Add files using upload-large-folder tool

preprint/preprint__27206ab5fd911ca9f9434dc07bf89032bd42e49f4bcd5d5e91e58a0a9bda02f7/preprint__27206ab5fd911ca9f9434dc07bf89032bd42e49f4bcd5d5e91e58a0a9bda02f7.mmd

@@ -0,0 +1,517 @@
+
+# An individualized protein-based prognostic model to stratify pediatric patients with papillary thyroid carcinoma  
+
+Zhihong Wang the First Hospital of China Medical University  
+
+He Wang Westlake University  
+
+Yan Zhou Westlake University  
+
+Lu Li Westlake University  
+
+Mengge Lyu Westlake University  
+
+Chunlong Wu Westlake Omics (Hangzhou) Biotechnology Co., Ltd.  
+
+Tianen He Westlake University  
+
+Lingling Tan Westlake Omics (Hangzhou) Biotechnology Co., Ltd.  
+
+Yi Zhu Westlake University https://orcid.org/0000- 0003- 0429- 0802  
+
+Tiannan Guo Westlake University https://orcid.org/0000- 0003- 3869- 7651  
+
+Hong- Kun Wu Zhejiang University  
+
+Hao Zhang The First Hospital of China Medical University  
+
+Yaoting Sun  
+
+sunyaoting@westlake.edu.cn  
+
+Westlake University https://orcid.org/0000- 0001- 7613- 648X
+
+<--- Page Split --->
+
+Keywords: Pediatric papillary thyroid carcinoma, Proteomics, Mass spectrometry, Machine learning, Random survival forest, Prognosis  
+
+Posted Date: June 9th, 2023  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 3015403/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: Yes there is potential Competing Interest. T.G. and Y.Zhu are shareholders of Westlake Omics Inc. C.W. and L.T. are employees of Westlake Omics Inc. Z.W., Y.S., H.Wang, H.Z. and T.G. have applied for a patent on this project. The other authors declare no competing interests in this paper.  
+
+Version of Record: A version of this preprint was published at Nature Communications on April 26th, 2024. See the published version at https://doi.org/10.1038/s41467-024-47926-w.
+
+<--- Page Split --->
+
+## 1 An individualized protein-based prognostic model to stratify pediatric  
+
+## 2 patients with papillary thyroid carcinoma  
+
+3 Zhihong Wang \(^{1\#}\) , He Wang \(^{2,3,4\#}\) , Yan Zhou \(^{2,3,4\#}\) , Lu Li \(^{2,3,4,5}\) , Mengge Lyu \(^{2,3,4}\) , Chunlong Wu \(^{6}\) , Tianen He \(^{2,3,4}\) , Lingling Tan \(^{6}\) , Yi Zhu \(^{2,3,4}\) , Tiannan Guo \(^{2,3,4}\) , Hongkun Wu \(^{7,8*}\) , Hao Zhang \(^{1*}\) , Yaoting Sun \(^{2,3,4*}\)  
+
+6 1Department of Thyroid Surgery, the First Hospital of China Medical University, 155 Nanjing 7 Bei Street, Shenyang 110001, China; 8 2iMarker lab, Westlake Laboratory of Life Sciences and Biomedicine, Key Laboratory of 9 Structural Biology of Zhejiang Province, School of Life Sciences, Westlake University, 18 10 Shilongshan Road, Hangzhou 310024, China; 11 3Institute of Basic Medical Sciences, Westlake Institute for Advanced Study, No. 18 Shilongshan 12 Road, Hangzhou 310024, China; 13 4Research Center for Industries of the Future, Westlake University, No. 600 Dunyu Road, 14 Hangzhou 310030, China; 15 5College of Pharmaceutical Sciences, Zhejiang University, No. 866 Yuhangtang Road, 16 Hangzhou 310058, China; 17 6Westlake Omics (Hangzhou) Biotechnology Co., Ltd., No. 1 Yunmeng Road, Hangzhou 18 310024, China; 19 7Department of Hepatobiliary and Pancreatic Surgery, the First Affiliated Hospital, Zhejiang 20 University School of Medicine, No. 79 Qingchun Road, Hangzhou 310003, China; 21 8Zhejiang Provincial Key Laboratory of Pancreatic Disease, No. 79 Qingchun Road, Hangzhou 22 310003, China.  
