusive to cancer cells, and has indeed been observed in several different types of tissue culture cells at frequencies of~40-45% [51,52,55,57,59].
Moreover, studies aimed at investigating various aspects of cell division provide useful information on the occurrence and relevance of this mechanism in contexts other than tissue culture. For example, in both the one-and two-cell stage of the Caenorhabditis elegans embryo, the centrosomes achieve opposing positions around the nucleus before the nuclear envelope breaks down [79]. Similarly, in the syncytial Drosophila embryo, the centrosomes always achieve diametric arrangement around prophase nuclei [80], and the nuclear envelope does not break down until the centrosomes have completed their movement around the nucleus [81]. This ability of cells in developing embryos to completely separate their centrosomes before NEB has also been observed in Drosophila melanogaster neurogenesis. Indeed, in both epidermoblasts and neuroblasts the centrosomes are completely separated before the onset of prometaphase [82]. Given the risk of chromosome mis-segregation associated with incomplete centrosome separation at NEB, it is not surprising that this defect is not observed during early development, as it would potentially lead to mosaic aneuploidy and possibly embryonic death.
The incidence of incomplete centrosome separation at NEB in normal adult tissues has not been investigated to date. However, a recent study showed that non-cancer RPE1 cells, which are known to maintain a stable karyotype with negligible rates of aneuploidy [72], always succeed to separate their centrosomes before NEB [83]. This observation suggests that centrosome separation in normal proliferating cells, like in developing embryo cells, may be better timed with NEB compared to cancer cells or certain stabilized tissue culture cells.
The open question, then, is whether incomplete centrosome separation at NEB may play a role in cancer cell CIN, and if so, to what extent. As discussed above, transient multipolarity is recognized as a major cause of chromosome mis-segregation in cancer cells [32,39,84]. However, some CIN cancer cell types cannot cluster the centrosomes of multipolar spindles very efficiently (Silkworth, Nardi, and Cimini, unpublished). Yet, these cells exhibit high rates of chromosome mis-segregation in the form of lagging chromosomes in bipolar anaphase cells (Silkworth, Nardi, and Cimini, unpublished), raising the possibility that incomplete spindle pole separation at NEB may play a role in promoting formation of kinetochore mis-attachments in these cells. Analysis of centrosome separation in early prometaphase shows that incomplete spindle pole separation at NEB can be observed in as many as 70% of the cells in those cancer cell lines that display inefficient centrosome clustering (Silkworth, Nardi, and Cimini, unpublished). These observations indicate that incomplete centrosome separation at NEB may play an important role in promoting CIN in certain cancer types.
Conclusions
Accurate partitioning of chromosomes to the daughter cells during mitosis is of utmost importance to ensure development and growth of all eukaryotic organisms. Defects of the mitotic spindle have a dramatic impact on chromosome segregation. However, the effect on the cell population can be even more dramatic if the spindle defects are only transient. Indeed, whereas permanent spindle defects typically lead to cell death in the progeny, transient spindle defects increase the rates of chromosome mis-segregation, but not to a level that would affect cell viability, thus ultimately being more threatening to the overall cell population and/or the organism. Here, we have discussed two types of transient mitotic spindle defects that are associated with increased rates of kinetochore mis-attachment formation and chromosome mis-segregation. One of them, transient multipolarity, is currently recognized as a common mechanism of CIN in cancer cells [32,39,84]. Conversely, the incomplete centrosome separation (at NEB) mechanism appears to occur at moderate levels in many different types of cancer cells, but it also occurs at very high frequencies in cells from certain types of cancers. This is a very interesting observation that needs further investigation. For example, it would be interesting to study whether the incidence of this mechanism relates to cancer progression or whether it is typical of cancer cells from specific sites. Moreover, the fact that incomplete centrosome separation at NEB is also observed in several non-transformed cell types suggests the possibility that this mechanism may occur in normal proliferating somatic cells and may, through its ability to promote chromosome mis-segregation, play a role in tumor initiation. This possibility undoubtedly deserves consideration in the near future. Future studies should also be focused on the causes of incomplete centrosome separation at NEB. It is plausible to imagine that mis-regulation of key motor proteins may be at the basis of this defect. An alternative hypothesis is that the ability of the centrosomes to separate in a timely fashion is dictated by signals from and/or physical interactions with transmembrane elements. For example, it is widely acknowledged that dividing cells within polarized epithelia rely on such mechanisms to orient the mitotic spindle [85][86][87]. Similar mechanisms may dictate not only the exact positioning of the centrosomes, but also the timing of centrosome separation. One last possibility is that centrosome separation per se is not impaired in cells with incomplete centrosome separation at NEB, but the exact timing is inaccurate, so that centrosome separation and NEB are no longer coordinated. It is possible that such coordination is finely regulated at the early stages of development [50,80,81,88], but is lost in the adult. In fact, optimal timing between centrosome separation, NEB, and chromosome segregation is very important during the early stages of development to ensure the chromosomal stability necessary to the development of a healthy adult organism. In conclusion, incomplete centrosome separation at NEB is a newly characterized mechanism of CIN for which we still have numerous open questions, and as such it will likely become the focus of many studies in the near future.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions WTS and DC conceived the ideas and wrote the paper. Both authors read and approved the final manuscript.
Authors' information WTS was a PhD student in the Department of Biological Sciences at Virginia Tech, and completed his dissertation work in July 2012 under the guidance of Dr. Daniela Cimini. DC is an Associate Professor in the Department of Biological Sciences at Virginia Tech. | 4,935 | 2012-08-11T00:00:00.000 | [
"Biology",
"Materials Science"
] |
Mammalian exocyst complex is required for the docking step of insulin vesicle exocytosis.
Glucose stimulates insulin secretion from pancreatic beta cells by inducing the recruitment and fusion of insulin vesicles to the plasma membrane. However, little is currently known about the mechanism of the initial docking or tethering of insulin vesicles prior to fusion. Here, we examined the role of the SEC6-SEC8 (exocyst) complex, implicated in trafficking of secretory vesicles to fusion sites in the plasma membrane in yeast and in regulating glucose-stimulated insulin secretion from pancreatic MIN6 beta cells. We show first that SEC6 is concentrated on insulin-positive vesicles, whereas SEC5 and SEC8 are largely confined to the cytoplasm and the plasma membrane, respectively. Overexpression of truncated, dominant-negative SEC8 or SEC10 mutants decreased the number of vesicles at the plasma membrane, whereas expression of truncated SEC6 or SEC8 inhibited overall insulin secretion. When single exocytotic events were imaged by total internal reflection fluorescence microscopy, the fluorescence of the insulin surrogate, neuropeptide Y-monomeric red fluorescent protein brightened, diffused, and then vanished with kinetics that were unaffected by overexpression of truncated SEC8 or SEC10. Together, these data suggest that the exocyst complex serves to selectively regulate the docking of insulin-containing vesicles at sites of release close to the plasma membrane.
Insulin is accumulated in large dense core vesicles within the pancreatic  cell and is released by exocytosis when glucose concentrations rise (1). Triggering of secretion appears to involve increases in intracellular ATP concentrations (or ATP/ ADP ratio), closure of ATP-sensitive K ϩ channels, and Ca 2ϩ influx through voltage-sensitive Ca 2ϩ channels (1). Because secretion in response to nutrient secretagogues is usually biphasic, with a rapid early phase followed by a subsequent second or sustained phase (2), it has been proposed that functionally distinct groups of vesicles may exist and represent "reserve," "readily releasable," and "intermediate" pools (3,4). At present, however, the mechanisms through which vesicles are recruited to the readily releasable pool, proposed to correspond to a subset of the "morphologically docked" (4) vesicles at the cell surface, are still incompletely understood (1).
Soluble N-ethylmaleimide-sensitive fusion protein attachment protein receptors (SNAREs) 1 are believed to be required for the trafficking of insulin and other secretory hormones from the endoplasmic reticulum to the Golgi apparatus, Golgi and trans-Golgi network to the plasma membrane, and finally exocytosis of mature vesicles (5,6). Dense core vesicle trafficking to the plasma membrane can itself be divided into three stages: transport of insulin vesicles to the plasma membrane, their first interaction with the plasma membrane (docking or tethering), and then subsequent fusion at the plasma membrane (exocytosis). Although SNARE proteins are believed to regulate the final fusion step in  cells (7), the machinery involved in the initial docking or tethering steps and possibly in regulating the sustained phase of insulin secretion (1,8) is undefined (9).
The budding yeast Saccharomyces cerevisiae has provided the genetic identification of proteins required for vesicle transport (10,11). Six of the sec gene products, Sec3p, Sec5p, Sec6p, Sec8p, Sec10p, and Sec15p, have been described in the vesicle transport complex, and mammalian homologues have been identified (12)(13)(14). Furthermore, two novel gene products, Exo70p and Exo84p, have been shown to be part of the complex (13,15,16), now known as the SEC6-SEC8 or "exocyst" complex.
The exocyst complex is concentrated at the sites of active vesicle exocytosis in the yeast bud and is essential for secretion (16,17). Similarly, the mammalian exocyst is localized at the plasma membrane of nerve terminals (12) and, in differentiated PC12 cells, displays a punctate distribution at the termini of cell processes at or near sites of exocytosis (13) or on the secretory vesicle membrane (18). Such findings suggest that the exocyst complex may serve as part of vesicledocking machinery at sites of regulated as well as constitu-tive exocytosis (12,19,20). At present, however, functional data supporting or refuting such a role in mammalian cells is unavailable.
In this study, we examined the role of the exocyst complex in regulated insulin secretion. We show that SEC6 is concentrated on insulin-containing dense core vesicles, whereas SEC5 and SEC8 are concentrated in the cytoplasm and on the plasma membrane, respectively. Although dominant-negative suppression of SEC8 and SEC10 function decreased the density of vesicles at the plasma membrane, overexpression of truncated SEC8 or SEC10 did not affect the kinetics of single exocytotic events. These data are compatible with a model in which the exocyst complex (a) forms as dense core vesicles approach the cell surface, (b) is then involved in the docking of vesicles at the plasma membrane by presently undefined mechanisms, but (c) is not involved in the final fusion step.
Generation of Truncated sec6 and sec8 Adenoviruses-cDNAs encoding C-terminal truncation mutants of rat sec6 (residues 1-333, Ad-sec6 CT ) or rat sec8 (residues 1-426, Ad-sec8 CT ) were amplified by PCR, subcloned into pCR2 (Invitrogen), and sequenced on both strands. These mutants corresponded to those identified by Novick and colleagues (17,24,25), which abrogate function (i.e. secretion) in yeast. We therefore postulated that the constructs would function as dominant negatives in mammalian cells. The cDNAs were subcloned into the pShuttle-CMV vector, linearized with PmeI, and cotransformed with pAdEasy into Escherichia coli strain BJ5183 by electroporation. Recombinants were selected and amplified in E. coli DH5␣. The chosen clones were linearized with PacI to expose the inverted terminal repeats and allow viral packaging when transfected in HEK293 cells. Large scale amplification and titer of viral stocks was performed as described previously (26).
To monitor exocytosis of NPY-mRFP constructs at the single vesicle level, we used a TIRF microscope similar to that described previously (22,28,29) with minor modifications. We used an objective lens with a numerical aperture of 1.45 (PlanAPO ϫ100/1.45 TIRFM, Olympus) to monitor single vesicle exocytosis. To observe the pIRES2-EGFP and NPY-mRFP fluorescence images, a 488-nm laser (argon ion laser, 30 milliwatts, Spectra-Physics) was used for total internal reflection illumination with a long pass filter (515 nm for EGFP, 600 nm for mRFP). Images were collected with a cooled charge-coupled device camera (640 ϫ 480 pixels, IMAGO, Till Photonics; operated with TillvisION software, Till Photonics). Images were acquired every 300 ms with 50 ms of exposure. To analyze the data, fusion events were manually selected, and the average fluorescence intensity of individual vesicles in a 1 ϫ 1-m square placed over the vesicle center was calculated.
The critical angle of our 1.45 numerical aperture objective lens is 63.56°, and the widest possible angle is 71.39°(when measured at the rim of the lens). We introduced the laser beam in the middle of these angles. We calculated the decay length of evanescent field (d) theoretically using the formula, where is the angle of incidence, the wavelength of light, n 1 and n 2 the refractive indices of water (1.33) and glass (1.53). For ϭ 67.5°and ϭ 488 nm, decay constant (d) becomes 81.1 nm, and effectively, a layer of 100 -200 nm is the reach of the evanescent light.
Confocal Imaging-To assess the density of the vesicles in exocyst complex mutant-expressing cells, we employed a confocal microscope (Leica TCS-AOBS laser-scanning confocal microscope). Before monitoring vesicle density, the bottom of the cell was located by inspection in confocal mode. The focal plane was then moved up by 1 m using a piezo motor, which drove the objective lens as controlled by the system software.
Generation and Purification of Polyclonal Anti-SEC5 Antibody-Antibodies were raised in sheep against residues 1-100 of the human SEC5 (GenBank TM accession number: NP_060773) sequence, encompassing the Ral-binding domain. The mouse SEC5 protein is identical in this region apart from residue 83, where arginine is replaced by lysine in the murine sequence. The relevant part of the coding region of SEC5 was inserted as a BamHI/EcoRI fragment into the vector pET28a. The resultant fusion protein, bearing an N-terminal hexahistidine tag, was expressed in E. coli and purified by chromatography on nickelnitrilotriacetic acid-agarose. Antibodies were purified from the antiserum by chromatography on a column of the fusion protein, immobilized on CNBr-activated Sepharose CL-4B.
Insulin Secretion Assay-MIN6 cells (5 ϫ 10 5 cells/well) were seeded in 6-well microtiter plates and infected with null (control), truncated sec6, and truncated sec8 adenoviruses. Culture was continued for 24 h in Dulbecco's modified Eagle's medium containing 25 mM glucose and then at 3 mM glucose for another 16 h. Cells were then washed in phosphate-buffered saline and incubated in KRB medium and 3 mM glucose, 30 mM glucose, or 50 mM KCl. Incubations were performed for 30 min at 37°C in a water bath. Total insulin was extracted in acidified ethanol (30). Total and secreted insulin were measured using radioimmunoassay by competition with 125 I-labeled rat insulin (Linco Research, St. Charles, MO) according to the manufacturer's instructions.
Data Analysis-Data are given as means Ϯ S.E. Comparisons between means were performed using one-way analysis of variance with GraphPad Prism TM (GraphPad Software, San Diego, CA) software.
The Exocyst Complex Is Associated with Insulin-containing
Vesicles-The exocyst complex has been shown to be present in high concentrations in developing brain and is believed to play a role in synaptogenesis (12,31). To examine the cellular localization of the complex in pancreatic  cells, formaldehyde-fixed pancreatic MIN6  cells were probed with antibodies against the three exocyst complex subunits (SEC5, SEC6, and SEC8).
Confocal images revealed that SEC5, SEC6, and SEC8 immunoreactivity displayed a markedly different subcellular localization (Fig. 1). SEC5 immunofluorescence was observed largely as diffuse, cytosolic fluorescence, although some punctate staining, partially colocalized with insulin, was apparent (Fig. 1, A-C). In contrast, SEC6 was found as punctate labeling, closely colocalized with insulin-positive vesicles (Fig. 1, D-F). SEC8 was mainly found on the plasma membrane of the cells (Fig. 1, G-I). These observations are similar to those made previously in PC12 cells (18) and suggest that the exocyst complex might regulate vesicle docking or fusion in pancreatic MIN6  cells.
Truncated Exocyst Component Decreases the Number of
Docked Vesicles on the Plasma Membrane-To determine the significance of exocyst complex function in insulin vesicle recruitment or docking, MIN6 cells were cotransfected with NPY-mRFP and either pIRES2-EGFP-sec8-⌬N or pIRES2-EGFP-sec10-⌬C. The pIRES2-EGFP plasmid is a bicistronic vector that allows the expression of both EGFP and an exocyst subunit mutant as two separate proteins. NPY-mRFP was used here as a surrogate for insulin (22,29) because the latter targets relatively poorly to the dense core vesicle after fusion with fluorescent proteins 2 (28). Cotransfected MIN6 cells were identified as those containing diffuse cytosolic EGFP fluorescence and punctate NPY-mRFP fluorescence. We then counted the number of intracellular and plasma membrane-docked NPY-mRFP vesicles by TIRF and confocal microscopy, respectively (Fig. 2, A-F). All of the cells examined that overexpressed truncated exocyst components showed a significant 40 -50% decrease in the number of docked vesicles at the plasma membrane (control, 1.36 Ϯ 2.2 vesicles/m 2 , n ϭ 12; sec8-⌬N, 0.86 Ϯ 2.8 vesicles/m 2 , n ϭ 7; sec10-⌬C, 0.65 Ϯ 1.1 vesicles/m 2 , n ϭ 12) (Fig. 2G, open squares). In contrast, cells overexpressing sec8-⌬N or sec10-⌬C showed a non-significant ϳ10% decrease in the number of intracellular vesicles (Fig. 2G, filled squares).
We next measured the impact of overexpression of the exocyst complex on glucose-or high K ϩ (depolarization)-induced insulin secretion. Insulin secretion was stimulated ϳ2.5-fold by 30 mM (versus 3 mM) glucose and 4-fold by high (50 mM) K ϩ in control cells (Fig. 2H). In Ad-sec6 CT -or Ad-sec8 CT -infected MIN6 cells, glucose-or high K ϩ -induced insulin secretion was decreased by ϳ40% (Fig. 2H). Taken together, these data suggest that the exocyst complex inhibits the docking step of insulin vesicle delivery to the plasma membrane, thus reducing the number of vesicles available for immediate release.
Effect of Overexpression of Exocyst Complex on Vesicle Fusion-To determine whether the apparent inhibition of docking observed above might be due to the exocyst complex decreasing the rate at which individual vesicles fused, thus causing a potential "bottleneck" that prevented subsequent fusion events, the dynamics of single vesicle exocytosis were analyzed in single NPY-mRFP-expressing vesicles. Although as expected, exocytotic events were detected less frequently in sec8-⌬N-and sec10-⌬C-expressing cells than in cells transfected with empty vector control (Fig. 2E), the kinetics of individual fusion events were identical in each case (Fig. 3D). Thus, stimulation with 30 mM glucose caused NPY-mRFP-expressing vesicles to brighten and spread suddenly during the release of the fluorescent peptide (22,29,32) with an identical time course in control (Fig. 3A), sec8-⌬N-overexpressing cells (Fig. 3B), or sec10-⌬C-overexpressing cells (Fig. 3C) was observed in the mean values for the rise time, half-widths, or decay time of fluorescence intensity between control and sec8-⌬N-or sec10-⌬C-overexpressing cells (data not shown). These data thus indicate that the exocyst complex does not alter the kinetics of individual insulin exocytosis events but rather regulates the docking step of insulin vesicles to the plasma membrane.
Effect of Overexpression of Exocyst Complex on Total Number of the Vesicles on the Plasma Membrane-We next measured the dynamic changes in the number of insulin vesicles docked at the plasma membrane during stimulation with glucose. During glucose stimulation of control cells (Fig. 4, A and D, open circles), the total number of docked vesicles slightly increased by up to 110% of the initial value, consistent with previous observations (33). This apparent recruitment of new vesicles to docking sites might thus contribute to the refilling of the docked vesicle pool (possibly equivalent to a readily releasable pool) during the second phase of glucose-induced insulin secretion. In contrast, in either sec8-⌬N-or sec10-⌬C-overexpressing cells, the total number of docked vesicles during glucose stimulation remained low, with no appreciable appearance of newly recruited vesicles to the plasma membrane (Fig. 4, B and C). As a result, in cells expressing sec8-⌬N or sec10-⌬C, the total number of docked vesicles decreased by 10 -40% of the initial number after 10 min of stimulation with high glucose concentrations (Fig. 4D, closed squares and triangles).
Truncated Exocyst Component Does Not Affect Microtubule Dynamics-To determine whether the effects of the dominant-negative sec constructs on vesicle distribution may be due to alterations in microtubule dynamics, the latter were assessed by monitoring the distribution of ␣-tubulin in nullor sec mutant-expressing cells. No differences were apparent in ␣-tubulin distribution in cells expressing any of the interfering sec constructs as assessed by confocal microscopy (Fig. 5). Thus, microtubule rearrangements would seem unlikely to explain the decreased vesicle recruitment to the cell surface. DISCUSSION We show here that the exocyst complex is present in significant concentrations in pancreatic MIN6  cells. Unexpectedly, we observed very different patterns of immunofluorescence staining for intracellular SEC5, SEC6, and SEC8. Thus, SEC5 immunofluorescence was predominantly seen diffusely throughout the cytoplasm in MIN6 cells. In contrast, SEC6 and SEC8 immunofluorescence were predominantly localized to insulin vesicles and the plasma membrane, respectively (Fig. 1). Thus, it would appear that the exocyst complex does not exist chiefly as a preformed unit but that instead, formation of the holo-complex may accompany translocation of vesicles to docking (pre-exocytosis) sites at the plasma membrane. Several studies in epithelial cells have shown that a portion of the exocyst complex exists in a partially assembled state with some of the protein being found dissociated from SEC6 (34)(35)(36). Also, studies by Finger et al. (37) and Guo et al. (38) showed that the localization of SEC8 to the plasma membrane at sites of polarized exocytosis is dependent on the accumulation of all other protein subunits of the complex in yeast. It therefore seems possible that SEC8 may be recruited to sites on the plasma membrane, probably in conjunction with a subset of other exocyst components prior to assembly of the whole exocyst complex. This may occur during docking of insulin-containing vesicles bearing SEC6 and possibly other exocyst subunits. Recently, several studies have shown that the exocyst complex interacts with active Ral GTPase (39)(40)(41). Wang et al. (42) showed that point mutations in both RalA and the sec5 subunit, which abolish RalA-SEC5 interaction, also abolish the inhibition of GTP-dependent exocytosis in PC12 cells. In contrast, the cleavage of a SNARE protein by Botulinum neurotoxin blocks both GTP-and Ca 2ϩ -dependent exocytosis. Therefore, the above authors (42) concluded that RalA and the exocyst complex are essential for GTP-dependent exocytosis, whereas the final fusion steps are SNARE-dependent. These data are fully compatible with the present findings.
We also analyzed here the docking and fusion of insulin vesicles in pancreatic MIN6 cells overexpressing truncated exocyst components. Our results showed that: 1) inhibition of the exocyst complex inhibited insulin secretion in response to cell depolarization or to glucose; 2) the kinetics of insulin secretion were identical in either control or truncated exocyst component-overexpressing cells; 3) the supply of vesicles and/or the ability to retain them on plasma membrane were impaired, resulting in a decrease in the number of docked insulin vesicles (to ϳ50% of that in normal MIN6 cells; Fig. 2D).
In conclusion, we present direct evidence that the exocyst complex contributes to the docking step of insulin vesicles on the plasma membrane in pancreatic  cells and may be involved in the recruitment of vesicles from a reserve to a readily releasable pool at the plasma membrane. Changes in the efficiency of the docking process may therefore play a role in the normal control of insulin release under conditions where secretion is strongly stimulated, such as during chronic hyperglycemia or during treatment with sulfonylureas. Because such alterations might conceivably contribute to insulin secretory deficiency in some form of type 2 diabetes, exocyst complex components may provide potential therapeutic targets in this disease. | 4,075.4 | 2005-07-08T00:00:00.000 | [
"Biology"
] |
Study of $B_{(s)} \to (\pi\pi)(K\pi)$ decays in the perturbative QCD approach
In this work, we provide estimates of the branching ratios, direct $CP$ asymmetries and triple product asymmetries in $B_{(s)} \to (\pi\pi)(K\pi)$ decays in the perturbative QCD approach, where the $\pi\pi$ and $K\pi$ invariant mass spectra are dominated by the vector resonances $\rho(770)$ and $K^*(892)$, respectively. Some scalar backgrounds, such as $f_0(500,980) \to \pi\pi$ and $K^*_0(1430) \to K\pi$ are also accounted for. The $\rho(700)$ is parametrized by the Gounaris-Sakurai function. The relativistic Breit-Wigner formula for the $f_0(500)$ and Flatt\'e model for the $f_0(980)$ are adopted to parameterize the time-like scalar form factors $F_S(\omega^2)$. We also use the D.V. Bugg model to parameterize the $f_0(500)$ and compare the relevant theoretical predictions from different models. While in the region of $K\pi$ invariant mass, the $K^*_0(1430)$ is described with the LASS lineshape and the $K^*(892)$ is modeled by the Breit-Wigner function. We find that the decay rates for the considered decay modes agree with currently available data within errors. As a by-product, we extract the branching ratios of two-body decays $B_{(s)} \to \rho(770)K^*(892)$ from the corresponding four-body decay modes and calculate the relevant polarization fractions. Our prediction of longitudinal polarization fraction for $B^0\to \rho(770)^0 K^*(892)^0$ decay deviates a lot from the recent LHCb measurement, which should be resolved. It is shown that the direct $CP$ asymmetries are large due to the sizable interference between the tree and penguin contributions, but they are small for the tree-dominant or penguin-dominant processes. The PQCD predictions for the triple product asymmetries are small which are expected in the standard model, and consistent with the current data reported by the LHCb Collaboration.Our results can be tested by the future precise data from the LHCb and Belle II experiments.
In this work, we provide estimates of the branching ratios, direct CP asymmetries and triple product asymmetries in B (s) → (ππ)(Kπ) decays in the perturbative QCD approach, where the ππ and Kπ invariant mass spectra are dominated by the vector resonances ρ(770) and K * (892), respectively. Some scalar backgrounds, such as f0(500, 980) → ππ and K * 0 (1430) → Kπ are also accounted for. The ρ(700) is parametrized by the Gounaris-Sakurai function. The relativistic Breit-Wigner formula for the f0(500) and Flatté model for the f0(980) are adopted to parameterize the time-like scalar form factors FS(ω 2 ). We also use the D.V. Bugg model to parameterize the f0(500) and compare the relevant theoretical predictions from different models. While in the region of Kπ invariant mass, the K * 0 (1430) is described with the LASS lineshape and the K * (892) is modeled by the Breit-Wigner function. We find that the decay rates for the considered decay modes agree with currently available data within errors. As a by-product, we extract the branching ratios of two-body decays B (s) → ρ(770)K * (892) from the corresponding four-body decay modes and calculate the relevant polarization fractions. Our prediction of longitudinal polarization fraction for B 0 → ρ(770) 0 K * (892) 0 decay deviates a lot from the recent LHCb measurement, which should be resolved. It is shown that the direct CP asymmetries are large due to the sizable interference between the tree and penguin contributions, but they are small for the tree-dominant or penguin-dominant processes. The PQCD predictions for the "true" triple product asymmetries are small which are expected in the standard model, and consistent with the current data reported by the LHCb Collaboration. Our results can be tested by the future precise data from the LHCb and Belle II experiments.
Besides the direct CP asymmetries, there is another signal of CP violation in the angular distribution of B (s) → V 1 V 2 decays, which is called triple-product asymmetries (TPAs) [47][48][49][50][51][52][53][54][55]. These triple products are odd under the time reversal transformation (T ), and also contribute potential signals of CP violation due to the CP T theorem. TPAs have already been measured by BABAR, Belle, CDF and LHCb [40,43,[56][57][58][59][60]. It is known that a non-vanishing direct CP violation needs the interference of at least two amplitudes with a weak phase difference ∆φ and a strong phase difference ∆δ. The direct CP violation is proportional to sin ∆φ sin ∆δ, while TPAs go as sin ∆φ cos ∆δ. If the strong phases are quite small, the magnitude of the direct CP violation is close to zero, but the TPA is maximal. Hence direct CP violation and TPAs complement each other. Even if the effect of CP violation is absent, T -odd triple products (also called "fake" TPAs), which are proportional to cos ∆φ sin ∆δ, can provide useful complementary information on new physics [51]. TPAs are excellent probes of physics beyond the SM since most TPAs are expected to be tiny within the SM and are not suppressed by the small strong phases.
As is well known, the kinematics of two-body decays is fixed, while the multi-body decay amplitudes depend on at least two kinematic variables. Meanwhile, the multi-body decays not only receive the resonant and nonresonant contributions, but also involve the possible significant final-state interactions (FSIs) [61][62][63]. In this respect, multi-body decays are considerably more challenging than two-body decays, but provide a number of theoretical and phenomenological advantages. In two-body B decays, the measured CP violation is just a number while the CP asymmetry depends on the invariant mass of the twomeson pair and varies from region to region in the Dalitz plot [64,65] in the three-body modes [66]. In addition, strong phases in multi-body decays arise nonperturbatively already at the leading power, through complex phases in matrix elements such as F π ∼< 0|j|ππ > and so on. Since the B (s) → V 1 V 2 , V 1 S 2 , S 1 V 2 , S 1 S 2 decays are expected to proceed through V 1 (S 1 ) → P 1 P 2 and V 2 (S 2 ) → Q 1 Q 2 with P 1,2 (Q 1,2 ) denoting pseudoscalar meson π or K, it is meaningful to study such decays in the four-body framework, which provide useful information for understanding the CP -violation mechanisms.
As addressed above, multi-body decays of heavy mesons involve more complicated dynamics than two-body decays. A factorization formalism that describes a multi-body decay in full phase space is not yet available at present. It has been proposed that the factorization theorem of three-body B decays is approximately valid when two particles move collinearly and the bachelor particle recoils back [67,68]. More details can also be found in Refs. [69,70]. This situation exists particularly in the low ππ or Kπ invariant mass region ( 2 GeV) of the Dalitz plot where most resonant structures are seen. The Dalitz plot is typically dominated by resonant quasi-two-body contributions along the edge. This proposal provides a theoretical framework for studies of resonant contributions based on the quasi-two-body-decay mechanism. Several theoretical approaches have been developed for describing the three-body hadronic decays of B mesons based on the symmetry principles [71][72][73][74][75][76], the QCDF [77][78][79][80][81][82][83][84][85][86][87] and the PQCD approaches [88][89][90][91][92][93][94][95][96][97][98][99][100][101][102]. Recently, the localized CP violation and branching fraction of the four-body decayB 0 → K − π + π + π − have been calculated by employing a quasi-two-body QCDF approach in Refs. [103,104]. Similar to three-body B meson decays, four-body B (s) → R 1 R 2 → (P 1 P 2 )(Q 1 Q 2 ) decay modes (R 1,2 represents the vector or scalar intermediate resonance) are assumed to proceed dominantly with two intermediate resonances R 1 or R 2 each decaying to a pseudoscalar pair. As a first step, we can only restrict ourselves to the specific kinematical configurations in which each two particles move collinearly and two pairs of final state particles recoil back in the rest frame of the B meson, see Fig. 1. Naturally the dynamics associated with the pair of final state mesons can be factorized into a two-meson distribution amplitude (DA) Φ h1h2 [105][106][107][108][109][110][111]. Thereby, the typical PQCD factorization formula for the considered four-body decay amplitude can be described as the form of, where Φ B is the universal wave function of the B meson and absorbs the non-perturbative dynamics in the process. The Φ P1P2,Q1Q2 is the two-hadron DA, which involves the resonant and nonresonant interactions between the two moving collinearly mesons. The hard kernel H describes the dynamics of the strong and electroweak interactions in four-body hadronic decays in a similar way as the one for the corresponding two-body decays.
In this work, we study the four-body decays B (s) → (ππ)(Kπ) in the PQCD approach based on k T factorization with the relevant Feynman diagrams illustrated in Fig. 2. In the considered (ππ) invariant-mass range, the vector resonance ρ(770) is expected to contribute, together with the scalar resonances f 0 (500) and f 0 (980). The Kπ spectrum is dominated by the vector K * (892) resonance and the scalar resonance K * 0 (1430). Throughout the remainder of the paper, the symbol ρ is used to denote the ρ(770) resonance and K * is for K * (892) resonance. For a comparison with the LHCb experiment [43], the invariant mass of the ππ pair is restricted from 300 MeV to 1100 MeV and the range of invariant mass for Kπ pair varies from 750 MeV to 1200 MeV. We calculate the branching ratios and polarization fractions of each partial waves. Besides, a set of CP -violating observables is investigated using B (s) meson decays reconstructed in the (ππ)(Kπ) quasi-two-body final state. Particular emphasis is placed on the TPAs of the considered decays. It should be mentioned that there is a possibility of existing two identical final state pions. Taking the B + → (f 0 (980) →)π + π − (K * + →)K 0 π + decay for an example, in the experimental side, they see four final states π + 1 π − K 0 π + 2 first, and have to pair one of the positive charged pions with the π − . They have to try two possible combinations. From the theoretical point of view, we deal with B + → f 0 (980)K * + → π + 1 π − K 0 π + 2 in the quasi-two-body mechanism and see f 0 (980)K * + first. Then each quasi-two-body intermediate resonance decays into two pseudoscalars (f 0 (980) → π + 1 π − and K * 0 → K 0 π + 2 ), respectively. Each meson pair generates a smaller invariant mass and flies back-to-back as shown in Fig. 1. It is evident that the final states π + have been already specified unambiguously on the theoretical side. Certainly, it is experimentally difficult to identify which resonance a π + comes from in the nonresonant region on a Dalitz plot. From the theoretical viewpoint, this difficulty implies that the tail of our Breit-Wigner propagator may not describe the events in the nonresonant region correctly. However, it is known that the nonresonant region gives a minor FIG. 1: Graphical definitions of the helicity angles θ1, θ2 and φ. Taking the B 0 → R1R2 decay as an example, with each quasi-two-body intermediate resonance decaying to two pseudoscalars (R1 → π + π − and R2 → K + π − ), θ1(θ2) is denoted as the angle between the directions of motion of π + (π − ) in the R1(R2) rest frame and R1(R2) in the B 0 rest frame, and φ is the angle between the plane defined by π + π − and the plane defined by K + π − in the B 0 rest frame. [45,46]. [45,46].
contribution, so the misparing of a π + with other final states should not make a significant impact to our results. The layout of the present paper is as follows. In Sec. II, we give an introduction for the theoretical framework. The numerical values and some discussions will be given in Sec. III. Section IV contains our conclusions. The Appendix collects the explicit PQCD factorization formulas for all the decay amplitudes.
A. Kinematics
Consider a four-body B(p B ) → R 1 (p)R 2 (q) → P 1 (p 1 )P 2 (p 2 )Q 1 (q 1 )Q 2 (q 2 ) decay, we will work in the B meson rest frame and employ the light-cone coordinates for momentum variables. The B meson momentum p B , the total momenta of the [124] pion-pion and kaon-pion pairs, p = p 1 + p 2 , q = q 1 + q 2 , and the quark momentum k i in each meson are chosen as where m B is the mass of B meson. The momentum fractions x B , x 1 and x 2 run from zero to unity. The factors g ± and f ± can be obtained through the momentum conservation p B = p + q, and p 2 = ω 2 1 and q 2 = ω 2 2 , respectively. The explicit expressions of g ± and f ± are written in the following form with the factor η 1,2 = . If the meson pairs are in the P -wave configuration, the corresponding longitudinal polarization vectors are defined as which satisfy the normalization ǫ 2 p = ǫ 2 q = −1 and the orthogonality ǫ p · p = ǫ q · q = 0.
As usual we also define the momentum p 1,2 of pion-pion pair and q 1,2 for the kaon-pion pair as with ζ 1 + r1−r2 2η1 = p + 1 /p + and ζ 2 + r3−r4 2η2 = q − 1 /q − characterizing the momentum fraction for one of pion-pion (kaon-pion) pair, and the mass ratios r 1,2,3,4 = m 2 1,2,3,4 /m 2 B , m 1,2,3,4 being the masses of the final state meson. One can obtain the relation between ζ 1,2 and the polar angle θ 1,2 in the dimeson rest frame in Fig. 1, with the upper and lower limits of ζ 1,2 For the sake of simplicity, we generally use the factor in the following sections.
The differential branching fraction for the B (s) → (ππ)(Kπ) in the B (s) meson rest frame is expressed as where dΩ with Ω ≡ {θ 1 , θ 2 , φ, ω 1 , ω 2 } stands for the five-dimensional measure spanned by the three helicity angles and the two invariant masses, and is the momentum of the ππ pair in the B (s) meson rest frame. The explicit expression of kinematic variables k(ω) is defined in the h 1 h 2 center-of-mass frame with the Källén function λ(a, b, c) = a 2 + b 2 + c 2 − 2(ab + ac + bc) and m h1,h2 being the final state mass. The four-body phase space has been derived in the analyses of the K → ππlν decay [112], the semileptonicB → D(D * )πlν decays [113], semileptonic baryonic decays [114,115], and four-body baryonic decays [116]. One can confirm that Eq. (8) is equivalent to those in Refs. [114,116] by appropriate variable changes. Replacing the helicity angle θ by the meson momentum fraction ζ via Eq. (6), the Eq. (8) is turned into
B. Helicity amplitudes
One can disentangle the helicities of R 1 (→ ππ)R 2 (→ Kπ) final state via an angular analysis, depicted in Fig. 1. Taking the B 0 → R 1 R 2 → (π + π − )(K + π − ) decay as an example, the θ 1 is the angle between the π + direction in the (π + π − ) rest frame and the (π + π − ) direction in the B 0 rest frame, θ 2 is the angle between the K + direction in the (K + π − ) rest frame and the (K + π − ) direction in the B 0 rest frame, and φ is the angle between the (π + π − ) and (K + π − ) decay planes. A ππ(Kπ) pair can be produced in the S or P -wave configuration in the selected invariant mass regions.
One decomposes the decay amplitudes into six helicity components: h = V V (3), V S, SV , and SS, each with a corresponding amplitude A h . The first three, commonly referred to as the P -wave amplitudes, are associated with the final states, where both ππ and Kπ pairs come from intermediate vector mesons. In the transversity basis, a P -wave decay amplitude can be decomposed into three components: A 0 , for which the polarizations of the final-state vector mesons are longitudinal to their momenta, and A (A ⊥ ), for which the polarizations are transverse to the momenta and parallel (perpendicular) to each other. As the S-wave ππ(Kπ) pair arises from R 1 (R 2 ) labelled in Fig. 2(a), the corresponding single S-wave amplitude is denoted A SV (A V S ). The double S-wave amplitude A SS is associated with the final state, where both two-meson pairs are generated in the S wave. Specifically, these helicity amplitudes for the B (s) → (ππ)(Kπ) decay denote Relying on the Eq. (6), we get the full decay amplitude in Eq. (11) as a coherent sum of the P -, S-, and double S-wave components by including the ζ 1,2 dependencies instead of θ 1,2 and azimuth-angle dependencies, On basis of Eq. (11), we can obtain the branching ratio for each helicity state, where the invariant masses ω 1,2 are integrated over the chosen ππ and Kπ mass window, respectively. The coefficients C h are the results of the integrations over ζ 1 , ζ 2 , φ in terms of Eq. (14) and listed as follows, The CP -averaged branching ratio and the direct CP asymmetry in each component are defined as below, respectively, whereB h is the branching ratio of the corresponding CP -conjugate channel. For the V V decays, the polarization fractions f λ with λ = 0, , and ⊥ are described as with the normalisation relation f 0 + f + f ⊥ = 1.
C. Triple product asymmetries
In this work, we not only calculate the direct CP asymmetries, but also pay more attention to the TPAs. Consider a four-body decay B → R 1 (→ P 1 P 2 )R 2 (→ Q 1 Q 2 ), in which one measures the four particles' momenta in the B rest frame. We definê n Ri (i = 1, 2) is a unit vector perpendicular to the R i decay plane andẑ is a unit vector in the direction of R 1 in the B rest frame. Thus we haven R1 ·n R2 = cos φ,n R1 ×n R2 = sin φẑ, implying a T -odd scalar triple product One can define a TPA as an asymmetry between the number of decays involving positive and negative values of sin φ or sin 2φ, According to Eq. (6), the TPAs associated with A ⊥ for the considered four-body decays are derived from the partially integrated differential decay rates as [40,50] with the denominator It has been found that TPAs originate from the interference of the CP -odd amplitudes A ⊥ with the other CP -even amplitudes A 0 , A . The above TPAs contain the integrands Im[A ⊥ A * i ] = |A ⊥ ||A * i | sin(∆φ + ∆δ) with i = 0, , where ∆φ and ∆δ denote the weak and strong phase differences between the amplitudes A ⊥ and A i , respectively. As already noted, Im[A ⊥ A * i ] can be nonzero even if the weak phases vanish. Thus, it is not quite accurate to identify a nonzero TPA as a signal of CP violation. To obtain a true CP violation signal, one has to compare the TPAs in the B andB meson decays. The helicity amplitude for the CP -conjugated process can be inferred from Eq. (13) through A 0 →Ā 0 , A →Ā and A ⊥ → −Ā ⊥ , in which theĀ λ are obtained from the A λ by changing the sign of the weak phases. Thus, the TPAsĀ i T for the charge-conjugate process are defined similarly, but with a multiplicative minus sign. We denote byĀ 0 ,Ā ,Ā ⊥ transversity amplitudes for the CP -conjugate decaȳ B →R 1 (→P 1P2 )R 2 (→Q 1Q2 ) and the corresponding three angles will be denoted byθ 1 ,θ 2 , andφ. Obviously θ 1 =θ 1 , θ 2 =θ 2 and φ =φ. We then have the TPAs for the CP -averaged decay rates [50] A withΓ being the decay rate of the CP -conjugate process, for the CP -conjugate decay. It is shown that the term Im[A ⊥ A * 0, −Ā ⊥Ā * 0, ] is proportional to sin ∆φ cos ∆δ, which is nonzero only in the presence of the weak phase difference. Then TPAs provide an alternative measure of CP violation. Furthermore, compared with direct CP asymmetries, A (i)ave T (true) does not suffer the suppression from the strong phase difference, and is maximal when the strong phase difference vanishes. While for the term Im[A ⊥ A * 0, +Ā ⊥Ā * 0, ] ∝ cos ∆φ sin ∆δ, the A (i)ave T (fake) can be nonzero when the weak phase difference vanishes. Such a quantity is referred as a fake asymmetry (CP conserving), which reflects the effect of strong phases [50,51], instead of CP violation.
For a more direct comparison with the measurements from LHCb [43], the so-called true and f ake TPAs are then defined as where the true or f ake labels refer to whether the asymmetry is due to a real CP asymmetry or due to effects from final-state interactions that are CP symmetric. It should be stressed that two asymmetries defined in Eqs. (25) and (30) are different in the most case, as well as two asymmetries in Eqs. (26) and (31). They become equal when no direct CP violation occurs in the total rate, namely D =D.
D. Two-meson distribution amplitudes
S-wave two-meson DAs
Here we briefly introduce the S and P -wave two-meson DAs and the corresponding time-like form factors used in our framework. One can see that resonant contributions through two-body channels can be included by parameterizing the two-meson DAs. The S-wave two-meson DA is written in the following form [117], In what follows the subscripts S and P are always associated with the corresponding partial waves.
For the scalar resonances f 0 (500) and f 0 (980), the asymptotic forms of the individual DAs in Eq. (34) have been parameterized as [105][106][107][108] with the time-like scalar form factor F S (ω 2 ) and the Gegenbauer coefficient a S . While for the scalar resonance K * 0 (1430), we will adopt similar formulas as those for a scalar meson [118,119], bearing in mind large uncertainties that may be introduced by this approximation. The detailed expressions of DAs for various twists are as follows: where the Gegenbauer polynomials C 3/2 . ω is the Kπ invariant mass and m 01,02 are the running current quark masses. The Gegenbauer moments B 1,3 at the 1 GeV scale from Scenario I in the QCD sum rule analysis and the related running current quark masses can be found in Refs. [2,119,120].
P -wave two-meson DAs
The P -wave two-meson DAs related to both longitudinal and transverse polarizations are organized in analogy with those in Ref. [92]. The explicit expressions of the P -wave pion-pion (kaon-pion) DAs associated with longitudinal and transverse polarization are described as follows, The two-pion DAs for various twists are expanded in terms of the Gegenbauer polynomials: with the Gegenbauer coefficients a i 2ρ and the two P -wave form factors F P (ω 2 ) and F ⊥ P (ω 2 ). For the kaon-pion DAs, the various twist DAs φ i P have similar forms as the corresponding ones for the K * meson [121] by replacing the decay constants with the time-like form factors, with t = (1 − 2z). The Gegenbauer moments associated with longitudinal polarization a 1K * , a 2K * are determined in Ref. [102], while the Gegenbauer moments associated with transverse polarization a ⊥ 1K * , a ⊥ 2K * are adopted the same values as those longitudinal ones.
Time-like form factor
The strong interactions between the resonance and the final-state meson pair, including elastic rescattering of the final-state meson pair, can be factorized into the time-like form factor F S,P (ω 2 ), which is guaranteed by the Watson theorem [122]. We usually use the relativistic Breit-Wigner (BW) line shape for a narrow resonance to parameterize the time-like form factors F (ω 2 ) [123]. The explicit formula is expressed as [124], where the m i and Γ i are the pole mass and width of the corresponding resonances shown in Table III, respectively. The massdependent width Γ i (ω) is defined as The k(ω) is the momentum vector of the resonance decay product measured in the resonance rest frame, while k(m i ) is the value of k(ω) when ω = m i . L R is the orbital angular momentum in the ππ (Kπ) system and L R = 0, 1, ... corresponds to the S, P, ... partial-wave resonances. Due to the limited studies on the form factor F ⊥ (ω 2 ), we use the two decay constants f The BW formula does not work well for f 0 (980), since its pole mass is close to the KK threshold. For scalar resonance f 0 (980), we adopt the Flatté parametrization where the resulting line shape is above and below the threshold of the intermediate particle [128]. If the coupling of a resonance to the channel opening nearby is very strong, the Flatté parametrization shows a scaling invariance and does not allow for an extraction of individual partial decay widths. Thus, we employ the modified Flatté model suggested by Bugg [129] following the LHCb collaboration [125,130], .
The coupling constants g ππ = 0.167 GeV and g KK = 3.47g ππ [125,130] describe the f 0 (980) decay into the final states π + π − and K + K − , respectively. The phase space factors ρ ππ and ρ KK read as [125,128,131] If there are overlapping resonances or there is significant interference with a nonresonant component both in the same partial wave, the relativistic BW function leads to unitarity violation within the isobar model [132]. This is the case for the Kπ S-wave at low Kπ mass, where the K * 0 (1430) resonance interferes strongly with a slowly varying NR S-wave component. In this work, the time-like scalar form factor F S (ω 2 ) for the S-wave Kπ system is parametrized by using a modified LASS line shape [133] for the S-wave resonance K * 0 (1430), which has been widely used in experimental analyses [124], with where the first term in Eq. (59) is an empirical term from the elastic kaon-pion scattering and the second term is the resonant contribution with a phase factor to retain unitarity. Here m 0 and Γ 0 are the pole mass and width of the K * 0 (1430) state. The parameters a = (3.1 ± 1.0) GeV −1 and r = (7.0 ± 2.4) GeV −1 are the scattering length and effective range [124], respectively. The slowly varying part (the first term in the Eq. (59)) is not well modeled at high masses and it is set to zero for m(Kπ) values above 1.7 GeV [124].