+
+24 #Co- first authors; 25 \\*Corresponding author(s): Yaoting Sun (sunyaoting@westlake.edu.cn); Hao Zhang (haozhang@cmu.edu.cn); Hongkun Wu (wuhongkun@zju.edu.cn).
+
+<--- Page Split --->
+
+## 28 Abstract (149 words)  
+
+28 Abstract (149 words)Pediatric papillary thyroid carcinomas (PPTCs) are with high inter- tumor heterogeneity and currently lack widely adopted recurrence risk stratification criteria. Hence, we propose a machine learning- based objective method to individually predict their recurrence risk. We retrospectively collected and evaluated the clinical factors and proteomes of 83 pediatric benign (PB), 85 pediatric malignant (PM) and 66 adult malignant (AM) nodules, and quantified 10,426 proteins by mass spectrometry. We found 243 and 121 significantly dysregulated proteins from PM vs. PB and PM vs. AM, respectively. Function and pathway analyses showed the enhanced activation of the inflammatory and immune system in PM patients compared with the others. Nineteen proteins were selected to predict recurrence using a machine learning model with an accuracy of 88.24%. Our study generated a protein- based personalized prognostic prediction model that can stratify PPTC patients into high- or low- recurrence risk groups, providing a reference for clinical decision- making and individualized treatment.  
+
+## 41 KEYWORDS  
+
+42 Pediatric papillary thyroid carcinoma; Proteomics; Mass spectrometry; Machine learning; Random survival forest; Prognosis.
+
+<--- Page Split --->
+
+Papillary thyroid carcinoma (PTC) is one of the most common endocrine malignant tumors in children and adolescents, with an incidence rate increasing by \(4.4\%\) yearly<sup>1</sup>. About \(1.8\%\) of thyroid cancers occur in children and adolescents, and PTC accounts for more than \(90\%\) of the cases<sup>2</sup>. Compared with adult PTC, pediatric PTC (PPTC) tends to have a larger tumor size, more lymph node metastases, a greater extrathyroidal extension rate, a higher distant metastasis rate, and a higher recurrence rate, while the overall mortality rate is lower. The guidelines for pediatric differentiated thyroid cancer have gaps regarding individualized diagnoses, treatments, and prognosis evaluation strategies up to the time of writing<sup>2</sup>. Specifically, unlike adults, pediatric patients are not age- stratified and do not receive individualized treatments: a one- size- fits- all treatment strategy is adopted for all of them<sup>3</sup>. Although most PPTCs have a favorable prognosis, recurrence seriously affects patients' disease- free survival and quality of life. Because the risk factors of PPTC recurrence are not clearly identified, there is currently a lack of effective methods for evaluating the prognosis of PPTC patients and classifying them into high or low recurrence risk groups. Therefore, patients with a low recurrence risk may undergo aggressive surgical resections, which unnecessarily increases their risk of complications. On the other hand, patients with a high recurrence risk may receive insufficient preoperative evaluations and postoperative monitoring, resulting in worse prognosis.  
+
+To date, the studies on the molecular mechanism of PPTC have been mostly limited to the genetic level<sup>4-7</sup>. They mainly focused on analyzing the PPTC etiology and providing a benign versus malignant diagnosis but did not produce tools for a personalized prognosis evaluation. Compared with adult PTC, PPTC is characterized by a higher prevalence of gene rearrangements
+
+<--- Page Split --->
+
+and a lower frequency of point mutations in the proto- oncogenes implicated in PTC.  
+
+Specifically, BRAF mutations are rarer, while RET/PTC rearrangements and gene fusions are more common in pediatric than in adult PTC<sup>8- 10</sup>. Consequently, these differences may affect the efficacy of gene- based diagnosis and prognosis evaluations of pediatric thyroid cancer.  
+
+Compared to genes, proteins could provide a more valuable contribution to the prognosis evaluation of diseases because they are the final products of gene expression<sup>11</sup>. However, the proteomic changes caused by PPTC remain unknown. Our previous study showed the potential of a machine- learning assisted proteomic analysis to discriminate between benign and malignant thyroid nodules<sup>12,13</sup>. Additionally, as we used trace samples from formalin- fixed paraffin- embedded (FFPE) tissues<sup>14</sup>, we showed the feasibility of such a study using preoperative fine- needle aspiration (FNA) samples which contain only thousands of cells and are hard to be analyzed by ordinary proteomic technology<sup>12</sup>.  
+
+In this study, we profiled the proteomic characteristics of pediatric thyroid nodules (malignant and benign). We aimed to find an efficient way to stratify pediatric patients with malignant tumors into high or low recurrence risk groups.  
+
+## RESULTS  
+
+## Clinical characteristics of our study population  
+
+The overall study design was demonstrated in Figure 1A. We enrolled 85 PPTC patients (PM), and collected their clinicopathological features (Figure 1B and Supplementary Table 1). This
+
+<--- Page Split --->
+
+group included 23 males and 62 females, with a male- to- female ratio of 1:2.7 and an average age of \(15.7 \pm 2.4\) years (ranging from 8.0 to 18.0 years). All 85 patients were admitted to the hospital with a mass in the neck, and their average tumor size was \(2.4 \pm 1.3\) cm (ranging from 0.3 to 7.0 cm). Besides the neck mass, one patient lamented hoarseness and another neck pain. Additionally, preoperative pulmonary computed tomography (CT) showed that one patient had multiple metastases in the lung. All patients were surgically treated; 47 (55.29%) underwent lobectomy, and 38 (44.71%) had a total thyroidectomy. Prophylactic central neck lymph node dissections were performed in all patients, while modified lateral neck dissections were performed in 43 patients. We recorded 16 cases (18.82%) with multifocal disease, 69 (81.18%) with lymph node metastases and 43 cases (43/85, 50.59%) of lateral cervical lymph node metastases.  
+
+The median follow- up time was 71 months (interquartile range 48- 113), during which no death was reported. Lung metastasis was discovered in one patient before the operation, and no change was reported after the radioactive iodine (RAI) therapy. Postoperative recurrence occurred in 12 cases (average age of 14): ten ipsilateral cervical lymph node metastases and one contralateral cervical lymph node metastases. One case developed postoperative lung metastases. All the cases of lymph node metastases were reoperated. During the follow- up evaluations, we found that the lesions of the patients with lung metastases had shrunk after the RAI therapy, and no growth or mental retardation was detected in any patient. Finally, no hematological or other secondary solid primary tumors were found during the postoperative follow- up.  
+
+## Three clinical features are the risk factors of PPTC recurrence
+
+<--- Page Split --->
+
+To identify the clinical recurrence risk factors in our study cohort, we built a univariate Cox proportional hazard (CoxPH) model for each of the ten clinical features collected for the PM patients (N=85). Age ( \(P = 0.0214\) , hazard ratio (HR)=0.8002, 95% confidence interval (CI): 0.6618- 0.9676), total lymph node metastasis number (TLNN, \(P = 0.0223\) , HR=1.076, 95%CI: 1.011- 1.146), and lateral lymph node metastasis number (LLNN, \(P = 0.0111\) , HR=1.101, 95% CI: 1.022- 1.185) had \(P\) values smaller than 0.05. Next, we split the PM patients into two groups (< or \(\geq\) the median) based on each significant factor. Further analyses showed significant differences between the Kaplan- Meier survival curves of the two groups for these three features (Figure 2A). Moreover, when treated as a categorical variable (0 representing ages below the median (16- year- old), 1 otherwise), age was more significantly associated with recurrence ( \(P = 0.0302\) , HR=0.2645, 95% CI: 0.07944- 0.8804). Our results showed that age, TLNN, and LLNN may be risk factors for recurrence in pediatric patients.  
+
+To determine the form of the age variable, the ten clinical features were next used as the inputs of multivariate CoxPH models. In particular, we input age as either a continuous integer or a categorical variable (as described in the previous paragraph). Our forest plot showed the HRs for ten clinical features, indicating the positive or negative influence of each feature on the PPTC recurrence. Only categorical age was highly related to PPTC recurrence. The global \(P\) value (logrank), the Akaike information criterion (AIC), and the Concordance Index (C- Index) outperformed when age was used as a categorical variable (Figure 2B- C). Therefore, age was determined as a categorical variable for the downstream analyses.