In experimental investigations of three-body hadronic B meson decays, the wide ρ resonant contribution is usually parameterized as the Gounaris-Sakurai (GS) model [134] based on the BW function [123]. It is a way to interpret the observed structures beyond the ρ resonance in terms of higher mass isovector vector mesons. By taking the ρ − ω interference and the excited states into account, the form factor F (ω 2 ) can be written in the form of [127] F where s = m 2 (ππ) is the two-pion invariant mass squared, i = (ρ ′ (1450), ρ ′′ (1700), ρ ′′′ (2254)), Γ ρ,ω,i is the decay width for the relevant resonance, m ρ,ω,i are the masses of the corresponding mesons, respectively. The explicit expressions of the function GS ρ (s, m ρ , Γ ρ ) are described as follows [123] GS ρ (s, m ρ , with the functions where β π (s) = 1 − 4m 2 π /s. Since the process ω → π + π − suffers the G-parity suppression, we find that the interference between the ρ and ω does not significantly change the values in our calculations. Hence, it's reasonable to set c ω = 0 in our latter numerical calculations.
In contrast to the vector resonances, the identification of the scalar mesons is a long-standing puzzle. Scalar resonances are difficult to resolve because some of them have large decay widths, which cause a strong overlap between resonances and background. For a comparison, we here parameterize the f 0 (500) contribution in two different ways: the BW and the Bugg model [135], respectively. The form factor of f 0 (500) with the Bugg resonant lineshape is written in the following form [135] where s = ω 2 = m 2 (π + π − ), j 1 (s) = 1 π 2 + ρ 1 ln 1−ρ1 1+ρ1 , the functions g 2 1 (s), Γ i (s) and other relevant functions in Eq. (64 ) are the following For the parameters in Eqs. (64,65), we use their values as given in the fourth column of Table I in Ref. [135]: And the parameters ρ 1,2,3 in Eq. (65) are the phase-space factors of the decay channels ππ, KK and ηη respectively, and have been defined as [135] with m 1 = m π , m 2 = m K and m 3 = m η .
III. NUMERICAL RESULTS
In this section, we calculate the CP -averaged branching rations (B), direct CP -violating asymmetries (A CP ) and the polarization fractions f λ , as well as estimate the size of TPAs, respectively. The pole mass and decay width for the corresponding resonance have been summarized in Table III The values of the Wolfenstein parameters are adopted as given in Ref. [136]: A = 0.836 ± 0.015, λ = 0.22453 ± 0.00044, ρ = 0.122 +0.018 −0.017 ,η = 0.355 +0.012 −0.011 . The Gegenbauer moments are adopted the same values as those determined in Ref. [2,92,102,137] a S = 0.2 ± 0.2, B 1 = −0.57 ± 0.13, B 3 = −0.42 ± 0.22, a 0 2ρ = 0.08 ± 0.13, a s 2ρ = −0.23 ± 0.24, a t 2ρ = −0.354 ± 0.062, In our numerical calculations, the theoretical uncertainties quoted in the tables are estimated from three sources: the first theoretical uncertainty results from the parameters of the wave functions of the initial states, such as the shape parameter ω B = 0.40±0.04 GeV or ω Bs = 0.48±0.048 GeV in B (s) meson wave function. The second one is due to the Gegenbauer moments in various twist DAs of ππ and Kπ pair with different intermediate resonances. The last one is caused by the variation of the hard scale t from 0.75t to 1.25t (without changing 1/b i ) and the QCD scale Λ QCD = 0.25 ± 0.05 GeV, which characterizes the effect of the next-to-leading-order QCD contributions. The possible errors due to the uncertainties of CKM matrix elements are very small and can be neglected safely. The major uncertainty comes from the Gegenbauer moments, which amounts to 30% ∼ 50% in magnitude. These parameters need to be constrained more precisely in order to improve the accuracy of theoretical predictions for four-body B meson decays.
We perform an amplitude analysis of B (s) → (ππ)(Kπ) decays in the two-body invariant mass regions 300< m(ππ) <1100 MeV, accounting for the ρ, f 0 (500) and f 0 (980) resonances, and 750< m(Kπ) <1200 MeV, which is dominated by the K * (892) meson, but the S-wave resonance K * 0 (1430) is expected to contribute. From the numerical results, one can address some issues as follows. IV: CP -averaged branching ratios and polarization fractions of the two-body B (s) → ρK * decays in the PQCD approach compared with the previous predictions in the PQCD approach [18], the updated predictions in the QCDF [3,4], SCET [25] and FAT [26]. Experimental results for branching ratios are taken from Table I and for polarization fractions from [45]. The theoretical uncertainties are attributed to the variations of the shape parameter ωB (s) in the B (s) meson DA, of the Gegenbauer moments in various twist DAs of ππ and Kπ pair, and of the hard scale t and the QCD scale ΛQCD. The isospin conservation is assumed for the strong decays of an I = 1/2 resonance K * to Kπ, namely, where we assume the K * → Kπ branching fraction to be 100%. According to the relation of the decay rates between the quasi-two-body and the corresponding two-body decay modes with B(ρ → ππ) = 100%, we extract the two-body B (s) → ρK * branching ratios and summarize them in Table IV. The polarization fractions of the two-body B (s) → ρK * decays calculated in this work are also listed in Table IV. For a comparison, we show the updated predictions in the QCDF [3,4], the previous predictions in the PQCD approach [18], SCET [25] and FAT [26]. Experimental results for branching ratios are taken from Table I and for polarization fractions from [45]. One can see that, except for the colour-suppressed (" C ") decay B 0 s → ρ 0K * (892) 0 , our predictions of the branching ratios are in good agreement with those two-body analyses as presented in the PQCD approach [18], and also similar to those predicted in the QCDF approach [3,4], SCET [25] and FAT [26] within errors. However, the situation of the " C "-type decay B 0 s → ρ 0K * (892) 0 is more complicated: (a) as claimed in Ref. [88] that the widths of the resonant states and the interactions between the final state meson pairs will show their effects on the branching ratios, the new four-body prediction deviates from the previous calculations in the PQCD approach, but agrees well with the corresponding results in the QCDF approach, SCET and FAT within errors; (b) the transverse polarization contribution in the PQCD approach is comparable to the longitudinal one due to the chirally enhanced annihilation and the hard scattering diagrams, which is quite different from those predictions in the QCDF approach, SCET and FAT. More precise data from the future LHCb and Belle II experiments will help us differentiate these factorization approaches and understand the underlying mechanism of the multi-body B meson hadronic weak decays.
For the charmless B (s) → ρ(→ ππ)K * (→ Kπ) decays, it is naively expected that the helicity amplitudes satisfy the hierarchy pattern |A 0 | ≫ |A + | ≫ |A − |, which are related to the spin amplitudes (A 0 , A , A ⊥ ) in Appendix by while A 0 is common to both bases. The above hierarchy relation satisfies the expectation in the factorization assumption that the longitudinal polarization should dominate based on the quark helicity analysis [138,139]. However, large transverse polarization of order 50% is observed in B → K * φ, B → K * ρ and B s → φφ decays, which poses an interesting challenge for the theory. The interest in the polarization in penguin transition, such as b → s decays B → ρK * , is motivated by their potential sensitivity to physics beyond the SM. Measurements of the longitudinal polarization fraction in B → ρK * by BABAR [27,31,32] and Belle [33] reveal a large fraction of transverse polarization, indicating an anomaly of polarization. An angular analysis of the B 0 → ρ 0 K * 0 decay by LHCb measurement found an unexpectedly low longitudinal polarization fraction f 0 = 0.164 ± 0.015 ± 0.022 where the first uncertainty is statistical and the second systematic [43]. As shown in Table IV, the longitudinal polarization fraction f 0 for the B (s) → ρK * decays from the PQCD approach (including the present work) are around 50% to 80%, which are mostly greater than the transverse one f T = f + f ⊥ in contrast to observations. The QCDF [3,4], SCET [25] and FAT [26] yield the similar pattern f 0 ∼ f T in despite of large uncertainties. In the PQCD approach, the large transverse polarization fraction can be interpreted on the basis of the chirally enhanced annihilation diagrams, especially the (S − P )(S + P ) penguin annihilation, introduced by the QCD penguin operator O 6 [140], which is originally introduced in Ref. [141]. A special feature of the (S − P )(S + P ) penguin annihilation operator is that the light quarks in the final states are not produced through chiral currents. So, there is no suppression to the transverse polarization caused by the helicity flip. Then the polarization fractions satisfy f 0 ≈ f T . However, these effects are not able to fully account for the above polarization anomaly. Our predictions for the longitudinal polarization fractions agree with the previous PQCD calculations [18]. It is worth mentioning that we have employed the same Gegenbauer moments for the transversely polarized Kπ DAs as those for the longitudinal polarized ones (see Eq. (69)) in this work, together with the same Gegenbauer moments for ππ associated with the transverse polarizations from previous work [92]. The Gegenbauer moments from the twist-3 DAs of Kπ pair may make significant sense to the polarization fractions, which has been verified in Ref. [7]. To be honest, these Gegenbuaer moments should be fitted similarly as those in Ref. [102]. With more and more experimental measurements, we can determine the precise values of these Gegenbauer moments for transversely polarized DAs.
B. Branching ratios of B (s) → [SS, SV, V S] → (ππ)(Kπ) decays
In contrast to the vector resonances, the identification of the scalar mesons is a long-standing puzzle. Scalar resonances are difficult to resolve because some of them have large decay widths, which cause a strong overlap between resonances and background. In fact, compared with the B (s) → V V → (ππ)(Kπ) decays, there are much less experimental data for the B (s) → [SS, SV, V S] → (ππ)(Kπ) decays. Furthermore, the underlying structure of scalar mesons is not theoretically well established (for a review, see Ref. [46]). We hope that the situation can be improved using nonperturbative QCD tools including lattice QCD simulations. The f 0 (980) is strongly produced in D + s decay [142], which implies a largess component, assuming Cabibbo-favored c → s decay. Meanwhile, the prominent appearance of the f 0 (980) points to a dominant (ss) component in The experimental data are taken from [46]. The sources of the theoretical errors are the same as in Table IV. the semileptonic D s decays and decays of B (s) mesons. However, there also exists some experimental evidences indicating that f 0 (980) is not purely anss state. Ratios of decay rates of B and/or B s mesons into J/ψ plus f 0 (980) or f 0 (500) were proposed to allow for an extraction of the flavor mixing angle and to probe the tetraquark nature of those mesons within a certain model [143,144]. The phenomenological fits of the LHCb do neither support a contribution of the f 0 (980) in the B → J/ψππ [130] nor an f 0 (500) in B s → J/ψππ decays [125] by employing the isobar model. Hence the authors conclude that their data is incompatible with a model where f 0 (980) is formed from two quarks and two antiquarks (tetraquarks) at the eight standard deviation level. In addition, they extract an upper limit for the mixing angle of 17 • at 90% confidence level between the f 0 (980) and the f 0 (500) that would yield a substantial (ss) content in f 0 (980) [130]. But in fact a substantial f 0 (980) contribution is also found in the B-decays in a dispersive analysis of the same data that allows for a model-independent inclusion of the hadronic final state interactions in Ref. [145], which puts into question the conclusions of Ref. [130]. At this stage, the quark structure of scalar particles are still quite controversial. On the theory side, there are some studies on the f 0 (980) by assuming the f 0 (980) as a puress state. For example, the authors studied the B s → J/ψf 0 (980) with the light-cone QCD sum rule and factorization assumption in Ref. [146] and using generalized factorization and SU(3) flavor symmetry in Ref. [147]. In Ref. [148], the authors calculated theB s → f 0 (980) form factor from the light-cone sum rules with B-meson DAs, and investigated the S-waveB s → KK form factors to study the width effect, where the f 0 (980) is dominated by thess configuration. As a first approximation, we take into account the scalar meson f 0 (500), f 0 (980), K * 0 (1430) in theqq density operator with q = (u, d, s). The S-wave time-like form factor F S (ω 2 ) used to parameterize the S-wave two-pion and kaon-pion DAs have been determined in Refs. [117,137].
We list the branching ratios of the four-body decays B (s) → [SS, SV, V S] → (ππ)(Kπ) with experimental data [46] in Table V. So far, only five of them, say B + → (f 0 (980) →)π + π − (K * + →)K 0 π + , B 0 → (f 0 (980) →)π + π − (K * 0 →)K + π − , B 0 → (ρ − →)π − π 0 (K * 0 (1430) + →)K 0 π + , B 0 → (ρ 0 →)π + π − (K * 0 (1430) 0 →)K + π − , and B 0 → (f 0 (980) → )π + π − (K * 0 (1430) 0 →)K + π − , have been reported by experiments. It is shown that, except for the colour suppressed (" C ") decay B 0 → (ρ 0 →)π + π − (K * 0 (1430) 0 →)K + π − , our predictions of other four channels are consistent with the available experimental data within errors, with the remaining predictions awaiting for the examinations from future experimental measurements. However, the branching ratio B(B 0 → (ρ 0 →)π + π − (K * 0 (1430) 0 →)K + π − ) = (7.9 +4.9 −3.6 ) × 10 −6 estimated in this work is smaller than the experimental data B = (18 ± 4) × 10 −6 [46] by a factor of ∼ 2. Since only leading order contributions are considered in this work, it indicates that this decay mode might be more sensitive to next-to-leading order corrections, and it is similar to the situation of other " C "-type decays, such as B 0 → π 0 π 0 , ρ 0 ρ 0 . Besides, under the isospin limit, it is naively expected that which is not borne out by experiment and needs to be further studied in the future. Among the three different kinds of theoretical errors considered in our work, one can see that the most important theoretical uncertainties for the branching ratio are caused by the nonperturbative input parameters of the wave functions for some decay modes. Taking the decay B + → (f 0 (500) → )π + π − (K * + →)K 0 π + as an example, which is dominated by the B → (f 0 (500) →)ππ transition progress, its branching ratio is much more sensitive to the Gegenbauer moment a S from the S-wave DAs. In the PQCD approach, wave functions are the most important input parameters and the improved knowledge of them is expected to yield improved estimates of the branching ratios and other observables, which may lead to better consistency with the data. Since the f 0 (500) is very broad, we use BW formula and Bugg model to parameterize the f 0 (500) resonance and compare their results. It is found that the model-dependence of the decay rate is indeed not significant. The central values of PQCD predictions based on the Bugg model are consistent with the ones from the BW formula. Our prediction of B 0 → (f 0 (980) → )π + π − (K * 0 →)K + π − is consistent with the current data, and also comparable with that from Table III in [104] within errors. In order to compare our predictions with other theoretical results for decays involving f 0 (980), we use the B(f 0 (980) → π + π − ) = 0.50, which is taken from [6] and in agreement with the value of B(f 0 (980) → π + π − ) = 0.46 obtained in [131]. We can extract the branching ratios of the two-body decays B → f 0 (980)K * from the corresponding four-body decays B → f 0 (980)(→ ππ)K * (→ Kπ) in Table V under the narrow width approximation. Taking the decay B 0 → f 0 (980)K * 0 as an example, we obtain its branching ratio B(B 0 → f 0 (980)K * 0 ) = 8.7 × 10 −6 , which is in good agreement with previous two-body results in the QCDF approach [6] and PQCD approach [15]. Strictly speaking, the narrow width approximation is not fully justified since such approximation has its scope of application. As mentioned above, the nonperturbative input parameters from the wave functions make important sense to the branching ratios. We can fit the related Gegenbauer moments with abundant data to match the experiment in the future. However, the fact that their rates can be accommodated in the two-quark picture for f 0 (980) does not mean thatqq composition should be supported. It is too difficult to make theoretical predictions on these decay modes based on the four-quark picture for scalar resonances. We just assume they are constituted by two quarks at this moment.
The decays B → ρK * 0 (1430) have already been studied systematically in the two-body framework within the PQCD approach [16]. Taking the two measured channels B 0 → ρ − K * 0 (1430) + → (π − π 0 )(K 0 π + ) and B 0 → ρ 0 K * 0 (1430) 0 → (π + π − )(K + π − ) as examples, we have: The results from the previous PQCD work [16] are obtained by multiplying the relevant two-body branching ratios according to Eqs. (70)- (71). Since the width of the resonant state and the interactions between the final state meson pair will show their effects on the branching ratios, the new four-body predictions are relatively larger than the converted values from previous PQCD calculations, but more close to the experimental data. Therefore, it seems more appropriate to treat these decay modes as four-body decays.
C. Direct CP asymmetries
In Table VI, VI: Direct CP asymmetries (in units of %) for the B (s) → V V → (ππ)(Kπ) decays compared with the previous predictions in the PQCD approach [18], the updated predictions in the QCDF [3,4], SCET [25] and FAT [26]. Experimental results for branching ratios are taken from Table II. The sources of the theoretical errors are the same as in Table IV. Expt. 31 ± 13 Expt.
SCET [25]
and FAT [26] as well as the PQCD predictions in two-body framework [18] are also presented. Meanwhile, direct CP asymmetries for the four-body B (s) → [SS, SV, V S] → (ππ)(Kπ) decays are displayed in Table VII. As we know, the kinematics of the two-body decays is fixed, the decay amplitudes of the quasi-two-body decays depend on the invariant mass of the final-state pairs, which result in the differential distribution of direct CP asymmetries. The CP asymmetry in the four-body framework is moderated by the finite width of the intermediate resonance appearing in the time-like form factor F (ω 2 ). Thus, it is reasonable to see the differences of direct CP asymmetries between the two-body and four-body frameworks in the PQCD approach. By comparing the numerical results as listed in Table VI, due to the different mechanism and origins of the strong phase, one can see that the QCDF and SCET results for the direct CP asymmetries are quite different from ours for some decay modes. As is well known, besides the weak phase from the CKM matrix elements, the direct CP asymmetry is proportional to the strong phase. In the SCET, the strong phase is only from the nonperturbative charming penguin at leading power and leading order, while in the QCDF and PQCD approaches, the strong phase comes from the hard spectator scattering and annihilation diagrams respectively. Besides, the power corrections such as penguin annihilation, which are essential to resolve the CP Table IV. puzzles in the QCDF, are often plagued by the endpoint divergence that in turn break the factorization theorem [3]. In the PQCD approach, the endpoint singularity is cured by including the parton's transverse momentum. Anyway, since current experimental measurements still have relatively large uncertainties, we have to wait for more time to test these different predictions.
In Tables VI and VII, a large CP asymmetry can be understood due to the sizable interference between the tree and penguin amplitudes, while a small value of CP asymmetry is attributed to the dominant tree or penguin amplitudes. For example, among the six considered B (s) → ρK * → (ππ)(πK) decays as presented in Table VI, the CP asymmetries A CP for the two penguin-dominant processes B 0 → (ρ 0 →)π + π − (K * 0 →)K + π − and B + → (ρ + →)π + π 0 (K * 0 →)K + π − are indeed quite small: less than 1%. However, for the " Color-suppressed " decay B 0 s → (ρ 0 →)π + π − (K * 0 →)K − π + , due to the large penguin contributions from the chirally enhanced annihilation diagrams, the sizable interference between the tree and penguin contributions makes the direct CP asymmetries A CP as large as ∼ 30%. For four B +,0 → ρK * → (ππ)(Kπ) decays, our predictions of CP asymmetries are in agreement with observations within uncertainties. Moreover, a helicity-specific analysis would provide interesting further insights. Very recently, LHCb [43] has reported the CP asymmetry associated with longitudinal polarization A 0 ρK * = −0.62 ± 0.09 ± 0.09, where the first uncertainty is statistical and the second systematic. The data is much different from our prediction A CP 0 (B 0 → ρ 0 (→ π + π − )K * 0 (→ K + π − )) = 3.5%. Considering the branching ratio of the B 0 → ρ 0 (→ π + π − )K * 0 (→ K + π − ) decay associated with the longitudinal polarization, the contributions of the penguin diagrams (B = 2.98 × 10 −6 ) are larger than the tree ones (B = 5.74 × 10 −8 ) by roughly a factor of 52, which results in the smallness of direct CP asymmetries. The big gap between the theory and experiment should be resolved in the future.
In the limit of U -spin symmetry, some of B s decays can be related to B 0 ones. For B (s) → V V decays, it has been studied in [4,18] and seems to hold well. Since we have calculated the B and B s decays to V V in this work in the PQCD approach, we also check the U -spin symmetry in some decay modes studied in [4,18]: On basis of these U -spin relations along with the branching ratios, the lifetimes of B and B s mesons and the direct CP asymmetries in B decays, we can get the relevant direct CP asymmetries in B s decays. This can be then compared with the corresponding predictions in the PQCD approach to check whether the U -spin symmetry works well or not. We show this comparison in Table VIII, where the entries in the last two columns have to be compared with each other. It turns out that U -spin symmetry is in general acceptable within the calculational errors. Modes The predicted TPAs for the B (s) → (ππ)(Kπ) decays are displayed in Table IX. It is shown that our PQCD predictions of "true" CP -violating TPAs are very small in the SM, which makes the measurement of a large value for that TPA point clearly towards the presence of new physics. As "fake" TPAs are due to strong phases and require no CP violation, the large fake A 1,2 T-fake simply reflects the importance of the strong final-state phases.
Since the left-handedness of the weak interaction A − ≪ A + is expected, it implies A ≈ A ⊥ . The A 2 T term requires both transversely polarized components A and A ⊥ and the decay amplitude associated with transverse polarization is smaller than that for longitudinal polarization in the naive expectation. Hence A 2 T is power suppressed relative to A 1 T . Meanwhile, the smallness of A 2 T is also attributed to the suppression from the strong phase difference between the perpendicular and parallel polarization amplitudes, which was found in the PQCD framework [18] and supported by the LHCb Collaboration [43]. An observation of A 2 T with large values can signify physics beyond the SM. As mentioned above, (A 1,2 T +Ā 1,2 T )/2 = A (1,2)ave T (true) when the decay channel has a nonzero CP asymmetries. We find that the greater difference between the A 1,2 T (true) and A (1,2)ave T (true) appears with the larger direct CP asymmetry.
Recently, the measurements of "true" and "fake" TPAs for B 0 → ρ 0 K * 0 → (π + π − )(K + π − ) have been reported by LHCb Collaboration [43]. The PQCD prediction of A 1 T-true agrees well with the experiment A ρK * ,1 T-true = −0.0210 ± 0.0050 ± 0.0022, where the first uncertainty is statistical and the second systematic. While for "fake" TPAs, our predictions are a little larger than the measurements but compatible within large uncertainties. It should be stressed that there are large uncertainties in both experimental measurements and the theoretical calculations for TPAs, so the discrepancy between the data and the theoretical results could be clarified with the high precision both in experimental and theoretical sides. Since "fake" TPAs strongly affected by the strong phases, we lack a perfect knowledge of all the possible signals of the strong phases, such as final-state interactions. For this reason, we just estimate the size of the corresponding TPAs. We hope the future experiments can test our predictions.
IV. CONCLUSION
In this work, we have presented six helicity amplitudes of four-body decays B (s) → (ππ)(Kπ), where ππ invariant-mass spectrum is dominated by the vector ρ resonance and scalar f 0 (500), f 0 (980) resonances, and the vector K * resonance and scalar resonance K * 0 (1430) are expected to contribute in the Kπ invariant-mass range. We have examined the branching ratios, polarization fractions, direct CP asymmetries, triple product asymmetries in B (s) → [V V, SS, SV, V S] → (ππ)(Kπ) decays. In our numerical study, there exist many theoretical uncertainties in the calculation. The uncertainties of the nonperturbative parameters of the two-meson DAs and the variation of the hard scale provide the dominant theoretical errors to the theoretical predictions for branching ratios and other physical observables. Therefore, the relevant Gegenbauer moments should be further constrained to improve the precision of theoretical predictions and meet with future data. In addition, one should make a great IX: PQCD predictions for the TPAs (%) of the four-body B (s) → (ρ →)ππ(K * →)Kπ decays. The sources of theoretical errors are same as in Table IV but added in quadrature.
TPAs-2 Modes
A 2 effort to evaluate the higher-order contributions to four-body B meson decays in order to reduce the sensitivity to the variation of the hard scales.
We have extracted the branching ratios of two-body B → ρK * decays from the results for the corresponding four-body decays under the narrow-width approximation and shown the polarization fractions of the related decay channels. The obtained two-body branching ratios agree well with previous theoretical studies performed in the two-body framework within errors. The predicted hierarchy pattern for the longitudinal polarization fractions in the B (s) meson decays into the P -wave ππ and Kπ pairs is compatible with the data roughly. However, there is a big gap between our prediction of longitudinal polarization fraction for B 0 → ρ 0 K * 0 and the recent LHCb measurement, which should be resolved. In addition, we have calculated the branching ratios of the four-body decays B (s) → [SS, SV, V S] → (ππ)(Kπ). For the decays associated with scalar resonance f 0 (500), we have used the BW and Bugg models to parameterize the wide f 0 (500) meson respectively but found that the modeldependence of the PQCD predictions is not significant. The branching ratios of B 0 → (ρ − →)π − π 0 (K * 0 (1430) + →)K 0 π + and B 0 → (ρ 0 →)π + π − (K * 0 (1430) 0 →)K + π − decays, which are related to isospin limit, remain puzzling and need to be resolved.
We have calculated the direct CP asymmetries with each helicity state (A CP 0, ,⊥ ) for the four-body B (s) → V V → (ππ)(Kπ) decays, together with the direct CP asymmetries of B (s) → [SS, SV, V S] → (ππ)(Kπ) decays. The CP asymmetry in the four-body framework is dependent on the invariant mass of the final-state pairs, which results in the differences between the two-body and four-body frameworks in the PQCD approach. Meanwhile, we perform an angular analysis on four-body B (s) → ρK * → (ππ)(Kπ) decays to obtain the triple product asymmetries in detail. We found that most "true" TPAs are very small, which are consistent with the predictions of the standard model. A "true" TPA that is predicted to vanish provides an excellent place for looking for new physics because there is no suppression from the strong phases. | 13,839 | 2021-07-22T00:00:00.000 | [
"Physics",
"Art"
] |
STUDY ON ADAPTIVE PARAMETER DETERMINATION OF CLUSTER ANALYSIS IN URBAN MANAGEMENT CASES
The fine management for cities is the important way to realize the smart city. The data mining which uses spatial clustering analysis for urban management cases can be used in the evaluation of urban public facilities deployment, and support the policy decisions, and also provides technical support for the fine management of the city. Aiming at the problem that DBSCAN algorithm which is based on the density-clustering can not realize parameter adaptive determination, this paper proposed the optimizing method of parameter adaptive determination based on the spatial analysis. Firstly, making analysis of the function Ripley's K for the data set to realize adaptive determination of global parameter MinPts, which means setting the maximum aggregation scale as the range of data clustering. Calculating every point object’s highest frequency K value in the range of Eps which uses K-D tree and setting it as the value of clustering density to realize the adaptive determination of global parameter MinPts. Then, the R language was used to optimize the above process to accomplish the precise clustering of typical urban management cases. The experimental results based on the typical case of urban management in XiCheng district of Beijing shows that: The new DBSCAN clustering algorithm this paper presents takes full account of the data’s spatial and statistical characteristic which has obvious clustering feature, and has a better applicability and high quality. The results of the study are not only helpful for the formulation of urban management policies and the allocation of urban management supervisors in XiCheng District of Beijing, but also to other cities and related fields. * Corresponding author
INTRODUCTION
With the rapid growth of large-scale data processing and in-depth analysis of demand in all walks of life, data mining has become a hot area of research for many scholars (Genlin et al.,2014).Refinement, which is an important goal of the urban operation and development, provides the technical support for the delicacy management of city operation (Jing 2014).With the progress of society and science technology, all kinds of issues with respect to urban operation have appeared in succession.According to the city's report on the work of the government, the number of cases about urban management(ChengGuan case in short)also increases year by year, which has influence on the urban appearance and steady running of the city.Therefore, it is of great theoretical and practical value to use the spatial data mining technology to analyze the urban management cases and assist the government decision-making.
As a method of data mining, clustering analysis has been widely used.DBSCAN algorithm based on density clustering which has high speed, data adaptability, noise insensitive characteristics was studied by many researchers (Xinyan and Deren,2005).However, the DBSCAN algorithm needs to manually determine the parameters Eps and MinPts, and the values of these two parameters directly affect the quality of data clustering.In view of the problem of how to select the optimal parameters, a large number of literatures have proposed the method of assuming MinPts value and then determining the Eps value.
Although avoiding parameter determination artificially, these methods based on the premise of assumption of MinPts are still lack of adaptive parameters Such as Ren Xingping (Xingping et al.,2007)
The Study Area And Data Source
This paper is based on the Xicheng District of Beijing in 2010.
Xicheng District is located in the center of Beijing, which is a set of politics, economic, culture and tourism as one of core development area.It has higher requirements for urban management because of its special geographical position.The specific distribution shown in Figure 1.
Ripley's K Function
Ripley's K method is a representative spatial point pattern analysis method, which can quantitatively evaluate the spatial distribution characteristic of point patterns (Tang et al.,2015).
In this method, Ripley's K function is used to analyze the clustering degree of point datasets at different spatial scales in a certain confidence interval, then the maximum clustering scale of the best clustering effect is quantitatively analyzed according to the expected K value and observed K value.The spatial scale is calculated as follows: ) In the above formula, d represents the distance, A represents the total area of the area occupied by the feature set, Kij represents the spatial weight.
At a certain specific distance, when the observed K value is
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-2/W7, 2017 ISPRS Geospatial Week 2017, 18-22 September 2017, Wuhan, China larger than the predicted K value, the clustering degree of the distribution is higher than that of the random distribution of the scale.Therefore, this paper selects the spatial scale with the largest difference between observation K and prediction K as the value of parameter Eps, in order to get a better clustering effect, this paper sets the confidence level to 95% based on the data size of research.In view of this problem, some scholars have proposed an improved method (Yi et al.,2016).This paper calculates the point data size K for each point object which is in the Eps threshold, and then takes the max frequency of statistical analysis of the value of K as MinPts value.
Component correlation analysis
Overlapping analysis method has the characteristics of low 5.According to the quantitative analysis in the light of the above chart, when aggregation scale of the street order cases is greater than the maximum aggregation scale of 563 meters, the observed K value gradually approaches the prediction K value and they The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-2/W7, 2017 ISPRS Geospatial Week 2017, 18-22 September 2017, Wuhan, China almost keep parallel; When the aggregation scale of the urban environment cases is larger than the maximum aggregation scale of 792 meters, the observation K value is close to the forecast K value and almost parallel, which shows that as for these two cases, when the spatial scale is larger than the maximum aggregation scale, the degree of aggregation of the data set is decreasing.In view of the fact that the observed K value is higher than the predicted K value at a certain distance, the aggregation degree is higher.Therefore, this paper selects the maximum aggregation scale as the parameter Eps when the aggregation effect is best.The clustering results and the parameter values are statistically analyzed, as shown in Table 7.
et al take MinPts as 4. According to the forth nearest neighbor distance graph of the data object set, the value of Eps is taken as less than the percentage of noise level; Zhou Dong(Dong and Peng,2009)et al assume that MinPts is 3, and then according to K-dist curve to determine the value of Eps.Some scholars have done some research on the adaptive determination of global parameters Eps and MinPts.Among them, the majority are the research that under the premise of statistical analysis of the data set.For example Xia luning (Luning and Jiwu,2009) et al proposed to k-dist probability curve and statistical model fit peak to Eps, drew the Noise curve and its inflection point MinPts method to achieve the parameters of the adaptive determination, but the whole process is too cumbersome and calculation is large, and the practicability is weak; Li Zonglin(Zonglin and Ke,2016.)etal established a suitable mathematical model to determine the Eps and MinPts values adaptively by using the kernel density estimation theory, but this method is not suitable for data set with large density difference, and the computational complexity of the algorithm is high.There are some scholars to explore the data partition as the premise of the research methods, such as Stefanakis(Stefanakis, 2007), Pandey Abhilash Kumar(Pandey Abhilash Kumar and Dubey Roshni,2014) and others on first divided again clustering of data area.Huang Gang(Gang et al.,2015)et al reduced the number of regional query method to achieve high efficiency clustering algorithm by selecting the seeds on behalf of objects.In summary, the existing literature on spatial data and spatial statistical characteristics of the research is less.The DBSCAN clustering algorithm based on density still needs to study the data set to explore the statistical characteristics of the data and achieve high quality clustering.This paper uses the Ripley's function and K-D tree to analyze the statistical characteristics of urban management case data, and determine the parameters of DBSCAN algorithm adaptively.The optimized DBSCAN algorithm is used for data mining in typical urban management cases to provide auxiliary decision for urban management policy making, for urban management supervision staff scheduling to provide quantitative analysis support and enhance the city running fine management ability.
Figure2
Figure2.Data statistical histogram K-D tree is similar to the binary tree, it is a data structure with left and right subtrees.The biggest difference between K-D tree and the binary index tree is that K-D tree is stored in the K-dimensional point data.The K-D tree algorithm is composed of two parts, including tree-building and search.K-D tree is divided into left subtree and right subtree according to the max variance of data.In order to make sure the left subtree and the right subtree have the same length, K-D takes the median where array of attribute value as the partition axis.There are two kinds of search methods in K-D tree structure: range search and K-nearest neighbor search.The range search refers to searching the point data within the threshold range for a certain point object in the given searching threshold; K-neighbor search refers to specifying a point object, and then traversing the original data set to find the nearest point of the object K point data.This classical K-D tree algorithm can only be used to search the high efficiency K in the low dimensional case, and the efficiency of searching for high-dimensional data is very low.
Figure6.Spatial distribution of cluster results Figure 6 and Table 7 indicate that: (1) In the case of the street order case, there are 38 clusters under the condition of confidence 95%, scanning radius 563m and scanning density 110, which are mainly distributed in the east and west streets of the city center and north of the city.In addition to the less clustered cluster distribution on Yuetan and Xinjiekou streets, apart from Yuetan and Xinjiekou streets with less clusters, the clusters of The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-2/W7, 2017 ISPRS Geospatial Week 2017, 18-22 September 2017, Wuhan, China West Chang'an Street, Shichahai Street, Desheng Street and Exhibition Road are uniformly distributed.The northern part of the city has larger population mobility, and the majority of the floating population on the convenience, fast shopping needs for unlicensed operators and operators outside the shop to provide a market.And the requirement for convenient and fast shopping of the majority of the floating population provides the unlicensed operators and operators outside the shop with the market.So It is extremely necessary to strengthen the management and deployment of urban managers in the northern part of the city; (2)Under the conditions of confidence 95%, scan radius 792 m and scanning density 200, the cluster of urban environment cases is 44, which are mainly distributed in Yuetan Street, Financial Street and West Chang'an Street in the center of the city, the northern part of the city also has a small amount of distribution, such as Desheng Street, Xinjiekou Street and Shichahai streets, in which the east side of the city, the Yuetan street is especially concentrated, and the number of cluster types is as high as 25, compared with the other streets in Xicheng District, the number of communities (32) and population (150,000) is the largest in Yuetan district.The excessive communities and population make the number of long-term abandoned vehicles on the roadside and garbage piled up on the side of the road larger than other streets.Yuetan streets become a high incidence of urban environment class cases because of lacking of Sanitation facilities and the exposed garbage can not receive timely processing.So we should strengthen city management personnel in Yuetan Street inspections, in order to reduce the number of cases of the occurrence of such cases, improve the current analysis of clustering results In view of the street order case is mainly affected by unlicensed business operators, shop-outside operators and vagrant begging and other issues; City environment cases are mainly affected by the exposed garbage, unclean pavements, accumulation of waste residue and abandoned vehicles and other issues.Lacking of urban public infrastructure is the root cause of a large part of the city environment.Therefore, in order to verify the rationality of the clustering results, this paper choose 14 classes cases about urban public infrastructure and urban environment, such as dustbin, garbage bin, comfort station, storage frame and so on to make correlation analysis according to " digital city management information system _ second parts: 2013) ".If there are a lot of duplicate areas between the clusters and the concentrated areas of component cases, which means that urban management cases is unrelated to the component facility.Conversely, If there is only a small amount of overlap between the clusters and the concentrated areas of the component cases, it shows that the urban management case is related to the configuration of the components.The specific correlation analysis results are shown in Figure 8, in which different colors represent different clusters, transition which is from white to black region shows the nuclear density of components from small to large. | 3,091.2 | 2017-09-14T00:00:00.000 | [
"Computer Science",
"Engineering"
] |
HISFCOS: Half-Inverted Stage Block for Efficient Object Detection Based on Deep Learning
Recent advances in object detection play a key role in various industrial applications. However, a fully convolutional one-stage detector (FCOS), a conventional object detection method, has low detection accuracy given the calculation cost. Thus, in this study, we propose a half-inverted stage FCOS (HISFCOS) with improved detection accuracy at a computational cost comparable to FCOS based on the proposed half inverted stage (HIS) block. First, FCOS has low detection accuracy owing to low-level information loss. Therefore, an HIS block that minimizes feature loss by extracting spatial and channel information in parallel is proposed. Second, detection accuracy was improved by reconstructing the feature pyramid on the basis of the proposed block and improving the low-level information. Lastly, the improved detection head structure reduced the computational cost and amount compared to the conventional method. Through experiments, the proposed method defined the optimal HISFCOS parameters and evaluated several datasets for fair comparison. The HISFCOS was trained and evaluated using the PASCAL VOC and MSCOCO2017 datasets. Additionally, the average precision (AP) was used as an evaluation index to quantitatively evaluate detection performance. As a result of the experiment, the parameters were increased by 0.5 M compared to the conventional method, but the detection accuracy was improved by 3.0 AP and 1.5 AP in the PASCAL VOC and MSCOCO datasets, respectively. in addition, an ablation study was conducted, and the results for the proposed block and detection head were analyzed.
Introduction
Recently, computer vision has been studied in related applications, such as object detection [1,2], semantic segmentation [3,4], and super resolution [5], because of the development of convolutional neural networks (CNNs). Object detection is a fundamental task in computer vision that classifies the categories of objects after a regression process to determine the location of an object. It is a core technology in autonomous driving [6], monitoring systems [7], and face recognition [8].
Deep-learning-based object detection methods are classified into one-and two-stage methods. It is divided into anchor-based and anchor-free methods according to the method for defining the anchor in detail. The two-stage method uses a region proposal network (RPN) to find regions where objects are most likely to be present. On the basis of the selected region, classification is performed after localization. Representative methods include regionbased convolutional neural networks (R-CNNs) [9][10][11]. Unlike the two-stage method, the one-stage method simultaneously handles both localization and classification. Compared to the two-stage method, there is a trade-off between detection accuracy and detection speed. The one-stage method has the advantage of detection speed. Representative methods include the single shot multibox detector (SSD) [12] and you only look once (YOLO) [13,14]. Additionally, the one-stage method improves detection accuracy by extending the depth of the model to improve low accuracy and the feature pyramid [15] structure, which uses feature maps of different scales.
Object detection researchers are studying anchors that affect detection performance. An anchor-free method for predicting the class of an object and learning its location, similar to a fully convolutional network (FCN) [16], was studied rather than defining an anchor as in a conventional detector. The fully convolutional one-stage detector (FCOS) [17] is an example of a representative method.
In this paper, we propose HISFCOS with improved detection accuracy while retaining FCOS network complexity. In the conducted work, the accuracy of the proposed network was improved using three main modules (half-inverted stage (HIS) block, HIS feature pyramid, and lightweight detection head). First, the HIS block improved detection accuracy by minimizing feature loss through parallel operation for the proposed spatial and channel information. Second, low-level feature information loss caused by the conventional method was minimized by reconstructing the feature pyramid on the basis of the proposed block. Lastly, the amount of computation was minimized while maintaining accuracy by improving the conventional detection head structure. The proposed method showed high detection accuracy in experiments with similar computational costs comparable to the FCOS. Through an ablation study, we analyzed the contribution of the detection head structure and the HIS block in HISFCOS.
The main contributions of this work are as follows: • We propose an HIS block that reduces loss of spatial and channel information. • We improved detection accuracy by reconstructing the feature pyramid on the basis of the proposed block and improving the low-level information. • We propose a lightweight detection head that reduces the amount of computation by improving the structure of the conventional detection head.
Fully Convolutional One-Stage Detector
The FCOS is an anchor-free, one-stage detector network with a feature pyramid structure that uses ResNet [18] as its backbone network. Figure 1 shows the FCOS structure. This method was proposed to address the drawbacks of the conventional anchor-based method. The anchor-based method affected the detection performance according to the box design, such as the size, aspect ratio, and number of predefined anchors. Additionally, many bounding boxes are created to achieve high recall. A recall is the ratio of object detection when object detection is more important than precision. However, class-imbalance problems arise because many bounding boxes are assigned to the negative samples. Therefore, FCOS predicts the class of an object in pixel units, such as an FCN, without using an anchor. When the predicted sample is positive, a detector of the anchor-free method, which detects objects without using an anchor, is proposed using distance vector regression with the predicted box based on the center of the object. Additionally, a feature pyramid structure was used to improve the low correct recall problem caused by not using an anchor. Lastly, a centerness loss function was proposed to solve the problem of detection accuracy being reduced as a low-level bounding box with a low score moved away from the center. The centerness loss function improves detection accuracy by removing the low-level bounding box in a manner that gives weights when the distance from the center is considered in relation to the center distance of the object.
Depthwise Convolution
Depthwise convolution (Dwconv), unlike standard convolution, can extract the spatial information of each channel without being affected by all channels of the input image. In other words, calculations for each channel are performed in the spatial direction without the involvement of other channels. Thus, each kernel has a parameter for a single channel. Consequently, only spatial information unique to each channel can be learned, which is the same as in the special case where the number of groups in the group convolution equals the number of channels. On the basis of this structure, MobileNet [19,20] was proposed for limited environments such as embedded devices using depthwise separable convolution [21], which combines Dwconv and 1 × 1 convolution to exponentially reduce the amount of computation and enable real-time operation. Figure 2 shows example of depthwise convolution with kernel = 3, input tensor size 8 × 8 (C × H × W). The symbols C, H, W and are the channel height, width and matrix product. Depthwise convolution diagram. Convolution kernel separated for each channel, a 2D kernel of 3 × 3 is attached to each separated matrix of size 3 × 8 × 8, each convolution operation is performed, and then only channel wise spatial information is learned through recombination.
Channel Attention
Channel attention (CA) [22] is a technique that emphasizes a specific channel using the correlation between channels in the feature map. This process is illustrated in Figure 3. First, the input feature map contains a vector representing each channel as a vector of the same size as the channel size of the feature map through global average pooling (GAP). The vector is compressed into a vector with meaningful information using a fully connected (FC) layer. Nonlinearity is added to the compressed vector using a rectified linear unit (ReLU) activation function. The vector is compressed, and nonlinearity is added via the second FC layer. A compressed vector is generated using the sigmoid activation function to enhance the vector of the channel size with a value between zero and one. The emphasized vector is refined using the input feature map and element multiplication operation to refine the unnecessary feature map in each channel and generate the emphasized feature for the object in the channel. The channel attention is expressed using Equation (1): where F Input , σ 1 , σ 2 , and GAP are input feature map, ReLU, sigmoid, and global average pooling, respectively.
Proposed Method
This section describes the proposed method. HISFCOS consists of a backbone network, HIS feature pyramid, and a lightweight detection head. Figure 4 shows the proposed HISFCOS architecture. First, ResNet-50, the same as the conventional FCOS, was used as the backbone network. The feature map extracted from the backbone network was used as an input for the reconstructed feature pyramid on the basis of the HIS block that minimizes feature loss. In addition, a bottom-up path was added to the feature pyramid of the top-down path to improve the lack of low-level information in the conventional network. As a result, feature information is improved by combining the high-level feature map and the low-level information. Lastly, unlike the conventional detection head composed of standard convolution, an inverted residual block structure is applied to reduce the amount of computation. The proposed method detects large and small objects for each scale using five heads that had passed through a feature pyramid. Details are covered in Sections 3.1-3.3.
Half-Inverted Stage Block
Feature information in the FCOS feature pyramid is lost in the deep layer. Therefore, we propose an HIS block that reduces feature loss by simultaneously calculating spatial and channel information. Figure 5 shows the structure of the HIS block. The structure of the proposed HIS block is as follows. First, if a large channel is used, unnecessary features and computations are increased. Therefore, the input feature map is compressed to 1 2 channel size using a 1 × 1 convolution. Subsequently, to minimize the loss of spatial and channel information, the optimized operation for each operation was used in parallel. Dwconv, which extracts spatial information through spatial operations in channel units, and CA, which emphasizes important features such as objects in the channel space, are used. (Channel attention replaces the FC layer with a 1 × 1 convolution to prevent an increase in the amount of computation). To refine the feature information extracted by a parallel operation, it is first combined with a concatenation operation. Feature loss generated during the compression process was minimized by combining the compressed and refined features with the size of the first input channel 1 2 . Lastly, to improve the spatial information required for object detection, spatial information was improved using a dilated convolution with a wide receptive field and a computational cost comparable to the standard convolution. Equations (2) and (3) represent the spatial and channel feature combinations and refinement process with the HIS block, respectively.
where each symbol is Dilated r=2 3×3 , and concat represent a dilated convolution and concatenation operation with a kernel size of 3 and dilated ratio of 2, respectively.
HIS Feature Pyramid
FCOS is composed of a feature pyramid with a top-down path structure. Some deep layers do not combine with low-level features to recover the loss after upsampling, resulting in feature loss. Therefore, feature loss was prevented by rebuilding the feature pyramid using the proposed HIS block. First, by incorporating a bottom-up path into the conventional top-down path structure, low-level feature loss caused by the model depth is minimized by combining low-level information such as contours and edges with high-level information such as texture and shape in each feature map. Second, because there is no process in the conventional method for restoring information loss in some features of the feature pyramid, there may be a part-false-positive problem of detecting multiple objects owing to feature loss when extracting a large object. Therefore, by combining the features of the backbone network, the loss of features is minimized, allowing for the part-false-positive problem to be solved when large objects are detected.
Lightweight Detection Head
A conventional detection head has a structure in which each classification and regression branch repeat 3 × 3 standard convolution four times for feature refinement. The conventional structure improves accuracy in a detector with a shallow structure [17,23]. However, a detection head with a structure that repeats the standard convolution reduces detection accuracy compared with the high computational cost. Therefore, the proposed structure is applied to a conventional detection head using an inverted residual block [24] structure to lower the high computational cost. The computational cost of an ablation study was compared with that of a conventional detection head. Figure 6 shows the structures of the conventional and proposed detection heads.
First, the proposed detection head extends the channel using 1 × 1 convolution. Spatial information was extracted from the extended feature map using 3 × 3 Dwconv. The extracted spatial information was compressed using 1 × 1 convolution. By combining the extracted spatial information with the channel axis, features could be extracted at a lower computational cost compared with standard convolution. This structure is an inverted residual block. Features were then refined using 3 × 3 standard convolution. By replacing it with this structure, efficient detection was possible at a lower computational cost than that of the conventional method.