+
+<--- Page Split --->
+
+## More than 10,000 protein qualification with high quality  
+
+We collected 234 thyroid tissues from three groups: 85 PM, 83 pediatric benign (PB), and 66 adult malignant (AM) nodules. We randomly selected two samples and conducted one more replicate in each group. The resulting 240 tissues were randomized into 16 batches with 15 tissue samples each, and one pooled sample was used as a linker for the batches. Among them, we quantified 10,426 proteins.  
+
+To reduce the statistical bias, we removed 1272 ( \(\sim 12.2\%\) ) proteins with a missing value (NA) rate above \(85\%\) . This resulted in a final dataset including 9154 proteins. Quality control analysis showed the coefficient of variations (CVs) of the proteins across the 16 pooled samples were mainly between 0.0 and 0.2, with a median of 0.0493 (Figure S1A); the CVs of the proteins across each pair of replicates were mostly less than 0.2, with medians of 0.0662, 0.0947, 0.1238, 0.0890, 0.0645 and 0.1123 (Figure S1B). These results indicate the high stability of our instrument and the reliability of our data (Figure S1C- D).  
+
+To reduce the influence of NAs and batch effects, we imputed the NAs and then used ComBat to adjust for batch effects, and, by visualizing the resulting data using two- dimensional Uniform Manifold Approximation and Projection (UMAP), no noticeable batch effects could be detected (Figure S1E). After these preprocessing steps, there were 240 samples (87 PM, 85 PB and 68 AM) and 9154 proteins left (Figure S1F).  
+
+## Proteomic differences among pediatric malignant, pediatric benign and adult malignant thyroid nodules  
+
+To further explore the differences between PM and PB/AM, we determined the dysregulated proteins and generated two volcano plots showing 243 (PM vs. PB) and 121 (PM vs. AM)
+
+<--- Page Split --->
+
+differentially expressed proteins (DEPs) with fold change \((\mathrm{FC}) > 1.5\) and adjusted \(P< 0.05\) . By lowering our FC threshold to 1.2, the number of DEPs increased to 1548 (PM vs. PB) and 1629 (PM vs. AM) (Figure 3A- B). The DEPs with \(\mathrm{FC} > 1.5\) from the two pairwise comparisons were distributed in the scatter plot, and 25 proteins were co- up/downregulated in PM vs. AM and PM vs. PB (Figure 3C). Furthermore, the expression of 37 selected proteins showed in the heatmap, which was from the top- five up- and down- regulated proteins in the two pair- wise comparisons and 25 overlapped DEPs with duplicates removed (Figure 3D). According to the enrichment analysis of annotated keywords performed using STRING database, the most upregulated proteins in PM, compared to the other two groups, were involved in MHC- II and immunity. These results show that PPTC has a unique protein expression that differs from pediatric benign nodules and adult PTC.  
+
+Next, the functions and pathways enriched in the PM and the PB groups were almost all related to immune system regulation: mainly functions pertaining of T cells and natural killer (NK) cells (Figure 3E and 3F). Then the comparison of the PM with the AM group further showed pediatric thyroid cancer was associated with the regulation of inflammatory or immune- related pathways (Figure S2). These results suggest the development of PPTC is related to altered immune system functions.  
+
+## Immune infiltration and expression level of immune checkpoints in pediatric thyroid nodules  
+
+Since multiple dysregulated immune- related pathways and biological processes were enriched, we further conducted an analysis of immune infiltration in the pediatric samples using 'in- silico flow cytometry' CIBERSORTx<sup>15</sup>. Seven types of immune cells were imputed, and their relative
+
+<--- Page Split --->
+
+proportions are shown in Figure 4A. The fractions of \(\mathrm{CD8 + T}\) cells \((P = 3.7 \times 10^{- 12})\) , macrophages \((P = 0.031)\) , dendritic cells \((P = 1.4 \times 10^{- 5})\) and Treg cells \((P = 0.007)\) vary significantly. \(\mathrm{CD8 + T}\) cells and macrophages are increased in PM samples, while dendritic cells and Treg cells are reduced in PM samples. To validate immune infiltration results from in- silico analysis, we processed immunofluorescent staining for \(\mathrm{CD4 + }\) and \(\mathrm{CD8 + T}\) cells which marked \(\mathrm{CD3 + / CD4 + }\) and \(\mathrm{CD3 + / CD8 + }\) , respectively. Representative staining images of enriched \(\mathrm{CD8 + T}\) cells and decreasing \(\mathrm{CD4 + T}\) cells in the PMs are shown in Figure 4B.  
+
+To further explore tumor immune microenvironment, we compared the abundances of immune checkpoint proteins found PB vs. no- recurrence PM (PM- NR) and PM- NR vs. recurrence PM (PM- R). Among the 31 immune checkpoints quantified in our proteome data, poliovirus receptor (PVR) and interleukin 10 receptor B (IL10RB) had significantly lower levels in the most aggressive group PM- R (Figure 4C). No immune checkpoint proteins were found upregulated with increasing malignancy within PB, PM- NR and PM- R groups.  
+
+## Development of PPTC prognostic prediction models and individualized  
+
+## prognostic stratification  
+
+To predict the PTC recurrence risk of patients from the PM group, we developed five models based on two algorithms (Cox proportional hazard model and random survival forest) and two types of features (clinical features and proteins). Specifically, we developed the following models: two Cox proportional hazard models based on clinical features (CliCox) or protein features (ProtCox); three random survival forests based on clinical features (CliRsf), protein features (ProtRsf), or clinical and protein features (CliProtRsf). The final hyperparameter
+
+<--- Page Split --->
+
+settings of the five models are summarized in Supplementary Table 2. The ProtRsf model was the best- performing one as it achieved the highest C- index values: \(99.62\%\) , \(96.86\%\) , and \(84.95\%\)  
+
+on the training, the cross- validation, and the independent test sets, respectively (Figure 5A).  
+
+Notably, the combination of features used by CliProtRsf only contained 21 proteins without any  
+
+clinical features, which means the clinical features did not contribute to the model's prediction  
+
+significantly when protein features exist. The clinical features even interfered with the protein  
+
+features; thus, more proteins were needed to compensate for this effect. However, even with  
+
+more protein features selected for the model, CliProtRsf did not outperform ProtRsf (containing  
+
+19 proteins) in C- index. Therefore, we chose the ProtRsf model for our downstream analyses.  
+
+Then for each patient, we predicted his/her individualized survival curve firstly and deduced the  
+
+Crank risk score. Then, as shown in Figure 5B, using the training set, we determined the risk  
+
+stratification threshold according to the risk scores of the recurrent and the non- recurrent  
+
+patients. Therefore, the PM patients were classified as high or low- risk according to this  
+
+threshold. The Kaplan- Meier curves of the high- and low- risk patients differed significantly in  
+
+the training and independent test sets, indicating the strong generalization capability of our  
+
+model (Figure 5C).  
+
+## Analysis of 19 feature proteins  
+
+The random survival forest algorithm selected 19 proteins as features for the ProtRsf model; the  
+
+importance of these proteins is shown in Figure 5D. Of these 19 proteins, five have already been  
+
+reported in thyroid cancer studies, including galectin- 3 (LGALS3) \(^{16}\) , chromogranin- A  
+
+(CHGA) \(^{17}\) , collagen alpha- 3(VI) chain (COL6A3) \(^{18}\) , collagen alpha- 1(XXIII) chain  
+
+(COL23A1) \(^{19}\) , and integrin alpha- 4 (ITGA4) \(^{20}\) . Furthermore, myocilin (MYOC) has been linked
+
+<--- Page Split --->
+
+to the thyroid's function<sup>21</sup> (Supplementary Table 3). The remaining 13 proteins have not yet been associated with thyroid disease which are novel findings in this study.  
+
+Our network analysis showed that 13 of the 19 protein features were directly or indirectly connected. In particular, LGALS3, the hub protein, may perform a significant role in pediatric thyroid carcinoma (Figure 5E). The protein abundance of LGALS3 in four groups (PB, PM low- risk, PM high- risk, and AM) is shown in Figure 5F with Wilcoxon \(P\) values. LGALS3 has the lowest expression in the PB group, with significant differences compared to the expression in the other groups. In contrast, its expression was the highest in the PM high- risk group. These results show that a high LGALS3 expression may be associated with higher recurrence risk.  
+
+To explore if the 19 proteins selected by the ProtRsf model were related to immune system, we calculated Pearson correlations of immune cell fractions and the 19 proteins in PM high- and low- risk groups, respectively (Supplementary Figure 3). ITGA4 and GAL3ST4 were found positively correlated to \(\mathrm{CD8 + }\) T cells in both groups. For the 31 immune checkpoint proteins quantified, only the abundance of IL10RB was found decreased with the predicted recurrence risk and highest in PB samples (Supplementary Figure 4).  