Loss Function
The classification loss and bounding box regression loss functions are used as loss functions in the proposed method, in the same manner as the FCOS. We used focal loss [23] as the classification loss function. The cross-entropy loss function compares the proposed method with the ground truth, and outputs an error. However, when the standard crossentropy loss function is used, the sample, which is easily detected, dominates the overall loss function. In object detection, most samples have a larger ratio of background samples to foreground samples, which causes class-imbalance problems. The use of such unbalanced samples for training is inefficient. Therefore, focal loss, which improved the standard cross-entropy loss function, was used as the classification loss function. The cross-entropy and focal loss functions are expressed using Equations (4) and (5), respectively: where p and y are the output values of ground truth and model, respectively.
Focal loss becomes smaller than conventional cross-entropy loss when p t approaches one. Conversely, as p t approaches zero, loss increases. Here, α,γ are hyperparameters that control the contribution to the loss function. When γ = 0, it is the same as the conventional cross-entropy loss function. We use α,γ, which had the same value as FCOS α = 2, γ = 0.25.
GIoU(B, B gt
For the bounding box regression loss function, generalized intersection over union (GIoU) [25] was used as a regression loss function on the basis of the intersection of union (IoU) [26]. Equation (6) represents GIoU, where each symbol B, B gt , and C represents the predicted bounding box, ground truth, and the small size area covering the predicted bounding box and ground truth, respectively.
The centerness loss function [17], which assigns weight to the bounding box with a low score away from the center, is a loss function that determines whether an object exists in the center. Therefore, binary cross-entropy (BCE) loss, a special form of cross-entropy, is used as a function to compare the two cases. Equation (7) is a centerness loss function expressed as follows: where Y andŶ are output values of ground truth and model, respectively. The total loss function consists of classification loss, regression loss, and centerness loss functions. Total loss is expressed using Equation (8) as follows:
Implementation Details
In this experiment, the public datasets PASCAL VOC [27] and MS COCO2017 [28] were used for HISFCOS performance validation and evaluation. The backbone network was pretrained on the Imagenet-1K dataset before experimentation. The hyperparameters used in the experiment were as follows: stochastic gradient descent (SGD) as used was the optimizer, momentum was set at 0.9, and weight decay was set at 5 × 10 −3 . For each dataset, batches were trained using 32 and 10 epochs, and 50 and 30 epochs were used. Additionally, for the input resolution, PASCAL VOC was used in the same manner as 512 × 512 and the MSCOCO dataset was used in the same manner as the conventional FCOS. An initial learning rate of 1 × 10 −2 was used, and in the case of the PASCL VOC dataset, the learning rate decreased by 0.1 for each 2 and 2.1 K, and in the case of the MSCOCO, the learning rate decreased by 0.1 at 60 and 90 K. The data augmentation technique was as follows: random crops are randomly cropped according to the resolution of the input image. The color jitter randomly changes the brightness, saturation, and hue of the input image. Finally, a random rotation changed the height and width of the input image at random. The hardware and framework used in the experiment are listed in Table 1. The code is available at https://github.com/hby1320/pytorch_object_detection (accessed on 15 March 2022). In this study, detection accuracy was evaluated using the average precision (AP) [29], which is a dataset evaluation metric used in the field of object detection.The AP was calculated by averaging the maximal precision of the recall value on the precision-recall (PR) curve. Precision and recall were obtained using Equations (9) and (10), respectively, as follows: When calculating AP, if there is only one classification class, it is defined as Equation (11).
Recall levels 0.0, . . . , 1.0 calculated the AP of each class as the mean of the maximal precision value for 11 levels. it is necessary to calculate the mean value for the AP because the classification task in the public dataset has more than each class. Equation (12) defined the average of AP for all classes.
where TP, FP, FN, r, ρ interp , N, and AP i denote the true positive, false positive, false negative, recall, precision value of each recall, total number of classes, and AP value of i th class, respectively.
Dataset
In this study, training and tests were conducted using object detection datasets PAS-CAL VOC (07+12) and MSCOCO 2017. PASCAL VOC was trained and evaluated. When classified into a total of 20 classification categories, it was divided into 8324 train datasets, 11,227 validation datasets, and 4952 test datasets were used. In this study, evaluation and resection studies were conducted using 4952 tests and datasets after learning using the training and evaluation dataset. In addition, MSCOCO dataset has a total of 80 classification categories and consisted of 118,287 train datasets, 5000 validation datasets, and 4952 evaluation datasets. In this study, it was used for comparison with the conventional network and other networks.
PASCAL VOC
We compared other networks with HISFCOS using the PASCAL VOC dataset, which is widely used in object detection. At this time, a fair comparison was carried out using input resolution of 512 × 512, which is frequently used in the one-stage detector in the PASCAL VOC dataset. Table 2 shows the results of comparing the detection accuracy (mAP) and number of parameters with those of other networks in the PASCAL VOC 2007 test dataset. The proposed method has parameters similar to those of the conventional method, and detection accuracy was improved by approximately 3.0%. Additionally, compared with R-FCN, which is a two-stage detector more specialized in detection accuracy than in speed, detection accuracy of approximately 0.9% was achieved. This demonstrates the usefulness of the proposed HIS block and lightweight detection head. Figure 7a shows that FCOS is difficult to detect overlapping objects due to spatial and low-level feature information loss during feature extraction. In addition, the bottle class at the bottom of Figure 7 is difficult to detect when there is a feature loss. The proposed method uses HIS blocks to reduce spatial and channel loss. In addition, by reconstructing the feature pyramid structure, low-level information was improved to improve the detection performance of difficult-to-detect overlapping objects. In Figure 7b, compared to the conventional method, the proposed method showed improved detection results for objects that are difficult to detect such as overlapping objects.
MSCOCO
Unlike the PASCAL VOC dataset, the MSCOCO dataset was evaluated using the average of values between 50% and 95% as the IoU threshold. Table 3 shows a comparison of detection performance between the MSCOCO 2017 minival dataset method and other object detection methods. HISFCOS was tested with the same input resolution of 800 × 1333 pixels as the FCOS. The proposed method achieved a detection accuracy of approximately 38.9% using ResNet-50 as the backbone network in the MSCOCO dataset. Compared with FCOS, detection accuracy was improved by approximately 1.5%. Additionally, compared with the two-stage method, the proposed method showed similar detection accuracy. Compared with MaskR-CNN, which is a two-stage method, the proposed method has approximately four times fewer parameters and 0.5% higher detection accuracy. Figure 8 shows the detection results compared with the conventional method for the MSCOCO dataset. Figure 8a shows the top of the image, where false detection for the knife class was confirmed. Owing to the loss of low-level information in the conventional method, similar feature information was misidentified when detecting an object, and a knife class that did not exist in the input image was erroneously detected. Additionally, the lower image in Figure 8a was a part-false-positive problem for a training class with a large aspect ratio. In the conventional method, some object information was lost during the upsample process in the feature pyramid, resulting in the incorrect detection of a large object. However, because the proposed method minimized the loss of features, it was possible to accurately detect it without false detection, as shown in Figure 8b.
Evaluation HISblock Analysis
First, to validate the effectiveness of the proposed method, an experiment was conducted using the PASCAL VOC 2007 test dataset. Table 4 shows the detection accuracy between each class of HISFCOS and FCOS in the PASCAL VOC 07 test dataset. The proposed method improved detection accuracy for all objects compared with the conventional method. Detection accuracy was improved by minimizing the feature loss by reconstructing the feature pyramid structure on the basis of the proposed HIS block. Additionally, the detection performance for the table, bottle, plant, and chair classes was significantly improved, which are difficult to detect because they overlap with other objects in the conventional method. Furthermore, in the case of objects with a large aspect ratio, such as trains and sofas, the conventional method has the problem of part-false-positive sample detection. However, HISFCOS minimizes the loss in each feature map through feature pyramid reconstruction based on the HIS block. The improvement in detection accuracy for previously difficult-to-detect objects was confirmed by resection studies.
Lightweight Detection Head Analysis
The structure of the proposed detection head was compared with that of the FCOS. Table 5 presents a comparison of the structure of the proposed detection head. The base network is based on a feature pyramid reconstructed with the proposed HIS block. When the conventional detection head was used, detection accuracy was reduced by approximately 0.7%. In shallow structures such as the FCOS structure, the process of refining unnecessary features is insufficient. Therefore, stagnating the features using convolution iterations in the detection head is effective. However, as the network deepens in sufficiently refined features, unnecessary operations that repeat many convolutions simply increase the computational cost. Furthermore, it was confirmed that the detection accuracy was reduced. Most object detection networks have a structure in which the standard convolution is not repeated more than twice in the detection head of the network. Therefore, the proposed method is an efficient detection head structure that can improve accuracy while lowering computational cost.
Conclusions
In this study, we proposed HISFCOS, an efficient object detection network that achieves high accuracy while maintaining a computational cost comparable to that of FCOS. First, the proposed method had high prediction accuracy. On the basis of the HIS block, feature loss that occurs during feature extraction was reduced, and spatial information was improved. By rebuilding the feature pyramid on the basis of the proposed block, detection accuracy was improved by minimizing feature loss in the FCOS and improving the low-level information required for each feature map. Second, by improving the structure of the detection head, the calculation cost was maintained while accuracy was improved.
By applying the inverted bottleneck structure to the section where the internal convolution of the conventional detection head structure was repeated, accuracy was improved while maintaining comparable computational cost to that of the conventional method. The proposed method showed high detection accuracy with similar computational cost compared with the conventional detection results because of experiments on the PASCAL VOC and MSCOCO2017 datasets. In the future, we plan to conduct weight reduction research that can be applied to any industrial application field. Data Availability Statement: Data presented in this study are openly available online: http://host. robots.ox.ac.uk/pascal/VOC/voc2007/index.html, http://host.robots.ox.ac.uk/pascal/VOC/voc2 012/index.html (accessed on 26 May 2020), reference number [27], and https://cocodataset.org/ #home (accessed on 26 May 2020), reference number [28]. | 5,811.4 | 2022-04-01T00:00:00.000 | [
"Computer Science"
] |
Study of thermal and electrical parameters of workpieces during spray coating by electrolytic plasma jet
In this paper the results are presented of thermal and electrical parameters of products in the system bottom layer - intermediate layer when applying protective coatings of ferromagnetic powder by plasma spray produced in an electric discharge with a liquid cathode, on steel samples. Temperature distribution and gradients in coating and intermediate coating were examined. Detailed descriptions of spray coating with ferromagnetic powder by plasma jet obtained in electrical discharge with liquid cathode and the apparatus for obtaining thereof is provided. Problem has been solved by using of Fourier analysis. Initial data for calculations is provided. Results of numerical analysis are provided as temporal functions of temperature in contiguity between coating and intermediate coating as well as temporal function of the value Q=q-φ; where q is density of heat current directed to the free surface of intermediate coating, φ is density of heat current in contiguity between coating and intermediate coating. The analysis of data given shows that in the systems of contact heat exchange bottom layer-intermediate layer with close values of the thermophysical characteristics of constituting materials is observed a slow increase of the temperature of the contact as a function of time.
Introduction
Durability and surface hardening of protective coating applied using plasma spray produced in the electric discharge containing a liquid cathode [1], and its adhesion to the bottom layer is largely dependent on the presence of residual stresses in the coating -bottom layer system. Residual tension in welded joints is extensively studied worldwide [2]. Residual tension has considerable influence on operational properties of machine parts subjected to cyclic mechanical and thermal action.
Nowadays spray coating is widely used in manufacturing of those machine parts [3]. Advantages of this process include quality and precision of resulting surface but this process also creates stretching residual tension. To reduce the level of residual stresses one may by applying an additional intermediate layer to the bottom layer [4]. Since the bottom layer and the intermediate layer have different thermal properties, then with the thermal effect on the system bottom layer -intermediate layer the thermal stresses arise in coating [5].
Therefore, in the development of the coating technology (bottom layer + intermediate layer) the information becomes important about the temperature distribution and temperature gradients in the bottom layer and the intermediate layer, and in contact area.
Statement of the problem
In this paper the results are presented of thermal and electrical parameters of products in the system bottom layer -intermediate layer when applying protective coatings of ferromagnetic powder by plasma spray (Figure 1) produced in an electric discharge with a liquid cathode, on steel samples [6 -8]. Ferromagnetic powder is obtained in the high-voltage electric discharge with voltage U = 1000 -1300 V, current I = 0.6 -10 A, between the anode of steel and liquid cathode at an interelectrode gap S = 5 -10 mm, at atmospheric pressure. We can assume that the distribution of temperature fields in this system is carried out by heat conduction as follows: ), 0 The challenge is to find the temperatures
Research method
In solving the problem (1) - (7) we use the Fourier method [9]. First, we find the solution of equation (1) satisfying the conditions of (3), (5), (6), temporally considering ) ( as known. This solution we obtain in the form are defined by following formulas: Similarly, we find the solution of equation (2) satisfying the conditions of (3), (5), (7), which has the form: (8), (9) we substitute in the condition (4). As a result, to determine the density ) ( we obtain the Volterra integral equation of the first kind [10] of the form where ) (x -theta function, defined by the formula The integral equation (10) we solve computationally [11]. Assume we obtain the recurrent equations (11) The analysis of data given shows that in the systems of contact heat exchange bottom layerintermediate layer with close values of the thermophysical characteristics of constituting materials is observed a slow increase of the temperature of the contact as a function of time t 1 . In such systems, heat flow density in the region of contact of the intermediate layer with the bottom layer (φ) differs slightly from the density of the flow (q) causing heating of the free surface of the intermediate layer, and with time φ and q tend to equalization (corresponding to falling curves Q=q-φ in Figure 3). The specified time laws t 1 (τ) and Q(τ) have caused by the condition that, in this case, the thickness of the bottom layer is much greater than the thickness of the intermediate layer.
Conclusion
Thus, in the process of applying protective coatings on the two-component facilities the solution of the problem of the regulated control of thermal parameters is directly dependent on the appropriate selection of components of a system in terms of optimizing the ratio between their sizes and thermal properties. | 1,164.8 | 2016-01-14T00:00:00.000 | [
"Engineering",
"Physics",
"Materials Science"
] |
A user-friendly plug-and-play cyclic olefin copolymer-based microfluidic chip for room-temperature, fixed-target serial crystallography
An improved design of an all-polymer microfluidic ‘chip’ for fixed-target serial crystallography is presented that can easily be fabricated in-house, is inexpensive and is highly modifiable to meet broad user needs for room-temperature serial crystallography at both synchrotron and XFEL light sources.
Introduction
X-ray crystallography has been the gold standard for protein structure determination.While data-collection limitations from large protein molecules are being eliminated by recent advances in cryoEM, X-ray crystallography techniques have been limited due to the need to grow large and well diffracting protein crystals.X-ray free-electron lasers (XFELs), light sources with a maxiumum brightness ten orders of magnitude higher than traditional synchrotron sources, offer a method to obtain protein structures from small and poorly diffracting crystal samples using the 'diffract before destroy' technique.Emerging serial crystallography techniques have correspondingly driven innovation in sample-delivery methods by creating the need to deliver up to millions of microcrystals at ambient temperature (Gru ¨nbein & Kovacs, 2019;Cheng, 2020).Several novel sample-delivery techniques have been developed to meet these requirements, with the most advanced being flow-based systems such as gas-focused dynamic virtual nozzle (GDVN) jets and high-viscosity extrusion, which are used to flow pre-loaded protein crystals past the X-ray beam interaction point (Echelmeier et al., 2019;Martiel et al., 2019).Alternatively, fixed-target systems, in which the crystals are fixed inside a sample holder which is rastered during data collection, have emerged as a viable alternative to flow-based target-delivery methods.Fixedtarget sample delivery offers some significant advantages compared with other methods, including improved ease of use, reduced sample consumption, the ability to screen different sample conditions rapidly with potentially higher hit rates, and adaptability for varying experimental conditions and samplegeometry requirements (Echelmeier et al., 2019;Ebrahim et al., 2019;Horrell et al., 2021).
Early fixed-target systems for XFEL serial crystallography were silicon-based chips with thin windows or pores that were fabricated using lithography or etching techniques and required crystal loading into specific window positions (Hunter et al., 2014;Martiel et al., 2021;Roedig et al., 2016).These silicon-based chips offer ultralow-background data collection but require costly and difficult manufacturing processing and long lead times.Samples must also be freshly loaded immediately before measurements and cannot be stored.Polymer-based fixed-target systems such as nylon or polyimide loops, various polymer meshes and other fully polymer-based chips have been developed as an alternative (Feld et al., 2015;Park et al., 2020;Lee et al., 2019Lee et al., , 2020)).Although amorphous scattering from the polymer material and sample buffer increases the background scattering (Sui & Perry, 2017), polymer-based fixed-target systems are more accessible due to their low cost, ease of use and flexibility in geometry, fabrication and sample handling, including on-chip crystallization and sample storage (Huang et al., 2020;Gicquel et al., 2018;Gilbile et al., 2021;Lyubimov et al., 2015;Doak et al., 2018).However, few designs can achieve ease of fabrication, enable in situ crystallization and slurry loading, and provide long-term stability in the same system.
Here, we describe a user-friendly, inexpensive, polymerbased microfluidic fixed-target system for protein crystal sample delivery for both serial femtosecond crystallography (SFX) and synchrotron serial crystallography (SSX) that dramatically improves on the design described in Gilbile et al. (2021).The original polymer chip was a robust, easy-to-use, low-background, fixed-target system and, importantly, was capable of in situ crystallization, with easy sample monitoring and stable, long-term storage.However, the fabrication process relied on lithography techniques, limiting its ease of production.As a result, changing the dimensions of the chip was very costly and time-consuming, limiting its adaptability and adoption by the broad structural biology community.Furthermore, the supporting framework limited the sample X-ray imaging area by 70%, obscuring many of the protein crystals in the sample layer.This work presents an improved design and methodology to address these problems, with a freestanding window area as large as 3.5 � �30 mm and minimal sample dead space.Additionally, the entire chip can be realistically designed, fabricated, assembled and loaded within a single day using widely accessible equipment, and is also compatible with standard magnetic bases for goniometer mounting.The simplification of the design enables rapid modifiability, including changes in chip dimensions, imaging area shape, loading techniques and even material, making it possible to tailor the new fixed-target chip for various beamline requirements, experiments or protein crystal types.
Chip fabrication and assembly
A schematic of the four different layers and assembly of the microfluidic chip is shown in Fig. 1(a).The enclosed X-ray imaging areas are made of thin-film COC (layer 3) prepared by spin-coating solutions of COC onto silicon wafers.The solutions were prepared by dissolving 15 wt% COC in secbutylbenzene at 120 � C overnight or until fully dissolved.Varying film thicknesses from 500 nm to 5 mm are possible depending on the spin speed and solution concentration.Fig. 2 shows spin curves for 6017 COC.While thicker films provide better support and lower water permeability, users should select the smallest possible film thickness to ensure a low background.To improve the delamination of the COC thin film from the silicon wafer, a water-soluble sacrificial layer of 9 wt% PVA in Milli-Q water was first spun onto a clean, UVozone-treated silicon wafer before COC film deposition.A 15 min UV-ozone treatment was used to improve the surface wettability of the silicon wafer.The PVA sacrificial layer was baked on a hotplate at 120 � C for a few minutes to fully evaporate residual water.Afterwards, warm COC solution (>80 � C) was spun on top of the dried PVA layer at 1000 rev min À 1 for 60 s.
Supporting frames (layer 2) were constructed from 1 mm PMMA with polyester adhesive transfer tape applied on one side of the PMMA sheet before cutting.The PMMA with polyester laminate was CO 2 laser-cut; the cutout areas define the sample-imaging window.A CO 2 laser was also used to cut the 48 mm double-sided acrylic adhesive spacer (layer 4).For a single flow channel with 2.7 mm 6017 COC film and 48 mm spacer, we found that a width of 3.5 mm is close to the maximum, with larger widths sometimes resulting in thin-film contacts across the air gap before sample loading.These dimensions are widely compatible with most crystallization solution viscosities and crystal slurries with a largest dimension of less than 30 mm.For different beamline needs, the dimensions of both the overall chip size and imaging windows are easily customizable.For an improved signal-to-noise ratio, the thickness of the sample flow layer should be matched to the crystal size to minimize the background from the crystallization solution.Adhesive spacers from 25 to 150 mm are commercially available.
The PMMA frame (layer 2) was placed onto the COC film (layer 3) with the polyester adhesive side down on the spuncoated COC film to assemble the chip.By pressing firmly on all parts of the PMMA frame, strong adhesion is formed between the supporting frame and the COC film.Multiple frames can be pressed on a single wafer.Afterwards, the adhered assembly was soaked in Milli-Q water until the frames with attached COC film were delaminated from the wafer.This process takes up to 36 h, but can be accelerated by using a razor blade to cut around the outer edges of the PMMA frame before soaking.The delaminated chip halves were then rinsed with Milli-Q water and dried using a gentle stream of N 2 gas.Only one of the PMMA frames has inlet/ outlet holes (Fig. 1a, top versus bottom).An adhesive spacer (layer 4) was aligned and placed onto one of the framesupported COC windows to complete the construction.A second completed half (layers 2 and 3) was then placed on the spacer, forming the chip (Fig. 1b).Detailed fabrication steps and vector files for laser cutting can be found in the supporting information.
Sample loading
The chips are compatible with direct crystal slurry loading and in situ, on-chip crystallization (micro-batch or vapordiffusion conditions) by directly pipetting either crystal slurry or protein/precipitant solution into an inlet hole (Fig. 1b).Sample solution flows into the imaging channels through capillary action (Sui et al., 2021).Once samples are loaded, Crystal Clear tape (layer 1) can be wrapped around the chip to reduce water loss during batch crystallization to improve sample stability or it can be left unwrapped for vapor-diffusion conditions.In both cases, in situ crystallization times are typically longer due to the high aspect ratio of the sample volume and the limited area for evaporation.
Lysozyme and thaumatin were used as model proteins to demonstrate chip performance.Commercially available lyophilized samples of each protein were dissolved in Milli-Q water to produce protein solutions of 50 mg ml À 1 lysozyme and 25 mg ml À 1 thaumatin.Stock solutions of 2 M NaCl with 0.1 M sodium acetate buffer pH 4.6 and 1 M l-sodium potassium tartrate with 0.1 M ADA buffer pH 6-6.5 were used as the precipitant solutions for lysozyme and thaumatin, respectively.For lysozyme, we also explored the use of hydrogels during crystallization to further demonstrate the utility of the chip for different crystallization conditions.In this case, 2 wt% low-melting-point agarose (purchased from Hampton Research) was heated to 85 � C, cooled to 40 � C and then added to the precipitant solution and mixed with the protein solution.The final mixture with 0.3 wt% agarose was pipetted above 30 � C into an inlet of the chip until the channel was filled.The total sample volume in each lane is 4 ml.For synchrotron measurements, larger protein crystals were desired and were crystallized in situ under micro-batch conditions and typically sealed with Crystal Clear tape to improve hydration stability.Supplementary Fig. S1 shows optical microscopy images of lysozyme crystals grown in situ inside a chip that was kept under ambient conditions.Over ten days, no significant dehydration was observed under ambient conditions.The crystals showed no directional preference when grown in situ.
Protein diffraction measurements
Two model proteins, lysozyme and thaumatin, were crystallized on a chip and measured at ambient temperature on beamline 12-1 at the Stanford Synchrotron Radiation Lightsource (SSRL).A screw-tightened, slotted holder with a magnetic base was used to hold the PMMA frame portion of the chip during diffraction experiments (Fig. 3b).The magnetic base was mounted onto the goniometer at the beamline.Diffraction data were collected at a wavelength of 0.9794 A ˚with a beam size of 0.05 � 0.04 mm using an EIGER X 16M detector (Dectris AG) at a detector distance of 0.2 m.The beamline sample-holder translational motors were used to align and center individual single crystals to the beam path, using inline high-resolution cameras to identify each crystal.Data sets were collected from these centered single crystals over 40 � wedges.Diffraction data from 15 and 14 individual crystals (50 mm in diameter on average) were merged to give complete data sets for lysozyme and thaumatin, respectively.An exemplar diffraction pattern from thaumatin using the improved chip is shown in Fig. 3(a).
Lysozyme crystals were grown in situ and measured on the Macromolecular Femtosecond Crystallography (MFX) beamline at the Linac Coherent Light Source (LCLS).A different crystallization protocol was applied to produce smaller (10 mm on average) dispersed lysozyme crystals for comparative XFEL measurements (50 mg ml À 1 lysozyme with a 1:1 ratio of protein solution to mother liquor: 1 M NaCl, 0.1 M sodium acetate pH 4.6).XFEL diffraction data were collected at ambient temperature in a helium-rich ambient (HERA) environment to reduce background from air scattering.A 1 00 � 1 00 chip with four 3.5 � 18 mm channels was fabricated to match the MFX sample-stage displacement range.A 3D-printed chip holder was made to attach the chip to the sample stage (Fig. 3c).Diffraction data were collected at 1.253 A ˚and 11% transmission using the SLAC ePix10k2M detector.The beamline sample holder translational motors were used to align and continuously raster the chip at 120 Hz.Variable shot spacings between 25 and 200 mm were possible at 120 Hz depending on the linear translational motor speed, and data collection was primarily at 50 mm displacements.The Figure 3 (a) X-ray diffraction pattern obtained from a thaumatin crystal at 10% transmission at SSRL.A 1 00 � 1 00 chip design with four independent volumes is mounted on (b) beamline 12-1 at SSRL and (c) the MFX beamline at LCLS (in a red 3D chip holder).
inline camera resolution was granular, so larger raster scans were used to optimize the data-collection efficiency to ensure that the sample window area was fully imaged.
Improved chip fabrication and assembly
As detailed in Table 1, the new chip generation (Chip 2.0) provides significant advancements in manufacturing time and cost, along with a larger continuous imaging area and reduced sample volume:area ratio.More shots per chip are possible by having a large, continuous window area, maximizing sampleto-data efficiency.The ease of fabrication and improvements in turnaround times indicate that chips can be mass fabricated and quickly adjusted for specific beamline requirements.Users can easily make their own chips based on their specific needs, requiring only a spin coater and a CO 2 laser cutter.The ability to optimize chips to match specific crystal types and measurement requirements enhances the chip versatility and its applicability for a range of experiment types.
Microfluidic chip background scattering
Fig. 4 presents the background radial scattering from two chips with different COC window thicknesses and a 48 mm spacer on beamline 12-1 at SSRL. Background scattering was measured for empty and buffer-filled chips.The majority of the background scatter from the chip was from the aqueous buffer, with a small diffuse peak from the COC films at approximately q = 1.2A ˚À 1 , consistent with previous studies (Martiel et al., 2021;Ren et al., 2018;Gilbile et al., 2021).For the chip demonstration measurements, the same channel geometries were used for consistency.The smaller crystals used at XFELs therefore have a relatively thicker layer of buffer surrounding them, resulting in a higher background.The chip can be further optimized and tailored to different beamline needs.Users are encouraged to choose the smallest channel thickness for an optimized signal-to-noise ratio.Conversely, higher intensity experiments can be paired with a slightly thicker window and more support structures.
Comparing the background scattering from the two different COC window thicknesses, a decrease in the COC thickness from 2.7 to 1.7 mm decreased the peak COC scattering contribution by approximately a factor of four.Overall, using a dramatically reduced COC film thickness (1-5 mm) contributes much less background scatter compared with similar systems based on thick (600-700 mm) COC sheets (Pinker et al., 2013;Vasireddi et al., 2022).
Synchrotron measurements.
High-resolution diffraction data sets were collected at room temperature.For lysozyme the structure was refined to R work and R free values of 0.185 and 0.212, respectively, at a resolution of 1.45 A ˚. Furthermore, the structure of thaumatin was refined to R work and R free values of 0.139 and 0.153, respectively, at a resolution of 1.48 A ˚.The thaumatin structure resolution achieved is comparable to those from other fabricated and commercially available fixed-target systems grown and loaded under similar conditions at synchrotron sources (PDB entries 3zej, 5a47 and 6xbx; Pinker et al., 2013;Zander et al., 2015;Gavira et al., 2020).The lysozyme and thaumatin synchrotron data have been deposited in the PDB (as PDB entries 8scy and 8fzw).Compared with the previous generation of chips (Gilbile et al., 2021), the diffraction resolution for lysozyme was improved in 0.3 wt% agarose solution.
One point to emphasize is that a single chip was sufficient for data collection, requiring only 0.6 and 0.3 mg of lysozyme Radial X-ray scattering background with air scattering subtracted measured on beamline 12-1 at SSRL.The water solvent ring at q = �1.8A ˚À 1 is evident in the filled chip (black data points).While the background scatter from the COC windows at q = �1.2A ˚À 1 is effectively controllable based on the film thickness (green versus red data points), it may exhibit minor fluctuations (green versus black data points) due to small variations in the film thickness across the film and among different chips.
and thaumatin, respectively.In addition to lysozyme and thaumatin, diffraction data from slurry-loaded, wild-type Nsp15 endoribonuclease (NendoU; Jernigan et al., 2023) was collected using the chip.These measurements demonstrate that a range of samples and conditions can be efficiently run during a single measurement period.Additionally, the lysozyme and thaumatin samples were crystallized in situ days in advance.We have maintained samples on chips for weeks with no observable decrease in sample quality.Optical microscopy can be used to pre-select optimal chips and sections of chips before beamtime, significantly increasing operational efficiency and dramatically reducing user stress during beamtime.
XFEL measurements.
To better represent XFEL samples, conditions that encouraged nucleation over crystal growth using in situ crystallization were selected.Microscopy images of the two chips used for data collection are shown in Fig. 5(a).Diffraction from small, dispersed lysozyme crystals of about 10-15 mm in the largest dimension with randomized orientations was obtained.An exemplar diffraction pattern is shown in Fig. 6 without any image processing.The merging statistics obtained from the data sets collected are shown in Table 2.For these first XFEL demonstration measurements, the transmission was 11% and the full repetition rate of 120 Hz was used.The inline camera image quality was relatively poor.Instead of carefully aligning the imaging X-ray window for the raster scan, the chips were crudely aligned and rastered over a wider area to ensure that the entire sample window was measured.Because the chips are entirely polymer and sample, there is no issue with rastering the frame or any region of the chip through the X-ray beam, eliminating the need for careful positioning or precision rastering.X-ray shots on the thick PMMA supporting frames have a much higher mean detector intensity and are easily excluded with an intensity cutoff when analyzing the data, as shown in Fig. 7.In this case, we ran a partially filled lysozyme chip as well to provide a rough estimate of the background of the chip relative to the sample.
Using a wide scan area, data collection from two chips took about 1 h, yielding 47 948 hits, from which 29 215 lattices were indexed.Two imaging windows were re-run to exhaust the remaining crystals, and the total indexed hit rate was 13.5%.Data were merged and refined to R work and R free values of 0.2512 and 0.2814, respectively, at a resolution of 1.70 A ˚. Fig. 5(b) shows an image of one of the chips after X-ray rastering.No damage to or significant dehydration of the microfluidic channel was observed, even after multiple scans.The chip was imaged five days after XFEL measurements, demonstrating the robustness of the chip and stability against dehydration.With the higher magnification available at this time, it was possible to detect patterns of tiny vapor bubbles with a 50 mm pitch in some regions, presumably adhered to the hydrophobic COC window film, as well as some vapor bubbles that diffused.Notably, the COC film and chip were still fully intact.
An additional chip with different protein screening samples was run at a range of transmissions.No decrease in hit rates was observed with up to 63% transmission.However, at 100% transmission diffraction from salt crystals was observed, which progressively increased over time.Dehydration was visible when the chip was removed from the beamline.Fig. 8 shows an optical image of a chip during the transition from 63% to 100% transmission.The 50 mm shot-spacing pattern is clearly evident, and a dramatic pattern in the COC film is observed at 100% transmission.Subsequent imaging at four times higher magnification did not detect perforations in the COC film.However, the film must have minor defects along the striated pattern, resulting in sample dehydration during measurement.Future work will carefully study the chips at high transmission to optimize for these conditions.In particular, the use of a 25 mm adhesive layer thickness will dramatically decrease the buffer background and water adsorption, while a highhumidity helium chamber may allow full operation without any changes by preventing sample dehydration.
Conclusions
Highly efficient, low-sample-consumption and easy-to-use sample-delivery methods are crucial to maximize the potential of SFX and SSX techniques for serial crystallography.This paper presents a methodology for fabricating an inexpensive, low-background and highly versatile design for fixed-target SFX and SSX sample delivery without needing lithography or etching steps.Since chip fabrication requires only a spin coater and a CO 2 laser cutter, our approach makes customizable fixed-target devices available to a broader community.The chips are compatible with different sample-loading modalities, including crystal slurry loading, micro-batch crystallization and vapor-diffusion crystallization.Diffraction data from two model proteins crystallized on-chip demonstrate that highquality diffraction data can be obtained at ambient temperature using our device at both synchrotron and XFEL light sources.In addition, the ability to rapidly reconfigure the chip Chip after 63% (top) and 100% (bottom) transmission.The 50 mm shot spacing is clearly visible in both cases, but no change in hit rate or dehydration was found at 63% transmission.At 100% transmission dehydration and diffraction from salt crystals were found, demonstrating that the X-ray shots likely led to cracks in the COC film or pinhole defects.
Figure 7
Mean detector intensity plot per XFEL shot while rastering a half-filled chip.The gray areas are where PMMA frames are scanned, the yellow areas are samples (buffer and diffracting crystals) and the blank areas are polymer chips with air only.X-ray shots that hit the PMMA frame are easily excluded from analysis using a mean intensity cutoff.The intensity from the samples was roughly three times the blank air background.
geometry and dimensions allows the user to customize the chip to match their specific sample, for example by (i) tailoring the spacer thickness to match the crystal size and reduce background scattering from the buffer, (ii) changing the film thicknesses or the grade of COC for different water-loss profiles, (iii) changing the imaging window dimensions to match sample-volume limitations and (iv) altering the overall chip size and shape for unique beamline fixed-target mounting configurations.
Compared with the previous-generation device reported in Gilbile et al. (2021), the new chip design and improved fabrication method offer comparable in situ crystallization conditions, crystal slurry loading and sample stability, while providing a dramatic improvement in the ease of fabrication, an increase in X-ray imaging window size with a concomitant decrease in dead volume, and rapid modifiability.Looking forward, many alterations/additions are currently being developed, including (i) removing excess buffer before data collection to reduce background contributions, (ii) controlling crystal nucleation density using electric fields (Alexander & Radacsi, 2019), (iii) controlling crystal nucleation locations using polymer brushes on the COC imaging window, (iv) design modifications for use with membrane proteins in the highly viscous lipidic cubic phase and (v) maintaining sample stability for measurements under vacuum.
Figure 2
Figure 2Spin curves showing COC film thickness as a function of spin speeds for different concentrations of COC 6017 dissolved in sec-butylbenzene.
Figure 1 (
Figure 1 (a) A schematic of the improved chip construction layers.Note that 'top' frames (layer 2) have inlet holes while 'bottom' frames do not.(b) Image of the assembled chip with sample X-ray imaging areas filled with colored solution.The dashed red line demarks the active X-ray imaging area.The sample is loaded by manual pipetting into one of the inlet/outlet holes.Each sample-imaging window is independent.
Figure 5 (
Figure5(a) Representative images of lysozyme crystals and the variation in crystal density in the two chips before XFEL measurements (the scale bar applies to both parts of the figure).(b) Lysozyme crystals five days after XFEL measurements at 50 mm shot spacing, 11% transmission and 0.9-1.0mJ pulse energy.Small vapor bubbles of a few micrometres in size were observed in some areas of the microfluidic channel where the beam was rastered, as highlighted by the dashed circles and the inset.
Figure 6 X
Figure 6 X-ray diffraction pattern obtained from a lysozyme crystal at 11% transmission on the MFX beamline at LCLS.Raw figure without any background subtraction.
Table 1
Advancement in the new generation, Chip 2.0, compared with the previous generation, Chip 1.0.
Table 2
Crystallographic statistics obtained for lysozyme and thaumatin.Values in parentheses are for the highest resolution shell. | 5,517.4 | 2023-09-25T00:00:00.000 | [
"Materials Science"
] |
Paraphrasing vs Coreferring: Two Sides of the Same Coin
We study the potential synergy between two different NLP tasks, both confronting predicate lexical variability: identifying predicate paraphrases, and event coreference resolution. First, we used annotations from an event coreference dataset as distant supervision to re-score heuristically-extracted predicate paraphrases. The new scoring gained more than 18 points in average precision upon their ranking by the original scoring method. Then, we used the same re-ranking features as additional inputs to a state-of-the-art event coreference resolution model, which yielded modest but consistent improvements to the model’s performance. The results suggest a promising direction to leverage data and models for each of the tasks to the benefit of the other.
Introduction
Recognizing that mentions of different lexical predicates discuss the same event is challenging (Barhom et al., 2019). Lexical resources such as WordNet (Miller, 1995) capture such synonyms (say, tell) and hypernyms (whisper, talk), as well as antonyms, which can be used to refer to the same event when the arguments are reversed ([a] 0 beat [a] 1 , [a] 1 lose to [a] 0 ). However, WordNet's coverage is insufficient, in particular, missing contextspecific paraphrases (e.g. (hide, launder), in the context of money). Conversely, distributional methods enjoy broader coverage, but their precision for this purpose is limited because distributionally similar terms may often be mutually-exclusive (born, die) or may refer to different event types which are only temporally or causally related (sentenced, convicted).
Two prominent lines of work pertaining to identifying predicates whose meaning or referents can be matched are cross-document (CD) event coreference resolution and recognizing predicate para-Tara Reid has checked into∨ Promises Treatment Center. Actress Tara Reid entered∨ well-known Malibu rehab center. Lindsay Lohan checked into × rehab in Malibu, California.
Director Chris Weitz is expected to direct∨ New Moon. Chris Weitz will take on∨ the sequel to "Twilight". Gary Ross is still in negotiations to direct × the sequel. Table 1: Examples from ECB+ (a cross-document coreference dataset) that illustrate the context-sensitive nature of event coreference. The illustrated predicates are co-referable, and hence may be used to refer to the same event in certain contexts, but obviously not all their mentions corefer.
phrases. The former identifies and clusters event mentions, across multiple documents, that refer to the same event within their respective contexts. The latter task, on the other hand, collects pairs of event expressions that, at the generic lexical level, may refer to the same event in certain contexts. Table 1 illustrates this difference with examples of co-referable predicate paraphrases, while their mentions obviously do not always co-refer.
Cross-document event coreference resolution systems are typically supervised, usually trained on the ECB+ dataset, which contains clusters of news articles on different topics (Cybulska and Vossen, 2014). Recent systems rely on neural representations of the mentions and their contexts (Kenyon-Dean et al., 2018;Barhom et al., 2019), while earlier approaches leveraged WordNet and other lexical resources to obtain a signal of whether a pair of mentions may be coreferring (e.g. Bejan and Harabagiu, 2010;Yang et al., 2015).
Approaches for acquiring predicate paraphrase, in the form of a pair of paraphrastic predicates or predicate templates, were based mostly on unsupervised signals. These included similarity between argument distributions (Lin and Pantel, 2001;Berant, 2012), backtranslation across languages (Barzilay and McKeown, 2001;Ganitkevitch et al., 2013;Mallinson et al., 2017), or leveraging redundant news reports on the same event, which are hence likely to refer to the same events and entities using different words (Shinyama et al., 2002;Shinyama and Sekine, 2006;Barzilay and Lee, 2003;Zhang and Weld, 2013;Xu et al., 2014;Shwartz et al., 2017). In some cases, the paraphrase collection phase includes a step of validating a subset of the paraphrases and training a model on these gold paraphrases to re-rank the entire resource (Lan et al., 2017).
In this paper, we study the potential synergy between predicate paraphrases and event coreference resolution. We show that the data and models for one task can benefit the other. In one direction (Section 3), we use event coreference annotations from the ECB+ dataset as distant supervision to learn an improved scoring of predicate paraphrases in the unsupervised Chirps resource (Shwartz et al., 2017). The distantly supervised scorer significantly improves upon ranking by the original Chirps scores, adding 18 points to average precision over a test sample.
In the other direction (Section 4), we incorporate data from Chirps, represented in the Chirps re-scorer feature vector, into a state-of-the-art event coreference system (Barhom et al., 2019). Chirps has a substantial coverage over the ECB+ coreferring mention pairs, and consequently, the incorporation yields a modest but consistent improvement across the various coreference metrics. 1
Background and Motivation
In this section we provide some background about the cross-document coreference resolution and paraphrase identification (acquisition) tasks, which is relevant for our approaches for synergizing these two tasks.
Event Coreference Resolution
Event coreference resolution aims to identify and cluster event mentions, that, within their respective contexts, refer to the same event. The task has two variants, one in which coreferring mentions are within the same document (within document) and another in which corefering mentions may be in different documents (cross-document, CD), on which we focus in this paper.
The standard datasets used for CD event coreference training and evaluation are ECB+ (Cybulska and Vossen, 2014), and its predecessors, EECB (Lee et al., 2012) and ECB (Bejan and Harabagiu, 2010). ECB+ contains a set of topics, each containing a set of documents describing the same global event. Both event and entity coreferences are annotated in ECB+, within and across documents.
The current state-of-the-art model from Barhom et al. (2019) iteratively and intermittently learns to cluster events and entities. A mention representation m i consists of several components, representing both the mention span and its surrounding context. The interdependence between clustering event vs. entity mentions is encoded into the mention representation, such that an event mention representation contains a component reflecting the current entity clustering, and vice versa. Using this representation, the model trains a pairwise mention scoring function that predicts the probability that two mentions refer to the same event.
Paraphrase Identification and acquisition
Paraphrases are differing textual realizations of the same meaning (Ganitkevitch et al., 2013), typically phrases or sentences (Dolan et al., 2005). A prominent approach for identifying and collecting paraphrases, backtranslation, assumes that if two (say) English phrases translate to the same term in a foreign language, across multiple foreign languages, this indicates that these two phrases are paraphrases. This approach was first suggested by Barzilay and McKeown (2001), later adapted to acquire the large PPDB resource (Ganitkevitch et al., 2013), and was also shown to work well with neural machine translation (Mallinson et al., 2017).
Paraphrase Identification through Event Coreference. An alternative approach for paraphrase identification, on which we focus in this paper, leverages multiple news documents discussing the same event. The underlying assumption is that such redundant texts may refer to the same entities or events using lexically-divergent mentions. Coreferring mentions are identified heuristically and extracted as candidate paraphrases. When long documents are used, the first step in this approach is to align each pair of documents by sentences. This was done by finding sentences with shared named entities (Shinyama et al., 2002) or lexical overlap (Barzilay and Lee, 2003;Shinyama and Sekine, 2006), and by aligning pairs of predicates or arguments (Zhang and Weld, 2013;Recasens et al., 2013). In more recent work, Xu et al. (2014) and Lan et al. (2017) extracted sentential paraphrases from Twitter by heuristically matching pairs of tweets discussing the same topic.
Predicate Paraphrases. In contrast to sentential paraphrases, it is also beneficial to identify differing textual templates of the same meaning. In this paper we focus on binary predicate paraphrases such as ("[a 0 ] quit from [a 1 ]", "[a 0 ] resign from [a 1 ]").
Earlier approaches for acquiring predicate paraphrases considered a pair of predicate templates as paraphrases if the distributions of their argument instantiations were similar. For instance, in "[a 0 ] quit from [a 1 ]", [a 0 ] would typically be instantiated by people names while [a 1 ] by employer organizations or job titles. A paraphrastic template like "[a 0 ] resign from [a 1 ]" is hence expected to have similar argument distributions, and can thus be detected by a distributional similarity approach (Lin and Pantel, 2001;Szpektor et al., 2004;Berant, 2012). Yet, as mentioned earlier, predicates with similar argument distributions are not necessarily paraphrastic, which introduces a substantial level of noise when acquiring paraphrase pairs using this approach.
In this paper, we follow the potentially more reliable paraphrase acquisition approach, which tries to heuristically identify concrete co-referring predicate mentions. Identifying such mention pairs, detected as actually being used to refer to the same event, can provide a strong signal for identifying these predicates as paraphrastic (vs. the quite noisy corpus-level signal of distributional similarity). In particular, we utilize the Chirps paraphrase acquisition method and resource, which follows this approach as described next in some detail.
Chirps: a Coreference-Driven Paraphrase Resource. Chirps (Shwartz et al., 2017) is a resource of predicate paraphrases extracted heuristically from Twitter. Chirps aims to recognize coreferring events by relying on the redundancy of news headlines posted on Twitter on the same day. It extracts binary predicate-argument tuples from each tweet and aligns pairs of predicate mentions whose arguments match, by some lexical matching criteria. The matched pairs of arguments are termed supporting pairs, e.g. ( . This score is proportional to the number of supporting pair instances in which the two templates were paired (n), as well as the number of different days in which such pairings were found (d), where N is the number of days the resource is collected. The Chirps resource provides the scored predicate paraphrases as well as the supporting pairs for each paraphrase.
Chirps has acquired more than 5 million distinct paraphrase pairs over the last 3 years. Human evaluation showed that this scoring is effective and that the percentage of correct paraphrases is higher for highly scored paraphrases. At the same time, due to the heuristic collection and scoring of predicate paraphrases in Chirps, entries in the resource may suffer from two types of errors: (1) type 1 error, i.e., the heuristic recognized pairs of non-paraphrastic predicates as paraphrases. This happens when the same arguments participate in multiple different events, as in the following paraphrases: "[Police] 0 arrest [man] 1 " and "[Police] 0 shoot [man] 1 "; and (2) type 2 error, when the scoring function assigned a low score to a rare but correct paraphrase pair, as in "[a 0 ] outgun [a 1 ]" and "[a 0 ] outperform [a 1 ]", for which only a single supporting pair was found.
Chirps*: Leveraging Coreference Information for Paraphrasing
Our goal in this section is to improve paraphrase scoring, in the context of Chirps, while leveraging available information and methods for event cross-document coreference resolution. To that end, we introduce Chirps*, a new supervised scorer for Chirps candidate paraphrases, whose novelties are two fold. First, we extract a richer feature representation for a candidate paraphrase pair (Section 3.1), which is fed into a supervised classifier for the candidates. Second, we collect, semi-automatically, distantly supervised training data for paraphrase classification, which is derived from the ECB+ cross-document coreference training set, leveraging the close relationship between the two tasks (Section 3.2). Finally, we provide some implementation details (Section 3.3).
Features
As described above, the original heuristic Chirps scorer relied only on a couple of features to score a candidate paraphrase pair. Our goal is to obtain a richer signal about the likelihood of a candidate predicate pair to indeed be paraphrastic. To that end, we collect a set of features from the available data, with a focus on assessing whether the instances from which the candidate pair was extracted indeed constitute cross-document coreferences.
Each candidate paraphrase pair consists of two predicate templates p 1 and p 2 , accompanied by the n supporting pair instances for the pair, each consisting a pair of argument terms, associated with this predicate paraphrase pair: support-pairs(p 1 , p 2 ) = {(t 1 1 , t 1 2 ), ..., (t n 1 , t n 2 )}. Each tweet included in Chirps links to a news article, whose content we retrieve. When representing a pair of predicate templates, we include both local features (based on a single supporting pair) and global features (based on all supporting pairs). Table 2 presents our 17 features, yielding a feature representation f p 1 ,p 2 ∈ R 17 for a paraphrase pair, grouped by different sources of information. The first group includes features derived from the statistics provided by the original Chirps resource. The other 4 sources of information are described in the following paragraphs.
Named Entity Coverage While the original Chirps method did not utilize the content of the linked article, we find it useful to retrieve more information about the event. Specifically, it might help mitigating errors in Chirps' argument matching mechanism, which relies on argument alignment considering only the text of the two tweets. We found that the original mechanism worked particularly well for named entities while being more error-prone for common nouns, which might require additional context.
Given (t i 1 , t i 2 ) ∈ support-pairs(p 1 , p 2 ), we use SpaCy (Honnibal and Montani, 2017) to extract sets of named entities, N E 1 and N E 2 , from the first paragraph of the news article linked from each tweet, respectively. We define a Named Entity Coverage score, NEC, as the maximum ratio of named entity coverage of one article by the other: We manually annotated a small balanced training set of 121 tweet pairs and used it to tune a score threshold T = 0.26, such that pairs of tweets whose NEC is at least T are considered coreferring. Finally, we include the following features: the number of coreferring tweet pairs (whose NEC score exceeds T ) and the average NEC score of these pairs.
Cross-document Coreference Resolution
We apply the state-of-the-art cross-document coreference model from Barhom et al. (2019) to data constructed such that each tweet constitutes a document and each pair of tweets corresponding to t j 1 and t j 2 in support-pairs(p 1 , p 2 ) forms a topic, to be analyzed for coreference. As input for the model, in each tweet, we mark the corresponding predicate span as an event mention and the two argument spans as entity mentions. The model outputs whether the two event mentions corefer (yielding a single event coreference cluster for the two mentions) or not (yielding two singleton clusters). Similarly, it clusters the four arguments to entity coreference clusters.
Differently from Chirps, this model makes its event clustering decision based on the predicate, arguments, and the context of the full tweet, as opposed to considering the arguments alone. Thus, we expect it not to cluster predicates whose arguments match lexically, if their contexts or predicates don't match (first example in Table 3). In addition, the model's mentions representation might help to identify lexically-divergent yet semantically-similar arguments (second example in Table 3).
For a given pair of tweets, we extract the following binary features with respect to the predicate mentions: Event Perfect when the predicates are assigned to the same cluster, and Event No Match when each predicate forms a singleton cluster. For argument mentions, we extract the following features: Entity Perfect if the two a 0 arguments belong to one cluster and the two a 1 arguments belong to another cluster; Entity Reverse if at least one of the a 0 arguments is clustered as coreferring with the a 1 argument in the other tweet; and Entity No
# Supporting pairs
The total number of support pairs of p1 and p2 across the template variants.
# Days
The total number of days d that p1 and p2 was matched in Chirps across the template variants.
# Available supporting pairs
The number of support pairs of p1 and p2 across the template variants that were still available to download.
# Days of available pairs
The total number of days d in which the support pairs above occurred in the available tweets.
Score
The maximal Chirps score across the template variants.
Named Entity Coverage
# NEC above threshold Number of pairs with NEC score of at least T .
Average above threshold
Average of NEC scores for pairs with a score of at least T .
# Event Perfect
Number of event pairs with perfect match.
# Event No Match
Number of event pairs with no match.
# Entity Perfect
Number of entity pairs with perfect match.
# Entity Reverse
Number of entity pairs with reverse match.
# Entity No Match
Number of entity pairs with no match.
# Perfectly Clustered + NEC
The number of pairs with NEC score of at least T and perfect clustering for event coreference resolution.
# Connected components
The number of connected components in Gp 1 ,p 2 .
Average component size
The average size of connected components in Gp 1 ,p 2 .
# In Clique
The number of pairs in support-pairs(p1, p2) that are in a clique. Match otherwise. Also, we extract Perfectly Clustered with NE Coverage that combines both Named Entity coverage and coreference-resolution, which count the number of pairs their events are perfectly clustered and with NEC score of at least T.
Connected Components
The original Chirps score of a predicate paraphrase pair is proportional to two parameters: (1) the number of supporting pairs; (2) the ratio of number of days in which supporting pairs were matched relative to the entire collection period. The latter lowers the score of paraphrase pairs which might have been mistakenly aligned on relatively few days (e.g. due to misleading argument alignments in particular events). The number of days in which the predicates were aligned is taken as a proxy for the number of different events in which the predicates co-refer. Here, we aim to get a more reliable partition of tweets to different events by constructing a graph of tweets as nodes, with supporting tweet pairs as edges, and looking for connected components.
To that end, we define a bipartite graph G p 1 ,p 2 = (V, E) for a candidate paraphrase pair, where V = tweets(p 1 , p 2 ) contains all the tweets in which p 1 or p 2 appeared, and E = support-pairs(p 1 , p 2 ). We compute C, the number of connected components in G p 1 ,p 2 , and define the following group: ConComp = {c ∈ C : |c| > 2}, which represents the number of connected components with size greater than 2. From this group we derive two features #connected(p 1 , p 2 ) = |ConComp| which represents the number of the connected components and avg connected(p 1 , p 2 ), which is the average size of the connected components in the graph. A larger number of connected components indicates that the two predicates were aligned across a large number of likely different events.
Clique We similarly build a global tweet graph for all the predicate pairs, G all = (V , E ), where V = ∪ (p 1 ,p 2 ) tweets(p 1 , p 2 ), and E = ∪ (p 1 ,p 2 ) support-pairs(p 1 , p 2 ). We compute Q, the set of cliques in G all of size greater than 2. We assume that a pair of tweets are more likely to be coreferring if they are part of a bigger clique, whereas if they were extracted by mistake they wouldn't share many neighbors. We extract the following feature of clique coverage for a candidate paraphrase pair: CLC(p 1 , p 2 ) = |{t j 1 , t j 2 ∈ support-pairs(p 1 , p 2 ) : ∃q ∈ Q such that t j 1 ∈ q ∧ t j 2 ∈ q}|.
Distantly Supervised Labels
In order to learn to score the paraphrases, we need gold standard labels, i.e., labels indicating whether a pair of predicate templates collected by Chirps is indeed a paraphrase. Instead of collecting manual annotations for a sample of the Chirps data, we chose a low-budget distant supervision approach.
To that end, we leverage the similarity between the predicate paraphrase extraction and the event coreference resolution tasks, and use the annotations from the ECB+ dataset. Our dataset consists of the predicate paraphrases from Chirps that appear in ECB+ (denoted ch-ECB+). As positive examples we consider all pairs of predicates p 1 , p 2 from Chirps that appear in the same event cluster in ECB+, e.g., from {talk, say, tell, accord to, confirm} we extract (talk, say), (talk, tell), ..., (accord to, confirm).
Obtaining negative examples is a bit trickier. We consider as negative example candidates pairs of predicates p 1 , p 2 from Chirps, which are under the same topic, but in different event clusters in ECB+, e.g., given the clusters {specify, reveal, say}, and {get}, we extract (specify, get), (reveal, get), and (say, get).
Note that the ECB+ annotations are contextdependent. Thus a pair of predicates that are in principle coreferable may be annotated as non- (2017), we annotated the templates while presenting 3 argument instantiations from their original tweets. Thus, we only included in the final data predicate pairs with at least 3 supporting pairs. We required that workers have 99% approval rate on at least 1,000 prior tasks and pass a qualification test. Each example was annotated by 3 workers. We aggregated the per-instantiation annotations using majority vote and considered a pair as positive if at least one instantiation was judged as positive. The data statistics are given in Table 4. This validation phase balanced the positive-negative proportion of instances in the data, from approximately 1:7 to approximately 4:5.
Model
We trained a random forest classifier (Breiman, 2001) implemented by the scikit-learn framework (Pedregosa et al., 2011). To tune the hyperparameters, we ran a 3 fold cross-validation randomized search, yielding the following values: 157 estimators, max depth of 8, minimum samples leaf of 1, and min samples split of 10. 2
Evaluation
We used the model for two purposes: (1) classification: determining if a pair of predicate templates are paraphrases or not; and (2) ranking the pairs based on the predicted positive class score. We consider the ranking evaluation as more informative, as we expect the ranking to reflect the number of contexts in which a pair of predicates may be coreferring. That is, predicate pairs that are coreferring in many contexts will be ranked higher than those that are coreferring in just a few contexts. We compare our model with two baselines: the original Chirps scores, and a baseline that assigns each pair of predicates the cosine similarity scores between the predicates using GloVe embeddings (Pennington et al., 2014). 3 For the classification decisions made by the two baseline scores (Chirps score and cosine similarity for Glove vectors), we learn a threshold that yields the best accuracy score over the train set, above which a pair of predicates is classified as positive. Table 5 displays the accuracy, precision, recall and F 1 scores for classification evaluation and the Average Precision (AP) for ranking evaluation. Our scorer dramatically improves upon the baselines in all metrics.
To show that the improved scoring generalizes beyond examples that appear in the ECB+ dataset, we selected a random subset of 500 predicate pairs with at least 6 support pairs from the entire Chirps resource and annotated them in the same method described in Section 3.2. The ranker evaluated on this subset gained 8 points in AP, relative to the original Chirps ranking. All results are statistically significant using bootstrap and permutation tests with p < 0.001 (Dror et al., 2018). Table 6 exemplifies highly ranked predicate pairs by our Chirps* scorer, the original Chirps scorer and the GloVe scorer, which illustrates the improved ranking performance of Chirps* (as measured in table 5 by the AP score).
Ablation Test To evaluate the importance of each type of feature, we perform an ablation test. Table 7 displays the performance of various ablated models, each of which with one set of features (Section 3.1) removed from the representation. In the classification task, removing the named entity coverage features somewhat improved the performance, mostly by increasing the recall. However, in terms of the (primary) ranking evaluation, each set of features contributed to the performance, with the full model performing best.
Leveraging a Paraphrasing Resource to Improve Coreference
In Section 3 we showed that leveraging CD event coreference annotations and model improves predicate paraphrase ranking. In this section, we show that this co-dependence can be used in both directions, and that using Chirps* as an external resource can improve the performance of a CD model. As a preliminary analysis, we computed Chirps' coverage of lexically-divergent pairs of co-referring event mentions in ECB+. We found approximately 30% coverage overall and above 50% coverage for coreferring verbal mentions. 4 This indicates a substantial coverage of the lexically-divergent positive coreferrability decisions that need to be made in ECB+. In absolute numbers, Chirps covers
Integration Method
The state-of-the-art CD coreference resolution model, by Barhom et al. (2019), trained a pairwise mention scoring function, M LP scorer (m i , m j ), which predicts the probability that two mentions m i , m j refer to the same event. The mention representation includes a lexical component (GloVe embeddings) as well as a contextual component (ELMo embeddings, Peters et al., 2018). The mention pair representation v i,j , which is fed to the pairwise scorer, combines the two separate mention representations. We extended the model by changing the input to the pairwise event mention scoring function to include information regarding the mention pair from Chirps*, as illustrated in Figure 1. We defined v i,j = [ v i,j ; c i,j ], where c i,j denotes the Chirps* features, computed in the following way: is the feature vector representing a pair of predicates (m i , m j ) for which there is an entry in Chirps, otherwise the input is a zero vector. M LP ch is an MLP with a single hidden layer of size 50 and output layer of size 100, which is used to transform the discrete values in f m i ,m j into the same embedding space of v i,j . The rest of the model remains the same, including the model architecture, training, and inference. 5
Evaluation
We evaluate the event coreference performance on ECB+ using the official CoNLL scorer (Pradhan et al., 2014 Barhom et al. (2019), and the left one is our Chirps* extension, which is transformed through M LP ch into the same embedding space. The two vectors are concatenated to form the mention pair representation, which is fed to the scoring function M LP scorer .
We compare the integrated model to the original model and to the lemma baseline which clusters together mentions that share the same mentionhead lemma. The results in Table 8 show that the Chirps-enhanced model provides an improvement of 3.5 points over the lemma baseline and a small improvement upon Barhom et al. (2019) in all F 1 score measures. The greatest improvement is in the link-based MUC measure, which counts the number of corresponding links between the mentions. The Chirps component helps link more coreferring mentions (improving recall) and prevents the linking of some wrong mentions (improving precision).
Although the gap between our model and the original model by Barhom et al. (2019) is statistically significant (bootstrap and permutation tests with p < 0.001), it is rather small. We can attribute it partly to the coverage of Chirps over ECB+ (around 30%), which entails that the majority of event mention pairs still have the same representation as in the original model. We also note that ECB+ suffers from annotation errors, as was observed by Barhom et al. (2019) and others.
Conclusion and Future Work
We studied the synergy between the tasks of identifying predicate paraphrases and event coreference resolution, both concerned with matching the meanings of lexically-divergent predicates, and showed that they can benefit each other. Using event coreference annotations as distant supervision, we learned to re-rank predicate paraphrases that were initially ranked heuristically, and managed to increase their average precision substantially. In the other direction, we incorporated knowledge from our re-ranked predicate paraphrases resource into a model for event coreference resolution, yielding a small improvement upon previous state-of-theart results. We hope that our study will encourage future research to make further progress on both tasks jointly. | 6,700.4 | 2020-04-30T00:00:00.000 | [
"Computer Science",
"Linguistics"
] |
AOA-Based Three-Dimensional Positioning and Tracking Using the Factor Graph Technique
In this paper, an angle-of-arrival (AOA)-based algorithm is proposed for tracking the position of an anonymous target in three-dimensional (3D) space. Distributed sensors are deployed, which can measure both the azimuth and elevation angles of the AOAs. Assuming the target movement is non-linear, the extended Kalman filter (EKF) is applied, where the observation process is realized by a practical AOA-based position detector, to form a unified factor graph (FG) framework. Moreover, the variance of observation errors, which is needed by EKF, is estimated in real time by using both the AOA measurements and the predicted target state. Such a dynamic estimating approach exhibits higher performance robustness compared to the conventional method, especially when the sensing environment is unstable. Additionally, the predicted target state is also used as the a priori information of the system, in order to reduce the impacts of burst sensing errors. According to the simulations, the proposed system is shown to achieve less root mean squared errors (RMSE) in different evaluation scenarios, with fast convergence behavior.
Introduction
Future wireless networks should not only realize communication, but also support new services, such as connected transportation systems, smart logistics, unmanned factories, etc. [1][2][3][4]. As one of the enabling techniques, position detection is becoming increasingly important. Its technical evolution can be found in discussions within the Third-Generation Partnership Project (3gpp), the Bluetooth Special Interest Group (SIG), and other standard development organizations. Such applications should not only be confined to two-dimensional (2D) planes, but also be implementable to three-dimensional (3D) spaces.Moreover, communication may also rely on positioning techniques. For example, in multiple-input multiple-output (MIMO) systems, beamforming is well utilized to focus the signal power towards a specified receiver. However, conventional approaches for deciding the beam direction may not satisfy the future requirement, especially when the target is moving. Hence, a robust and low-complexity tracking solution should be considered.
To track non-linear target movement in practice, the extended Kalman filter (EFK) is used. One of the factors that influences the accuracy of EKF is the variance estimation of observation errors [5,6]. In this paper, the observation process in EKF is realized by a specific location detector, instead of a theoretical model in many previous works [7,8]. Usually, it is difficult to estimate the variance of detection errors in real-time scenarios. Such estimation may get even worse when the environment changes. For example, in connected car networks, car sensing may not only be performed by fixed infrastructures on the roads, but also be conducted by surrounding vehicles, which forms a simultaneous localization and mapping (SLAM) problem [9,10]. In SLAM, the sensing environment may change as devices move, and the variance of detection errors may become unstable. Hence, it is a crucial challenge for the tracking system to perform accurate estimations of such variance in real time.
In SLAM, multiple sensors can be used for detecting a single target. Actually, since the network density continues to grow, the utilization of multiple distributed sensing devices will become more convenient in the future. Therefore, this paper considers a scenario with distributed sensors. To perform a location detection with distributed sensors, many algorithms have been proposed. For example, in [11], the least squares (LS) method was applied for indoor positioning, and a hybrid system combining time-difference-of-arrival (TDOA) and angle-of-arrival (AOA) measurements was considered. In [12], a factor-graph (FG) method [13] was proposed for location detection using time-of-arrival (TOA) measurement. The performance was shown to achieve the maximum likelihood (ML) bound with relatively low complexity. Due to such advantages, many extensions of the FG algorithm to other applications can be found in [14][15][16] for received-signal-strength (RSS), TDOA, and AOA, respectively. This work will also focus on FG for both location detection and tracking.
A location detection and tracking algorithm using TOA was proposed in [17]. However, the use of TOA always requires strict time synchronization, because the signal emitting time should be known. Such a limitation will increase the difficulties of implementing such a technique in practice. Moreover, for an anonymous target that may not provide any transmission information to sensors or the fusion center, TOA cannot be used. There are still some limitations for RSS-based detection [14]. For example, RSS is also regarded as a technique only for a known target, because the training of received signal strength has to be conducted before the real deployment. To deal with the drawbacks mentioned above, TDOA-based techniques can be used, which does not require synchronization between the target and sensors, as well as off-line training. However, the performances are too sensitive to the measurement errors, due to the light speed.
In contrast, the performances of AOA-based techniques are not so sensitive to measurement errors, while still enjoying the same advantages as TDOA for detecting anonymous targets. Therefore, this paper will focus on AOA. In [16,18,19], AOA-based location detection and tracking was studied. However, the previous works only considered the scenarios in 2D. Actually, the use cases in 3D are currently increasing. For example, it can be applied to aviation control for airports, where not only the scheduled aircraft, but also the anonymous invaders can be detected and tracked. Moreover, live TV broadcasts via cameras on drones will become more popular, which requires accurate drone control in 3D by the fusion center. Additionally, future intelligent transportation systems may realize 3D detection capability to assist the navigation of vehicles running in the complex multi-layer overpass roads. Note that the measurement of AOA samples, including azimuth and elevation angles in 3D, may require more complicated soft/hard designs, such as MUSIC and ESPRIT [20,21], but they are out of the scope of this paper.
For location detectors that work in a relatively stable condition, the variance of detection errors will also be stable. In this case, a fixed variance value can be fed to EKF for estimating the performance of the observer. Usually, such a value can be calculated from off-line training. However, this paper considers a dynamic system, which indicates that the measuring abilities of sensors may vary through the tracking process. Therefore, a real-time tuning approach of the observation variance is required. This paper proposes an estimation method by using an approximated Cramer-Rao bound (ACRB). Specifically, the Cramer-Rao bound (CRB) provides a low bound on the detection performance according to the measurement of indirect parameters. For the AOA-based location detection problem, the CRB can be calculated by using the variances of AOA measurements and the real position of the target. However, the latter parameter is unavailable in practice. Instead, it is possible to get a predicted target position with EKF, with which the ACRB can be calculated as the estimation of the observation variance. Such a process can work in dynamic environments and has been shown to achieve a robust tracking performance in this paper.
The main contribution of this paper can be summarized as follows.
1. An integrated FG structure is formulated to realize 3D location detection and tracking with AOA measurements.
2.
It has been proven that the information from the elevation angle can help to improve the detection of 2D space, from both theoretical and simulation analyses.
3.
The proposed technique has exhibited robust performances even with an unstable sensing environment.
4.
By utilizing the ACRB as side information (SI) to the detector, the proposed system can greatly reduce the impacts of burst sensing errors.
The organization of this paper is as follows. The proposed system model is described in Section 2. Then, the proposed FG algorithms for realizing the location detector and tracking are provided in Sections 3 and 4, respectively. After that, the derivations of CRB and ACRB are given in Section 5. The simulation results that prove the superiority of the proposed technique are provided in Section 6. Finally, the conclusion is given in Section 7 with some remarks.
System Model
In this paper, a discrete and non-linear state space model (SSM) is considered. A single target is to be detected, and its state at time index k can be denoted by the 3D location s k = [x k , y k , z k ] T , with k = {1, 2, ..., K}. Moreover, there are N distributed sensors deployed in this system, and the fusion center is assumed to know their position (X n , Y n , Z n ), with n = {1, 2, ..., N}. The azimuth and elevation angles, denoted by φ and θ, respectively, are shown in Figure 1. For the sake of performing location detection, the relative distance ∆d n from the n-th sensor to the target is computed by: (1) x y z Figure 1. 3D coordinate with azimuth (φ) and elevation (θ) angles.
The measured AOA at each sensor is comprised of both azimuth φ and elevation θ angles. The definition of φ and θ in the spherical coordinate exactly follows the standard way shown in Figure 1. The relative distances calculated above can be connected by the true AOAs as follows, where the subscripts are omitted for simplicity. ∆x = ∆z · tan (θ) · sin (φ) (2) = ∆y · tan (φ) , As shown in Equations (3)-(6), the calculations of relative distances are self-contained. In order to obtain a converged result, iterative detection is performed based on an FG framework, which is known as the message passing algorithm. Note that there are always two ways to calculate the relative distance, which indicates that there will be two factor nodes included in the proposed FG structure, connecting to the variable node of the relative distance. A detailed discussion of the FG structure will be given in the next section. Moreover, it is assumed that for every time index, sensors are able to measure L samples of AOA from the target, and the variable of measured AOA can be described as a Gaussian model as: where l = {1, 2, ..., L}, and the noise components n φ and n θ follow N 0, σ 2 φ and N 0, σ 2 θ , respectively. Even though the quantization errors may break the Gaussianity of the measured AOAs, they are assumed to be very small in this paper, such that the input of the FG detector can be still regarded as Gaussian distributed variables. Moreover, it is assumed that the measured AOAs can be sent from sensors to the fusion center with error-free transmissions. Due to the Gaussian assumption of the measured AOAs at each sensor, they can be described by only parameters of the Gaussian probability density function (PDF), instead of L snapshots. By omitting the indexes n and k for simplicity, mφ and σ 2 φ are used to represent the mean and the variance ofφ, respectively. Similarly, mθ and σ 2 θ are used to represent the mean and the variance ofθ, respectively. Clearly, when L becomes larger, the means and variances obtained from AOA measurements will be closer to the real values.
To describe the tracking model, the state process equation in EKF is first given by: where w k = w x,k , w y,k , w z,k T denotes the noise vector, and each of the elements follows a Gaussian distribution with N 0, σ 2 w . As a non-linear function, f (·) determines the movement of the target. To realize EFK, the conventional way is to approximate , according to the Taylor expansion of the first order. Unfortunately, in this paper, the target's movement is unknown, which should be dynamically adjusted at each time index. Therefore, this work introduces an α such that f (α) = s k−1 holds. Therefore, (10) can be expressed by: where v k−1 = f (α) (s k−1 − α), known as the adjustment term, which should be updated dynamically at every time of the tracking process. Obviously, the accuracy of such a model can be guaranteed only if the target does not move far between two adjacent time indexes. Then, the observation equation can be also expressed as a function of the real target state, as: where the observation noise vector is represented by e k . Since s k is fully determined by the true AOAs φ and θ, g (s k ) can be also expressed byg (φ k , θ k ). Note thatg (φ k , θ k ) is not theoretically formulated in this paper, but realized by a practical 3D location detector. Within the proposed FG structure, all messages passed through are approximated to be Gaussian, including the noise of observation. Therefore, it is assumed that each element from e k follows N 0, σ 2 e . Unfortunately, σ 2 e , which is needed to perform EKF, cannot be directly measured at the sensors. The theoretical bound of the observation result is determined by the CRB. In other words, the smallest achievable σ 2 e can be calculated based on the variances of AOA measurements.
Location Detector
This section introduces the FG structure in detail for AOA-based 3D location detection, as shown in Figure 2. It is also known as the observer in EFK. Note that in the following discussion, the indexes of the sensor are ignored from the variables for an easier expression. First of all, the variable nodes (A φ , A θ ) are used, which contain the means and variances of the measured AOAs at each sensor. The outputs of (A φ , A θ ), i.e., (m φ , σ 2 φ ) and (m θ , σ 2 θ ), are then sent to the factor nodes (T A , T B , T C ) for trigonometric calculations, which triggers the FG iterations. As shown in Figure 2, T A only takes the information from A φ and then feeds its outputs to the variable nodes ∆x and ∆y, which forms a completed 2D location detection as in [16]. The calculations within the factor node T A are referred to (3) and (5). In order to further fulfill 3D location detection, the information from A θ should also be utilized. Therefore, factor nodes T B and T C are introduced, which connect both A φ and A θ and feed to (∆x, ∆z) and (∆y, ∆z), respectively. The operations within T B and T C are based on (2), (4), (6), and (7). As a result, each of the relative distance variables is linked to two different trigonometric factor nodes, as shown in Figure 2.
It should also be noted that in T A , the multiplication of two independent Gaussian variables is needed, while in T B and T C , the multiplication of three independent Gaussian variables is required. Then, the means and variances of such multiplication results have to be calculated, assuming all messages are Gaussian.
Specifically, given two independent variables a ∼ N (m a , σ 2 a ) and b ∼ N (m b , σ 2 b ), the product a · b can be expressed by the mean and the variance as: If a third variable c ∼ N (m c , σ 2 c ) is introduced, the product of them a · b · c can be expressed in a similar way, as: However, the original trigonometric calculations shown in (3)- (6) are not linear, which may break the Gaussianity of the output of the factor nodes. To deal with this problem, the first order TSexpansion is used to approximate the non-linear function f (α) as: where the linear approximation is assumed to be around α = m α . As a result, three constant values f (m α ), f (m α ), and m α are obtained. Based on (17), the approximated mean and variance of f (α) can be expressed by: Based on (18) and (19), it is straightforward to calculate the means and the variances of particular trigonometric functions of (3)-(6), which are summarized in Table 1.
Approximated Mean Approximated Variance
Given the mathematical preliminaries above, the proposed FG detection will be explained in detail as follows. First of all, messages containing means and variances of the measured AOAs will be fed to the trigonometric factor nodes T A , T B , and T C . After that, the output messages go to relative distance variable nodes ∆x, ∆y, and ∆z. As can be seen in Figure 2, each of the relative distance variable nodes connect to two trigonometric factor nodes, which requires the multiplication of two independent Gaussian PDFs as, Based on (20), particular calculations of the means and variances of the iterative messages are made possible, from the relative distance variable nodes to the relative distance factor nodes, which are expressed by, m ∆y→D B = m T A →∆y ·σ 2 T C →∆y +m T C →∆y ·σ 2 and: On the other hand, the means of messages passing from the trigonometric factor nodes to the relative distance variable nodes are calculated as: In addition, the variances passing from the factor node T A to the variable node ∆x can be calculated by: and the variance passing from the factor node T B to the variable node ∆x is given by: However, calculating the rest of the variances required in the proposed FG detection is omitted, due to the limited space. Actually, the way of such calculation is very similar to (33) and (34). Even though many calculations are involved in the equations mentioned above, some of them are reusable, which guarantees a low implementation complexity.
Finally, according to (1), messages passing from the factor nodes of a relative distance to different axes can be expressed by: and: In the equations shown above, the variable subscripts indicating the sensor index are ignored for simplicity, because such steps are conducted for all sensors in parallel. At the variable nodes of estimated target position, messages coming from different sensors will be exchanged in order to make the FG iteration work. Therefore, the variable subscripts indicating the sensor index will be used when calculating the means and variances, as [13]: According to the design of real systems, the FG iteration can be performed until some particular conditions are satisfied. For example, if the gap of detected positions between two adjacent time indexes is less than a pre-defined threshold, or the maximum allowed iteration number is simply reached, the FG iteration can stop. Finally, the target location detected by the observer can be obtained by (50)-(52), where σ 2 x , σ 2 y , and σ 2 z are calculated by (47)-(49), respectively. (52)
FG-EKF
The FG structure-based EKF is explained in this section. Based on (11) and (12), the goal of the tracking system is to obtain the values of s k and v k , which allow the a posteriori probability p (s k , v k |o 1:k ) to be maximized, with (·) 1:k denoting time indexes. The following equation calculates the marginal function of s k and v k , denoted byp (s k , v k ), as: where ∼ is the operator for exclusion. Factorization of the conditional PDF function of (53) can be given by: where (54) is calculated by following Bayes' rule, and (55) is computed since o k only depends on s k . Moreover, assuming that s k only depends on s k−1 and v k−1 , while v k only depends on v k−1 . Then, by substituting (56) into (55), p (s 1:k , v 1:k |o 1:k ) can be expressed by: where the denominator of (56) is omitted, i.e., p (o k |o k−1 ). It should be noted that p (s 1:k−1 , v 1:k−1 |o 1:k−1 ) in (56) denotes the output of FG-EKF from the previous time index. Hence, ∏ is used in (57) considering the case with the time index from one to k. Given (57), the FG-EKF algorithm used in this paper can be detailed in two steps as follows.
Prediction of the State
At the beginning, state prediction of the current time index is conducted based on the FG-EKF outputs at the previous time index. According to Figure 3, the message µ c (s k|k−1 ) indicating the state prediction is given by: where the message flows µ a (s k−1 ) and µ b (v k−1 ) are taken from the FG-EKF outputs at time index
State Refinement
Then, the predicted state s k|k−1 can be further refined by the detected target position of the observer, yielding the FG-EKF results at the current timing k. According to Figure 3, the message flow of the FG-EKF output µ e (s k ) is given by: where µ d (o k ) represents the message flow of the observed target state, computed by the proposed 3D location detector.
ACRB Derivation
In this section, the proposed ACRB for the DOA-based 3D location detector is derived. Based on the result in [16], the conventional CRB is given by: where FI M represents the Fisher information matrix and s denotes the real target position. Given the PDF of the random variableÂ, denoted by p(·), the Fisher information matrix can be expressed by: where represents the measured DOAs with L samples. Then, the CRB of the DOA-based 3D location detector can be expressed by: where J represents the Jacobian function, which can be further expressed by: where ∆x n = X n − x, ∆y n = Y n − y and ∆z n = Z n − z, ∆xy n = (X n − x) 2 + (Y n − y) 2 , and ∆xyz n = Finally, the ACRB at time index k can be expressed by: where the expression of J k|k−1 is equal to J, but with ∆x n = X n − x k|k−1 , ∆y n = , and ∆xyz n =
Results
In this section, simulation results are provided in order to prove the accuracy, convergence, and robustness of the proposed technique. The real path of the moving target is formulated by functions with regard to the time index k, as: where φ = π/60 and s 0 = [x 0 , y 0 , z 0 ] T = [0, 0, 0] T indicating the initial location. According to the specific simulation scenarios, the setting of sensor positions, FG iteration time, and AOA measurement variances will be different.
Accuracy Evaluation
In this subsection, the accuracy of the proposed tracking technique is evaluated by simulations. Three distributed sensors are deployed on the ground, with the positions at [100, −10, 0] T , [−10, −10, 0] T , and [−50, 100, 0] T , respectively. Such a deployment is only based on practical considerations. Actually, the optimal sensor location cannot be guaranteed due to the mobility of the target, as in (65) and (66). The FG iteration number of the detector is fixed at 15, and 100 snapshots are assumed for the detection of each time index.
First of all, the real path of the target and tracking trajectories are shown in Figure 4. Note that the standard deviations of AOAs from both the azimuth and elevation angles are set equal in the simulations, denoted by σ φ,θ , in order to evaluate the general tendency of the system behavior. However, in practice, σ φ and σ θ can be different. The tracking results with σ φ = 5 • can be seen in Figure 4a. In order to evaluate the accuracy of the proposed tracking technique, RMSEs are calculated. Specifically, the observer, realized by the 3D location detector described in Section 3, achieves RMSE = 1.6 meters, while this value is reduced to 1.4 with the proposed EKF. The RMSE values of the observer and the EKF are 6.15 and 4.6 in Figure 4b, which are higher than Figure 4a, because of a larger σ φ,θ = 20 • is assumed in the simulation. Moreover, another simulation was conducted to evaluate whether the 2D tracking performance can be improved with the help of the measured elevation angles. Actually, in many applications, the importance of 2D location accuracy is more than that of 3D. Therefore, it will be very beneficial to see if the 2D detection can be enhanced with additional assistance from elevation measurement. As shown in Figure 5, the RMSE values achieved by using only the azimuth angle φ for 2D detection are plotted, versus the standard deviation σ φ . Interestingly, considering both φ and θ with 3D detection, the RMSEs in 2D are found smaller than the case where only φ is used, which are closer to the theoretical CRB bound. This finding proves the fact that the information carried by the elevation angle θ also contributes to the detection in the 2D plane, which was ignored in the previous works [16]. Therefore, the introduction of the elevation angle is not only for realizing 3D detection, but also for enhancing the entire system performance. The mathematical proof of such improvement in terms of a lower CRB is left as future work. Only φ is used Both φ and θ are used CRB 2D Figure 5. Improvement of 2D location with elevation measurement.
Convergence
In this subsection, the system convergence behavior in terms of the FG iteration number is evaluated. The sensor deployment is assumed to be the same as the previous simulation, while both σ φ,θ = 5 • and σ φ,θ = 20 • are set for comparison. According to Figure 6, the tracking errors in terms of RMSEs can be much reduced only after five FG iterations. The performances will be further improved and stable with 10 iterations. Moreover, even the standard deviations of the AOA measurement are different, and the convergence behavior is found to be very similar.
Robustness Analysis
In this subsection, the robustness of tracking performance is evaluated, which is also regarded as the main advantage of the proposed technique. First of all, the effect of sudden sensing errors is considered, such as the false alarm, which may cause unreasonable measurement AOAs. In contrast to the previous simulations, four sensors are used with their positions at [−100, −10, 0] T , [−10, −10, 0] T , [−50, −100, 0] T , and [−50, −5, 10] T . The reason why more sensors are utilized is to avoid the case in which not enough sensors can work normally when some suffer from significant sensing errors. The FG iteration number is fixed at 15, while 100 snapshots are assumed for measurement at each time index. The real path of the target is set the same as in the previous simulations.
Specifically, at each time index, one of the four sensors is assumed to encounter sudden errors with a probability equal to 0.2, besides the true AOA measurement. Without the help of SI, the fusion center can simply discard the data of a sensor if AOAs from more than one origin are detected. However, the system performance will also decrease when fewer sensors are used. In the proposed system, a predicted AOA will be calculated for each sensor by the fusion center, based on the predicted sensor location by EKF. If the measured AOA is close to the predicted one within a predefined threshold, set by 10 • in the simulation, it will be used for location detection at the current time index. Otherwise, the measured data will be discarded. By doing this, sudden errors such as false alarms can be eliminated while still keeping the desired measurement, in order to achieve a better tracking performance.
The performance improvement with SI is shown in Figure 7. Actually, SI is regarded as the a priori information generated from ACRB, which helps the fusion center make a pre-decision to eliminate the impacts of burst sensing errors. In other words, ACRB enables the prediction of the expected AOA measurements. For example, in this simulation of false alarms, if the measured mean AOAs are larger than the predicted values beyond a threshold, i.e., 5 • , they will be directly eliminated. It can be seen that the RMSEs achieved with SI are always smaller than the cases without SI. The improvement is found to be more significant when the standard deviation of the AOAs becomes larger. Moreover, simulations are conducted to evaluate the robustness of the system against the dynamic measurement variances. In this case, σ φ,θ is assumed to be randomly chosen from a set of {5 • , 10 • , 15 • , 20 • }, for each time index. After location detection, the observation variance is calculated based on the proposed ACRB, which is dynamically fed to the EKF. As a comparative study, another scheme assuming a fixed σ φ,θ equal to two is also simulated. According to Figure 8, the RMSE achieved by the proposed technique is 4.3, which is smaller than the comparative scheme with the RMSE being 5.4. Obviously, the two local maximum RMSEs appeared for the comparative scheme happen when σ φ,θ = 20 • was chosen. However, such errors are much reduced with the proposed technique. Therefore, the robustness of the proposed system is proven.
Conclusions
This paper provided an AOA-based tracking technique in 3D with an FG algorithm. The system accuracy was proven via simulations even with large measurement variances. It can be found that the introduction of elevation angels not only realized 3D detection, but also enhanced the detection accuracy in 2D. Moreover, the convergence performance of the proposed FG algorithm was evaluated by simulations, which was shown to be achieved within around 10 iterations despite a large measurement variance. Finally, the predicted target state by EKF was used to eliminate sudden sensing errors and estimate the observation variance through calculating the ACRB, which significantly enhanced the robustness of the tracking performances. | 6,958 | 2020-08-01T00:00:00.000 | [
"Computer Science"
] |
Varying pre-plasma properties to boost terahertz wave generation in liquids
Laser-driven nonlinear phenomena can both reveal the structural features of materials and become the basis for the development of various translated technologies, including highly intense terahertz sources. Here we realize a modified single-color double-pulse excitation scheme for enhancing the terahertz wave generation in flat liquid jets, and we show that the pre-ionization effect is crucial for finding the optimal input conditions. The experimental results, being supported by numerical simulations, reveal the preference for longer pre-pulses to induce the effective ionization process and shorter signals for the strong laser-plasma interaction. In addition to the identified features of the terahertz wave energy enhancement with respect to the duration change for both pulses and their ratio variation, we state the possibility of achieving the optical-to-THz conversion efficiency value up to 0.1% in the case of double-pulse excitation of an α-pinene jet. Terahertz frequency radiation provides a powerful tool for the investigation of matter from the life science to the solid state and plasmas. The authors experimentally and numerically present enhanced terahertz wave generation by single-color double-pulse excitation in flat liquid jets, providing a deeper understanding of the mechanism that underpins the terahertz generation in multi-pulse experiments.
O scillating with a fundamental period of about 1 ps, terahertz frequency radiation can be a perfect tool for the investigation of protein vibrational modes, small molecule rotations, solid-state and gaseous plasma. For over a couple of decades, this electromagnetic range has been considered universal and sufficiently safe for medical applications, products, and environmental ecology quality control 1 . Despite the rapid development of terahertz science, both from the fundamental and applied aspects, the problem of developing a highly efficient source of terahertz radiation cannot be considered completely resolved.
As the most relevant approaches to the generation of terahertz radiation, it is fair to highlight those which are based on optical rectification (OR) in crystals 2 , free-electron acceleration (FEL) 3 , and laser-driven plasmas. All methods have their advantages and limitations, such as, for instance, low damage threshold or the relativistic input energy requirement. The approach based on laser-matter interactions (gas, plasma, solid-state, or liquid) has gained popularity due to its relatively simple experimental implementation, rather a high conversion efficiency values, and wide spectral coverage of the output terahertz field. Further work has focused on the improvement and expansion of this approach. One of the first attempts was to use an external electrostatic field to increase the energy of the generated terahertz waves 4 . Further breakthrough and future work of many scientific groups were associated with the implementation of the so-called two-color scheme, where both fundamental and second harmonics are used as a pump. Two-color filamentation in gases allows obtaining the optical-to-THz conversion efficiency of the order of 0.01% 5 .
The main directions of the laser-driven plasma terahertz source development are the search for an optimal generation medium 6 (with a high damage threshold, weak terahertz wave absorption, pronounced nonlinear effects) and input conditions variation, e.g., changing pump radiation parameters or modifying the experimental scheme configuration. The former includes the research on the generation of terahertz waves in a wide variety of gases 7 , metals 8 , as well as relatively new works on liquid media [9][10][11] . As an example, for input conditions variation, it was shown that the terahertz wave pulse energy could be more than 5fold enhanced by applying the abruptly autofocusing beam instead of the usual Gaussian one under the same conditions 12 . Another promising approach is the pump radiation wavelength shift towards the mid-infrared (IR) range. Thus, the authors of ref. 13 demonstrated the possibility of obtaining an optical-to-THz conversion efficiency value up to 2.36% during the two-color filamentation of femtosecond laser pulses at 3.9 μm. Such an impressive value was explained through the stronger photocurrents, longer plasma channels, and additional field symmetry breaking by the generated high harmonics.
In the present work, we investigate the modification of the experimental scheme due to the introduction of double-pulse liquid jet irradiation and bring out the features that distinguish this approach from the single-pulse configuration. Although the double-pump technique itself is not a recent development 14,15 , the induced pre-plasma effect that manipulates the output energy of the generated terahertz radiation is of considerable interest, motivating further laboratory and theoretical experiments with various input conditions. Thereby, in this study we provide a comprehensive analysis, considering both polar and nonpolar liquid media together with varying pre-pulse and signal pulse duration with their mutual ratio. All this allowed us to achieve the maximum optical-to-THz conversion efficiency value up to 0.1%.
Results
In this work, we use an experimental scheme with a double-pulse laser excitation of a flat liquid jet. Femtosecond p-polarized laser radiation with a central wavelength of 800 nm and a 1 kHz pulse repetition rate is divided by a beam splitter (BS1) into the reference and signal beams. In contrast to work 14 , where a Michelson interferometer was used, in this research a scheme with a Mach-Zehnder interferometer is presented (Fig. 1). This modification is introduced to study the effect of various pulse durations of the reference and signal beam. Their durations are changed by using a dispersion medium of 2-10 cm thicknesses (fused silica). Using the same dispersion medium (QP1, QP2) in both interferometer arms, we avoid the discrepancy in the reference and signal pulse energy characteristics. Thereby, the duration is varied by dispersion broadening and is controlled (measured) by a second-order autocorrelator. The energy of the reference and signal pulses is 0.45 mJ, and their durations are varied from 60 to 350 fs. One of the interferometer arms controls the time delay between pulses from 0 to 30 ps.
A sequence of two collinear pulses falls on a parabolic mirror (PM) with a focal length of 5 cm. The pre-pulse (reference) and signal are focused on a flat liquid jet (see "Methods" section for details regarding liquid jet preparation) at the optimal angle to obtain the most efficient optical-to-THz conversion 16 . The estimated laser spot size (FWHM) is 124 μm (the laser intensity is then 2.5 ⋅ 10 13 W/cm 2 for 150 fs pulse). A filter of black Teflon (F) is used to remove visible and IR radiation. Terahertz radiation is then collimated by a TPX lens (L). We register terahertz waves with a standard electro-optical scheme (EOS) based on a 1 mm ZnTe crystal, which allows detecting a signal up to 3 THz.
The inset of Fig. 1 demonstrates the temporal and spectral structures of the generated terahertz field for a single-color double-pulse water jet optical excitation. In this case, the temporal delay is 2 ps, and each pulse duration is equal to 150 fs. The increase in the terahertz field signal can be clearly seen. The "Methods" section provides more details on terahertz radiation energy estimation and pulses separation.
Laser pulse duration effect on the terahertz energy enhancement. Firstly, we compare the optimal value for the pump pulse duration in the case of single-and double-pulse excitation. We conduct experiments to study the terahertz field enhancement considering 2 ps temporal delay 14 and varying pre-pulse and signal pulse durations from 60 to 250 fs (Fig. 2). As a result, in contrast to the pump pulse duration maximum of about 200 fs obtained for the case of single-pulse excitation (Fig. 2a), using a double-pulse scheme, the optimal value shifts to 100-150 fs (Fig. 2b). Moreover, in this case, we obtain a 14-fold enhancement compared to a 4-fold one from the previous results 14 .
It turns out that the first observation is the advantage of shorter pulses in double-pulse excitation experiments. The presence of an optimal duration value in single-pulse experiments was earlier interpreted through the combined effect of the electron density exponential growth due to the cascade ionization and pulse energy attenuation with the increase in pulse duration 11 . The double-pump case is indeed different; it seems reasonable to assume that the optimal pulse duration will experience a shift towards lower values since the important moment is not the effective ionization of the molecules, but rather the strong interaction with a pre-ionized medium.
Optimal liquid for a double-pulse excitation scheme. Since the optimal value for the pump pulse duration was revealed in the previous section, here we establish which of the studied liquids is the most optimal for a double-pulse excitation scheme. Figure 3 shows the results of the signal pulse enhancement in different liquid media excited by pre-pulse and signal of 150 fs, normalized to the maximum value under the same conditions for water. The curve form, where the preference of a picosecond temporal delay for the liquid media excitation is pronounced, was also revealed by the authors of ref. 17 and explained by the authors of ref. 14 . Moreover, similar nature of the dependence was obtained in the experimental measurements of X-ray generation in water 18 .
In this work, the enhancement range (indicated by the arrows in Fig. 3), presumably corresponding to the plasma lifetime, is revealed to be different for these liquids. In the "Methods" section, we provide additional measurement results on the plasma lifetime by the third-harmonic (TH) enhancement method, which appear to be in good agreement with this conclusion. Thereby, the double-pump method could shed light on the differences in the molecular response of various liquids in the field of ultrashort IR-pulses. We compare the terahertz pulse energy peak values in the case of each liquid medium and obtain 0.45, 0.37, and 0.33 μJ for α-pinene, ethanol, and water, respectively. Hence, the maximum optical-to-THz conversion efficiency is achieved for an α-pinene jet double-pulse excitation, which is about 0.1%. Similar to previous work, which considered liquid media singlepulse excitation 9 , nonpolar α-pinene demonstrated the highest optical-to-THz conversion due to negligible terahertz absorption and low ionization potential, which leads to the more efficient plasma formation process. Fig. 1 The double-pump experimental scheme based on the Mach-Zehnder interferometer. Using the Mach-Zehnder interferometer it is possible to vary the duration of the reference and signal pulses in its arms. This variation is produced by passing the interfering pulses through the quartz plates QP1, QP2 of equal, or different thicknesses (from 2 to 10 cm). An optical delay line is introduced in one of the interferometer arms to control the time delay between the pulses. A sequence of two collinear pulses is then focused by a 5 cm parabolic mirror (PM) on the liquid jet at the optimal angle of incidence. The generated terahertz radiation, collected and collimated by TPX lens L and filtered by a Teflon filter F, is registered on the standard electro-optical scheme EOS. The inset demonstrates the temporal and spectral structures of the generated terahertz field. The terahertz wave energy dependence on the pre-pulse/signal pulse duration ratio. Finally, we present the results (Fig. 4) on the pre-pulse and signal pulse duration ratio variation to achieve the terahertz wave maximum energy and thus, define the optimal conditions. The measured terahertz pulse energy duration dependence after the signal beam excited pre-ionized liquid medium is shown in Fig. 4a.
The figure reveals a bright enhancement area around the prepulse-to-signal pulse duration ratio of 150:150 fs. Moreover, it demonstrates the advantage of longer pre-pulses to induce the initial electron density and then shorter signal pulses for achieving the maximum energy of a terahertz wave (see the characteristic dashed area in the bottom part of Fig. 4). In order to investigate this result in more detail, we will further study the described dependences using numerical simulation methods and discuss the possible physical mechanisms, which are beyond the experimental results.