+
+## Overall and individualized performance of the 19-protein model  
+
+We next evaluated the efficacy of our ProtRsf model in stratifying PPTC patients into groups with a high or low risk of recurrence. The model could correctly predict the prognosis of 75 cases of our 85 PM patients with an accuracy of \(88.24\%\) (Figure 6A). However, ten patients were wrongly classified: two were false negatives and eight were false positives. The predicted prognostic survival curves of each misclassified patient are shown in Figure 6B- C.
+
+<--- Page Split --->
+
+Then, we carefully analyzed the ten wrong predictions. The two false negative events corresponded to patients that underwent a recurrence but were classified as the low- risk group by the model. However, their recurrences were detected after 104 and 116 months, which were much longer than the median follow- up time (71 months) (Figure 6B). For the false- positive patients, the follow- up times (14, 17, 17, 25, 30, 48, 64, and 67 months) were all shorter than the median follow- up time (71 months) (Figure 6C). These patients were only follow- up for a short period when we start this study, which means they may go through recurrence in the future.  
+
+## DISCUSSION  
+
+PPTC is the most common endocrine malignant tumor in pediatric patients, which exhibits different clinical characteristics from adult PTC. There is still no effective strategy for evaluating the recurrence risk of pediatric thyroid carcinoma. In our study, we collected 85 PM, 83 PB, and 66 AM thyroid nodule tissues from 234 patients. Using labeled quantitative proteomic technology, we measured 10,426 proteins, to our knowledge, which is the first large- scale proteomic study on pediatric thyroid cancer patients. It is also a valuable data with considerably deep quantifications (more than 10,000 proteins) in thyroid nodules comparing with previous studies which detected thousands of proteins \(^{12,22,23}\) . We next found that immune processes were upregulated in PM nodules. Finally, we generated a model capable of predicting the recurrence risk of PM patients.  
+
+From our clinical data, we found age and lymph node metastases were important prognostic indicators of PPTC which are matched with previous findings \(^{24 - 27}\) . In our PM group, the age of 16 was the cutoff for predicting recurrence- free survival (RFS), as nine of the 12 recurrent
+
+<--- Page Split --->
+
+patients were younger than 16. Furthermore, 69 cases (81.18%) from our PM group had total lymph node metastases. Additionally, we found that TLNN and LLNN correlate with RFS. However, unlike previous studies<sup>25</sup>, the lymph node metastasis rate of our study cohort did not suggest recurrence, which may be due to the different number of lymph node dissections. Although several factors were shown related to poor prognosis, however, we found that the risk factors derived from clinical indicators are only suggestive of clinical phenomena and are, thus, insufficient for formulating prognostic predictions and risk stratification by the model. It has become a trend for molecular detection to apply to tumor risk stratification according to the latest version of the World Health Organization published in 2022<sup>28</sup>. Many studies have suggested that gene expression and clinical features of PTC were different between children and adults which is related to different prognosis<sup>6,29</sup>. Whereas gene correlation analyses can explain, to a certain extent, the difference in clinicopathological features between pediatric and adult thyroid cancers, current genomics studies have a limited role in the risk stratification of PPTC. Therefore, we chose to use proteomics data as the base of our predictions since proteins are the biological activity effectors. Using proteomic data, our predicting model achieves higher performance in predicting recurrence risk. Even when we combined proteins with clinical features as candidates, the model did not select any clinical features. The panel of proteins evolved by the model is significantly more accurate for predicting PPTC prognoses than clinical features.  
+
+Among the 19 proteins, CHGA<sup>17</sup>, COL6A3<sup>18</sup>, COL23A1<sup>30</sup>, ITGA4<sup>20</sup> and LGALS3<sup>16</sup> been reported to be associated with thyroid cancer, and MYOC<sup>21</sup> related to thyroid function. Notably, LGALS3 is an important marker located in the core of the network (Figure 5E). Its inhibitor inhibits apoptosis resistance and the invasion of thyroid cancer cells through the AKT/β- catenin
+
+<--- Page Split --->
+
+pathway<sup>31</sup>. In agreement with these previous findings, in our study, the expression of LGALS3 in the PM high- risk group was significantly higher than in the PM low- risk group and the AM group. The high expression of LGALS3 might promote cancer invasion and impede the function of immune system to make the cell apoptosis, leading to cancer recurrence.  