Discussion
In this section, we numerically investigate the generation of terahertz radiation upon the excitation of liquid media by two collinear laser pulses to obtain a clearer picture of the pulse duration-dependent features. The equations of strong-field dynamics in the isotropic dielectric medium for a three-band system are used for this purpose 19 (see "Methods" section for some additional details). They can be reduced to one field equation, which here is given in its normalized form: Zτ0 À1Ẽ 2 exp À ðτ 0 Àτ 00 Þ hωiτ p ! dτ 00 ¼ 0: ð1Þ where E is the electric field of radiation, E 0 is the peak amplitude of the spectral-limited pulse at the input surface, and ξ = τ pulse /τ SL is the ratio of the chirped pulse to a spectral-limited pulse duration to consider the change in a peak amplitude with varying pulse duration; z̃= za 〈ω〉 3 , z is the propagation coordinate, a is normal dispersion coefficient, 〈ω〉 is central radiation frequency with wavelength λ 0 = 800 nm;τ = τ〈ω〉, where τ = t − zn 0 /c is time in the moving frame of reference, c is the speed of light in vacuum, n 0 is a linear refractive index. Coefficientg = gE 2 0 =ðahωi 2 Þ, where g = 2n 2 /c and n 2 is the nonlinear refractive index, characterizes low-inertia cubic nonlinear response of the medium,g p = g p E 4 0 =ðahωi 3 Þ, where g p ¼ 2π cn 0 αβ, is introduced to describe the plasma nonlinearity, where αβ characterizes the efficiency of the electrons transition to the quasi-free states. τ c and τ p are the times of collision relaxation and relaxation of highly excited states. All medium parameters are taken according to ref. 14 .
To match the experimental conditions, the normalized radiation field, yielded by two collinear chirped Gaussian pulses with the same energy of 0.45 mJ, a pulse delay of 3 ps and variable pulse duration in the range of 50-300 fs, is used: whereτ pulse1;2 ¼ τ pulse1;2 hωi are the normalized durations for reference and signal pulses, respectively; A 1,2 represents a frequency modulation and is chosen such that the width of the chirped pulse spectrum fit the width of the spectrum for the 35 fs spectral-limited pulse; Δτ ¼ Δτhωi is a normalized temporal delay between the pulses.
To begin with, we demonstrate that the results of numerical simulation are in fairly good agreement with experimental studies of the terahertz wave generation during single-and double-pulse excitation of a liquid medium (see Fig. 2, solid lines). The numerical simulation of the terahertz energy enhancement during the liquid medium irradiation with a signal pulse, depending on the mutual ratio of the pre-pulse and signal pulse durations is not Fig. 3 Terahertz wave enhancement dependences on the temporal delay between two collinear 150 fs pulses. The measurements are implemented for water (blue), ethanol (yellow), and α-pinene (red). The accuracy of Golay cell measurements is depicted by error bars and is estimated to be 7-10%. The lines here are introduced as a guide for an eye. The arrows correspond to the temporal enhancement range for each liquid. that clear (Fig. 4b). The maximum of terahertz radiation energy is obtained for 130-130 fs and 150-100 fs cases with a little dominance of the first one. However, it demonstrates the same trend of long pre-pulse/short signal preference.
Since the generation efficiency of terahertz radiation is varied by the photocurrent, which in turn depends on the induced electron density, we propose an analysis of the terahertz wave energy enhancement during the double-pulse liquid jet excitation mechanism.
Both plasma and Kerr nonlinearities contribute to the terahertz emission. However, with our input conditions, the terahertz emission is dominated by the effects of induced plasma. The research on the comparison of plasma and Kerr nonlinearity contributions in the single-pulse excitation case was already done in ref. 19 . Numerical simulation based solely on considering the third-order Kerr nonlinearity leads to the negligible terahertz generation, while the plasma effect seems to determine this process significantly. Thereby, the first stage in the case of doublepulse excitation can be represented as a microplasma formation and terahertz generation due to the photocurrent and bound electron nonlinearity. This is followed by the terahertz field enhancement stage due to the laser-plasma interaction. Generation of a strong photocurrent is more critical at this stage; thereby, pulses of shorter duration with a higher peak power are preferable. This is consistent with the time evolution of the current density ∂j ∂τ þ j The current is proportional to E 3 and to the number of quasi-free electrons ρ; consequently, the dependence on the peak intensity is stronger.
The theoretical model based on the density matrix formalism (Eq. (3)) considers no possible reflection of the pump radiation. Thereby, a short signal pulse will only enhance the shielding effect 20 and reduce the energy contribution of the pump radiation. This may be a reason for the discrepancy between the experimental and theoretical results in Fig. 4a, b. Moreover, here we can also talk about the possibility of energy redistribution since in the experiment, generation occurs in water and additional nonlinear absorption also happens. Decreasing the pulse duration, we increase the nonlinear absorption, reducing the energy contribution directly in the terahertz generation process.
In conclusion, we propose an effective modified double-pulse excitation scheme with a Mach-Zehnder interferometer to investigate the pre-plasma effect on the terahertz wave energy enhancement in flat liquid jets, and thereby, achieve new conversion efficiency values. Both experimental and theoretical results reveal that longer pre-pulses are preferable to induce the effective ionization process, while shorter signals-for the strong laser-plasma interaction. This is in agreement with the presence of a pump pulse duration optimal value for the single-pulse excitation case, which origin is determined by the simultaneous cascade ionization processes enhancement and the peak intensity attenuation with the increase in the pump pulse duration. Furthermore, we provide plasma lifetime measurements, which are consistent with THz enhancement traces. In this work, the possibility of achieving the optical-to-THz conversion efficiency value up to 0.1% in the case of double-pulse excitation of an αpinene jet is demonstrated. It is important to emphasize, that this result, obtained for an 800 nm laser pump, can be scaled further when shifting the central wavelength to a mid-IR region 13 and reach even more impressive values of about 1-5%. Thus, revealing the optimal conditions for the highly efficient terahertz wave generation in liquids, our study brings the development of powerful and economical terahertz sources closer to reality.
Methods
Liquid jets creation. As a source of a flat liquid jet, we use a system that allows pumping liquid under pressure at 1 m/s speed and forms a plane-parallel flow through the nozzle 21 . In all experiments, a liquid jet with a thickness of 100 μm is studied. As a generation medium, we choose polar water and ethanol, and nonpolar α-pinene to investigate the effect of the liquid media properties on the optical-to-THz conversion efficiency, which was already essential in the singlepulse case 9 .
Terahertz energy estimation and pulses separation. To estimate the terahertz energy and optical-to-THz conversion efficiency, we use two methods of registration, both EO, and the Golay cell. The latter becomes necessary since it is impossible to separate the pulses with a temporal delay of Δτ < 2 ps using EO detection. To obtain the output energy, we integrate the square of the terahertz field amplitude. The conversion efficiency is evaluated when the radiation reaches the detector. Therefore, we consider that after passing the Teflon filter and focusing system, the radiation losses of 10% occur.
To obtain an additional estimate of the optical-to-THz conversion efficiency, we further compare the temporal and spectral structures of the terahertz field generated by α-pinene double-pulse excitation and by the common two-color air filamentation technique (Fig. 5a, b).
The experimental scheme is not significantly modified in this case. For comparison, the duration of the pump pulse is taken to be optimal for the case of two-color filamentation in air and is equal to 35 fs. The filament is generated by focusing the femtosecond pump radiation with a 5 cm lens. The β-BBO crystal with 300 μm thickness placed behind the lens focus is used for the second-harmonic generation (with 10% conversion efficiency). In our case, we determine the optimal parameters for the β-BBO crystal position when a maximum terahertz signal was observed. In a result, comparing the integrated spectral power for both methods, power of terahertz radiation generated during two-color air filamentation and double-pulse liquid jet excitation. The double-pulse technique is implemented for α-pinene liquid for the optimal 150 fs duration (magenta), while in the case of two-color air filamentation we use 35 fs pulse (blue). The optimal parameters are determined for the β-BBO crystal position when a maximum terahertz signal was observed.
we obtain more than 10-fold enhancement in the case of α-pinene jet double-pulse excitation.
Plasma lifetime measurements. To measure the plasma lifetime in liquids, we apply the third-harmonic (TH) enhancement method, which is proposed and fully described in ref. 22 . We use an experimental setup analogous to the one from this work. Two filaments are formed by focusing the pump and the probe beam by a 5 and 10 cm lens and crossed perpendicularly to each other. The probe beam is collimated by a quartz lens with a focal length of 10 cm. The pulse energy in each channel is 0.9 mJ, and the pulse duration is about 40 fs. The pump beam is moved by a motorized delay line with a step of 1 μm. The spectrum is measured on ASP 100 (Avesta) spectrometer with 1 μm resolution in the range from 190 to 1100 nm. We use ultraviolet (UV) filters UG-1, UG-5 (Standa), which allow us to obtain the spectral ranges of 240-420 nm, 250-420 nm, respectively. The liquid jet is placed under the angle of 45°relative to the pump and probe beam. The scanning is implemented 140 times for each liquid, after which the data is averaged.
We measure the plasma lifetime for water and ethanol. Using α-pinene jet we get strong evaporation with further damaging of optical elements. These results in a low signal-to-noise ratio. The typical TH spectrum along with the spectrum of enhanced TH is demonstrated in Fig. 6a.
The change in the intensity of TH emission (λ TH = 265 nm) from water and ethanol depending on the temporal delay is shown in Fig. 6b. Plasma lifetime, which is determined as decay from the peak value to the 1/e level, indeed coincides with the terahertz energy enhancement range demonstrated in Fig. 3.
Theoretical model features and approximations. To numerically investigate the generation of terahertz radiation upon double-pulse excitation, we consider the equations of strong-field dynamics in the isotropic dielectric medium for the threeband system 19 : The first equation of this system describes the dynamics of the radiation field, considering the dispersion of the linear refractive index and non-inertial cubic nonlinearity, as well as the effect of the induced plasma. The second equation characterizes the evolution of quasi-free electrons current density under the influence of a radiation field, while the third one describes the change in the concentration of electrons of excited energy states, the transition to which is allowed from the ground.
When deriving these equations, the radiation is assumed to be linearly polarized. Since the current density at the output is represented as the time derivative of the induced polarization of the medium, formally, the currents that are perpendicular to the axis of radiation propagation are considered.
In addition, since we use thin liquid jets of 100 μm thickness for all experimental measurements and the focal length of the effective positive lens induced by a thin water flow is f ¼ n 0 w 2 0 =4n 2 I 0 L $ 17 mm, we neglect spatial effects influence on the pulse in Eq. (3).
Data availability
The data used in this manuscript are available from the corresponding author upon reasonable request. | 5,539.8 | 2021-01-04T00:00:00.000 | [
"Physics"
] |
The Application of Gray Model and Support Vector Machine in the Forecast of Online Public Opinion
The forecast of online public opinion is a kind of complex forecasting problem with information, small sample and uncertainty. In order to improve the accuracy for the forecast of online public opinion, a new forecasting method based on a gray model and a support vector machine is proposed. The method comprises the steps of clustering the text, extracting the hotspots, aggregating the data and implementing other pretreatments of the network data, then creating a model GM (1, 1) for the time series of online public opinion, correcting the forecasting results of the model GM (1, 1) with a support vector machine, and then testing through a simulation experiment. The experimental results show that compared with traditional forecasting methods, the application of gray model and support vector machine improves the accuracy for the forecast of online public opinion. Moreover, a new method for the forecast of online public opinion is presented to some extent.
INTRODUCTION Online public opinion, which is also known as network public opinion, refers to the opinions or remarks with a certain influence and tendentiousness of netizen to the social public affairs, especially the hot social focuses through the Internet [1][2]. With the rapid development of Interne in China, network has become one of the main carriers for the reflection of social public opinion [3]. At present, the economic and social development of China is in a crucial stage, in which the various deeply rooted contradictions and problems arise day by day, so the hotspots of online public opinion are emerged one after another, which involve broad regions as well as extensive contents. In such a situation, the negative online public opinion will have great negative impact on the national security and social stability, if the online public opinion cannot be guided and supervised correctly [4]. Therefore, it has become a hotspot of research at present to forecast the trend of the development of online public opinion accurately.
In recent years, there are more and more studies focusing on the forecast of online public opinion, which can basically be divided into two categories: traditional forecasting method and modern forecasting method. According to the traditional forecasting method, the data of online public opinion is converted into time series, and the model is created by using the forecasting methods of autoregressive moving average, exponential smoothing and other time series. This method is simple and easy to be carried out. However, it assumes online public opinion is changed linearly, which is inconsistent with the actual changing characteristics, and therefore the results of forecast are not ideal. As for the modern forecasting method, the model is created on the basis of nonlinear theory. Compared with traditional forecasting method, the accuracy for the forecast of online public opinion is improved correspondingly, and the main forecasting models include Hidden Markov Model [5], G (Galam) [6], intuitionistic fuzzy reasoning [7], support vector machine [8][9], etc. Online public opinion is a kind of uncertain forecasting problem with information and small samples. In order to improve the accuracy of forecast further, some scholars have proposed some assembled forecasting models for the online public opinion based on the combination optimization theory and the advantages of each single model. For example, Zhang Jue put forward the online public opinion forecasting model based on ARIMA and BP neural network and achieved good forecasting results [10].
Gray forecasting theory [11] is proposed for the first time by domestic scholar, Deng Julong, in 1982, which studies "small sample" and "poor data information" uncertain system of "partial known information and partial unknown information. GM (1, 1) model, the important component of gray forecasting theory, is featured with less data required by model establishment. Support vector machine (SVM), referring to a modern machine learning algorithm specially for the small sample and uncertainty forecasting problems based on the statistical learning theory (SLT), is widely applied in the study of the field of nonlinear time series forecasting.
In the study, the grey model is attempted to be combined with the support vector machine model and applied in the forecast of online public opinion. Firstly, GM (1, 1) is used to establish the forecasting model of online public opinion. Secondly, the forecasting result of GM (1, 1) is modified by the support vector machine. At last the performance of the model is verified by simulation experiment. to collect the various information sources of online public opinion thereon [12][13]. However, the messy and disordered data of online public opinion will be acquired, which should be converted into the related data by text clustering treatment.
A. Text Clustering
The hierarchical clustering algorithm [14][15][16] is used in the study to cluster the data of network pubic opinion, and the advantage and disadvantage of clustering is evaluated on the basis of the purity index. After the text clustering, the purity index for the clustering r is defined as follows: In the formula, r n is the number of documents in the r (th) clustering category, and i r n is the number of texts belonging to the predefined category i , but distributed into the r (th) clustering category by mistake.
So, all the purity indexes of the text clustering result are defined as follows:
B. Hotspot Acquisition
The hot topic of network means the information set regarding the network as the communications media, paid attention by a certain of crowd widely and continuously and capable of reflecting the situation of online pubic opinion [17][18]. The process of hotspot acquisition is as follows: (1) The reporting frequency, continuous reporting time and network click rate of the topic are adopted as the characteristics of the hotspot topic, which are performed statistics.
(2) The values of media attention and public attention are calculated.
(3) The specific gravity balance factor and threshold value are set and the pubic attention is calculated (4) If the public attention is more than the threshold value, it shows that the topic is the hotspot topic.
C. Data Aggregation
The collected online public opinion information with different vectors is organized by data aggregation and converted into the discrete-time series of hotspot topic by the data aggregation software.
III. FORECASTING MODEL OF ONLINE PUBLIC OPINION BASED ON GM (1, 1) AND SUPPORT VECTOR MACHINE
The gray model is capable of revealing the development trend of the data, but is not suitable for the forecast of time-invariance and nonlinear data, while the support vector machine is applied to describe the nonlinear and small sample data series. Thus, the forecasting model of online public opinion based on GM (1, 1) and support vector machine can be established by combining the advantages of the both.
A. GM (1, 1) Model
In recent years, the gray model GM (1, N) is widely applied and studied, wherein GM (1, 1), referring to the most common and simplest gray model and the model composed of differential equation only including single variable, is a special case of GM (1, M). Assume that the original data series is the model establishment series 0 X of GM (1,1), that is to say: The original data series is accumulated by the accumulation and generation method, and 1-AGO series generated by accumulation for one time is as follows: In the formula, So, the gray differential equation model of GM (1, 1) is as follows: is put into the above formula and the formula is obtained as follows: The above equation can be converted into the matrix equation as follows: In the formula, B is the data matrix, N Y is the data vector, P is the parameter vector, that is to say: The solution is carried out by least square method to obtain the formula as follows: The obtained coefficient is put into the formula (6), and then the differential equation is solved to obtain the expression of gray GM (1, 1) intrinsic model as follows: In the formula, is the residual; that is to say: The residual is inversely proportional to the accuracy of model. For the general requirements, % 20 ) ( k , and the best condition is % 10
B. Model of Support Vector Machine
The complexity of corresponding quadratic programming problem solved by the support vector machine is inversely proportional to the calculating speed. The least squares support vector machine (LSSVM), modifies the model of support vector machine and reduces the complexity of solution, so it has the advantages of less calculation resources as required as well as fast solution speed and convergence speed. Therefore, LSSVM is adopted as the forecasting mode in the study. For the time series of online public opinion, the regression function of LSSVM is as follows: In the formula, w is the weight vector and b is the bias constant.
According to the inductive principle of structure risk minimization, the model of least squares support vector machine for solving the regression problem is as follows: The constraint condition is as follows: In the formula, is the regularization parameter and i is the slack variable.
Lagrange multiplier is introduced to obtain the formula as follows: In the formula, ) , , is Lagrange multiplier.
The following formula is obtained according to KKT (Karush-Kuhn-Tucker) condition in the optimization theory: So, the last solution can be obtained as follows: In the formula, . According to the Mercer condition, the kernel function is defined as follows: is introduced to convert the formula (17) to obtain the forecasting model of LSSVM as follows: The radial basis function (RBF) is featured with good universality and better expression for processing the time series problem than that of other kernel functions, so in this paper, the radial basis function is used as the kernel function of LSSVM, and the expression is as follows: In the formula, 2 is the kernel width of RBF.
IV. EXPERIMENTAL RESULT AND ANALYSIS In order to verify the function of the gray model and support vector machine in the forecast of online public opinion, in the environment of Intel Core i5 3.2G CPU, 4GB RAM and hardware having Microsoft Windows Sever 2003 as the operating system, the implementation algorithm is realized by programming via MATLAB. A certain hot topic on the internet is forecasted and 30 data of the amount of the relevant posts is obtained, which is shown in Figure 1 for detail. In order to quicken the training speed of the model to reflect the variation trend of the online public opinion better, the time series of online public opinion is pretreated, and the normalized data is shown below: In the formula, i x is the data after normalization, max x and min x represent the maximum and minimum of the time series of online public opinion, respectively.
The data is divided into two parts. The former 22 data is used as the training sample set, and the later 8 data as the test sample set. And then the test sample set is forecasted by several models respectively. The obtained forecasting result is shown in Figure 2.
It is shown in the analysis for the value result of the test sample of online public opinion by the forecasting model in Figure 2 results of the single gray model and support vector machine model and the actual value is large, and the errors are quite high, which indicates that a single model only can explain the fragment and part of variation rules of the complex online public opinion, but difficult to describe the laws of time invariance and nonlinear variation of online public opinion completely and accurately. However, the advantages of the forecasting models of the GM (1, 1) and support vector machine are combined with the advantages of the gray model and support vector machine to overcome the defect of the single model, which not only can explain the characteristics of set information and small sample of the online public opinion, and accurately forecast the rules of time invariance and uncertainty variation, but also capture the variation trend of the online public opinion, so as to improve the forecasting accuracy of it.
V. CONCLUSIONS Online public opinion is a complex and time-varying system with larger burstiness and volatility. If it can be forecasted accurately, particularly, those hot public opinion which can arouse the attentions of most netizens, it will help the relevant departments to find out the potential risks timely, research and respond to the online public opinion actively, improve the ability to communicate with the public, and lead the online public opinion to the healthy development. In order to improve the accuracy for the forecast of online public opinion, a forecasting model of online public opinion based on the combination of a gray model and a support vector machine was proposed, aiming at the characteristics of online public opinion by taking advantages of the gray model and support vector machine. The experimental results show that the application of gray model and support vector machine can not only improve the accuracy for the forecast of online public opinion effectively and make up for the deficiency of a single forecasting model, but also provide a new idea for the study on the forecast of online public opinion. | 3,101 | 2013-03-24T00:00:00.000 | [
"Computer Science"
] |
Vapor Phase Sensing Using Metal Nanorod Thin Films Grown by Cryogenic Oblique Angle Deposition
We demonstrate the chemical sensing capability of silver nanostructured films grown by cryogenic oblique angle deposition (OAD). For comparison, the films are grown side by side at cryogenic (∼100 K) and at room temperature (∼300 K) by e-beam evaporation. Based on the observed structural differences, it was hypothesized that the cryogenic OAD silver films should show an increased surface enhanced Raman scattering (SERS) sensitivity. COMSOL simulation results are presented to validate this hypothesis. Experimental SERS results of 4-aminobenzenethiol (4-ABT) Raman test probe molecules in vapor phase show good agreement with the simulation and indicate promising SERS applications for these nanostructured thin films.
Introduction
The detection and identification of hazardous chemical and biological agents is important for several areas of defense and security as well as in other industries that deal with hazardous chemicals [1].Gas chromatography has the advantage of providing quick and accurate detection capability; however cost, size, and lack of portability have limited its widespread use [2].Ion-mobility spectrometer is another popular technique for chemical sensing; however the level of information that can be extracted is not comparable to most vibrational spectroscopy techniques [3].Surface enhanced Raman scattering (SERS) has shown the ability to detect the presence of very low concentrations of chemical agents quickly [1,[4][5][6].Several groups have also demonstrated a portable Raman setup for chemical and biosensing applications [7,8].In many of these applications, colloidal silver nanoparticles are used as the SERS substrates, which limits these to only liquid phase applications [9].On the other hand, silver nanostructured SERS substrates have the flexibility to work with liquids or vapors.Some of the earlier literature on SERSbased vapor sensing includes simulants for highly toxic chemicals such as nerve and mustard agents [7].Most of these used electrochemically roughened silver substrates or silver film over nanosphere substrates.All of these techniques are limited by the available surface area for the vapor molecules to bind and adsorb onto.On the other hand, nanorod-based substrates can offer a significantly greater surface area for the same foot print area.
Oblique angle physical vapor deposition technique has led to the evolution of a new class of thin films with very large effective surface areas.The technique is based on atomistic level self-shadowing.Whereas in typical thin film deposition setups the substrate is held normal to the vapor source, in the OAD setup the substrate is held at a very large oblique angle with respect to the incoming vapor flux.This angle creates shadows behind each condensing atom preventing subsequent atoms from condensing in the shadowed areas.Instead, they land on the previously nucleated sites resulting in the evolution of columnar morphology.A detailed overview and potential applications of these thin films are given in the review paper by Hawkeye and Brett [10].Even though OAD has been demonstrated to work well with metals like titanium (Ti), chromium (Cr), nickel (Ni), and so forth, it has shown limited success with soft metals like silver (Ag), gold (Au), and copper (Cu) [11,12].Most of the results with the latter metals show lower aspect ratios and/or collapsed columns, especially when the substrate is held at ambient temperatures during the growth.
In this study we show the results of silver nanorod (AgNR) thin films grown by OAD at cryogenic temperatures (∼100 K) compared to room temperature (∼300 K).The AgNR substrates were then incubated with vapors of 4-ABT and characterized for SERS measurements.Experimental results show a significant improvement in vapor phase detection capability of the cryogenically grown films compared to the room temperature grown films.These results are also confirmed by COMSOL MultiPhysics simulations of the plasmonic field enhancement in the AgNR.
Experimental Details
2.1.Fabrication.All of our depositions were done using an MDC evap-4000 e-beam evaporator with a throw distance of approximately 0.5 m.A substrate mounting plate was designed and built to carry liquid nitrogen in an open-cycle Dewar configuration.The samples were held by angled blocks made from Cu and Teflon with a preset angle of 88 ∘ with respect to the incident flux and mounted to this cold plate.This allows concurrent growth of a cryogenically cooled sample and an ambient temperature sample on the Cu and Teflon blocks, respectively.Separate thermocouples were attached to the blocks to monitor the substrate temperatures.During growth, the Cu block was at approximately 100 K and the Teflon block was close to 300 K. Further details of this experimental setup are given in [13].The substrates used for all the experiments were <100> silicon and the films were grown at a rate of 3 Å/s.The film thicknesses were measured using a quartz crystal microbalance at normal incidence (i.e., not at a glancing angle).
Vapor Phase Deposition of 4-Aminobenzenethiol (4-ABT)
Test Probes.4-ABT molecules contain sulfur which binds very effectively with the silver surface.The compound is highly volatile due to the presence of thiol and amino groups.In the setup shown in Figure 1, approximately 10 milligrams powder of 4-ABT was loaded in a 10 mL glass beaker.A microscope glass slide with double sided tape was used to attach the AgNRs samples.A plain untreated silicon sample was also incubated with the 4-ABT vapor as the background control sample.The samples were incubated for different time periods to vary the concentration of the vapor being deposited on the substrate.Due to the sensitivity of the 4-ABT molecule to the ambient moisture, the incubation was done in a glove box setup, purged three times with high purity N 2 prior to allowing the evaporation to occur.The samples were exposed to the 4-ABT vapor for different times starting with 60 min exposure followed by 30, 20, and 5 min exposure.
Micro-Raman and SERS Measurements.
The SERS measurements were obtained using a Raman measurement tool (LabRamHR 800 system).A 632.8 nm HeNe laser, with an unattenuated output power of 15 mW, was used as the excitation source.The backscattered Raman and SERS signals were collected using a thermoelectrically cooled CCD detector.The focused laser beam spot size on the AgNR surface was about 1 m.Under these experimental conditions, the spectral resolution was about 1 cm −1 .Point measurements were conducted to determine the appropriate integration times, and optical density filters were used to reduce the incident laser power.A 5 s acquisition time and signal averaging over 3 cycles were used to acquire the spectra with adequate signal to noise ratio.SERS spectra at random locations were collected across the AgNR and the silicon substrates.
Results and Discussion
3.1.SEM Results. Figure 2 shows the SEM cross section views of the AgNR thin films grown at 100 K and 300 K substrate temperatures.The distinct difference in the morphology between the two films can be immediately seen from these images.Table 1 shows quantitative differences between the two AgNRs.The AgNRs grown at 100 K was 28% smaller in diameter and 56% longer compared to the 300 K sample, resulting in an overall increase in their aspect ratios.The 100 K AgNRs also shows a curvature, or sag, in its structure, which we have also observed in other soft metals (and not observed in metals like Cr, Ni, and Ti).The 100 K AgNRs also shows a 39% increase in the rod density per unit substrate area.All of these morphological differences in the 100 K AgNRs effectively increase the surface area available for molecular adsorption.This change in morphology is attributable to the reduction in surface diffusion and mobility of the silver atoms at cryogenic temperatures.At cryogenic temperatures, the atoms freeze on impact at the landing site.This results in very small clusters of atoms on which the columns evolve.At higher temperatures, atoms have enough energy to migrate some distance away from their landing sites resulting in larger nucleation clusters.Even after the column growth has begun, higher growth temperatures can induce a greater atomic diffusion on the top of each column.This effect is often observed as an enlargement of the column diameter with increasing height, like a mushroom structure.The increase in surface area of the 100 K grown sample was estimated to be about 214% compared to room temperature sample.Further experimental results of the temperature effects can be found in [13].
COMSOL MultiPhysics Modelling.
For computing the SERS enhancement factor a three-dimensional full wave simulation using COMSOL MultiPhysics was performed.In the simulation, one unit cell which consists of one individual nanorod is modeled by setting a periodic boundary condition.633 nm wavelength was used as the excitation light source, which is the same as the one used in the subsequent experimental SERS measurements.The optical properties of silver were taken from the experiment data of Johnson and Christy [14].The normal incident plane wave has the polarization direction along the direction of the nanorod array as shown in Figure 3. Figure 4 shows normalized electric field distribution of the 100 K and 300 K AgNRs.The scale bar on the right represents the normalized amplitude of the scattered electric field with respect to the incident electric field amplitude.
The interaction between the nanostructured metallic surface and the incident light source generates a localized surface plasmon resonance (LSPR) [15].This causes an enhancement in the electric field distribution along the nanorod surface for both 100 K and 300 K AgNRs.As seen in Figure 4(c), the LSPR forms a standing wave along each nanorod [16,17].Due to this standing wave, the hot spots were not only observed on the tip of individual nanorods but were also observed along the sidewalls of the nanorods.
Based on the electric field enhancement observed in the 100 K grown AgNRs compared to the 300 K AgNRs, a parametric study was done by varying the spatial distances ( ) between the adjacent nanorods for the 100 K configuration. Figure 5 shows the results of this parametric study.The change in spatial distance affects the plasmonic resonance between the two adjacent nanorods along the direction.This effect causes a change in the amplitude and frequency of the standing waves along the nanorod direction, which causes a change in the location and magnitude of the hot spots.
The electric field distribution along the nanorods surface was used to calculate the surface average SERS enhancement factor ( SERS ) [ where and stokes , pumps , and 0 are the enhanced Raman, enhanced incident electric field, and the incident electric field, respectively. is the localized SERS enhanced factor. 0 is the corresponding unstructured area ( × ).The calculated enhancement factor for the 100 K grown AgNRs was SERS = 1.85 × 10 6 and that of 300 K grown AgNRs was SERS = 6.22 × 10 5 .Hence the enhancement factor is larger by approximately a factor of 3 for the 100 K sample compared to the 300 K sample.
SERS Measurement Results
. Figure 6 shows the SERS spectra collected from 100 K and 300 K AgNRs that were incubated with 4-ABT vapor for 5, 20, 30, and 60 minutes.It is clear that 4-ABT vapor was detected for all exposure times and the SERS signal intensities from 100 K AgNRs are significantly larger than the 300 K AgNRs.During the sample incubation with 4-ABT vapor, a plain silicon substrate was also incubated.Raman spectra from the silicon sample showed no detectable presence of 4-ABT molecules.The key difference between the AgNRs at 100 K and 300 K is the ability to resolve and detect all representative 4-ABT peaks.For samples that were exposed for 30 and 60 minutes, it appears that the SERS spectra from 100 K AgNRs distinctly detect and resolve all 4-ABT representative peaks.This behavior may be explained by the increase in nanostructured surface area available in the 100 K grown AgNRs compared to the 300 K sample.Also the increased plasmonic activity and electric field enhancement provides another cause for the increase in molecular detection sensitivity in the 100 K AgNRs compared to the 300 K AgNRs.The 300 K AgNRs have a collapsed physical structure as observed in Figure 2(b).This will prevent the molecules from being absorbed along the sidewall because less surface area is presented to the molecules.The SERS signal intensity from the 5-minute exposed sample appears higher and has shown a different baseline compared to the 20-, 30-, and 60minute exposure samples.This was presumably due to the 4-ABT molecules degradation or reactivity with the ambient prior to adsorption on the AgNRs surface.Nevertheless, it is clear that even a 5-minute exposure to 4-ABT vapor has been detected with a significantly stronger SERS response with 100 K grown AgNRs compared to that of the 300 K sample.
Based on the SERS response from 5-minute-exposed sample, it might be possible to improve the lower limit of detection by decreasing the exposure time even further and still be able to acquire SERS spectra with high signal to noise ratio.However, we have not explored this lower limit in this study.SERS spectra from unexposed 100 K and 300 K samples were also acquired prior to incubation with 4-ABT vapor.These spectra typically have shown varied amounts of Raman peaks indicative of unidentified organics present on the surface of AgNRs.The intensity of these Raman peaks was higher on 100 K AgNRs samples compared to 300 K samples.The two possible sources for these organics were the substrate cleaning using organic solvents (acetone, isopropyl alcohol, and methanol) prior to the growth and the exposure of samples to the ambient environment.However the level of these Raman peaks was buried in the baseline of the Raman spectra once the samples were exposed to 4-ABT.This indicates that the contribution of noise due to the unidentified organic peaks to the acquired 4-ABT spectra is practically negligible.Raman spectra from the silicon samples incubated concurrently with the AgNRs samples showed no evidence of 4-ABT molecules on the surface.The spectra showed a strong peak at 521 cm −1 indicative of the silicon substrate.Considering the minimal interference of background Raman spectra from AgNRs samples, 1080 cm −1 SERS peak was chosen as the representative peak to compare the integrated area intensities.An average of 282% improvement in SERS response was observed in 100 K grown AgNRs compared to that of 300 K grown AgNRs.In the present study, the AgNRbased nanostructured surface has shown high selectivity for detecting 4-ABT molecule even in the presence of low concentrations of organics and other impurities in the ambient environment.Furthermore, we have not examined the reusability of the substrates, but we speculate that it can potentially be done by heating the substrate to drive off the adsorbed molecules.
Conclusions
In summary, the improved morphology of Ag nanorods thin films grown at 100 K temperature has a strong influence on the enhanced plasmonic activity compared to room temperature grown samples.We have demonstrated this with 4-ABT test probe molecules and also validated the results with FEM simulation results.Although we only demonstrated this with 4-ABT molecules, any organic molecules that adsorb easily to the AgNRs surface and have distinct Raman spectra are effective candidates for detection using this nanostructured surface.
Figure 1 :
Figure 1: Schematic showing experimental setup used for Ag nanorods thin films incubation with 4-ABT vapors.
Figure 2 :
Figure 2: Side view of SEM results of (a) AgNRs at 300 K and (b) AgNRs at 100 K.
Figure 3 :
Figure 3: 3D schematic showing physical parameters of silver nanorods used in the simulation.
Figure 4 :
Figure 4: 3D normalized electric field distributions of nanorod grown at (a) 100 K and (b) 300 K; 2D electric field distributions of nanorod grown at (c) 100 K and (d) 300 K.
Figure 5 :
Figure 5: SERS enhancement factor as a function of different spatial distances ( ) between the adjacent nanorods grown at 100 K.
Table 1 :
Morphological comparison of silver nanorods thin films grown at 100 K and 300 K. 18] | 3,617.6 | 2013-12-23T00:00:00.000 | [
"Physics"
] |
FLUKA studies of hadron-irradiated scintillating crystals for calorimetry at the High-Luminosity LHC
Calorimetry at the High-Luminosity LHC (HL-LHC) will be performed in a harsh radiation environment with high hadron fluences. The upgraded CMS electromagnetic calorimeter design and suitable scintillating materials are a focus of current research. In this paper, first results using the Monte Carlo simulation program FLUKA are compared to measurements performed with proton-irradiated LYSO, YSO and cerium fluoride crystals. Based on these results, an extrapolation to the behavior of an electromagnetic sampling calorimeter, using one of the inorganic scintillators above as an active medium, is performed for the upgraded CMS experiment at the HL-LHC. Characteristic parameters such as the induced ambient dose, fluence spectra for different particle types and the residual nuclei are studied, and the suitability of these materials for a future calorimeter is surveyed. Particular attention is given to the creation of isotopes in an LYSO-tungsten calorimeter that might contribute a prohibitive background to the measured signal.
Motivation
The Compact Muon Solenoid (CMS [1]) detector will be upgraded for the High-Luminosity LHC (HL-LHC). The HL-LHC is scheduled for 2022 and will provide a five times higher instantaneous luminosity (L = 5 · 10 34 cm −2 s −1 ) than the LHC. Calorimetry at the HL-LHC will be performed in a harsh radiation environment with high hadron fluences, where typical doses in the forward region will be around 30 Gy/h at |η| = 2.6. The current forward electromagnetic calorimeter (ECAL endcaps, EE) is designed for LHC running with an integrated luminosity of 500 fb −1 . In the forward region, this calorimeter consists of 2 × 61200 Lead Tungstate (PbWO 4 ) crystals and covers a region of 1.48 < |η| < 3.0. The PbWO 4 crystals show a cumulative hadroninduced damage, which degrades the energy resolution [3] and makes a replacement of the EE for the HL-LHC desirable. One option for an upgraded forward ECAL would be a sampling calorimeter with alternating layers of absorber (tungsten), and active plates of a heavy inorganic scintillator. Suitable candidates are cerium fluoride (CeF 3 ), cerium-doped lutetium orthosilicate (Ce 2x (Lu 1−y Y y ) 2(1−x) SiO 5 -LYSO) and cerium-doped yttrium orthosilicate (Ce 2x Y 2(1−x) SiO 5 -YSO). These sampling calorimeter designs have to be studied in terms or radiation properties under HL-LHC conditions, using the Monte Carlo simulation program FLUKA ( [6], [7]) which is optimized for such studies.
Comparison of FLUKA simulation to proton irradiated crystals
One goal of the following FLUKA radio-activation simulation is to validate the FLUKA description of the respective materials with measurements. The results of this study are Figure 1. Measurements of induced ambientdose equivalent rateḢ * (10) ind at 5.7 cm from the side of the crystals as a function of time after irradiation, compared to FLUKA simulations for a fluence Φ p = 10 13 p/cm 2 . The measurements are affected by a relative scale uncertainty of 7 % due to the precision of fluence determination [4]. described in detail in [4]. A LYSO sample with the dimension of 25 × 25 × 100 mm 3 , corresponding to 8.8 X 0 in length, was irradiated with 24 GeV/c protons up to a fluence of Φ p = (8.85 ± 0.62) × 10 12 cm −2 at the CERN PS T7 beam line. Since the fluence in this test has been delivered to the samples over few hours, the rate values measured here are not representative for what would be observed in situ [4]. However, the comparison between different crystal types is crucial to anticipate the expected exposure. The induced ambient-dose equivalent rate ("dose") H * (10) ind has been measured at 5.7 cm distance from the middle of the LYSO sample at various times after irradiation. The measurements are compared with results from a FLUKA simulation (version 2011.2b.3), after rescaling to Φ p = 10 13 p/cm 2 . A very good agreement is observed in Fig. 1, between FLUKA simulations and measurements, over one year and over two orders of magnitude in dose, with the simulation results falling slightly on the low side. Uncertainties in the comparison can arise from the limited knowledge of Lutetium cross-sections in FLUKA, and from non-uniformities of the beam intensity during irradiation combined with the self-shielding of the crystal. The comparison benchmarks the reliability and predictive power of FLUKA for such LYSO crystals. In a second step, FLUKA simulations have been performed to compare the expected remnant radioactivity among crystals of similar dimensions in terms of radiation lengths (26 X 0 ), while keeping the granularity the same (24×24 mm 2 ). The description of CeF 3 , based on [5], was added to the comparison. The simulation results for PbWO 4 are compared to existing measurements [3]. Accordingly, the crystals were assumed to be proton-irradiated with a 20 GeV/c squared beam profile for an integrated fluence of Φ p = 10 13 p/cm 2 . TheḢ * (10) ind for each crystal was recorded as described in [4] and is depicted in Fig. 2. The FLUKA simulations for LYSO show a remnant dose similar to the one of PbWO 4 , with the LYSO results being roughly a factor two above the PbWO 4 results. This is an indication that LYSO might become slightly more radioactive than PbWO 4 in a 26 X 0 deep calorimeter. As the transverse dimensions are kept the same, CeF 3 has effectively a smaller density. This smaller amount of material leads to aḢ * (10) ind which is a factor two lower. In conclusion, the remnant dose rate for LYSO and CeF 3 is expected to be similar to PbWO 4 in a 26 X 0 transversely extended calorimeter.
FLUKA studies for a forward ECAL for CMS at the HL-LHC
This section studies the behavior of an electromagnetic calorimeter in the forward region of CMS at the HL-LHC. The simplified FLUKA geometry is optimized to study the overall behavior of the sampling calorimeter with respect to radiation properties.
Implementation of CMS ECAL in FLUKA simulation
The implementation of a sampling calorimeter in the EE is based on the available baseline FLUKA geometry version 1.0.0.0, corresponding to the situation prior to the long shutdown 1 Figure 2. Induced ambient dose equivalent ratė H * (10) ind at distance 5.7 cm from the side of the crystal as a function of time calculated with FLUKA for a LYSO and CeF 3 crystal exposed to a fluence Φ p = 10 13 p/cm 2 of 20 GeV/c protons.These simulations are compared to data and simulations for PbWO 4 from [3] for the same irradiation conditions (all crystals 26 X 0 long). The uncertainties are discussed in the text. Based on Fig.8 in [4] with updated results from [5]. Figure 4. Electron and positron fluence in the EE for an integrated luminosity of 3000 fb −1 as a function of depth expressed in X 0 using FLUKA. The fluence is averaged over a region of 48 cm < R < 109 cm, corresponding to 1.8 < η < 2.6 for z = 320 cm. The average fluence is similar for all options.
Particle fluence in the sampling calorimeter at the end of HL-LHC
The primary proton-proton collisions with an energy of 7 TeV per beam were simulated with the DPMJetIII event generator and normalized to an inelastic collision cross section of 80 mb, using FLUKA version 2011.2b.5. Precise thresholds for the particle transport were applied. The transport and production thresholds were set to 100 keV for electrons and positrons, and to 50 keV for photons in the EE. Every other particle was transported down to 100 keV and the low energy neutron transport down to thermal energies (10 −5 eV) was included. The electron and positron fluence in the EE for 3000 fb −1 as a function of depth, expressed in X 0 , is depicted in Fig. 4. The fluence is averaged over a region of 48 cm < R < 109 cm, corresponding to 1.8 < η < 2.6 for z = 320 cm. The z-coordinate was converted to X 0 by scaling with the effective radiation length X 0ef f , using a bin size of 10 % of X 0ef f . For all options the average fluence is similar and comparable to the values of the PbWO 4 crystals. Most of the electromagnetically interacting particles in a minimum bias environment have an energy of a few MeV up to 1 GeV. The fluence reflects the development of the electromagnetic shower of such particles.The maximum electromagnetic particle flux develops in the first 7 X 0 , particularly in the second scintillator layer. The effect due to the difference in density of the scintillator materials is clearly visible up to 11 X 0 . In conclusion, the predicted electron-positron fluence will be comparable to the one of the current PbWO 4 crystals after 3000 fb −1 . Further studies show that the magnitude and overall shape of other particle fluences are also similar among the different sampling calorimeter options.
Absorbed dose in the sampling calorimeter at the end of HL-LHC running
The absorbed dose is the deposited energy per unit mass of the medium. Fig. 5 shows the dose absorbed in the EE for an integrated luminosity of 3000 fb −1 as a function of the radiation length X 0 in the same region as above. Comparing the CeF 3 and YSO sampling calorimeters with the PbWO 4 calorimeter, the absorbed dose is overall comparable. The LYSO layers however show a significantly higher absorbed dose. This can be explained by the high thermal neutron capture cross section of 175 Lutetium (σ th.n. = 23.1 barn [2]). In Fig. 6 an estimation of the maximum neutron fluence energy spectrum, which occurs in the second scintillator layer at 2.2 < η < 2.6 (see Fig. 4) after 3000 fb −1 is depicted. In the region of low energy neutrons, the number and width of the bins is given by the thermal neutron cross section library in FLUKA. Thermal neutrons carry an energy up to a few eV, followed by a resolved resonance region. This region shows different resonances for each option due to the different elements in the materials. Because of the high thermal neutron capture cross section of Lutetium, the LYSO/W option shows a significantly lower fluence for thermal neutrons. This explains the higher absorbed dose in the LYSO layers in Fig. 5. Further FLUKA simulations show that the thermal neutron fluence in the low density regions adjacent to both sides of the sampling calorimeter is higher than the one in the scintillator materials. The enhancement is likely also due to the polyethylene layers at both sides, which moderate neutrons down to thermal energies, but do not absorb them. Due to the high σ th.n. of Lutetium, the absorbed dose rate in LYSO increases significantly at both ends of the calorimeter and in the scintillator layers. A thin absorber layer made out of boron polyethylene at both ends is advisable to moderate neutrons and absorb the thermal ones.
Study of a potentially prohibitive background due to decaying isotopes
The energy emitted by decaying isotopes might lead to a prohibitive background to the main electromagnetic signal. The LYSO/W calorimeter shows the highest dose values and is thus surveyed in a conservative scenario. Fig. 7 compares the absorbed dose rate (Ḋ) during HL-LHC running (1.44 × 10 13 pp-int./h) with theḊ by decaying particles at the irradiation stop after 2475 fb −1 . The irradiation schedule follows the official recommendations for a radioactivation simulation from LS3 until LS5. For the activation study, coalescence and evaporation of heavy fragments were enabled.Ḋ is visualized for two regions in R: 48 cm -71 cm and 109 cm -135 cm, corresponding to η = 1.6 -1.8 and 2.2 -2.6 at z = 320 cm. The inelastic cross section of the prompt electromagnetic shower is normalized to running conditions whereas the decay products are normalized to 100 mb, a conservative limit adopted for activation studies. Figure 7.
Absorbed dose rate due to the prompt electromagnetic shower and the decay products for an LYSO/W sampling calorimeter using FLUKA. The absorbed dose rate is projected to the calorimeter length in z, from 320 cm -332.7 cm.
In Fig. 7, the dose rateḊ in the higher η−region is on average one order of magnitude higher than the one of the lower η−region. In comparison, the dose rate of the decay products is at a constant 1 % level of the prompt particles. This offset of deposited energy by decaying isotopes is still at a manageable level for HL-LHC running. The overall activity of the created residual nuclei after 2475 fb −1 in the first scintillator layer at 2.2 < η < 2.6 is A = (4.53 ± 0.07) × 10 10 Bq. A percentage of 96 % of this activity comes from residual nuclei created by low energy neutrons. Furthermore, the dose rate of the decay products is dominated by short lived radio nuclides. The deposited energy by the decaying isotopes after 10 h is one third of the one after 2475 fb −1 . Responsible for this slight build-up is mainly the long lived isotope 177 Lutetium with an activity of A = (1.91 ± 0.03) × 10 10 Bq after 2475 fb −1 and a half life of τ = 6.64 d.
Conclusions
A possible forward electromagnetic sampling calorimeter for the CMS experiment at HL-LHC was studied in terms of radiation exposure. The dedicated FLUKA study shows that the average particle fluence and the absorbed dose are similar for all options considered. As the ionizing dose in a LYSO-tungsten sampling calorimeter increases significantly towards both ends, a thin boronpolyethylen layer at both ends would be advisable. The deposited energy by decaying isotopes is at a constant 1 % level of the prompt electromagnetic shower and is dominated by short-lived radioisotopes. All three designs, using tungsten as an absorber and CeF 3 , LYSO or YSO as the sensitive material are suitable candidates for a future sampling calorimeter at the HL-LHC in terms of activation levels. | 3,306.8 | 2015-02-13T00:00:00.000 | [
"Physics"
] |
StarDist Image Segmentation Improves Circulating Tumor Cell Detection
Simple Summary Automated enumeration of circulating tumor cells (CTC) from immunofluorescence images starts with a selection of areas containing potential CTC. The CellSearch system has a built-in selection algorithm that has been observed to fail in samples with high cell density, thereby underestimating the true CTC load. We evaluated the deep learning method StarDist for the selection of possible CTC. In whole blood sample images, StarDist recovered 99.95% of CTC detected by CellSearch and segmented 10% additional CTC. In diagnostic leukapheresis (DLA) samples, StarDist segmented 20% additional CTC and performed well, whereas CellSearch had serious failures in 9% of samples. Abstract After a CellSearch-processed circulating tumor cell (CTC) sample is imaged, a segmentation algorithm selects nucleic acid positive (DAPI+), cytokeratin-phycoerythrin expressing (CK-PE+) events for further review by an operator. Failures in this segmentation can result in missed CTCs. The CellSearch segmentation algorithm was not designed to handle samples with high cell density, such as diagnostic leukapheresis (DLA) samples. Here, we evaluate deep-learning-based segmentation method StarDist as an alternative to the CellSearch segmentation. CellSearch image archives from 533 whole blood samples and 601 DLA samples were segmented using CellSearch and StarDist and inspected visually. In 442 blood samples from cancer patients, StarDist segmented 99.95% of CTC segmented by CellSearch, produced good outlines for 98.3% of these CTC, and segmented 10% more CTC than CellSearch. Visual inspection of the segmentations of DLA images showed that StarDist continues to perform well when the cell density is very high, whereas CellSearch failed and generated extremely large segmentations (up to 52% of the sample surface). Moreover, in a detailed examination of seven DLA samples, StarDist segmented 20% more CTC than CellSearch. Segmentation is a critical first step for CTC enumeration in dense samples and StarDist segmentation convincingly outperformed CellSearch segmentation.
Introduction
The circulating tumor cell (CTC) load detected by the CellSearch system reflects the state of disease [1,2]. Accurate enumeration of CTC is important for its use as a biomarker in patient risk assessment and to evaluate treatment response [3,4] or disease progression [5,6] during the course of the disease by true changes in CTC counts. After sample processing and imaging, CTCs are identified in 140 four-channel immunofluorescence images. This process starts with the identification of events of interest in the images through segmentation [7]. Based on this segmentation, CTC candidates are selected and the corresponding thumbnail images are presented to the user. Ideally, the task of segmentation is to identify every single event of interest inside a sample. Segmentation should not join multiple events together, nor split single events into multiple ones. As cell density increases, the segmentation task becomes more challenging. In samples derived from diagnostic leukapheresis (DLA) and run on the CellSearch system [8,9] we noticed that in these cell dense samples the segmentation used by CellSearch failed to identify single objects, leading to artificially low CTC counts. A better segmentation was achieved by the open-source image analysis program ACCEPT [10]. However, ACCEPT employs an active contour method for segmentation and joins objects that are in contact with each other into a single event, thus failing to identify single objects in denser samples. StarDist is a Deep Learning based method that describes all events with a star-convex shape and which is quite effective in segmenting single cells in tissue sections [11]. Here, we evaluate StarDist for the segmentation of all objects as single events in CellSearch images corresponding to peripheral blood and DLA samples. Using StarDist, we demonstrate a clear improvement in the segmentation of CellSearch images leading to a more accurate CTC count.
Sample Archives to Evaluate
For the comparison of CellSearch and StarDist segmentations, 1163 CellSearch archives of previously scored CellSearch images were used. This set included peripheral blood samples of 90 healthy donors (NCT00133913 [12]), 442 castration-resistant prostate cancer patients (NCT00133900 [13]), and 601 DLA samples of prostate (n = 24), breast (n = 49), and non-small cell lung cancer patients (n = 528) [8,14,15]. The time between sample collection and sample preparation was known for 436 whole blood prostate cancer samples and for 322 of the DLA non-small cell lung samples. All study participants had signed informed consent forms per the Helsinki declaration and all protocols were approved by the Ethics Committees of the respective studies.
Stardist Segmentation
The StarDist method from github (https://github.com/stardist/stardist, accessed on 1 June 2021) was applied with a few non-default parameters. Specifically, for network "n_channel_in" was set to 3 and "u_net_n_depth" was set to 4, for optimization "train_background_reg" was set to 0.0004, "train_learning_rate" was set to 0.001 and train_batch_size" was set to 8, and for post-processing "nms_thresh" was set to 0.3 and "prob_thresh" was set to 0.3. The input images were locally contrast-enhanced. Training data were generated in QuPath 0.2.3 [16] using the procedure described in the StarDist documentation.
Performance in High Density Samples
The cell density of CellSearch samples varies greatly from sample to sample, with occasional high densities in blood samples and more often in DLA. To assess the performance of StarDist in higher density samples, we created samples with increasing amounts of magnetically labeled white blood cells and a fixed amount of tumor cells. For this, we magnetically labeled, separated, and fluorescently stained white blood cells from blood using the CellSearch system with the standard reagents, except for the EpCAM-ferrofluid, which was replaced by streptavidin ferrofluid (Biomagnetic solutions, State College, PA, USA) coupled to CD45-biotin. Cells of the prostate cancer cell line LNCaP were processed using the regular CellSearch reagents. The cell concentrations of the resulting samples were determined and an increasing number of white blood cells together with~6000 cells from the prostate cancer cell line LNCaP were spiked into a mixture of PBS and CS-fixative to reach a final volume of 325 µL. The spiked samples were manually placed in CellSearch cartridges and scanned using the CellTracks Analyzer II system. In our experience, an increase in the time between blood draw and sample prep causes a higher white blood cell carry-over during enrichment and therefore a higher cell density in the resulting samples, while as mentioned DLA samples show an even higher cell density. To test this assumption, we evaluated the number of segmented events by StarDist using 436 whole blood and 322 DLA samples for which the time between sample collection and sample prep was known.
Recovery of CellSearch CTC by StarDist and Potential Gain in CTC
As CellSearch is the gold standard in the enumeration of CTC, we compared the StarDist segmentations to those from CellSearch. The CellSearch segmentation draws rectangular segmentations with a 10-pixel margin around the event, whereas StarDist draws a star-convex segmentation without any margin. The vast majority of StarDist segmented events are cytokeratin-PE negative white blood cells, whereas CellSearch only segments events that express some cytokeratin-PE and some DAPI. A direct comparison of segmentations is therefore not meaningful. To ensure that the StarDist algorithm does not cause any losses of events that were scored by a reviewer as CTC in CellSearch, we selected CellSearch segmentations in which a CTC was detected (with a maximum of 20 from a single sample). Together with the corresponding StarDist segmentation, these were shown to a reviewer, who evaluated whether (1) the CTC found by CellSearch were segmented by StarDist, and (2) whether the segmentations were correct. A correct segmentation outlines the whole event and only the event, so it does not split a single event into two segmentations and it does not join multiple events together. For CTC enumeration by a human reviewer, a CTC just needs to be included in the segmentation; however, for automated enumerations, a correct outline is also important.
To assess whether StarDist segments any CTC that were not detected by CellSearch, we presented to four reviewers all events detected in 442 blood samples as well as seven DLA samples that were CTC candidates based on their staining properties, but which were not selected by the CellSearch segmentation and therefore had not been evaluated by the original reviewer. For this purpose, a possible CTC was defined as an event meeting all of the following requirements: cytokeratin-PE intensity maximum >75 and mean >45, DAPI intensity maximum >60 and mean >50, CD45-APC mean intensity <50 or mean PE intensity >1000 (because some crosstalk exists from PE to the APC channel), and a total event area between 36 and 1000 pixels, with a stained area in PE of at least 30 pixels. These boundaries were loosely based on an existing ACCEPT CTC definition [17]. We do not expect that this definition will include all CTC, and we know that the majority of events included by this definition are not CTC. To assess if the more sensitive StarDist segmentation would lead to an increase in false positives, we also performed this selection and evaluation for 93 healthy donor samples for which the CellSearch results were reported previously [9].
Code Environment
All trainings and evaluations were performed in Python 3.7, utilizing StarDist 0.6.2 [11]
Performance of Segmentation Algorithms on Dense CellSearch Images
Segmentation algorithms developed for CellSearch images include the built-in algorithm and the active contour method employed by ACCEPT. However, both the ACCEPT as well as the CellSearch segmentation methods were developed for CTC samples in which the cell density is such that adjacent cells do not touch each other. In high cell density samples, including most DLA samples and some whole blood samples, the segmentation algorithms join all cells that are in contact with each other into a single event, as shown in Figure 1. In addition, CellSearch seems to miss cells that should have been presented to a reviewer. In contrast, StarDist achieved segmentation of almost all single events, as illustrated in Figure 1. An overview of the capabilities of the different methods is given in Table 1. shown in Figure 1. In addition, CellSearch seems to miss cells that should have been presented to a reviewer. In contrast, StarDist achieved segmentation of almost all single events, as illustrated in Figure 1. An overview of the capabilities of the different methods is given in Table 1. The second column shows the same image in black and white with the yellow rectangular Cell-Search segmentation areas. In the third and fourth columns, the ACCEPT and StarDist segmentations are depicted as a separate color for each segmented event. StarDist clearly performs best in segmenting all single events separately, especially in areas containing a high cell density. The second column shows the same image in black and white with the yellow rectangular CellSearch segmentation areas. In the third and fourth columns, the ACCEPT and StarDist segmentations are depicted as a separate color for each segmented event. StarDist clearly performs best in segmenting all single events separately, especially in areas containing a high cell density.
Extent of the Problem in CellSearch Segmentation
In the CellSearch segmentation algorithm, a single cell of 20 µm diameter will result in a segmentation of 2500 pixels. Therefore, we reasoned that those segmentations larger than 2500 pixels, the equivalent of 32 × 32 µm 2 , are likely to contain more than one cell. For these sizes, it becomes probable that multiple CTCs are contained within a single segmentation, leading to an undercount of CTCs. Furthermore, when the segmentations get even larger, the review becomes more difficult due to presentation issues, likely leading to misclassifications. Presentation issues become noticeable when segmentations are around 70 × 70 µm 2 (equivalent to 11,881 pixels) and become much worse for larger segmentations. The occurrence of 'regular' (≤2500 pixels), larger than a cell (2500 to 25,000 pixels), too large to review (25,000-100,000), and very large segmentations (>100,000 pixels), as well as the area of the cartridge and the segmented area consisting of these segmentation sizes, are listed in Table 2. Table 2. CellSearch segmentation sizes in blood and DLA samples. The percentage of samples with different segmentation sizes present in whole blood and DLA samples, the mean percentage of the total cartridge area covered, and the mean percentage of the total segmented area by the different CellSearch segmentation sizes. The ranges are shown in parentheses. Overall, in whole blood of cancer patients, 40% of segmentations are larger than a large single cell, and 0.2% of segmentations are so large that review is possibly affected. In whole blood of healthy donors, 35% of segmentations are larger than a large single cell, and 0.07% of segmentations are so large that review is possibly affected. Very large segmentations of more than 100,000 pixels were observed in 9% (55 of 601) of DLA samples, comprising up to 70% of the segmented area. One out of 442 whole blood samples from cancer patients also had very large segmentations, which comprised 0.7% of the segmented area. For these cartridges, the accuracy of the CTC count is compromised. To illustrate these sizes, Figure 2 shows an example for a segmentation of approximately 2500, 25,000, and 160,000 pixels.
Diagnostic
For these sizes, it becomes probable that multiple CTCs are contained within a single segmentation, leading to an undercount of CTCs. Furthermore, when the segmentations get even larger, the review becomes more difficult due to presentation issues, likely leading to misclassifications. Presentation issues become noticeable when segmentations are around 70 × 70 µ m 2 (equivalent to 11,881 pixels) and become much worse for larger segmentations. The occurrence of 'regular' (<=2500 pixels), larger than a cell (2500 to 25,000 pixels), too large to review (25,000-100,000), and very large segmentations (>100,000 pixels), as well as the area of the cartridge and the segmented area consisting of these segmentation sizes, are listed in Table 2. Table 2. CellSearch segmentation sizes in blood and DLA samples. The percentage of samples with different segmentation sizes present in whole blood and DLA samples, the mean percentage of the total cartridge area covered, and the mean percentage of the total segmented area by the different CellSearch segmentation sizes. The ranges are shown in parentheses. Overall, in whole blood of cancer patients, 40% of segmentations are larger than a large single cell, and 0.2% of segmentations are so large that review is possibly affected.
Diagnostic
In whole blood of healthy donors, 35% of segmentations are larger than a large single cell, and 0.07% of segmentations are so large that review is possibly affected. Very large segmentations of more than 100,000 pixels were observed in 9% (55 of 601) of DLA samples, comprising up to 70% of the segmented area. One out of 442 whole blood samples from cancer patients also had very large segmentations, which comprised 0.7% of the segmented area. For these cartridges, the accuracy of the CTC count is compromised. To illustrate these sizes, Figure 2 shows an example for a segmentation of approximately 2500, 25,000, and 160,000 pixels.
Cell Density
To assess the performance of StarDist in cartridges with various cell densities, spikein experiments were performed. The number of cells spiked was approximately 6000
Cell Density
To assess the performance of StarDist in cartridges with various cell densities, spike-in experiments were performed. The number of cells spiked was approximately 6000 LNCaP cancer cells together with zero, 50-, 100-, 150-, 200-, 300-and 400-thousand white blood cells. In a CellSearch cartridge and scan, 400,000 cells correspond to~3000 cells per image and about 50% of the surface area covered by cells. Larger cell concentrations are not evaluable as the CellTracks Analyzer II fails to perform autofocus in denser samples. Figure 3 shows the result of these spike-in experiments. It can be seen that at very low concentrations, StarDist detects an excess of approximately 30.000, mostly small, events, while a good correlation is seen up to the maximal density. Although useful to evaluate the ability of StarDist to segment high-density samples, these titrations do not contain the clumped and broken cells often seen in high-density DLA samples.
image and about 50% of the surface area covered by cells. Larger cell concentrations are not evaluable as the CellTracks Analyzer II fails to perform autofocus in denser samples. Figure 3 shows the result of these spike-in experiments. It can be seen that at very low concentrations, StarDist detects an excess of approximately 30.000, mostly small, events, while a good correlation is seen up to the maximal density. Although useful to evaluate the ability of StarDist to segment high-density samples, these titrations do not contain the clumped and broken cells often seen in high-density DLA samples. To assess the dependence between white blood cell carry over, causing high cartridge densities, and the time between sample collection and sample preparation, we calculated the number of events in 436 blood as well as 322 DLA samples. In Figure 4 the empirical cumulative distribution functions (CDF) for the number of StarDist events per sample is shown for blood as well as DLA samples with one, two, three, and four days between sample collection and sample prep. There is a large variation in the number of events per sample and it can be seen that for blood the total number of events increases when there is more than one day between the sample collection and sample prep. In fact, for whole blood, there is first-order stochastic dominance of the empirical CDF for one day between collection and prep and the empirical CDFs for two, three, and four or more days. The median number of events for whole blood samples with one day between collection and prep was 13-thousand, in contrast to 28-, 37-and 54-thousand events for two, three, and four or more days respectively. DLA samples have a much less pronounced dependence on the time between sample collection and sample prep. For DLA samples, the median number of events was 108-, 136-, 138-and 156-thousand for one, two, three, and four or more days respectively. To assess the dependence between white blood cell carry over, causing high cartridge densities, and the time between sample collection and sample preparation, we calculated the number of events in 436 blood as well as 322 DLA samples. In Figure 4 the empirical cumulative distribution functions (CDF) for the number of StarDist events per sample is shown for blood as well as DLA samples with one, two, three, and four days between sample collection and sample prep. There is a large variation in the number of events per sample and it can be seen that for blood the total number of events increases when there is more than one day between the sample collection and sample prep. In fact, for whole blood, there is first-order stochastic dominance of the empirical CDF for one day between collection and prep and the empirical CDFs for two, three, and four or more days. The median number of events for whole blood samples with one day between collection and prep was 13-thousand, in contrast to 28-, 37-and 54-thousand events for two, three, and four or more days respectively. DLA samples have a much less pronounced dependence on the time between sample collection and sample prep. For DLA samples, the median number of events was 108-, 136-, 138-and 156-thousand for one, two, three, and four or more days respectively.
Recovery of CellSearch CTC by StarDist and Potential Gain in CTC
We evaluated 1948 CTC identified by reviewers after CellSearch segmentation to verify that these CTC were also segmented by StarDist and assessed whether the segmentation was within a few pixels of the true outline. We found that one CTC was not segmented by StarDist, resulting in a StarDist segmentation of 99.95% of CTC segmented by CellSearch. Of the segmentations, 14 (0.7%) were split into multiple events, while 8 (0.4%) segmentations included only part of the event. In 11 (0.6%) segmentations, the CTC was merged with (part of) an adjacent object into a single event. Such splitting and merging errors are not problematic for a human reviewer, but may cause automated enumeration methods to exclude such events, leading to a lower recovery of CTC.
Recovery of CellSearch CTC by StarDist and Potential Gain in CTC
We evaluated 1948 CTC identified by reviewers after CellSearch segmentation to verify that these CTC were also segmented by StarDist and assessed whether the segmentation was within a few pixels of the true outline. We found that one CTC was not segmented by StarDist, resulting in a StarDist segmentation of 99.95% of CTC segmented by CellSearch. Of the segmentations, 14 (0.7%) were split into multiple events, while 8 (0.4%) segmentations included only part of the event. In 11 (0.6%) segmentations, the CTC was merged with (part of) an adjacent object into a single event. Such splitting and merging errors are not problematic for a human reviewer, but may cause automated enumeration methods to exclude such events, leading to a lower recovery of CTC.
To determine the extent to which CTCs were missed by the CellSearch segmentation, we reviewed a total of 36,996 events from 442 whole blood image archives that were CTC candidates based on their staining properties, but had not been reviewed in the original review because they were not segmented by CellSearch. Of these events, 762 were CTC by consensus of a panel of four reviewers. Ordinary least-squares linear regression showed an average increase of 8.7% of the number of CTC found in the original review plus 0.214 CTC (CTCStarDist = 1.087 × CTCCellSearch + 0.214, R 2 = 0.99). Of the 218 samples that originally had zero CTC, 20 samples (9.2%) gained at least one CTC after implementing StarDist segmentation and CTC review. In two of the 442 samples, the increase in CTC led to a conversion from a 'lower risk' (<5 CTC) to 'a higher risk' (≥5 CTC) group, see Figure 5. Additionally, in seven image archives of prostate cancer DLA samples we reviewed a total of 4666 events that were CTC candidates based on their staining properties, but were not reviewed because they were not segmented by CellSearch. Of these events, 219 were CTC by consensus of the reviewers. Ordinary least squares linear regression showed an average increase of 19.8% of the number of CTC found in the original review plus 1.136 CTC (CTCStarDist = 1.198 × CTCCellSearch + 1.136, R 2 = 0.99). To determine the extent to which CTCs were missed by the CellSearch segmentation, we reviewed a total of 36,996 events from 442 whole blood image archives that were CTC candidates based on their staining properties, but had not been reviewed in the original review because they were not segmented by CellSearch. Of these events, 762 were CTC by consensus of a panel of four reviewers. Ordinary least-squares linear regression showed an average increase of 8.7% of the number of CTC found in the original review plus 0.214 CTC (CTC StarDist = 1.087 × CTC CellSearch + 0.214, R 2 = 0.99). Of the 218 samples that originally had zero CTC, 20 samples (9.2%) gained at least one CTC after implementing StarDist segmentation and CTC review. In two of the 442 samples, the increase in CTC led to a conversion from a 'lower risk' (<5 CTC) to 'a higher risk' (≥5 CTC) group, see Figure 5. Additionally, in seven image archives of prostate cancer DLA samples we reviewed a total of 4666 events that were CTC candidates based on their staining properties, but were not reviewed because they were not segmented by CellSearch. Of these events, 219 were CTC by consensus of the reviewers. Ordinary least squares linear regression showed an average increase of 19.8% of the number of CTC found in the original review plus 1.136 CTC (CTC StarDist = 1.198 × CTC CellSearch + 1.136, R 2 = 0.99).
In image archives from 90 healthy donor samples, two events segmented by StarDist but not by CellSearch were selected by the reviewers as CTC. Human reviewers had previously identified three CTC in these samples. Taken together, 5.4% of the healthy donors had one (false positive) CTC, while none had two or more. This percentage is in line with the 5.5% of healthy donor samples containing one CTC-like event as found in the original CellSearch study [24] and does not indicate a large number of false positives as a result of the segmentation.
In image archives from 90 healthy donor samples, two events segmented by StarDist but not by CellSearch were selected by the reviewers as CTC. Human reviewers had previously identified three CTC in these samples. Taken together, 5.4% of the healthy donors had one (false positive) CTC, while none had two or more. This percentage is in line with the 5.5% of healthy donor samples containing one CTC-like event as found in the original CellSearch study [24] and does not indicate a large number of false positives as a result of the segmentation. As a sample archive consists of a number of adjacent images, some events will be present on the image edges. The segmentation of events on an edge is more challenging for segmentation algorithms that employ pixels surrounding the event, for example in the local contrast enhancement employed in CellSearch. Additionally, the non-uniformity of the sample illumination means events near the edge have lower signal to noise, which could reduce the likelihood of being segmented. To assess if the events missed by Cell-Search are predominantly present on the edges of the image, we created a heat map consisting of 25 × 31 bins (~25 × 25 µ m per bin) displaying the locations of the CTC segmented As a sample archive consists of a number of adjacent images, some events will be present on the image edges. The segmentation of events on an edge is more challenging for segmentation algorithms that employ pixels surrounding the event, for example in the local contrast enhancement employed in CellSearch. Additionally, the non-uniformity of the sample illumination means events near the edge have lower signal to noise, which could reduce the likelihood of being segmented. To assess if the events missed by CellSearch are predominantly present on the edges of the image, we created a heat map consisting of 25 × 31 bins (~25 × 25 µm per bin) displaying the locations of the CTC segmented only by StarDist. In the heatmap, as shown in Figure 6, it can be seen that although additional events are found throughout the image, there is a higher incidence at the edges, with 36.6% of events being present in an edge bin, compared to 14.5% expected for a uniform distribution. The highest concentration of missed events is on the left and top edges, which could be attributed to the implementation details of the CellSearch algorithm. only by StarDist. In the heatmap, as shown in Figure 6, it can be seen that although additional events are found throughout the image, there is a higher incidence at the edges, with 36.6% of events being present in an edge bin, compared to 14.5% expected for a uniform distribution. The highest concentration of missed events is on the left and top edges, which could be attributed to the implementation details of the CellSearch algorithm.
Discussion
For the classification of objects in fluorescence images, single objects need to be identified and presented to a reviewer, an algorithm, or a combination of both. During the development of the CellSearch system, an algorithm segmenting image sections containing both DAPI and CK-PE staining was chosen to facilitate CTC enumeration. These sections are presented to a reviewer who decides which ones contain a CTC. Failure to segment a CTC means this CTC will not be presented to the reviewer, and thus, not accounted for in the final CTC count. Furthermore, incorrect segmentations can contain two or even more clearly separated CTC. Another possibility is that the segmentations are too large to be reviewed effectively, as they can encompass a whole image or more. Such failures can also lead to underestimation of the real CTC load, albeit less dramatically than when the CTC is not presented at all.
Here, we evaluated StarDist as an alternative to the CellSearch segmentation and demonstrated that it would lead to the segmentation of on average 8.7% + 0.21 more CTC in whole blood, while barely missing any (0.05%) of the CTC detected using the CellSearch segmentation. In DLA samples, the advantage of StarDist is even more profound, since we detected on average 19.8% + 1.14 more CTC, albeit in a small sample size of only seven DLA samples. For the current evaluation, a pre-selection of CTC candidates was made that likely missed some CTC. The average 10.5% CTC gain found with this selection thus represents a lower bound for the number of CTC not segmented by the original CellSearch segmentation. Using a deep learning classification algorithm currently in development [25], we estimate the upper bound of the average CTC gain due to improved segmentation to be 38%. Considering the large number of events segmented per sample (mean 54 • 10 3 ) using the StarDist method, a review of all segmented events is neither feasible nor meaningful. If StarDist was applied in the current CellSearch workflow, a pre-selection of events to be presented to the reviewer would be needed to reduce the number of candidate CTC. Fully automated classification could be performed using a fully designed gating strategy as applied in ACCEPT [17], or through a deep learning approach as presented previously [25]. The main advantage of the ACCEPT approach is that it is relatively simple and thus interpretable. The main downside is that it has a limited ability to encode more complex classification rules. A deep learning approach would be able to encode very complex classification rules, and the major downside is that these classification rules are typically not interpretable. While deep learning classifiers have been shown to outperform
Discussion
For the classification of objects in fluorescence images, single objects need to be identified and presented to a reviewer, an algorithm, or a combination of both. During the development of the CellSearch system, an algorithm segmenting image sections containing both DAPI and CK-PE staining was chosen to facilitate CTC enumeration. These sections are presented to a reviewer who decides which ones contain a CTC. Failure to segment a CTC means this CTC will not be presented to the reviewer, and thus, not accounted for in the final CTC count. Furthermore, incorrect segmentations can contain two or even more clearly separated CTC. Another possibility is that the segmentations are too large to be reviewed effectively, as they can encompass a whole image or more. Such failures can also lead to underestimation of the real CTC load, albeit less dramatically than when the CTC is not presented at all.
Here, we evaluated StarDist as an alternative to the CellSearch segmentation and demonstrated that it would lead to the segmentation of on average 8.7% + 0.21 more CTC in whole blood, while barely missing any (0.05%) of the CTC detected using the CellSearch segmentation. In DLA samples, the advantage of StarDist is even more profound, since we detected on average 19.8% + 1.14 more CTC, albeit in a small sample size of only seven DLA samples. For the current evaluation, a pre-selection of CTC candidates was made that likely missed some CTC. The average 10.5% CTC gain found with this selection thus represents a lower bound for the number of CTC not segmented by the original CellSearch segmentation. Using a deep learning classification algorithm currently in development [25], we estimate the upper bound of the average CTC gain due to improved segmentation to be 38%. Considering the large number of events segmented per sample (mean 54 × 10 3 ) using the StarDist method, a review of all segmented events is neither feasible nor meaningful. If StarDist was applied in the current CellSearch workflow, a pre-selection of events to be presented to the reviewer would be needed to reduce the number of candidate CTC. Fully automated classification could be performed using a fully designed gating strategy as applied in ACCEPT [17], or through a deep learning approach as presented previously [25]. The main advantage of the ACCEPT approach is that it is relatively simple and thus interpretable. The main downside is that it has a limited ability to encode more complex classification rules. A deep learning approach would be able to encode very complex classification rules, and the major downside is that these classification rules are typically not interpretable. While deep learning classifiers have been shown to outperform fully designed classifiers in many classification tasks, deep learning classifiers require a large number of annotated segmented image sections before they are reliable and can be quirky when trained on insufficient data. For the identification of clusters, it is expected that regardless of the used method also the properties of the surrounding events will need to be taken into account during classification, as StarDist segments adjacent cells into separate segmentations.
Recent data revealed that the enumeration of tdEV in the original CellSearch images further improves the prognostic stratification of CTC [26]. The CellSearch segmentation algorithm however does not present the tdEV for manual review. This shortcoming was overcome by the introduction of ACCEPT which allows gating and enumeration of tdEVs [26]. ACCEPT employs a Bregman active contour method that finds the outline of each event. The event outline permits quantitative characterization of the event in terms of signal intensity (e.g., mean intensity and max intensity) and morphology (e.g., area and eccentricity). These extracted values are useful for the identification of CTC [17] as well as for the quantification of marker expression on the CTC [27] which can for instance be used to identify epithelial to mesenchymal transition (EMT). We expect that this identification and characterization will also be possible with StarDist segmentations as they are sufficiently accurate event outlines, and the segmented events include tdEV, bare nuclei, and white blood cells. Furthermore, in high cell densities, as found in some whole blood samples and most DLA samples, StarDist continues to perform well while ACCEPT fails to find the outline of single events.
The StarDist network was optimized to detect all cell types as well as tdEVs in the CellSearch archives, because we wish to investigate the prognostic potential of all sample constituents. Another possible approach for CellSearch samples would be to train two networks separately for the identification of CTC and tdEV. This may be a more performant approach if the aim is only to enumerate CTC or tdEV. However, the current training allowed us to look at the impact of sample age on the total number of cells in a cartridge.
Here we found that for whole blood the number of events in a cartridge is on average more than two-fold higher when more than one day has passed between sample collection and preparation. No relationship was observed between sample age and the total number of CTC nor on the number of CTC segmented only by StarDist. This suggests non-specific binding for white blood cells is increased in older samples. For DLA samples, a similar, albeit relatively smaller, effect could be observed.
The star-convex model that is applied by StarDist also has its limitations, some examples of which are shown in Figure 7A-C. Panel A shows two cells that are difficult to describe with a star-convex model, and StarDist splits these cells into multiple segmentations. Such cells are relatively rare in CTC samples, but do mean that StarDist could not be applied for the segmentation of circulating endothelial samples. Furthermore, Panel B shows very faint events that are close to bright events. These are false negative events in both StarDist and CellSearch. This behavior could be improved for StarDist by reducing the local window size used in pre-processing, but this would result in false-negatives for areas with densely packed cells such as CTC clusters. Panel C shows the plastic edge of a CellSearch cartridge, where StarDist segments small variations in autofluorescence.
In the aforementioned cases, the true segmentation can be easily identified by the human reviewer, but there are also instances in which the true segmentation is difficult to ascertain. Panel D in Figure 7 shows some examples of events where StarDist does perform a segmentation, but the human reviewer cannot determine whether this segmentation is correct or not.
Conclusions
Here, we demonstrated the occurrence of critical failures in samples segmented by the built-in CellSearch segmentation algorithm. The ACCEPT algorithm developed for CellSearch samples did not perform well in dense (DLA) samples, because it joined nearby objects together into a single event. To overcome this we evaluated the StarDist deeplearning-based segmentation method and found it outperforms both ACCEPT and the current CellSearch segmentation in both whole blood as well as DLA samples. The StarDist method segments individual outlines up to the maximal cell density that can be scanned using the CellTracks system while also segmenting tdEV. The StarDist segmentations closely follow the cell outline in most cases, enabling precise quantification of signal intensities. These intensities subsequently enable quantitative phenotypic characterization of the segmented events. Furthermore, we also found that StarDist segmented at least an additional 10% of CTC in CellSearch whole blood samples, and an additional 20% of CTC in CellSearch DLA samples, while recovering 99.95% of all CellSearch selected CTC.
Institutional Review Board Statement:
The study was conducted in accordance with the Declaration of Helsinki, and approved by the Institutional Review Boards of all institutes participating in NCT00133913, NCT00133900, Cancer-ID studies.
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.
Data Availability Statement:
The data presented in this study are available on request from the corresponding author.
Conclusions
Here, we demonstrated the occurrence of critical failures in samples segmented by the built-in CellSearch segmentation algorithm. The ACCEPT algorithm developed for CellSearch samples did not perform well in dense (DLA) samples, because it joined nearby objects together into a single event. To overcome this we evaluated the StarDist deeplearning-based segmentation method and found it outperforms both ACCEPT and the current CellSearch segmentation in both whole blood as well as DLA samples. The StarDist method segments individual outlines up to the maximal cell density that can be scanned using the CellTracks system while also segmenting tdEV. The StarDist segmentations closely follow the cell outline in most cases, enabling precise quantification of signal intensities. These intensities subsequently enable quantitative phenotypic characterization of the segmented events. Furthermore, we also found that StarDist segmented at least an additional 10% of CTC in CellSearch whole blood samples, and an additional 20% of CTC in CellSearch DLA samples, while recovering 99.95% of all CellSearch selected CTC.
Institutional Review Board Statement:
The study was conducted in accordance with the Declaration of Helsinki, and approved by the Institutional Review Boards of all institutes participating in NCT00133913, NCT00133900, Cancer-ID studies.
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.
Data Availability Statement:
The data presented in this study are available on request from the corresponding author. | 9,032.8 | 2022-06-01T00:00:00.000 | [
"Medicine",
"Computer Science"
] |
Parametric and Experimental Modeling of Axial-Type Piezoelectric Energy Generator with Active Base
: A computational and experimental approach to modeling oscillations of a new axial-type piezoelectric generator (PEG) with an attached mass and an active base is considered. A pair of cylindrical piezoelements located along the generator axis is used as an active base. Plate-type piezoelectric elements, made in the form of two bimorphs on an elastic PEG base, use the potential energy of PEG bending vibrations. Energy generation in cylindrical piezoelectric elements occurs due to the transfer of compressive forces to the piezoelectric element at the base of the PEG during excitation of structural vibrations. The active load scheme is selected separately for each piezoelectric element. Numerical simulation was performed in the ANSYS FE analysis package. The results of modal and harmonic analysis of vibrations are presented. A technique for experimental analysis of vibrations is presented, and a laboratory test setup is described. Numerical and experimental results are presented for the output characteristics of a piezoelectric generator at a low-frequency load. For one of the versions of the generator and a certain displacement amplitude for a frequency of 39 Hz, in the results of a comparative experimental analysis at a load of 10 k Ω , the maximum output power for each cylindrical piezoelectric element was 2138.9 µ W, and for plate-type piezoelectric elements, respectively, 446.9 µ W and 423.2 µ W.
Introduction
At present, in connection with their development and introduction into production, renewable electrochemical batteries are used to the greatest extent. These electric batteries have the properties of cyclic recovery during a certain resource period and are limited in use when a certain finite operating time is reached [1,2].
Along with renewable batteries, promising devices are used that extract electrical energy from the environment. One of the directions is the extraction of electric energy using solar energy. Some approaches can be presented as examples described in works using solar batteries [3,4]. The energy of the environment in the form of vibration entering compact devices can be extracted using the energy of the movement of air masses (wind), including using energy generation devices in the form of piezoelectric transducers [5,6]. The use of the mechanical energy of the movement of water masses in narrowed volumes, sea tides and tides, and the movement of water masses of rivers and dams, can be represented by the example of [7]. The use of devices for converting thermal energy into electrical energy is presented in [8]. Devices for converting mechanical vibrations of structural elements and soils are presented in the reviews of works [9,10].
Devices for generating electrical energy, the so-called power converters of mechanical energy into electrical energy, using piezoelectric elements, are called piezoelectric generators (PEGs). Their development requires further modernization and improvement. This is possible only when designing new devices and analyzing their operation under various conditions of dynamic loading. As a review of the current use of PEG shows, such devices require further development due to the low output power and voltage generated.
Basic information about energy generators, as well as problems arising in the development of energy storage devices using piezoelectric elements, were discussed in review articles [11][12][13][14][15]. The primary analysis of review articles shows that models are considered that use piezoelectric elements operating in compression and buckling with tension. Piezoelectric elements can be built into various mechanisms using both pressure and rotational loads. The reviews present the output characteristics of PEG energy of various types of manufacture.
Monographs [16,17] present theoretical works by some authors. Methods for conducting experimental studies are given.
As the literature reviews [12][13][14][15][16][17] show, the most common element in the generator is a cantilever or a linear type element that is fixed at one or two points. There are schemes with fastening of bending elements in the form of a truss structure. A mass or element in the form of a magnetic attachment is attached to some end or intermediate part of such a device. An electrical load in the form of capacitance C, resistance R or inductance L can be applied to each piezoelectric element separately. Typically, a single-component electrical load is used, which has a certain resistance. In [18], the main topologies of load-relief circuits in the form of electrical energy with the simplest composite circuit of a voltage rectifier in the form of a diode bridge are given. Within the framework of the presented work, it is supposed to measure the voltage at individual electrodes. Thus, under dynamic loading, it is possible to investigate the phase characteristics of the removed potential and the potential on each individual piezoelectric element.
The research in paper [19] proposed a novel hybridization scheme with electromagnetic transduction to enhance the power density of PEHs. The hybrid energy harvester was designed based on the BC-PEH. To compare the power density of the BC-PEH and the hybrid energy harvester, we built a prototype and conducted many experiments. The generator delivers a high voltage of 21.9 V at a drive acceleration of 0.3 g using an array of variable magnets. A peak output power of 103.53 mW is obtained.
In article [20], a cantilever-type generator with an active base was considered. Four cylindrical piezoelements were used in the base. An experimental approach was presented, showing that the maximum output power was 41.8 µW and 7.42 µW for the two types of piezoelectric elements, respectively.
In [21], modeling of the cantilever-type PEG with symmetrical and asymmetrical location of proof mass was presented, and the linear theory of elasticity and electrodynamics was used, taking into account the dissipation of energy as well as the equations of motion in the acoustic approximation. The built FE model was numerically realized in the ANSYS software.
Reference [22] presented a simulation of two generators having a modification with mass elements symmetrically attached, with respect to the main axis of the generator. It was shown that the maximum output power was 6.97 mW.
In [23], FE modeling of a two-axis PEG is considered. The piezoelectric elements were fixed in the form of a bimorph on two cantilevers at the base of the PEG. The calculated output power was 720 µW.
The research in [24] considered a cantilever-type PEG having an aluminum base, with piezoelectric elements in the pinched area and an attached mass at the end. The peculiarity of this generator is that there are stopper devices as a component of the device. The authors fixed the movements of the generator base bar. Analytical modeling of the generator was given, and the results of the nonlinear analysis of the output energy characteristics are presented. The output power of the generator was obtained at a level of up to 4.95 mW. Frequencies were considered up to 40 Hz.
The article [25] considered a broadband PEG with several degrees of freedom based on five resonant frequencies that were fairly close to each other. The PEG consisted of five end masses, two U-shaped cantilever beams and a straight beam. The selection of resonant frequencies was realized due to the special design of the parameters. The electrical characteristics of the PEG were analyzed through simulation and experiment, confirming that the PEG can effectively expand the operating bandwidth and collect vibration energy at low frequency. Experimental results showed that PEG has five low-frequency resonant frequencies, which are 13, 15, 18, 21 and 24 Hz; under the action of acceleration of 0.5 g, the maximum output power is 52.2, 49.4, 61.3, 39, 2 and 32.1 µW, The article [26] considered micro-MEMS PEG, which has a wide frequency band with included stoppers; one and two sides are carefully studied. The results of the experiment showed that the operating band is extended to 18 Hz (30-48 Hz) and the corresponding optimal power ranges from 34 to 100 nW, with a base acceleration of 0.6 g. Mathematical modeling based on the application of differential equations of motion was carried out.
In [27], mathematical modeling of PEG was carried out based on the application of Lagrange's electromechanical coupling equations. A PEG prototype with two internal single arms was designed, manufactured and experimentally tested. The accuracy of the proposed mathematical modeling was verified by finite element modeling and experimental results. When tested with low harmonic amplitude, the PEG generated 2.48V, 6.21V and 1.55V at three resonant frequencies between 15 and 30Hz, respectively.
A brief analysis of the modifications and studies of PEG shows the following. The authors of the article [28] optimized the power generator and the software for the analysis of FE, performed using the software ANSYS and ACELAN. The optimal design was based on matching the resonant frequency of the device with the excitation frequency of the environment [28]. In the article [29], a packet-type piezoelectric energy generator was studied. An experimental setup was used and measurements of the response of a multilayer piezoelectric stack in an energy harvester were described. Paper [30] proposed a single-degree fractal structure system for energy collection. The authors optimized the design and experimentally evaluated the performance of the system.
In [31], a piezoelectric power generator was developed with rotation amplitude limitation to avoid resonance conditions. The radially entrained magnetic force was used to collect energy. In [32], the authors considered and analyzed the collection of piezoelectric energy from the characteristics of compliant mechanisms. The authors divided configurations into monostable, multistable, multiple degrees of freedom, frequency upconversion and under load optimization. The authors also introduced a normalized power density to compare the power generation capabilities of a power generator [32].
In [33], the authors designed an energy harvester to extract energy from a smart road. The authors investigated the number of stacks layers, influences of connection mode, number of units and ratio of height to cross-sectional area [33]. Few studies have considered, depending on the area of application, a different design for the piezoelectric generator, in which a direct piezoelectric effect is used when the excitation in the sensitive element vibrates longitudinally (d33), by bending (d31) and by sharing [34][35][36]. In [37], a three-dimensional finite element analysis is presented for a cantilever plate structure excited by piezoelectric drive sections. The paper considered the modeling of actuators of the optimal configuration of actuators for selective excitation of the modes of a cantilever plate structure. Such elements can be used for technical analysis of vibrations of various structures using MEMS technologies [38].
The aim of this research is to carry out computational and experimental studies for estimation of the output parameters of a new type of piezoelectric energy generator of axial type, which has bimorph active structures on the base bar and symmetrically fixed piezoelectric cylinders at the base, located co-axially to bar; this is a new design of axial type piezoelectric generator. This design harvests energy from d 31 and d 33 simultaneously. In this design, proof mass can be mounted on the duralumin beam in between piezoelectric patches and the screw side. This variation in fixing the proof mass endows flexibility onto the natural frequencies of the PEGs. As per the mechanical vibration input, the first natural frequency can be adjected within the limit for high power output. This design provides flexibility and enhances the output power options compared to the other previous designs. The mechanical vibrations are used as an input parameter. This study concerns the new design of a generator, in which proof mass plays a key role in achieving higher power output. The analysis is also carried out on the electrical load dependency.
The paper is organized as follows: Section 2 provides a description of model parameters and the electric scheme presents a description of axial-type PEG elements, a schematic description of the structure and a description of the parameters of materials used in both numerical and experimental modeling. Section 3 presents a description of the theoretical approach to the study of composite elastic, electroelastic and acoustic media in FE modeling. Section 4 covers the numerical simulation of the generator. The results of modal analysis and some output data of PEG parameters as a result of harmonic modeling are presented. Section 5 outlines a description of an experimental setup for testing the operation of PEG under a certain loading. A description of the results of testing PEG under dynamic loading in a certain frequency range is presented. Finally, Section 6 provides the conclusion.
Description of Model Parameters and Electric Scheme
An axial piezoelectric transducer for converting mechanical energy into electrical energy contains a base plate (7) in the form of a rigid beam structure made of elastic material, on which piezoelectric bimorph (2, 3) elements are glued (see Figure 1). One end of the beam structure is bolted (9) to the base (4) by Fixing supports (5). To the other end, at the base, piezoelectric elements of the cylindrical type (1) are fixed coaxially with the base bar through an L-shaped bar (6). An additional proof mass M (8) is located between the edge of the base, fixed with a bolt (9) and the piezoelectric elements (3). Thin symmetric piezoelectric elements (PE) are polarized in thickness. They are glued to the base console and are arranged in a row (see Figure 2). Characteristics of the dimensions of the PEG elements are presented in the Table 1, the properties of the elements are presented in the Tables 2-4 show the mechanical properties of PE materials. In the full-scale model, we used PE made of CTS-19 material produced at Piezopribor LLC of the Southern Federal University (Rostov-on-Don, Russia). In the simulation, the parameters of materials with equivalent properties presented in [39] are used. Table 3. Elastic moduli C E pq (×10 10 Pa), piezoelectric coefficients e kl (C/m 2 ) and relative permittivity ε ξ kk/ε0 of piezoceramics (based on measurements at room temperature). Working Principle of the Model
Piezoelement
The proposed model works as follows: When the rigid base (8-10) is exposed to external mechanical forces such as shocks and vibrations, the vibrations are transmitted to the base plate (7), affecting the piezoelectric elements, in which alternating deformations of compression (in cylinders) and tension-compression (in plates) due to the reaction of the supports occur.
Due to the direct piezoelectric effect on the electrodes of additional piezoelectric elements, AC voltage is generated and, therefore, additional electrical energy. The combined use of such elements allows you to increase the output power and conversion efficiency of the converter efficiency. This AC voltage and additional electrical energy can be converted using bridge rectifiers to DC voltage, which is stored in batteries using harvesting energy systems. Piezoelectric elements can be connected in parallel, in series or separately. The choice between the types of connections of elements depends on the device that needs to be powered: if a higher output voltage is required, then a serial connection should be selected, and if a higher output current is required, then a parallel connection.
A schematic electrical diagram of a PEG connection with a resistive load is shown in Figure 2. The resistive load is supplied to each PE individually. Voltage is found at the contact points of the resistor. In numerical simulation, voltage is calculated as the difference of its amplitudes at the nodes of the FE element presented, with an option indicating the type of resistor. The added weight can vary from 3 to 25 g. In the experiment, we used the masses M = 3.71; 6.13 g. The proof mass was located at a distance of Lm = 150 to 230 mm to the left edge of the generator (Figure 1).
Theoretical Description of the Model of Composite Elastic, Electroelastic and Acoustic Media by FE Simulation
The energy storage PEG is a composite elastic and electroelastic body. It is assumed that the device performs elastic small oscillations in a moving coordinate system. The rectilinear vertical motion of this system in the area of fixation is given by the law for steady oscillations: Under these conditions, a sufficiently adequate mathematical model of the operation of the device is the initial-boundary value problem of the linear theory of electroelasticity [40,41].
In the general formulation, the equations for a piezoelectric medium are written as: where ρ is the material density; c ijkl are the components of the fourth rank tensor of the elastic moduli; e ikl are the components of the third rank tensor of piezoelectric coefficients; ε kl are the components of strain tensor; E k are the components of the electric field vector; ϕ k is the electric potential; ik are the components of the dielectric constants tensor; α, β, ζ are non-negative damping coefficients (the value of ζ d is used in ANSYS software).
For elastic medium, we have: Since harmonic analysis is used in the calculations, the following actions are performed for the corresponding components of the equations: To solve the problem, the following mechanical and electrical boundary conditions are accepted.
Boundary conditions are given in the form of a displacement field on the boundary S u Boundary conditions in the form of a vector of surface stresses p The boundary conditions on the electrodes of the piezoelectric element S E = k S E k are given as Boundary conditions on non-electrode sections S_D, at the intersection of the corresponding areas S = S E S D D n | S D = 0 The attenuation coefficient parameters are between the frequencies и f r1 and и f r2 . It is assumed that within the framework of the experiment, the change in the damping parameters α and β will be minimally changeable. In the FE package ANSYS, the damping parameters were described in the form From the experiment, the quality factor Q is found from the expression where ∆ω is the width of the resonance curve. This is found at the corresponding resonance ω.
The solution of the set system of Equations (1)-(8) is solved, taking into account the initial boundary conditions for non-stationary problems [40]. The elements of solving the system of Equations (1)-(8) within the framework of modeling can be implemented in the Ansys complex. Within the framework of the presented studies, a number of direct calculations were carried out in the form of a modal analysis with obtaining natural modes of oscillations and frequency characteristics. When carrying out harmonic analysis, the electric potential on the electrodes was calculated relative to the zero potential on certain surfaces. The displacement of the base by 0.01 mm was taken as the perturbation parameter.
Modeling
Within the framework of the presented studies, modal and harmonic analyses of PEG oscillations were carried out during modeling. When modeling, the value of the attached mass M= {0.05; 1; 3; 5; 7; 9} gr, as well as its position Lm = {150; 170; 190; 210; 230} mm on a flexible base plate is used. The value of the active load was taken as R = 1000 Ω. At the stage of harmonic analysis, the vibration amplitudes were recorded at a displacement of the PEG rigid fastening zone by {0.001...0.01} mm.
The next task was to consider the harmonic modeling of PEG oscillations with a constant displacement of the generator housing attachment zone by 0.01 mm and variations in material properties (duralumin: E 1 = 0.7 10 11 , Pa, ρ 1 = 2600 kg/m 3 , ν 1 = 0.33; and fiberglass: E 2 = 0.06 10 11 , Pa, ρ 2 = 1600 kg/m 3 , ν 2 = 0.33) and corresponding thickness options h = {1; 2} mm base plate. In this case, the attached mass was fixed in the position L m = 150 mm and its value was M = 0.5 gr, 30 gr. As an active load in the form of resistance, R for each piezoelectric element was taken, equal to 1000 Ω.
The natural frequencies, the oscillation amplitude of the center of the flexible base plate of the generator and the output voltage taken from the piezoelectric elements were considered as output parameters.
The simulation was carried out in FE software ANSYS. Figure 3 shows the FE model of the PEG. PEG consists of a base in the form of a strip, clamped on one of the sides. During the modeling, elements of the SOLID92 type with a tetraidal structure were used, which facilitated the partition of the model. When modeling the thin walls of the planks, it was assumed not to consider deformations over the thickness of the structure. As a result, the value of the smallest edge of the FE in the form of a tetrahedron was taken, equal to the wall thickness of the structural element. Modeling of piezoelectric elements was carried out using FE type SOLID5. Modeling of elements in the form of active resistance was carried out using the FE CIRCU94 with the resistor option. The PE of a three-dimensional structure in the form of piezocylinders (PE PC) was divided by a FE mesh with an edge size multiple of 0.15 of the piezocylinder height. In a PE with a planar structure (PE PS), the size of the partitioning of the FE mesh by thickness was a multiple of the value equal to the thickness of the PE. The division was in the form of triangular prisms. The direction of polarity in a cylindrical PE was taken along the main axis of the PE, coaxially with the PEG, in a plate-shaped PE along the thickness. The polarization vector in the bimorph for the upper and lower PEs was directed along the normal to the surface. The number of FE elements when splitting the model was more than 33,000 nodes greater than 62,500. At the first stage, modal PEG analysis was carried out. Figure 4 shows the four vibration modes of the model. With this configuration, the model had a first design frequency of 259 Hz. The vibration mode for the first mode had a prevailing bending character for the generator warp bar. Vibration modes 1, 4 and 8 correspond to (1,2,3) flexural vibration modes of the PEG base plank in the vertical direction. The second mode shape was obtained at 525 Hz. It was assumed that bending deformations of the bar in the smallest plane of rigidity (in the vertical direction) would excite the highest output stress in all PEs. Investigations of oscillations of axial-type PEG under harmonic action were carried out. The harmonic action was calculated under the action of a uniformly applied acceleration of 10 m/s 2 for all structural units at the corresponding harmonics. For the first and second vibration modes, the dependences of their value on the mass and location of the load were calculated. The corresponding graphs ( Figure 5) show that the lowest frequency values are in the range of 212-220 Hz for the first vibration mode and values of 501-510 Hz can be achieved with the central location of the load and its maximum mass of 9 g. This, at the point of attachment of the load, will be maximum, which is shown in Figure 6. Figure 7 shows the calculated parameters of the dependence of the output voltage at the electrodes of piezocylinders (U 1 ) and piezoelectric elements in the form of plates (U 2 and U 3 ), respectively, with their calculated arrangement from left to right (see Figure 1). In the calculations, the value of the active load of the corresponding PE was taken to be 1000 Ω for one vibration mode.
A numerical experiment was set up to establish the dependence of the first oscillation frequency on the associated proof mass 8 (Figure 1). The variable parameters were the modulus of elasticity E i of the base 7 ( Figure 1) and its thickness h i . The proof mass varied within M = 0.5 gr, 30 gr. The place of fixation-proof mass L m = 150 mm is the point in the middle of the base. Accordingly, at fixed dimensions of the load, the specific density of the material was calculated. It was assumed that the most sensitive characteristic for changing the natural frequency of PEG vibrations is the value of the mass-proof mass. The modulus of elasticity of the volume-proof mass was taken equal to the properties of the base. The base modeling was considered, using the following properties of duralumin for calculations: E 1 = 0.7 10 11 , Pa, ρ 1 = 2600 kg/m 3 , ν 1 = 0.33; and using the properties of fiberglass: E 2 = 0.06 10 11 , Pa, ρ 2 = 1600 kg/m 3 , ν 2 = 0.33. The thickness of the base (7) was assumed to be h 1 = 2 mm and h 2 = 1 mm for calculations. Thus, four options for the layout of parameters for modeling were used: 1-E 1 , h 1 ; 2-E 1 , h 2 ; 3-E 2 , h 1 ; 4-E 2 , h 2 . The results of numerical calculations obtained on the basis of the modal analysis carried out are presented on Figure 8. An analysis of the obtained frequency dependences shows the following: with an increase in mass for all design simulation options, the first natural frequency decreases. Therefore, with a conditionally small mass of 0.5 gr, the first natural frequency for options was 1-287.4 Hz; 2-154 Hz; 3-122.8 Hz; 55.3 Hz. In this case, the first natural frequency with a mass of 30 gr was, respectively, 1-143.4 Hz; 2-59.7 Hz; 3-50 Hz; 18.5 Hz. In a comparative analysis for all calculation options, the first frequency changed, respectively, for variations, more than 1-2 times, 2-2.57 times, 3-2.45 times and 4-2.97 times. Thus, with various initial parameters of the properties of the base model, the use of this PEG is possible in various loading ranges, both in the low-frequency region up to 50 Hz and in the region of higher frequencies, using only the first oscillation mode up to 287 Hz. The loading mode in the region of more than 50 Hz involves the use of devices for mechanical excitation of oscillations, for example, rotary motors with magnetic media. In addition, these modes of operation of the PEG can be used as the use of the PEG in the form of vibration sensors of the impulse action on the structure in a certain fixed frequency range.
Experimental Probe
To conduct field studies of the operation of an axial-type PEG and obtain primary results for assessing its output parameters, a laboratory test setup, LTS -01, was created with certified testing devices. A structural diagram of its work was built and a description of the setup and a research methodology was prepared.
The LTS -01 laboratory test setup for studying the PEG output characteristics is shown in Figure 9. The direct piezoelectric effect on the electrodes of additional piezoelectric elements and AC voltage was generated and, therefore, additional electrical energy. The combined use of such elements allows you to increase the output power and conversion efficiency of the converter efficiency. This AC voltage and additional electrical energy can be converted using bridge rectifiers to DC voltage, which is stored in batteries using harvesting energy systems. Piezoelectric elements can be connected in parallel, in series or separately. The LTS -01 laboratory test setup for studying the output characteristics of an axial-type PEG consists of an exciter of mechanical vibrations-an electromagnetic vibrator VEB Robotron 11,077 (4), on a work Table 5, on which the studied PEG sample was installed. The sample under study is a plate (14) fixed on the base (12). On the plate (12), there are two pairs of piezoelectric elements (17) made in the form of a bimorph. At the left end of the plate base, there were cylindrical-type piezoelectric elements having a longitudinal arrangement along the main PEG axis. The other end of the generator base was rigidly bolted, thereby clamping and effectively fixing the opposed generator base with cylindrical piezoelectric elements. There was a proof mass on the base bar (14), with the help of which it was possible to easily adjust the frequency characteristics of the generator in certain ranges. With the help of optical sensors of mechanical displacements (6) and (8) and their matching devices (7) and (9), data on the vibration amplitudes of the corresponding PEG points were transmitted to the computer. The optical sensor (6) of the REF603 type was located above the point in the center of the rigid base of the PEG and transmitted information about the vibrations of the working table plate of the VEB Robotron 11,077 electromagnetic vibrator (4). The optoNCDT optical linear displacement transducer (7) was located above the PEG base plate and could transmit a voltage linearly, proportional to the oscillation amplitudes at predetermined fixed points of the base, through a conductive loop. Within the framework of the experiment, the vibrations of the central region of the base plate were recorded.
The measuring part of the LTS -01 consisted of an analog-to-digital converter-an external ADC/DAC E14-440D module from L-Card (10) and a personal computer (PC) (1). This external module was used to record the voltage at the electrical contacts of the piezoelectric elements under an active electrical load Rl (11), as well as the voltage at the electrical contacts of the laser displacement sensors (6) and (8) and the AFG 3022B Tektronix master signal generator (2).
The Principle of Operation of the Laboratory Test Setup LTS -01
The process of measuring the frequency response of the oscillatory process of the PEG model and recording its output characteristics was as follows and is shown in Figure 10. On a personal computer (1), two programs were launched: (i) a program for recording voltage parameters on an E14-440D ADC/DAC module from L-Card (10) PowerGraph and (ii) a swept signal generation program with its own developed software [42]. The signal had a fixed output oscillation amplitude with a voltage of 2 V. The process of sweeping (enumerating) the signal was carried out by alternating the excited frequencies from 1 Hz to 1000 Hz. Excitation of each vibration frequency was carried out for 0.5 sec with a pause of 0.5 sec between the transition to another frequency, with a step between frequencies of 1 Hz. The signal could be transmitted via the audio path or USB channel to the AFG 3022B Tektronix signal generator (2), thereby being recorded or generated on it. The PowerGraph program allows high-quality recording and unloading of the measured signal voltage to the ADC module (10). The generated signal was fed through the current-carrying path to the electromagnetic forced oscillation exciter. Voltage was supplied from the oscillator of the sweeping frequency of the device with a stable amplitude and a linearly varying frequency. The AC voltage value was amplified by the power amplifier (3) and supplied to the vibrator (4); the resulting output voltage from the piezoelectric generator was loaded with AC electrical resistance and supplied to one of the ADC channels (10) of the external module. The values of these voltages were reproduced on the computer monitor screen (1) in the form of an amplitude-time characteristic (ATC). On the working Table 5 of the electromagnetic forced oscillation exciter VEB Robotron 11,077 (4), the investigated PEG sample of the axial type was fixed with a rigidly bolted connection. When voltage was applied to the vibrator (4), predetermined amplitude displacements occured in the vertical direction of the working Table 5. The vibration amplitude in a certain frequency range was linearly proportional to the vibration excitation frequency. The resulting vertical displacements of the working Table 5 set the PEG in motion. Mechanical vibrations of various parts of the generator generated voltage across all piezoelectric elements. The piezoelectric elements were connected to the grounding path (15) and in parallel with the active electric load resistance bank (11) of the corresponding rating. There were six piezoelements and, accordingly, each of them had its own individual electrical path. The voltage excited on the piezoelectric elements was recorded using the ADC module (10). Using the ADC module (10), the voltage values from the optical displacement sensors (6) and (8) were also recorded. Subsequent signal processing was performed by software using an external ADC/DAC E14-440 module from L-Card, a personal computer and PowerGraph software. Observation of the shape of the exciting signal and signals from the sensors was performed on the monitor screen and can be duplicated using a digital oscilloscope. Accurate measurement of the signal frequency was performed using the Fourier analysis module in the software. Adjustment and calibration of receiving sensors were carried out using a measuring microscope. Test results are stored in the PC's memory in digital and text format; signal elements can be duplicated in various text editor modules.
Main technical characteristics of the LTS -01 test setup: (i) a range of measurable lateral displacements of the PEG substrate from 0 to 5 mm; (ii) frequency of forced oscillations from 1 to 1000 Hz; (iii) linear range of forced vibration amplitudes from 20 to 1000 Hz; (iv) the sensitivity limit of the optical displacement sensor is not less than 5 µm; (v) electric voltage at the input of the electromagnetic exciter of oscillations from 0.1 to 10 V.
PEG Results Validation
The numerical simulations of harmonic analysis of the PEGs were carried for various configurations. The experimental results were taken care of for PEGs. The simulation and experimental results are shown in Figure 11 and summarized in Table 5. The simulation and experimental results were carried in case the proof mass had been located at 230 mm. The analysis of the comparison of the oscillation parameters of harmonic analysis showed a satisfactory coincidence of experimental data on oscillations and the results of numerical studies of their own frequencies and corresponding oscillation amplitudes for the first two modes within 6%.
Method of Natural Modeling of the Oscillatory Processes of PEG
The sequence for determining the parameters of PEG characteristics was as follows. At the first stage, a preliminary assessment of the difference in natural frequencies obtained in the experiment and simulation was carried out. By formula (11), the discrepancy of natural frequencies was calculated: (11) Table 6 shows a comparison of the first and second natural harmonics of oscillations, with an experimental value of the mass of the load of 3.71 g at its various locations. The active load for each PE was taken, both in the experiment and in the calculation of 10 kΩ. The analysis of the obtained results showed a satisfactory coincidence of the values of the natural frequencies (the difference was no more than 5%). First, the frequency response of the piezoelectric generator was recorded, then the dependence of the output voltage and power of the PEG at the resonant frequency on the value of the active electrical load and proof mass on the value of the output voltage of the PEG and frequency characteristics was investigated. In the test PEG sample, fixed on a vibrating plate, transverse bending vibrations were sequentially excited in the frequency range from 0 to 1000 Hz (the generator voltage amplitude did not change in this case) value of electrical load resistance Rl from 10 kΩ to 2 MΩ. The output power was calculated for the obtained voltage characteristics for 1, 2 and 3 PE. Figure 12 shows examples of visualization of the process of measuring the output characteristics of a PEG during sweeping. This analysis reveals the parameters of the output voltage on the PE electrodes at different loads. During the sweep, two points were chosen to obtain the estimated parameters of the PEG operation: (i) 39 Hz and (ii) 107 Hz. At a frequency of 107 Hz, the oscillation frequencies of the system elements were found, causing a resonant increase in oscillations. For a frequency of 39 Hz, the oscillation amplitude of the rigid base of the generator (4) was 0.266 mm, while the oscillation amplitude of the base (7) center was 1.683 mm. For a frequency of 107 Hz, the oscillation amplitude of the rigid base of the generator (4) was 0.035 mm, while the oscillation amplitude of the base (7) center was 0.356 mm. The results are shown in Tables 7 and 8. Data analysis shows that the maximum output voltage of 1.31 V for cylindrical PE occurs at a frequency of 39 Hz and a load of 2 MΩ. At the same time, for PE in the form of plates, the highest voltage is achieved at an active load of 150 kΩ and a frequency of 39 Hz. Figures 13 and 14 show the graphs of the dependence of the output voltage and peak output power for all types of PE used. Analysis of the output power shows that with a large value of resistance, it drops. In this case, the maximum output power can be obtained at a load of up to 51 kΩ.
Conclusions
As a result of the research carried out, a finite element, by using ANSYS software and experimental approaches to modeling oscillations of a new type of piezoelectric generator (PEG) with proof mass, and an active base were developed; as for the latter, cylindrical piezoelements located along the axis of the generator were used. The results of modal and harmonic analysis of oscillations are presented. In this case, the generation of energy in the cylindrical-type piezoelectric elements occurred due to the transfer of compression forces to the piezoelectric element at the base of the PEG upon excitation of structure vibrations. The piezoelectric elements of the plate type, made in the form of a bimorph, used the potential energy of bending vibrations of the PEG on the PEG base bar. The presented technique for the experimental analysis of vibrations, as well as the laboratory setup developed, made it possible to obtain experimental results of the output characteristics of the piezoelectric generator under low-frequency loading, which differ from the finite element results within 5%.
The presented PEG has frequencies of the first vibration mode in the range from 210 Hz to 300 Hz with the corresponding proof mass position. The scope of this generator can be attributed not only to high-frequency loading ranges. On the experimental setup, created within the framework of the research, a variant of low-frequency loading at frequencies of 39 and 107 Hz was considered, at which the highest output voltage for frequency 39 Hz for a cylindrical PE was 0.14 V and the output power was P 1 2138.9 µW. For the PE plate type for this frequency 39 Hz, the maximum peak power was P 2 446.9 µW and P 3 423.2 µW. When conducting a comparative analysis with literature data, the analysis shows that the output parameters of the generator are at an average level and may require further modification of the design.
The analysis of the simulation of this PEG has shown the possibility of working in different loading ranges, both in the low frequency range up to 50 Hz and in the higher frequency range when using only the first oscillation mode up to 287 Hz. The loading mode in the region of more than 50 Hz involves the use of devices for mechanical excitation of vibrations, for example, rotary motors with magnetic carriers. Furthermore, these operating modes of the PEG can be used as the use of the PEG in the form of vibration sensors of pulse action on the structure in a certain fixed frequency range.
The considered PEG device can be upgraded and studied in more detail under other types of loading. | 9,182.6 | 2022-02-07T00:00:00.000 | [
"Engineering",
"Physics"
] |
Exendin-4 Ameliorates Lipotoxicity-induced Glomerular Endothelial Cell Injury by Improving ABC Transporter A1-mediated Cholesterol Efflux in Diabetic apoE Knockout Mice*
ATP-binding cassette transporter A1 (ABCA1), which promotes cholesterol efflux from cells and inhibits inflammatory responses, is highly expressed in the kidney. Research has shown that exendin-4, a glucagon-like peptide-1 receptor (GLP-1R) agonist, promotes ABCA1 expression in multiple tissues and organs; however, the mechanisms underlying exendin-4 induction of ABCA1 expression in glomerular endothelial cells are not fully understood. In this study we investigated the effect of exendin-4 on ABCA1 in glomerular endothelial cells of diabetic kidney disease (DKD) and the possible mechanism. We observed a marked increase in glomerular lipid deposits in tissues of patients with DKD and diabetic apolipoprotein E knock-out (apoE−/−) mice by Oil Red O staining and biochemical analysis of cholesterol. We found significantly decreased ABCA1 expression in glomerular endothelial cells of diabetic apoE−/− mice and increased renal lipid, cholesterol, and inflammatory cytokine levels. Exendin-4 decreased renal cholesterol accumulation and inflammation and increased cholesterol efflux by up-regulating ABCA1. In human glomerular endothelial cells, GLP-1R-mediated signaling pathways (e.g. Ca2+/calmodulin-dependent protein kinase, cAMP/PKA, PI3K/AKT, and ERK1/2) were involved in cholesterol efflux and inflammatory responses by regulating ABCA1 expression. We propose that exendin-4 increases ABCA1 expression in glomerular endothelial cells, which plays an important role in alleviating renal lipid accumulation, inflammation, and proteinuria in mice with type 2 diabetes.
Diabetes mellitus is a chronic disease that has become a global epidemic. One of the most common microvascular complications of diabetes mellitus is diabetic kidney disease (DKD), 2 a progressive fibrotic kidney disease that is the lead-ing cause of end-stage renal disease (1). DKD begins with normoalbuminuria, microalbuminuria, and macroalbuminuria before progressing to end-stage renal disease (2). Although several factors involved in the occurrence, progression, and outcome of DKD have been described, the exact pathogenic mechanism remains unclear. Renal lipotoxicity (i.e. dyslipidemia and renal lipid accumulation) is known to contribute to kidney damage in chronic kidney disease (3). The accumulation of intrarenal lipid droplets appears to accelerate podocyte dysfunction, glomerulosclerosis, and interstitial fibrosis through lipid infiltration, induction of oxidative stress, and up-regulation of proinflammatory and profibrotic cytokines and growth factors (3). Increasing evidence has shown that renal lipotoxicity also plays an important role in the pathogenesis of DKD (4 -6).
Intracellular cholesterol accumulation in the kidney is a characteristic of high fat-induced kidney damage (7), which may be caused by alterations in cholesterol intake, intracellular synthesis, esterification, and efflux. The ATP-binding cassette transporter A1 (ABCA1), which is responsible for intracellular cholesterol efflux, along with ABCG1 and scavenger receptor class B member 1 (SR-B1), mediates intracellular cholesterol export from endothelial cells, -cells, adipocytes, and podocytes. Excessive intracellular cholesterol is transferred to lipidpoor apolipoproteins such as apoA-I to generate HDL particles (8 -11). ABCA1 also inhibits the expression of inflammatory cytokines IL-1, IL-6, and TNF-␣ in macrophages (12)(13)(14)(15), suggesting a potential role in suppressing inflammation.
Lipid Deposits Increased in Glomerular Endothelial Cells in
Patients with Early and Advanced DKD and Diabetic ApoE Ϫ/Ϫ Mice-We evaluated renal lipid accumulation in glomerular endothelial cells from patients with early and advanced DKD and diabetic apoE Ϫ/Ϫ mice. Electron microscopy examination of kidney tissues from DKD patients showed a widening of the glomerular basement membrane, mesangial expansion, podocyte foot process effacement, and extensive accumulation of lipid droplets in fenestrated endothelial cells (Fig. 1A). Oil Red O staining also showed lipid droplets in glomerular endothelial cells of human renal biopsies (Fig. 1B) and apoE Ϫ/Ϫ diabetic mice (Fig. 1C). We found evidence of cholesterol accumulation by filipin staining in diabetic apoE Ϫ/Ϫ mice (Fig. 1D). Oil Red O and CD31 staining confirmed lipid accumulation in the glomerular endothelial cells of diabetic apoE Ϫ/Ϫ mice (Fig. 1E). Biochemical analysis of lipid composition in the renal cortex of diabetic apoE Ϫ/Ϫ mice indicates that accumulation of neutral lipid deposits corresponds to significantly increased cholesterol content ( Fig. 2A).
Exendin-4 Decreased Blood Lipid Levels and Alleviated Kidney Damage in apoE Ϫ/Ϫ Mice-Histological analysis of the kidneys of apoE Ϫ/Ϫ diabetic mice show that exendin-4 improves glomerular hypertrophy, basement membrane thickening, and mesangial expansion (Fig. 2, B and C) as well as HbA1c (glycated hemoglobin) level, creatinine clearance, albuminuria, urinary albumin excretion, the ratio of kidney weight to body weight, total cholesterol, and LDL level. In addition, exendin-4-treated diabetic apoE Ϫ/Ϫ mice had a slightly higher plasma HDL level compared with untreated diabetic apoE Ϫ/Ϫ mice (Table 1), whereas serum triglycerides did not differ significantly between these two groups.
Exendin-4 Attenuated Lipid Droplets Accumulation and Inflammatory Injury in Glomerular Endothelial Cells of Diabetic ApoE Ϫ/Ϫ Mice by Increasing ABCA1 Expression-Oil Red O staining, filipin staining, and biochemical analysis of lipid composition revealed a greater accumulation of lipid droplets in the glomeruli of diabetic apoE Ϫ/Ϫ mice than in glomeruli of nondiabetic apoE Ϫ/Ϫ mice, which was improved by exendin-4 treatment (Fig. 1, C and D, and Fig. 2A). The results of double immunofluorescence staining showed that GLP-1R and ABCA1 were expressed in glomerular endothelial cells of nondiabetic apoE Ϫ/Ϫ mice (Fig. 3, A and B). ABCA1 expression in the kidneys of diabetic apoE Ϫ/Ϫ mice was lower than that of nondiabetic apoE Ϫ/Ϫ mice ( Fig. 3C) but was increased by exendin-4 treatment. Real-time PCR and Western blotting results also showed that ABCA1 expression in glomeruli from diabetic apoE Ϫ/Ϫ mice was lower than in nondiabetic apoE Ϫ/Ϫ mice ( Fig. 3E) but was significantly increased by exendin-4 treatment (Fig. 3, D and E). We next assessed the effect of exendin-4 on inflammation in diabetes. Immunohistochemistry results showed that TNF-␣ and IL-6 protein levels were higher in the glomeruli of diabetic apoE Ϫ/Ϫ mice compared with nondiabetic apoE Ϫ/Ϫ mice, and these cytokines were down-regulated by exendin-4 (Fig. 4A). The results of real-time PCR and Western blotting confirmed the down-regulation of TNF-␣ and IL-6 ( Fig. 4, B and C) in exendin-4-treated diabetic apoE Ϫ/Ϫ mice.
Exendin-4 Ameliorated Inflammation and Up-regulated ABCA1-mediated Cholesterol Efflux in HRGECs by Activating the CaMKK/CaM Kinase Type IV (CaMKIV), cAMP/PKA, and PI3K/AKT Pathways and Inhibiting the ERK1/2 Pathway-We first examined the effect of exendin-4 on the expression of ABCA1, which is critical for regulating cellular cholesterol homeostasis. The results obtained from Western blotting and real-time PCR analysis (Fig. 5, A and B) showed decreased ABCA1 expression in HRGECs cultured under high glucose and high cholesterol conditions; however, exendin-4 treatment increased ABCA1 expression in these cells. We next examined the effect of exendin-4 on apoA-I-specific cholesterol efflux in HRGECs by high-performance liquid chromatography and cholesterol efflux assay. Our results showed that treatment with exendin-4 decreased the cholesterol content (Tables 2 and 3) and increased the cholesterol/apoA-I ratio (Fig. 5C). These results indicate that exendin-4 up-regulates apoA-I-specific cholesterol efflux and ABCA1 expression in HRGECs.
Double immunostaining showed that CaM kinase kinase (CaMKK) was expressed in glomerular endothelial cells of nondiabetic apoE Ϫ/Ϫ mice (Fig. 5D). The CaMKK/CaMKIV signaling pathway was investigated as a possible mechanism underlying the effect of exendin-4 on ABCA1 expression. The results of real-time PCR and Western blotting analysis (Fig. 5, A and B) showed that exendin-4 increased ABCA1 expression in HRGECs under high glucose and high cholesterol conditions, but this effect was suppressed by the CaMKK inhibitor STO-609. We also examined the effect of exendin-4 on CaMKIV activity in HRGECs and found increased CaMKIV phosphorylation in exendin-4-treated cells; this effect was also inhibited by STO-609. Similarly, siRNA-mediated knockdown of CaMKIV suppressed the increased effect of ABCA1 mRNA and protein levels induced by exendin-4 (Fig. 5, E and F), whereas ABCA1 expression increased in cells transfected with scrambled siRNA. These findings demonstrate the involvement of CaMKK/CaMKIV signaling in exendin-4-induced ABCA1 expression.
To further investigate the mechanism underlying the effect of exendin-4 on ABCA1-mediated cholesterol efflux in HRGECs, we analyzed GLP-1R-mediated signaling pathways, such as the cAMP/PKA pathway, which plays a role in cholesterol efflux (18,19). Intracellular cAMP activity was assessed by ELISA, and PKA activity was assessed by the Kemptide phosphorylation assay. We found that cAMP and PKA activity were markedly decreased under high glucose and high cholesterol conditions compared with normal glucose conditions and that exendin-4 increased both cAMP and PKA activity (Fig. 6, A and B). The results of real-time PCR and Western blotting analysis showed that PKA siRNA blocked the effect of exendin-4 on ABCA1 expression (Fig. 6, C and D). Taken together, these findings suggested the involvement of cAMP/PKA signaling in the up-regulation of ABCA1 by exendin-4.
Previous studies suggested that PI3K/AKT signaling contributes to the regulation of ABCA1 expression (20). Therefore, we examined PI3K/AKT activation in exendin-4-treated HRGECs. We found that exendin-4 increased PI3K and AKT phosphor-ylation (Fig. 7A), an effect that was attenuated by the PI3K inhibitor LY294002 (Fig. 7A). In addition, LY294002 inhibited exendin-4-induced up-regulation of ABCA1 mRNA and protein levels (Fig. 7). These results suggest that exendin-4 regulates ABCA1 expression by activating the PI3K/AKT pathway. We also investigated the effect of exendin-4 on ERK1/2 activation in HRGECs by real-time PCR and Western blotting analysis. Our results showed that ERK1/2 phosphorylation was increased under high glucose and high cholesterol conditions, but phosphorylated ERK1/2 levels were significantly decreased by exendin-4 and by the selective ERK1/2 inhibitor PD98059 (Fig. 8A). We then examined the effect of PD98059 on ABCA1 expression and found that, similar to exendin-4, PD98059 upregulated ABCA1 expression (Fig. 8, A and B). These results suggest that exendin-4 regulates ABCA1 expression by inhibiting the ERK1/2 pathway.
Real-time PCR and Western blotting analysis also showed that TNF-␣ and IL-6 levels were significantly increased in HRGECs under high glucose and high cholesterol conditions, but this effect was inhibited by exendin-4. However, exendin-4 did not decrease TNF-␣ and IL-6 mRNA and protein levels in cells transfected with ABCA1 siRNA (Fig. 8, C and D). Furthermore, TNF-␣ and IL-6 expression in exendin-4-treated cells was increased by pretreatment with STO-609, CaMKIV siRNA, PKA siRNA, or LY294002 and decreased by PD98059.
Discussion
Excess lipid accumulation in the liver, heart, and pancreas is associated with lipotoxicity, inflammation, and fibrosis. In addition, renal accumulation of lipids is thought to play a role in the pathogenesis of DKD (21,22). DKD is associated with excessive generation of reactive oxygen species and endoplasmic reticulum stress, damage to podocytes and interstitial tubular cells, mesangial expansion, and inflammatory cell infiltration, ultimately leading to alterations in the glomerular filtration barrier and to renal failure (23)(24)(25). In our study, we observed lipid deposition and cholesterol accumulation in glomerular endothelial cells of patients with early and advanced DKD, as well as in apoE Ϫ/Ϫ diabetic mice. We also showed that decreased ABCA1 expression in the glomerular endothelial cells of diabetic mice is the key factor linking lipid accumulation to the development of nephropathy. Treatment with the GLP-1R agonist exendin-4 improved ABCA1 expression, ABCA1-mediated cholesterol efflux, and inflammation in glomerular endothelial cells.
A novel finding of our study is that the exendin-4-mediated up-regulation of ABCA1 in glomerular endothelial cells alleviates diabetic kidney damage. This effect of exendin-4 on ABCA1 levels was shown previously to play an important role in cholesterol efflux in adipocytes (26), hepatocytes (27), and pancreatic cells (28). Increasing evidence also suggests that glomerular endothelial cells, which are an integral part of the glomerular filtration barrier, and cross-talk between glomerular endothelial cells and podocytes are important to the development and progression of DKD (29 -31). Endothelial cells express several genes involved in reverse cholesterol transport, such as ABCA1, ABCG1, and SR-B1 (32-34), which encode transporters and receptors that promote the cellular efflux or uptake of cholesterol. Previous studies report that endothelial expression of ABCA1 protects against endothelial dysfunction (35,36).
Defects in ABCA1, an ABC transporter superfamily member, are the cause of Tangier disease (37), which is characterized by decreased blood HDL levels. Efflux of free cholesterol from cells, an early step in reverse cholesterol transport, is mediated by ABCA1 (11), which interacts directly with apolipoproteins and promotes the solubilization of lipids and their release from cells (38). In the mouse model of type 1 diabetes mellitus, ABCA1 expression is down-regulated in both kidneys and circulating macrophages (39,40). In humans, ABCA1 genetic variants are strongly associated with the risk of coronary artery disease (41), which is inversely correlated with the cholesterol efflux capacity of macrophages (42). Renal lipid accumulation is due to increased cholesterol and fatty acid synthesis, as well as reduced ABCA1-mediated cholesterol efflux (43) in glomerular mesangial cells and tubular cells (44). However, the level of ABCA1 expression in glomerular endothelial cells and its role in glomerular endothelial cell injury was unclear. Inducing diabetes with a high-fat diet and streptozotocin significantly decreased renal expression of ABCA1, an alteration that appears to precede the development of diabetic nephropathy. We also showed that ABCA1 is expressed in glomerular endothelial cells; this expression was down-regulated in apoE Ϫ/Ϫ diabetic mice and under high glucose conditions. Taken together, these findings suggest that defects in ABCA1 promote cholesterol accumulation in renal cells. Cholesterol accumulation is closely associated with chronic metabolic inflammation, including atherosclerosis, obesity, and other metabolic diseases (45). ABCA1 is reported to have anti-inflammatory effects, suggesting a putative mechanism and potential target for protection against atherosclerotic cardiovascular disease (46 -48). Accordingly, humans with dysfunctional ABCA1 and familial HDL deficiency tend to have chronic low-grade inflammation (49,50). The decreased ABCA1 expression and lipid accumulation in glomerular endothelial cells in DKD appears to promote inflammatory injury. Thus, exendin-4 treatment may prevent glomerular endothelial cell damage by increasing ABCA1 expression.
TABLE 1 General and metabolic parameters
These parameters were set after 8 weeks of study in apoE Ϫ/Ϫ mice in the presence and absence of diabetes with and without exendin-4 treatment (n ϭ 6/group).
Variable
Control Body weight (g) 25 tion (54,55). In our study, we observed that ABCA1 and GLP-1R were co-expressed in glomerular endothelial cells. Similar to previous studies, we found that exendin-4 has renoprotective and anti-inflammatory effects in diabetic mice with nephropathy. Exendin-4 also decreased lipid droplet accumulation in vivo and increased glomerular endothelial cell cholesterol efflux in vitro. Additionally, exendin-4 inhibited the expression of the proinflammatory cytokines IL-6 and TNF-␣. The siRNA-mediated knockdown of ABCA1 increased inflammatory cytokine production in exendin-4-treated HRGECs, indicating that ABCA1 plays a crucial role in the anti-inflammatory activity of exendin-4.
Exendin-4 Improves ABCA1-mediated Cholesterol Efflux
din-4 promotes ABCA1 expression via the CaM kinase pathway in glomerular endothelial cells. In our study CaMKK expression was observed in the glomerular endothelial cells of mice in vivo, and CaMKIV phosphorylation did not differ significantly under normal glucose, high glucose, or high cholesterol conditions in vitro. However, exendin-4 treatment increased CaMKIV phosphorylation, an effect that was inhibited by the CaMKK inhibitor STO-609. In addition, STO-609 or CaMKIV siRNA suppressed exendin-4-induced ABCA1 expression, ABCA1-mediated cholesterol efflux, and anti-inflammatory effects. Taken together, our findings indicate that exendin-4 increases the phosphorylation of CaMKIV by CaMKK, which up-regulates ABCA1 expression, thereby enhancing cholesterol efflux and ameliorating inflammation.
Upstream mechanisms involved in GLP-1-mediated renal protection in diabetes are unclear but may involve activation of cAMP/PKA (52), which is a positive regulator of ABCA1 (18). Our results suggest that cAMP/PKA signaling contributes to the exendin-4-mediated restoration of ABCA1 expression under high glucose and high cholesterol conditions in HRGECs. This conclusion is supported by the fact that exendin-4-induced up-regulation of ABCA1 was abrogated by inhibiting the cAMP/PKA pathway. PI3K/AKT and ERK1/2 pathways also play key roles in GLP-1R action. For example, ABCA1 expression is regulated by the PI3K/AKT pathway (32,59). In addition, the GLP-1R agonist liraglutide suppresses damage in microvascular endothelial cells through the PI3K/ AKT pathway (33). In this study, our results showed that exendin-4 up-regulates ABCA1 expression through phosphorylation of PI3K and AKT to increase cholesterol efflux and decrease inflammation in HRGECs. In animal studies, ERK1/2 signaling is activated under hyperglycemic conditions in aortic vascular smooth muscle cells in vivo (34,60) and under high glucose conditions in cultured vascular smooth muscle cells. Similar to our findings, previous studies have shown that exendin-4 inhibits ERK1/2 signaling (53,61). Moreover, the ERK1/2 inhibitor PD98059 increased ABCA1 expression in HRGECs similar to exendin-4.
In summary, we have provided new evidence of lipid droplet accumulation and lipotoxicity in glomerular endothelial cells in a mouse model of DKD and in the kidneys of patients with DKD. We found that diabetes and hyperglycemia significantly decrease ABCA1 expression in glomerular endothelial cells both in vitro and in vivo. Exendin-4 significantly decreases lipid droplet accumulation, promotes cholesterol efflux, and exerts anti-inflammatory effects in glomerular endothelial cells by upregulating ABCA1 expression via activation of the CaMKK/ CaMKIV, cAMP/PKA, and PI3K/AKT pathways and inhibition of the ERK1/2 pathway. Our findings demonstrate that lipid accumulation in glomerular endothelial cells plays a crucial role in the development and progression of DKD. Thus, therapeutic agents that enhance GLP-1R action may prevent glomerular endothelial dysfunction and slow the progression of diabetic nephropathy. DECEMBER 16, 2016 • VOLUME 291 • NUMBER 51
Experimental Procedures
Human Kidney Sample Preparation-This study was approved by the Institutional Ethics Committee of the West China Hospital of Sichuan University. Three patients with early and advanced DKD were included in this study, and kidney tissues were obtained from renal biopsies. Three other patients with renal carcinoma and estimated glomerular filtration rate Ͼ 60 ml/min/1.73 m 2 , but without hypertension, diabetes, or other comorbidities, served as controls. The clinical data and laboratory test results of these patients are shown in Table 4. Kidney specimens from the controls were obtained during surgical nephrectomy. Lipid droplet deposition in the kidney biopsy specimens was detected by transmission electron microscopy (JEOL, Ltd., Tokyo) in the West China Hospital pathology department. Frozen sections of the kidney tissues were prepared for Oil Red O staining.
Animal Model-Animal experiments were performed with the approval of the Sichuan University Animal Ethics Committee. Male apoE Ϫ/Ϫ mice and C57BL/6J mice (6 weeks old, Huafukang Animal Centre, Beijing, China) were housed at the Laboratory Animal Centre of West China Hospital and given unrestricted access to water and food. The mice were divided into the following groups (each group, n ϭ 6): 1) wild type C57BL/6J controls; 2) nondiabetic apoE Ϫ/Ϫ mice (ApoE Ϫ/Ϫ ); 3) diabetic apoE Ϫ/Ϫ mice (ApoE Ϫ/Ϫ DM); and 4) exendin-4treated diabetic apoE Ϫ/Ϫ mice (ApoE Ϫ/Ϫ DM ϩ Ex4). All groups of mice were fed a high-fat diet (regular diet plus 27.3% lard, 54.6% sucrose, 16.4% cholesterol, and 1.6% sodium cholate (w/w)) (Beijing Keao Xieli Feed Co., Ltd., Beijing, China). Dietinduced obesity causes peripheral insulin resistance in rodents (62), resulting in inflammation and lipid accumulation (63). In our study, high-fat diet-fed apoE Ϫ/Ϫ mice with high homeostatic model assessment of insulin resistance (HOMA-IR) values (64) (fasting plasma glucose: (mmol/liter) ϫ fasting insulin (mIU/liter)/22.5) were defined as insulin-resistant and injected with four doses of low-dose streptozotocin (55 mg/kg, Sigma-Aldrich) in citrate buffer (pH 4.5) after overnight fasting. These mice served as a model of accelerated renal damage associated with diabetes due to dyslipidemia, which is also a feature of diabetic nephropathy in humans. The apoE Ϫ/Ϫ mice with fasting blood glucose Ͼ 16.7 mmol/liter for 3 consecutive days, as determined by a glucometer (Hoffmann-La Roche), served as the diabetic group in this study. The wild type C57BL/6J mice were injected with citrate buffer (1 ml/kg). All mice were maintained on their respective diets until the end of the study. The group of exendin-4-treated diabetic apoE Ϫ/Ϫ mice received intraperitoneal injections (1.0 nmol/kg/day, Sigma) for 8 weeks after diabetes induction.
Biochemical Measurements-Prior to euthanasia, blood glucose was measured using a glucose analyzer after a 6-h daytime fast. Urine samples were collected after the mice were housed individually in metabolic cages for 24 h. Blood samples were obtained by cardiac puncture. Blood urea nitrogen, plasma creatinine, plasma total cholesterol, plasma triglycerides, and urinary protein excretion were measured on a biochemistry autoanalyzer (Cobas Integra 400 Plus, Hoffmann-La Roche) using commercial kits.
Oil Red O Staining-Sections (6 m) of renal biopsies from mice and patients with early and advanced DKD were stained with Oil Red O according to the manufacturer's protocol (Sigma-Aldrich). The frozen kidney sections were rinsed with distilled water and then with 60% isopropanol, stained for 15 min in the Oil Red O working solution, and returned to distilled water. The sections were counterstained with hematoxylin for 1 min and mounted in glycerin jelly.
Filipin Staining-Sections (4 m) of fixed frozen mouse kidneys were fixed with 4% paraformaldehyde for 30 min, washed with PBS three times, and then stained with freshly prepared filipin solution (125 g/ml, Sigma-Aldrich) for 30 min. Then, the slides were washed with PBS, and a drop of glycerol was added. The slides were mounted with coverslips and examined by confocal laser scanning microscopy. A semi-quantitative analysis of filipin-positive areas was performed using the software package Image-Pro Plus, version 6.0.
Isolation of Glomeruli-Glomeruli were isolated from the renal cortex of mice by Dynabead perfusion using a previously reported method (65) with some modifications. The kidneys were removed, minced into 1-mm 3 pieces, and digested with collagenase (1 mg/ml collagenase A and 100 units/ml deoxyribonuclease I in Hanks' balanced salt solution (HBSS)) at 37°C for 30 min with gentle agitation. The collagenase-digested tissue was pressed gently through a 100-m cell strainer (Sangon Biotech, RA441) with a flattened pestle, and the cell strainer was washed with 5 ml of HBSS. The filtered cells were passed through a new cell strainer without pressing, and the cell strainer was washed with 5 ml of HBSS. The resulting cell suspension was then centrifuged at 2000 ϫ g for 5 min. The supernatant was discarded, and the cell pellet was resuspended in 2 ml of HBSS. Finally, glomeruli containing Dynabeads were gathered by a magnetic particle concentrator and washed at least three times with HBSS. During the procedure, the kidney tissues were kept at 4°C, except for collagenase digestion at 37°C, and acridine orange staining was used to identify the isolated glomeruli. The glomeruli were analyzed by real-time PCR and Western blotting.
Cell Culture-Well characterized HRGECs purchased from ScienCell Research Laboratories were cultured in endothelial cell medium (ScienCell) containing 5% FBS and 1% endothelial Transfection of siRNA-HRGECs were seeded in 6-well plates and flasks and maintained in endothelial cell medium with 5% FBS. Silencer negative control siRNA and Silencervalidated siRNA against CaMKIV and ABCA1 were purchased from Biology Engineering Corp. (Shanghai, China). Cells were transfected with siRNA diluted in Opti-MEM I (100 nmol/liter) using HiPerFect transfection reagent (Invitrogen) according to the manufacturer's protocol. The culture medium was changed 24 h after transfection. After washing twice with PBS, cells were collected for RT-PCR and Western blotting analysis.
cAMP Assay-HRGECs (3 ϫ 10 5 ) were seeded in a 24-well plate, and after being subjected to the experimental treatments, the cells were collected and lysed by freeze and thaw cycle 5 times. The samples were centrifuged at 1300 ϫ g for 5 min at room temperature. Intracellular cAMP concentrations in these supernatants were determined by using an ELISA kit (Jiang Lai Biotechnology, Shanghai, China) according to the manufacturer's instructions. Absorbance at 595 nm was detected using a microplate reader (BioTek Instruments, San Jose, CA) (66).
PKA Activity Assay-After treatment, cells were collected and placed in PBS containing protease inhibitor mixture (1:500, Sigma-Aldrich) and phosphatase inhibitor mixture (1:200, Sigma-Aldrich). Cells were then subjected to sonication to disrupt the cell membrane, and cytosolic proteins were harvested by centrifugation. PKA activity was determined using the Kemptide phosphorylation assay as described previously (66).
Lipid Extraction and Measurement of Lipid Composition-Lipids were extracted from the renal cortex as described previously (21,22). The triglyceride and cholesterol content was measured using kits (Sigma-Aldrich) according to the manufacturer's instructions.
Cholesterol Efflux Assay-After incubation with various concentrations of glucose, HRGECs were labeled with [ 3 H]cholesterol (1 mCi/well, Abcam) for 24 h at 37°C. The cells were then equilibrated with 0.5% BSA for 4 h and treated with apoA-I (20 mg/ml, Abcam), which was used to induce cholesterol efflux from the labeled cells for 6 h, as described previously (44). Both the cells and the cell culture medium were collected, and radioactivity was quantified by liquid scintillation counting using a Packard scintillation counter (PerkinElmer Life Sciences). The percentage of cholesterol efflux from cells was calculated as the radioactivity in the medium divided by the total radioactivity (in both cells and medium).
High-performance Liquid Chromatography Assay-Cell sterol analyses were performed as described previously (19,67). Briefly, cells were washed three times with PBS and then sonicated using an ultrasonic processor for 2 min. An appropriate volume of 0.5% NaCl (usually 1 ml) was added to 50 -200 g/ml cellular protein. An 0.1-ml aliquot of the cell solution (containing 5-20 g of protein) was used to measure free cholesterol, and an 0.1-ml aliquot was used to measure total cholesterol. Free cholesterol was dissolved in isopropanol (1 g cholesterol/ml) and stored at Ϫ20°C as a stock solution. Cholesterol standard calibration solutions ranging from 0 to 40 g cholesterol/ml were obtained by diluting the cholesterol stock solution in the same cell lysis buffer. Then 0.1 ml of each sample (cholesterol standard or cell suspension) was supplemented with 10 ml of the reaction mixture (500 mM MgCl 2 , 500 mM Tris-HCl (pH 7.4), 10 mM dithiothreitol, and 5% NaCl). To determine free cholesterol, 0.4 unit of cholesterol oxidase in 10 ml of 0.5% NaCl was added to each tube. To determine total cholesterol, 0.4 unit of cholesterol oxidase and 0.4 unit of cholesterol esterase were added to each tube. The tubes were incubated at 37°C for 30 min, and the reaction was stopped by adding 100 ml of the methanol:ethanol mixture (1:1). The reactions were kept cold for 30 min to allow protein precipitation and then centrifuged at 1000 ϫ g for 10 min at 15°C. The supernatant (10 ml) was applied to the chromatograph system (PerkinElmer Life Sciences), which consisted of a PerkinElmer
Exendin-4 Improves ABCA1-mediated Cholesterol Efflux
series 200 vacuum degasser, pump, PerkinElmer series 600 LINK, PerkinElmer series 200 UV-visible detector, and a Discovery C-18 HLPC column (Supelco, Inc.). The column was eluted using an isopropanol:n-heptane:acetonitrile mixture (35:13:52) at a flow rate of 1 ml/min for 8 min. Absorbance at 216 nm was monitored. Data were analyzed using TotalChrom software (PerkinElmer Life Sciences). RNA Isolation and Real-time PCR Analysis-Total RNA from glomeruli or cultured cells was extracted using TRIzol reagent (Thermo Fisher Scientific), and its concentration was measured using a microspectrophotometer (Thermo Fisher Scientific). After confirming the RNA quality by agar gel electrophoresis, cDNA was synthesized. Real-time PCR was performed with the CFX96 TM real-time PCR detection system (Bio-Rad) using SYBR Premix Ex Taq TM (Tli RNase H Plus, Takara) as described previously (64). The gene encoding -actin was used as an internal standard. Relative mRNA levels were calculated as the ratio of target mRNA to -actin mRNA and expressed as means Ϯ S.E.
Western Blotting Analysis-The cells and isolated glomeruli were harvested, and the proteins were extracted with radioimmune precipitation lysis buffer, separated by SDS-polyacrylamide gel electrophoresis, and transferred to a PVDF membrane for Western blotting. Briefly, after blocking nonspecific binding with 5% BSA, the membranes were incubated with primary antibodies against ABCA1, phospho-CaMKIV Thr-196, phospho-ERK1/2, phospho-PI3K, phospho-AKT, TNF-␣ (Abcam), and IL-6 (Santa Cruz Biotechnology); -actin was used as an internal reference. Membranes were then incubated with a secondary antibody followed by an ECL reagent (Western blotting chemiluminescence luminol reagent, Santa Cruz Biotechnology) and exposed to Canon EOS 60D film. The signals were detected with an Odyssey infrared imaging System (Li-COR Biosciences, Lincoln, NE) and quantified using the NIH ImageJ program. Relative protein levels were calculated as the ratio of target protein to -actin and expressed as mean Ϯ S.E.
Statistical Analysis-Data are expressed as mean Ϯ S.E. Multiple groups were compared by one-way analysis of variance. Two-way comparisons were performed using two-sample and paired t tests. All analyses were performed with SPSS software (version 11.5, IBM Corp.), and p Ͻ 0.05 was considered significant.
Author Contributions-F. L. and Q. Y. designed and conducted the study. Q. Y., R. Z., and Y. W. participated in the data collection and analysis. The manuscript was prepared by Q. Y., R.Z., L. L., Y. W., J. Z., and L. B. and revised by J. L., J. C., P. F., and F. L. All authors participated in discussions about the manuscript and approved the final version. | 6,571.6 | 2016-10-26T00:00:00.000 | [
"Biology"
] |
Fluorinated Polyimide-Film Based Temperature and Humidity Sensor Utilizing Fiber Bragg Grating
We propose and demonstrate a temperature and humidity sensor based on a fluorinated polyimide film and fiber Bragg grating. Moisture-induced film expansion or contraction causes an extra strain, which is transferred to the fiber Bragg grating and leads to a humidity-dependent wavelength shift. The hydrophobic fluoride doping in the polyimide film helps to restrain its humidity hysteresis and provides a short moisture breathing time less than 2 min. Additionally, another cascaded fiber Bragg grating is used to exclude its thermal crosstalk, with a temperature accuracy of ±0.5 °C. Experimental monitoring over 9000 min revealed a considerable humidity accuracy better than ±3% relative humidity, due to the sensitized separate film-grating structure. The passive and electromagnetic immune sensor proved itself in field tests and could have sensing applications in the electro-sensitive storage of fuel, explosives, and chemicals.
Introduction
Humidity monitoring and regulation are of great significance in pharmacy, semiconductors, and costly facility maintenance and explosive storage [1][2][3]. Relative humidity, defined as the ratio of vapor content in the air, describes the present moisture content compared with what the air can hold at most [4]. Mechanical psychrometers and dew-point hygrometers are limited in humidity sensing due to their low accuracy and complexity [1]. Chemical methods mainly depend on the hygroscopic reaction, which has been found to be irreversible and nonreusable [5]. Thus, capacitance and resistance are selected as measurable signs of humidity change, which are accurate, fast-responding, and small-scaled [3]. However, for safety reasons, active electronic humidity sensors prove unsuitable for electro-sensitive applications such as explosives, chemicals and fuel storage [1][2][3][4].
Later, he demonstrated a humidity sensor with polyimide coating at the grating by measuring the moisture-induced expansion [22]. T. Yeo and K. Grattan analyzed the strain transfer between the polyimide coating and the fiber Bragg grating, and explained the relationship between humidity and wavelength shift in detail [23]. The linear sensitivity was estimated to be +4.5 pm/RH with an uncertainty of ±4% RH, while the polyimide coating resulted in humidity hysteresis and a response time over 40 min. Recently, similar attempts have been made for a higher sensitivity and a shorter breathing time, which are hard to achieve simultaneously with an ordinary polyimide film in a conventional structure. The reason lies in the tough hydrogen bonding in polyimide with a porous surface, which causes humidity hysteresis and even moisture agglomeration [24,25]. Moreover, the moisture exchange and humidity sensitivity are restricted with each other depending on the tradeoff between film thickness and surface, where it is difficult to balance the response time and accuracy. According to the aforementioned analysis, a humidity sensor with high performance still remains to be realized both in chemical modification and structural design.
In this paper, a fluorinated polyimide-film-based temperature and humidity sensor utilizing fiber Bragg grating is proposed and demonstrated. The sensor was constructed with an FBG and a 20 µm-thick fluorinated polyimide film by direct manual gluing. Moisture-induced film expansion or contraction causes an extra strain, which is transferred to the FBG and leads to a humidity-dependent wavelength shift. Compared with the coating method, the separate film-grating structure provides a large surface area for moisture exchange, an improved dynamic range, and considerable sensitivity. The -CF 3 modification in the monomer of the polyimide reduces its moisture capacity, restrains its humidity hysteresis, and provides a short moisture breathing time less than 2 min. Another cascaded FBG temperature sensor is used to exclude its thermal crosstalk, with a temperature accuracy of ±0.5 • C. Experimental monitoring over 9000 min revealed a relative humidity accuracy better than ±3% RH, while field testing for tobacco storage proved its stability and practicability as designed. The passive and electromagnetism-immune sensor with an accurate and fast response can be applied for the electro-sensitive storage of explosives, chemicals, and fuels.
Functional Material
For common polyimide films, the vapor in the air is easily captured by the hydrophilic hydrogen bonds and porous surface. The strong absorption causes a slow moisture release and may even lead to humidity hysteresis and agglomeration. Thus, chemical modification is necessary to improve the surface properties of the polyimide film. The fluorinated polyimide film was synthesized and provided by the group of Pro. L. Fan, Institute of Chemistry Chinese Academy of Sciences, and is commercially available with excellent hysteresis resistance [26]. As can be seen from the chemical formula and the molecular structure of the fluorinated polyimide shown in Figure 1a, the trifluoromethyl modification replaces the ether bond in the original aromatic monomer with a molecular weight of 630. In the detailed straight-chain polycondensation, the fluorine modification improves the hydrophobicity of the film [26,27]. With the increase in fluorine content from 15.32% to 30.16% in the monomer, the transparent film becomes an applicable candidate for a fast-response humidity sensor. Compared with a non-fluorinated film synthesized with a similar monomer with a molecular weight of 496 such as the dark one shown in Figure 1b, the trifluoromethyl acts in a moisture-proof manner and restrains its capacity from 5% to 1%, which is quantified by weight gain at saturated humidity. Thus, with such a hydrophobic modification, a small humidity hysteresis and a short breathing time can be expected when using a fluorinated polyimide film. As the moisture capacity is significantly reduced in a fluorinated polyimide film, the sensitivity to humidity declines sharply along with the insufficient expansion. The conventional coating method is no longer suitable for practical use, with which it is hard to achieve a fast response, a low hysteresis, and considerable sensitivity simultaneously.
Sensor Structure
Thus, a separate film-grating structure was specifically designed to maintain the dynamic range and the humidity sensitivity, as a result of the large cross-sectional ratio between the functional film and fiber grating. As shown in Figure 2a, the sensing unit was fabricated with a bare FBG of 125 μm in diameter and a film of 20 μm in thickness. The FBG and the film were assembled with an aluminous convertor and clamped between two brackets of a hollow frame by direct manual gluing. The unit was pre-stretched at 2000 με to ensure an approximately linear spectral response. Once the humidity changes, the film self-rebalances by exchanging moisture with that in the air. In this way, the expansion or contraction leads to an axial strain change transferred to the FBG. Finally, the humidity variation is visualized by tracing the wavelength shift, which indicates dynamic moisture absorption and release. Moreover, Figure 2b shows another cascaded FBG temperature sensor to exclude the thermal crosstalk and extract the pure relative-humidity changes from the co-effects on the wavelength shift, as circled with the red dotted line.
Working Principles
The humidity sensing mechanism is further discussed below. According to the equal stress in a rigid connection as in Equation (1), the strain transferred to the FBG is strictly associated with the expansion and contraction of the polyimide film, As the moisture capacity is significantly reduced in a fluorinated polyimide film, the sensitivity to humidity declines sharply along with the insufficient expansion. The conventional coating method is no longer suitable for practical use, with which it is hard to achieve a fast response, a low hysteresis, and considerable sensitivity simultaneously.
Sensor Structure
Thus, a separate film-grating structure was specifically designed to maintain the dynamic range and the humidity sensitivity, as a result of the large cross-sectional ratio between the functional film and fiber grating. As shown in Figure 2a, the sensing unit was fabricated with a bare FBG of 125 µm in diameter and a film of 20 µm in thickness. The FBG and the film were assembled with an aluminous convertor and clamped between two brackets of a hollow frame by direct manual gluing. The unit was pre-stretched at 2000 µε to ensure an approximately linear spectral response. As the moisture capacity is significantly reduced in a fluorinated polyimide film, the sensitivity to humidity declines sharply along with the insufficient expansion. The conventional coating method is no longer suitable for practical use, with which it is hard to achieve a fast response, a low hysteresis, and considerable sensitivity simultaneously.
Sensor Structure
Thus, a separate film-grating structure was specifically designed to maintain the dynamic range and the humidity sensitivity, as a result of the large cross-sectional ratio between the functional film and fiber grating. As shown in Figure 2a, the sensing unit was fabricated with a bare FBG of 125 μm in diameter and a film of 20 μm in thickness. The FBG and the film were assembled with an aluminous convertor and clamped between two brackets of a hollow frame by direct manual gluing. The unit was pre-stretched at 2000 με to ensure an approximately linear spectral response. Once the humidity changes, the film self-rebalances by exchanging moisture with that in the air. In this way, the expansion or contraction leads to an axial strain change transferred to the FBG. Finally, the humidity variation is visualized by tracing the wavelength shift, which indicates dynamic moisture absorption and release. Moreover, Figure 2b shows another cascaded FBG temperature sensor to exclude the thermal crosstalk and extract the pure relative-humidity changes from the co-effects on the wavelength shift, as circled with the red dotted line.
Working Principles
The humidity sensing mechanism is further discussed below. According to the equal stress in a rigid connection as in Equation (1), the strain transferred to the FBG is strictly associated with the expansion and contraction of the polyimide film, Once the humidity changes, the film self-rebalances by exchanging moisture with that in the air. In this way, the expansion or contraction leads to an axial strain change transferred to the FBG. Finally, the humidity variation is visualized by tracing the wavelength shift, which indicates dynamic moisture absorption and release. Moreover, Figure 2b shows another cascaded FBG temperature sensor to exclude the thermal crosstalk and extract the pure relative-humidity changes from the co-effects on the wavelength shift, as circled with the red dotted line.
Working Principles
The humidity sensing mechanism is further discussed below. According to the equal stress in a rigid connection as in Equation (1), the strain transferred to the FBG is strictly associated with the expansion and contraction of the polyimide film, where ε pi (ε G ) and E pi (E G ) refer to the strain and the elastic modulus of the film (grating), respectively. W pi and d pi describe the transverse section of film, while R G denotes the radius of the bare fiber. Besides, the polyimide film expansion strictly equals the contraction in the FBG, fixing an isolated length at a constant temperature in Equation (2).
α pi and H represent the hygroscopic expansion coefficient and the humidity variation, and L G and L pi stand for the lengths of the fiber grating and film available, respectively. After derivation calculus on Equation (2) and simplification with Equation (1), we deduce the differential sensitivity S H as the humidity-dependent wavelength shift, where k is defined as the transfer coefficient between the axial strain and wavelength shift, and ε 0 refers to the initialized strain in the FBG. The negative expression indicates a blue-shift direction with an alterable sensitivity, which can be enhanced by increasing the cross-sectional ratio and the length ratio between the film and FBG. A slight nonlinearity in the sensitivity is introduced at different humidity, which can be well corrected with a second-polynomial coefficient.
Results
The temperature and humidity sensors were calibrated with a highly accurate Humidity Generator (GEO Calibration, Model 2000), and their sensing performance was validated in humidity-mimicking enclosures created with different saturated solutions. As shown in Figure 3, the sensors were exposed in different humidity enclosures of 12% (LiCl), 33% (MgCl 2 ), 60% (NaBr), 75% (NaCl), and 98% (K 2 SO 4 ). The wavelengths measured at humidity of 10.4%, 34.2%, 57.6%, 76.1%, and 96.2% at 25 • C were distributed near the calibration fitting curve, similar with those at other temperatures.
Sensors 2020, 20, x FOR PEER REVIEW 4 of 9 where εpi (εG) and Epi (EG) refer to the strain and the elastic modulus of the film (grating), respectively. Wpi and dpi describe the transverse section of film, while RG denotes the radius of the bare fiber. Besides, the polyimide film expansion strictly equals the contraction in the FBG, fixing an isolated length at a constant temperature in Equation (2).
αpi and H represent the hygroscopic expansion coefficient and the humidity variation, and LG and Lpi stand for the lengths of the fiber grating and film available, respectively. After derivation calculus on Equation (2) and simplification with Equation (1), we deduce the differential sensitivity SH as the humidity-dependent wavelength shift, where k is defined as the transfer coefficient between the axial strain and wavelength shift, and ε0 refers to the initialized strain in the FBG. The negative expression indicates a blue-shift direction with an alterable sensitivity, which can be enhanced by increasing the cross-sectional ratio and the length ratio between the film and FBG. A slight nonlinearity in the sensitivity is introduced at different humidity, which can be well corrected with a second-polynomial coefficient.
Results
The temperature and humidity sensors were calibrated with a highly accurate Humidity Generator (GEO Calibration, Model 2000), and their sensing performance was validated in humiditymimicking enclosures created with different saturated solutions. As shown in Figure 3, the sensors were exposed in different humidity enclosures of 12% (LiCl), 33% (MgCl2), 60% (NaBr), 75% (NaCl), and 98% (K2SO4). The wavelengths measured at humidity of 10.4%, 34.2%, 57.6%, 76.1%, and 96.2% at 25 °C were distributed near the calibration fitting curve, similar with those at other temperatures. As assumed in theoretical analysis, the polynomial fitting curves exhibit an approximately linear dependence between the humidity and the wavelength shift. Additionally, the slight nonlinearity can be estimated with a second coefficient as the humidity varies from 12% RH to 98% RH. Its central wavelength shift towards the blue direction from 1535.1 to 1534.6 nm agrees well with the looseness of the polyimide film and decrement in axial strain. The temperature and humidity sensitivities were As assumed in theoretical analysis, the polynomial fitting curves exhibit an approximately linear dependence between the humidity and the wavelength shift. Additionally, the slight nonlinearity can be estimated with a second coefficient as the humidity varies from 12% RH to 98% RH. Its central wavelength shift towards the blue direction from 1535.1 to 1534.6 nm agrees well with the looseness of the polyimide film and decrement in axial strain. The temperature and humidity sensitivities were obtained as +5.033 pm/ • C and −6.09 pm/%RH, respectively, providing opposite spectral responses to better exclude the thermal crosstalk.
The spectral response to humidity change is compared with that of a CE314 electronic sensor, and the humidity hysteresis is discussed in Figure 4. Both the FBG sensor and the electronic sensor were exposed in the same enclosures of the LiCl (12% RH) and K 2 SO 4 (98% RH) solutions, as well as in the lab (62% RH). The red line above indicates the humidity recorded by CE314, while the blue line represents the wavelength shift determined with the FBG humidity sensor. As the humidity changes among 12% RH, 98% RH, and 62% RH, the central wavelength of the FBG exhibits a red-shift or blue-shift in response to a decrement or increment in the humidity, respectively. Due to the fluoride doping in the polyimide film, a short breathing time less than 2 min is observed with low hysteresis. Besides, the considerable sensitivity and accuracy are preserved, as a consequence of the separate film-grating structure. Thus, the humidity sensor is experimentally proved to be feasible with considerable sensitivity, a fast response, and low humidity hysteresis.
Sensors 2020, 20, x FOR PEER REVIEW 5 of 9 The spectral response to humidity change is compared with that of a CE314 electronic sensor, and the humidity hysteresis is discussed in Figure 4. Both the FBG sensor and the electronic sensor were exposed in the same enclosures of the LiCl (12% RH) and K2SO4 (98% RH) solutions, as well as in the lab (62% RH). The red line above indicates the humidity recorded by CE314, while the blue line represents the wavelength shift determined with the FBG humidity sensor. As the humidity changes among 12% RH, 98% RH, and 62% RH, the central wavelength of the FBG exhibits a red-shift or blueshift in response to a decrement or increment in the humidity, respectively. Due to the fluoride doping in the polyimide film, a short breathing time less than 2 min is observed with low hysteresis. Besides, the considerable sensitivity and accuracy are preserved, as a consequence of the separate film-grating structure. Thus, the humidity sensor is experimentally proved to be feasible with considerable sensitivity, a fast response, and low humidity hysteresis.
. In the experiment as shown in Figure 5, the sensors were applied for practical monitoring after aging for one month. The signals from the Amplified Spontaneous Emission (ASE) source were modulated by temperature and humidity sensors, harvested by an Optical Spectral Analyzing (OSA) module, and demodulated with pre-recorded calibration coefficients. The temperature and relative humidity visually appeared on the display and were continuously recorded backstage. Figure 6 shows the temperature varying from 17 to 35 °C for the FBG sensors and an electronic sensor CE314, which is consistent between each. The deviations between the black, red, green, and blue lines demonstrate the temperature differences over 9000 min of monitoring. Mostly, the temperature deviations are well restricted within ±0.5 °C, while such errors beyond tolerance occasionally occur at the perturbations as circled by the dashed lines above. In the experiment as shown in Figure 5, the sensors were applied for practical monitoring after aging for one month. The signals from the Amplified Spontaneous Emission (ASE) source were modulated by temperature and humidity sensors, harvested by an Optical Spectral Analyzing (OSA) module, and demodulated with pre-recorded calibration coefficients. The temperature and relative humidity visually appeared on the display and were continuously recorded backstage.
Sensors 2020, 20, x FOR PEER REVIEW 5 of 9 The spectral response to humidity change is compared with that of a CE314 electronic sensor, and the humidity hysteresis is discussed in Figure 4. Both the FBG sensor and the electronic sensor were exposed in the same enclosures of the LiCl (12% RH) and K2SO4 (98% RH) solutions, as well as in the lab (62% RH). The red line above indicates the humidity recorded by CE314, while the blue line represents the wavelength shift determined with the FBG humidity sensor. As the humidity changes among 12% RH, 98% RH, and 62% RH, the central wavelength of the FBG exhibits a red-shift or blueshift in response to a decrement or increment in the humidity, respectively. Due to the fluoride doping in the polyimide film, a short breathing time less than 2 min is observed with low hysteresis. Besides, the considerable sensitivity and accuracy are preserved, as a consequence of the separate film-grating structure. Thus, the humidity sensor is experimentally proved to be feasible with considerable sensitivity, a fast response, and low humidity hysteresis.
. In the experiment as shown in Figure 5, the sensors were applied for practical monitoring after aging for one month. The signals from the Amplified Spontaneous Emission (ASE) source were modulated by temperature and humidity sensors, harvested by an Optical Spectral Analyzing (OSA) module, and demodulated with pre-recorded calibration coefficients. The temperature and relative humidity visually appeared on the display and were continuously recorded backstage. Figure 6 shows the temperature varying from 17 to 35 °C for the FBG sensors and an electronic sensor CE314, which is consistent between each. The deviations between the black, red, green, and blue lines demonstrate the temperature differences over 9000 min of monitoring. Mostly, the temperature deviations are well restricted within ±0.5 °C, while such errors beyond tolerance occasionally occur at the perturbations as circled by the dashed lines above. Figure 6 shows the temperature varying from 17 to 35 • C for the FBG sensors and an electronic sensor CE314, which is consistent between each. The deviations between the black, red, green, and blue lines demonstrate the temperature differences over 9000 min of monitoring. Mostly, the temperature deviations are well restricted within ±0.5 • C, while such errors beyond tolerance occasionally occur at the perturbations as circled by the dashed lines above.
Sensors 2020, 20, x FOR PEER REVIEW 6 of 9 Figure 6. Temperature variations recorded by 3 FBG sensors and the CE314 sensor (a); temperature deviations between these 3 FBG sensors and the CE314 sensor (b). Figure 7 shows the humidity variation aligned with the monitoring timeline of the temperature. The humidity is also demonstrated with black, red, green, and blue lines, which are consistent with each other as well. Basically, the humidity deviations are restricted within ±3% RH, while the errors mainly occur at temperature jumping. Moreover, as shown in Figure 8, the practicability was verified through field tests in a tobacco warehouse, Fujian province, Southeast of China. As the tobacco standards require, the unprocessed tobacco should be stored at 20 to 30 °C, and 65% RH to 75% RH. Three FBG temperature/humidity sensors were placed in Storage 1, Storage 2, and the outdoor area, which were connected to Channel 1, Channel 2, and Channel 3 via commercial cables with an insertion loss of about 1.1 dB. The backward optical signals were harvested by a distant interrogator and demodulated to real-time temperature and humidity. During 18 h of discontinuous monitoring over 3 days, data were recorded by the FBG sensors and are expressed with solid lines, while those from the manual inspections with CE314 are depicted with scatter points. By comparison, the FBG sensors agreed well with the electronic sensor, whether in storage or outdoors. Since the circumstance in the storage was relatively stable, the temperature varied from 27.5 to 28.4 °C and the humidity changed from 68% RH to 74% RH. Due to the airflow perturbations outdoors, the temperature varied from 28.7 to 35.3 °C and the humidity changed from 46% RH to 73% RH. The deviation of T ≤ 0.3 °C and H ≤ 2.8% RH outdoors is larger than Figure 7 shows the humidity variation aligned with the monitoring timeline of the temperature. The humidity is also demonstrated with black, red, green, and blue lines, which are consistent with each other as well. Basically, the humidity deviations are restricted within ±3% RH, while the errors mainly occur at temperature jumping. Moreover, as shown in Figure 8, the practicability was verified through field tests in a tobacco warehouse, Fujian province, Southeast of China. As the tobacco standards require, the unprocessed tobacco should be stored at 20 to 30 °C, and 65% RH to 75% RH. Three FBG temperature/humidity sensors were placed in Storage 1, Storage 2, and the outdoor area, which were connected to Channel 1, Channel 2, and Channel 3 via commercial cables with an insertion loss of about 1.1 dB. The backward optical signals were harvested by a distant interrogator and demodulated to real-time temperature and humidity. During 18 h of discontinuous monitoring over 3 days, data were recorded by the FBG sensors and are expressed with solid lines, while those from the manual inspections with CE314 are depicted with scatter points. By comparison, the FBG sensors agreed well with the electronic sensor, whether in storage or outdoors. Since the circumstance in the storage was relatively stable, the temperature varied from 27.5 to 28.4 °C and the humidity changed from 68% RH to 74% RH. Due to the airflow perturbations outdoors, the temperature varied from 28.7 to 35.3 °C and the humidity changed from 46% RH to 73% RH. The deviation of T ≤ 0.3 °C and H ≤ 2.8% RH outdoors is larger than Moreover, as shown in Figure 8, the practicability was verified through field tests in a tobacco warehouse, Fujian province, Southeast of China. As the tobacco standards require, the unprocessed tobacco should be stored at 20 to 30 • C, and 65% RH to 75% RH. Three FBG temperature/humidity sensors were placed in Storage 1, Storage 2, and the outdoor area, which were connected to Channel 1, Channel 2, and Channel 3 via commercial cables with an insertion loss of about 1.1 dB. The backward optical signals were harvested by a distant interrogator and demodulated to real-time temperature and humidity. During 18 h of discontinuous monitoring over 3 days, data were recorded by the FBG sensors and are expressed with solid lines, while those from the manual inspections with CE314 are depicted with scatter points. By comparison, the FBG sensors agreed well with the electronic sensor, whether in storage or outdoors. Since the circumstance in the storage was relatively stable, the temperature varied from 27.5 to 28.4 • C and the humidity changed from 68% RH to 74% RH. Due to the airflow perturbations outdoors, the temperature varied from 28.7 to 35.3 • C and the humidity changed from 46% RH to 73% RH. The deviation of T ≤ 0.3 • C and H ≤ 2.8% RH outdoors is larger than that in storage, of T ≤ 0.2 • C and H ≤ 2.2% RH, which is also consistent with actual situations.
Conclusions
We propose and demonstrate a fluorinated-film-based FBG temperature and humidity sensor. The fluoride doping resolves the humidity hysteresis and provides a short breathing time less than 2 min. The calibrated sensors show different humidity corresponding to enclosures created with saturated solutions and good stability during 9000 min of monitoring in the lab. Additionally, the field tests prove the outstanding performance and practicability of the humidity sensors, but the consistency, size, weight, and cost of the humidity sensors can be further improved in the future, revealing potential for pharmacy, semiconductors, facility and equipment protection, and chemical and explosive storage.
Conclusions
We propose and demonstrate a fluorinated-film-based FBG temperature and humidity sensor. The fluoride doping resolves the humidity hysteresis and provides a short breathing time less than 2 min. The calibrated sensors show different humidity corresponding to enclosures created with saturated solutions and good stability during 9000 min of monitoring in the lab. Additionally, the field tests prove the outstanding performance and practicability of the humidity sensors, but the consistency, size, weight, and cost of the humidity sensors can be further improved in the future, revealing potential for pharmacy, semiconductors, facility and equipment protection, and chemical and explosive storage. | 6,153 | 2020-09-24T00:00:00.000 | [
"Materials Science",
"Engineering",
"Physics"
] |
Assessing Relative Performance of Econometric Models in Measuring the Impact of Climate Change on Agriculture Using Spatial Autoregression
Although econometric models have been widely used to measure the impact of climate change on agriculture, there exist differences among the modelers on which specification should be preferred. To help explain the discrepancies, this paper assesses four different econometric models, i.e. OLS, panel, and two spatial models using a South American agricultural household data. The relationship among the econometric specifications is examined in terms of the freedom given to a spatial autoregressive parameter. In spatial models, the spatial parameter is free within the model, but is fixed a priori in the aspatial models. Empirical results show a high correlation of the land values across South America. Spatial models result in somewhat lower climate change impact estimates than those from the aspatial models.
INTRODUCTION
Econometric models have been widely used to measure the impact of climate change on agriculture in the United States and around the world (Mendelsohn, Nordhaus, andShaw, 1994, 1996;Kumar and Parikh, 2001;Reinsborough, 2001;Timmins, 2005;Seo, Mendelsohn, and Munasinghe, 2005;Schlenker, Hanemann, and Fisher, 2005;Kurukulasuriya et al., 2006;Deschenes and Greenstone, 2007;Seo and Mendelsohn, 2008a).Using a cross-sectional data, researchers regressed land values or net revenues against climate variables after carefully controlling non-climate factors.Although many cross-sectional studies have been conducted, there still exist substantial discrepancies among the modelers on which specification should be preferred, i.e. whether they should use a weighted OLS (Mendelsohn, Nordhaus, and Shaw, 1994), or a panel model with year and county fixed effects (Deschenes and Greenstone, 2007), or a spatially correlated error model of heating degree days at the county level (Schlenker, Hanemann, and Fisher, 2006).There is a need in the climate research community to assess relative performance of these different econometric models.This task is especially important since previous econometric models resulted in different damage estimates ranging from a slight benefit from climate change (Deschenes and Greenstone, 2007) to a more than 50 percent loss of agricultural income by the mid century (Schlenker, Hanemann, and Fisher, 2006), 1 which lead to opposing policy recommendations.
+ The data used in this study came from a World Bank household survey in South America.I thank the editor and two anonymous reviewers for constructive comments.
Senior Fellow at the University of Sydney, Address: Faculty of Agriculture, Food and Natural Resources, The University of Sydney, NSW 2006, Australia.Email<EMAIL_ADDRESS>1 The large loss in U.S. agriculture was attributed by the authors to a proper accounting of irrigation system in the western US States.They limited their analysis to non-irrigated rain-fed farms.Following the authors, this paper will also provide a similar sensitivity analysis focusing on the rain-fed farms.
The Review of Regional Studies, Vol. 38, No 2, 2008 © Southern Regional Science Association 2010.
To explain different outcomes among the models, this paper pays particular attention to spatial characteristics of economic data.In contrast to experimental crop models, cross-sectional studies of land values are prone to spatial dependence.That is, land values are determined by their locations (von Thünen, 1826;Hartwick and Olewiler, 1986).For example, two neighboring plots are more likely to have similar values than are two agricultural plots that are further apart.Consequently, econometric models that do not address this spatial nature will result in biased climate parameter estimates because they wrongly assign the difference in land values resulting from spatial dependence to that from climate variation (Anselin, 1988;Dubin, 1988;Benirschka and Binkley, 1994;Olmo, 1995).
With these points in mind, this paper assesses past econometric models in terms of their accounting of spatial dependence in land values.We examine two spatial econometric models, spatial lag and spatial error models, and compare them directly with two aspatial models: OLS and panel models.A systematic comparison is made among the four econometric models on the basis of the freedom given to a spatial autoregressive parameter in each model.Different spatial weight matrices are examined.The four specifications are applied to the agricultural data in South America during the 2003-2004 agricultural seasons.Spatial models are then used to examine whether rain-fed non-irrigated farms are more vulnerable (Schlenker, Hanemann, and Fisher, 2005).The paper also tests whether the damage estimate would depend upon the capacity to adapt (Kelly, Kolstad, and Mitchell, 2005).
A CATEGORY OF ECONOMETRIC MODELS
The original modelers of the econometric approach applied to a cross-sectional data argued that the method provides an alternative approach to measure the impacts of climate change on agriculture that fully accounts for the possible substitution behaviors of the farmers when climate changes (Mendelsohn, Nordhaus, and Shaw, 1994).This approach has since been applied to the U.S., Brazil, India, Sri Lanka, the United Kingdom, Canada, Africa, and Latin America (Mendelsohn, Nordhaus, andShaw, 1994, 1996;Kumar and Parikh, 2001;Reinsborough, 2001;Timmins, 2005;Seo, Mendelsohn, andMunasinghe, 2005: Kurukulasuriya et al., 2006;Seo and Mendelsohn, 2008a).Recently, researchers analyzed annual net revenues over multiple years for U.S. agriculture to correct for year-county fixed effects (Deschenes and Greenstone, 2007).Another group of researchers analyzed the sensitivity of the land value in the U.S. to heating degree days while correcting for spatially correlated errors, but the authors relied on the county-level land values (Schlenker, Hanemann, and Fisher, 2006).
All of these models are based on the assumption that each farmer maximizes net revenue subject to the exogenous conditions of his farm.If he chooses optimal outputs (Q*) and inputs (M*), then the maximized net revenue will be a function of just the variables that are exogenous to the farmer: , where P Q and P M are output and input prices, C a set of climate variables, S a set of soil variables, W a set of water flow variables, and H a vector of socio-economic variables.Economists estimated Equation (1) by running a form of OLS regression over a set of independent variables as follows: (2) = + , y Xβ e © Southern Regional Science Association 2010.
where y is a vector of land value (or net revenue) per hectare of land, X a vector of regressors, β a vector of parameters, and ε = (ε 1 , ε 2 ,…, ε n ) T .It is assumed that the error term follows the classical assumption of an independently and identically distributed (iid) normal error distribution.
However, cross-sectionally collected land values are spatially dependent, i.e. land values are determined by its location (von Thünen, 1826;Hartwick and Olewiler, 1986).As the land is further apart from the town center, so does the value of the land decline steadily.An urban land use is no longer profitable at some location beyond which crop land use is more profitable.A land use for crops is no longer profitable at some location beyond which grasslands for livestock are more profitable (See Seo andMendelsohn [2008b, 2008c] for the discussion of livestock management).And, again, grassland livestock management is no longer profitable at some location beyond which farmers pick fruit and sell timber from the forests.
The spatial dependence of land values can also be thought of as spatially correlated land values.Two neighboring lands are more likely to have similar land values than the two agricultural plots that are farther apart.As the distance from one plot of land to another plot increases, the correlation between the two plots decreases.Farmland values within a village are more similar than they are within a country.Land values within a district are more similar than those within a province.
The existence of spatial dependence can be tested using spatial correlation test statistics.Let's partition a spatial sample of interest, South American farms in this study, into nonoverlapping subsets such that they are exhaustive and exclusive.If land values within each district are assumed to be correlated, then each subset will correspond to each district in the whole sample.Then the spatial lag operator is defined as: (3) where for each row i, the matrix elements ij w are only non-zero for those j in subset i and the ij w equal unity for all j =1 to N. Consequently, the spatial lag operator is interpreted as a weighted average of the neighboring land values or as a spatial smoother. 2The test of spatial auto correlation can be based on the above spatial lag operator.A most cited statistic, Moran's I, is defined as: where e is a vector of OLS residuals and to 1 and is close to zero when there is little spatial autocorrelation.Another statistic used as often is Geary's C, which turns out to be expressed as a function of Moran's I statistic.
The spatial dependence can be modeled explicitly (Anselin, 1988).A simplest form of a spatial model that uses a spatially lagged dependent variable, spatial lag model (SL model), is specified as: The Review of Regional Studies, Vol. 38, No 2, 2008 © Southern Regional Science Association 2010. ( where is a spatial autoregressive coefficient and u is a vector of error terms that are identically and independently normal distributed.The SL model is preferred when authors are concerned about the influence of the spatial dependence on the final outcomes.Note that even though the errors in Equation ( 5) are independent, land values are correlated, which can be seen by reorganizing the above equation as: The SL model is appropriate when researchers want to examine the effects of spatial dependence of farmland values on the estimates of climate change impacts.Another spatial process model, the spatial error (SE) model, is appropriate when researchers are interested in non-spherical error term in which the off-diagonal elements of the covariance matrix express the structure of spatial dependence: where all the terms are defined as previously.Using Equations ( 3) and ( 7), the SE model is defined as: The parameters in Equations ( 5) and ( 8) are estimated by maximum likelihood method or generalized methods of moment. 3Past econometric models can be understood as a special case of the SL or SE model by the degree of freedom given to the spatial parameter as follows: (9a) OLS model when First, note that the spatial model should perform best among the three models because the spatial autoregressive parameter ρ is not constrained a priori and determined optimally within the model.Second, the panel model in Equation ( 9b) is a district (or county in the U.S.) fixed effect model as there is no time component in the data.As the authors correctly stated in the paper, it is difficult to use time series data to measure climate change impacts because there is no variation in the climate year by year as climate is defined as the thirty-year average weather.Land values also cannot be used since there is no variation in the land values year by year (Deschenes and Greenstone, 2007).However, multiple year net revenues (or land values) can be still used to control for the county fixed effects using the SL model in Equation ( 5).In addition, it is possible to use climate data instead of the year by year weather data using the SL or SE model.Hence, the spatial model is a broader econometric model and should perform best among the econometric models currently used in the literature.
DATA
The three econometric models in Equations ( 9a), (9b), and (9c) were applied to South American agriculture data collected throughout the continent during the agricultural seasons © Southern Regional Science Association 2010.from 2003 to 2004 (Seo and Mendelsohn, 2008a).As part of a World Bank project, household surveys were conducted from the following seven countries: Argentine, Uruguay, Chile, Brazil, Venezuela, Ecuador, and Colombia.These countries were chosen to cover a broad range of climate zones.Within each country, districts were chosen to represent diverse climate zones within that country.In each country, between 15 and 30 districts were selected and 20 to 30 households were interviewed in each district.Within each county, cluster sampling was done to limit the cost of the survey.The surveys cover agricultural activities during the agricultural period from July 2003 to June 2004.
Temperature data were obtained from the satellites operated by the U.S. Department of Defense, which recorded surface temperature twice daily at the centroids of districts from 1988 to 2004(Bassist et al., 2001: Mendelsohn et al., 2007).Precipitation data were obtained from the weather station records by the World Meteorological Organization from 1960 to 1990 (WMO, 1989).Seasonal climates are defined by the three-month averages.For example, summer temperature is defined as the average of December, January, and February temperatures in the Southern Hemisphere and winter temperature as the average of June, July, and August temperatures.The seasons in the Northern Hemisphere are defined reversely by switching summer and winter.Soil data were obtained from the Food and Agricultural Organization (FAO) digital soil map of the world CD ROM (FAO, 2001).
EMPIRICAL RESULTS
The sample is summarized using descriptive statistics of several important variables in Table 1.Temperature is lowest at 9 degrees Celsius in Chile and highest at 20 degrees Celsius in Brazil, Colombia, and Venezuela.Chile is also the driest country with 63 mm of rainfall per month.Colombia and Venezuela are the wettest places with 132 mm of rainfall per month.The value of a hectare of land ranges from $1,000 to $2,000 US. 4 Number of collected surveys varied from fewer than 200 in Ecuador to 700 in Brazil to reflect the size difference of the countries.There was a relatively smaller number of surveys collected from Uruguay and Venezuela, so we added Uruguay to the Argentine sample and Venezuela to the Colombia sample.In both cases, they border each other and have similar geographic and climatic conditions as well as agriculture practices.As discussed in the theory section, the OLS regression can be seen as a special case of the SL or SE regressions in which the spatial autoregressive coefficient ρ is constrained to zero.We analyze the data using the four econometric specifications: two aspatial regressions and two spatial regressions.Across the four regressions, we keep the same independent variables. 5irst, we present two aspatial models from Equations (9a) and (9b).A standard OLS regression (Equation ( 2)) and a panel analysis with county fixed effects are shown in Table 2.6 A most noticeable difference between the two models lies in the significance of the estimated climate parameters.In the OLS, climate parameters are highly significant.But in the panel model, individual climate parameters are mostly not significant.The reason is obvious and was already stated by the researchers that the OLS attributes, wrongly, the difference in land values arising from county fixed effects to that from climate.Once county fixed effects are controlled, climate is less important in the determination of land values across the space (Deschenes and Greenstone, 2007).Soil variables are well identified.Clay soils reduce land values.Flat plains command a higher land value than do steep hills.Soil Phaeozems and Ferralsols increase the value of the land.Other socio-economic variables are also significant.The availability of electricity improves the value of land.Land values are lower in most countries relative to Brazil, which is the omitted category.
From the estimated coefficient itself, it is hard to see what difference it makes to specify the panel model with county fixed effects instead of using a simple OLS model.In Table 3, we calculate the marginal effects of temperature change and rainfall change to compare across the model.A clear difference is that the marginal effects from the panel model are only half of those from the OLS model.Summer and winter temperature effects are reduced by half with the panel model.Winter precipitation effects are also reduced by half.Marginal summer precipitation effects change signs from negative to positive.
To further compare the two aspatial models with two spatial specifications introduced in the theory section, two spatial regression results are shown in Table 4.The Moran's I and Geary's C statistics are shown at the bottom of the spatial regressions to test whether district land values are spatially correlated.The element of the weight matrix ij w is defined to be only non-zero for those j in district i and 1 . The Moran's I in Equation ( 4) tests whether land values in each district are spatially correlated.The Moran's I statistic and Geary's C statistic do not reject the null hypothesis of no spatial correlation.However, the spatial autoregressive parameters in both regressions are very high-0.85for both cases.
As in the panel analysis, individual climate parameter estimates are mostly insignificant when spatial dependence is explicitly modeled.In the spatial models, however, most non climate variables have also become insignificant, although the F-values are very high for the two regressions.© Southern Regional Science Association 2010.Note: Moran's I for a provincial weights matrix=0.33,ρ for a provincial weights matrix=0.74.
In Table 5, we show the estimated marginal effects and elasticities from the two spatial models.In comparison to the panel model estimates, marginal climate parameter estimates have become even smaller.The marginal temperature elasticities are an order of magnitude smaller than those from the OLS.The precipitation elasticities are also much smaller.Comparing the SE with the SL, marginal effects are even smaller under the SE specification in the case of temperatures, but larger in the case of precipitations.At the bottom panel of the table, we also test whether the above spatial regression results are dependent crucially upon the spatial resolution of the spatial weights matrix.It shows the marginal effects from an alternative spatial weights matrix, which assumes provincial land values, instead of district land values, are correlated.Under this weights matrix, both Moran's I and spatial correlation statistics are smaller than those from the spatial models with the district spatial weight matrix.The marginal temperature effects are twice as large as those of the SL and SE regressions.However, they are twice as small as those of the OLS regression.They are closer to those marginal effects from the panel model with county fixed effects, but marginal winter temperature effect is still much smaller than those of the panel model.Marginal rainfall effects are also smaller than those from the aspatial models.
We ask whether non-irrigated rain-fed farms are more vulnerable to climate change than irrigated farms (Schlenker, Hanemann, and Fisher, 2005).We run a separate regression only on the rain-fed farms using the SL model.We then calculate marginal effects in Table 6.The results reveal that the marginal effects on rain-fed farms are not noticeably different.Except for the winter temperature effects, which are relatively negligible, the estimates are quite close to those of the SL model results from the whole sample.Unlike the results in the U.S., rain-fed farms and irrigated farms in South America are not substantially different in their vulnerabilities to climate change.It is likely that irrigation means much more sophisticated technologies and government interventions in the U.S. than those in the developing countries.Moreover, it is likely that this paper captures irrigation decisions better than the U.S. studies because we rely on individual The Review of Regional Studies, Vol. 38, No 2, 2008 © Southern Regional Science Association 2010.household-level irrigation decisions, whereas previous studies could only discern irrigated and non-irrigated counties.
Another concern frequently raised by researchers is whether farmers will be restrained by their capacities in adapting to climate change in the future (Kelly, Kolstad, and Mitchell, 2005).In our sample, it is likely that the Andean countries have only limited resources to cope with climate change as many of them are small household farms in the high mountains of the Andes (Rosenzweig and Hillel, 1998).To test this hypothesis, we run a separate SL regression by limiting the sample to Argentina and Brazil, excluding the farms in the Andes.Table 6 shows the marginal effects from the two-country regression using the SL model.The results confirm that the two countries are indeed more equipped with the necessary capacity to cope (i.e.marginal effects of summer and winter temperature are smaller than those from the whole sample that includes the Andean countries).However, the difference between the non-Andean sample and the Andean sample does not seem to be very large.The marginal effects from the precipitation variables are not different from those of the whole sample indicating that both the non-Andean and the Andean countries are equally vulnerable to rainfall variations.
CLIMATE SIMULATIONS
We simulate the impact of climate change assuming several climate scenarios by the middle of this century using both a spatial model and an aspatial model.We utilize the following three Atmospheric Oceanic General Circulation Model (AOGCM) scenarios for 2060 that cover a broad range of climate predictions consistent with recent estimates from the Intergovernmental Panel on Climate Change (IPCC, 2007): the Canadian Climate Center (CCC, Boer, Flato, and Ramsden, 2000); the Center for Climate System Research (CCSR, Emori et al., 1999); and the Parallel Climate Model (PCM, Washington et al., 2000).These scenarios are summarized, by study, for South America in Table 7.The CCC is a hot and dry scenario with a 2.7-degree Celsius increase in temperature and a 10-percent decrease in precipitation.The PCM is a mild and wet scenario with a 1.3-degree Celsius increase in temperature and a 10-percent increase in precipitation.The CCSR predictions fall between the CCC and PCM predictions.
The impact of these climate scenarios on land values is calculated in Table 8 as the difference between the predicted value in 2060 under each climate scenario and the current value using the estimated parameters in Tables 2 and 4. The aspatial model, OLS regression, predicts © Southern Regional Science Association 2010.
the loss of land values by 20 percent under the CCC scenario, by 25 percent under the CCSR scenario, and by 8 percent by the PCM scenario.The SL model predicts much smaller damage to agriculture.Farmers lose by 6 percent under the CCC, by 4 percent under the CCSR, and by 1 percent under the PCM scenario.This finding is consistent with a recent study in the U.S. in which authors found that climate change impacts are much muted when county fixed effects are controlled (Deschenes and Greenstone, 2007).However, it is inconsistent with another recent study which found large damage from climate change, more than a 50 percent loss in agricultural income by the middle of the century in U.S. agriculture even with spatial corrections (Schenkler, Hanemann, and Fisher, 2006), but with the county-level aggregate data.
The results imply that the models that do not account for spatial dependence may overstate the damage of climate change on agriculture.To the degree that land values are determined by their spatial locations, spatial specifications should be preferred to aspatial specifications such as the frequently cited county fixed effect model.Spatial effects are more evident in microlevel household data than in county-level aggregate data.The Review of Regional Studies, Vol. 38, No 2, 2008 © Southern Regional Science Association 2010.
CONCLUSIONS
Although econometric methods have been widely used over the past decades to measure the impact of climate change on agriculture, there is little agreement among the modelers on which specification should be preferred.This paper assesses the four econometric specifications (i.e.OLS, panel, and two spatial models) based on how they treat spatial characteristics of crosssectional data using South American agricultural household data.
This paper emphasizes the need to model spatial dependence explicitly when researchers rely on cross-sectional net revenues or land values.In contrast to the previous studies, this paper analyzes farm-level land value data.The analysis shows that there exists a high spatial correlation of the land values within each district.Accordingly, spatial models result in substantially different impact estimates from those of the aspatial models.
The relationships among the four econometric specifications were examined in terms of the freedom given to the spatial autoregressive parameter.In the spatial models (both the SL and SE models), the spatial parameter is allowed to change freely and is determined in the model.However, in the aspatial models, the parameter is fixed a priori.In the OLS specification, the parameter is fixed to zero.In the panel model, the parameter is fixed to one.
This paper shows that spatial models have more freedom in capturing complex spatial dependence than OLS and county fixed effect models.Empirical analyses indicate that the panel model predicts slightly smaller climate damage to agriculture than the OLS model as was shown in the literature (Deschenes and Greenstone, 2007), but the spatial models predict much smaller climate damage than does the panel model.
The paper also tests whether accounting for irrigation is crucially important (Schenkler, Hanemann, and Fisher, 2005).Unlike the U.S. study, the impact estimates do not differ substantially between irrigated and non-irrigated farms in our study.It is likely because irrigation is more sophisticated and crucial in the west coast U.S. counties than it is in South America.Alternatively, an investigation at the county level may not capture the incentive at the farm level to adopt irrigation or not.
Agricultural experts may be surprised by the results presented in this paper because crops are believed to be highly sensitive to heat and rainfall.This paper, however, does not argue that individual crops are not sensitive to climate.They are indeed highly sensitive (Magrin et al., 2007).What this paper does reveal is that farmers have a large array of crop and livestock varieties to choose from.So if one crop fails under a warmer climate, they are likely to find a better crop or livestock type that can survive under such conditions.Across South America, farmers own a vast variety of assets that have different climate sensitivities such as cereals (wheat, maize, barley, rice, and oats); oil seeds (soybean, peanuts, and sunflower); vegetables/tubercles (potatoes and cassava); a variety of perennial grasses; specialty crops such as cotton, tobacco, tea, coffee, cacao, sugarcane, sugar beets; and tree/shrub crops such as fruits, oil palm, and others.On top of this, farmers own beef cattle across the continent as well as dairy, chickens, pigs, and sheep (Seo and Mendelsohn, 2008a).
On a more cautious note, this paper does not account for the effects of carbon fertilization on agriculture, which is generally believed to be beneficial to crops in the near term (Reilly et al., 1996) but with large uncertainty (Lobell and Field, 2008).Secondly, it also assumes fixed prices, no transition costs, and no technological advance (Cline, 1996: Kelly
TABLE 1 : Descriptive Statistics of the Sample
© Southern Regional Science Association 2010.
TABLE 2 : Aspatial Econometric Models of Land Values (USD/ha)
Note: T refers to temperature and P refers to precipitation.202TheReview of RegionalStudies, Vol.38,No 2, 2008© Southern Regional Science Association 2010.
TABLE 4 : Spatial Models of Land Values (USD/ha)
© Southern Regional Science Association 2010.
TABLE 6 : Sensitivity Analysis from Spatial Models of Land Values (USD/ha)
: ρ for the rain-fed farms=0.84,ρ for the limited sample=0.88. Note
TABLE 8 : Land Value Changes by Each Climate Scenario by 2060 (USD/ha)
Note: * statistically significant with a 5% level of confidence. | 6,109.6 | 2008-06-03T00:00:00.000 | [
"Agricultural and Food Sciences",
"Economics",
"Environmental Science"
] |