+
+of immune system to make the cell apoptosis, leading to cancer recurrence.  
+
+Based on the 19- protein panel, our ProtRsf model achieved accuracy of \(88.24\%\) in stratifying  
+
+PPTC patients into groups with a high or low risk of recurrence (Figure 6A). Although the high  
+
+performance we got, ten patients were wrongly classified. We next have carefully investigated  
+
+the mispredicted samples (Figure 6B- C). The eight false- positive samples are patients who are  
+
+predicted to relapse but have not yet done so. These patients have a relatively short current  
+
+follow- up period and, in terms of survival curves, each of them has a low risk of recurrence as of  
+
+the current follow- up time, but their probability of recurrence at five years after surgery will  
+
+increase substantially as time continues to progress shown in the predicted prognostic survival  
+
+curves (Figure 6C). It is therefore difficult to be sure that the model is predicting them  
+
+incorrectly at this time and close follow- up is still needed for these patients to allow time to give  
+
+us the true answer. For the two false negatives, the recurrence intervals are both more than 100  
+
+months which is much longer than the median follow- up time (71 months), to some extent, it  
+
+also represents relatively inert biological behavior.  
+
+The tumor immune microenvironment also plays a key role in the development and progression  
+
+of thyroid cancer<sup>32</sup>. Most studies of the immune microenvironment of PTC have focused on  
+
+adults rather than children and adolescents. In our study, we showed for the first time that the  
+
+243 DEPs between PM and PB patients are closely related to immune dysregulation.  
+
+Additionally, a high level of PD- L1 is associated with poor prognoses, such as an increased risk  
+
+of thyroid cancer recurrence and lymph node metastases<sup>33- 37</sup>. The results imply that dysregulated
+
+<--- Page Split --->
+
+immune cell compositions and altered immune monitoring may play crucial roles in papillary thyroid carcinoma genesis in pediatric patients.  
+
+CD8+ T cells recognize tumor cells which express tumor antigens and attack by inducing cell death<sup>38</sup>. In adult PTC, CD8+ T cells were found to have a higher frequency than in benign samples<sup>39</sup>, and the infiltration of CD8+ T cells was related to increasing disease- free survival<sup>40</sup>. Our data showed higher levels of CD8+ T cells infiltration in PM than in PB, which is consistent with adult patients. CD4+ T cells were not found to be significantly different between PB and PM. Similarly, the functions of these cells in tumor prognosis were not found<sup>40,41</sup>.  
+
+The findings of this study have to be seen in the light of some limitations. This is a retrospective study in a single center; therefore, future studies should collect preoperative samples in more centers prospectively. Also, our results need to be validated with a larger cohort and longer follow- up time to evaluate our model's generalization. Despite these limitations, we have shown the feasibility and importance of using proteomics data for the stratification and prognostic prediction of PPTC patients.  
+
+Proteomics offers, among others, the advantages of high- throughput quantification and microsampling, the latter enabling clinical applications with preoperative FNA samples. With this method, we can make high- and low- risk stratification assessments before the operation, guide the resection scope during the operation, evaluate the prognosis after the operation, and formulate individualized follow- up strategies.  
+
+In conclusion, we generated a protein- based personalized prognostic prediction model that could stratify pediatric patients with papillary thyroid carcinoma, providing a reference for clinical decision- making and individualized treatment.
+
+<--- Page Split --->
+
+## METHODS  
+
+## Study population and samples collection  
+
+In this retrospective study, we evaluated pediatric patients ( \(\leq 18\) years) with thyroid nodules, including 85 PM and 83 PB thyroid nodules, who underwent surgery in the First Hospital of China Medical University between November 2007 and April 2021. The exclusion criteria were the following: (a) with a history of radiation exposure or family history, (b) with subtypes of highly invasive diseases, i.e., tall- cell variant, columnar, and poorly differentiated PTC, (c) loss of follow- up or incomplete clinical data, and (d) non- primary operation. We also included 66 AM patients with PTC to compare pediatric and adult thyroid cancer proteomic profiling. Lobectomy and ipsilateral central lymph node dissections were performed in unilateral PTC. Total thyroidectomy was performed in patients with ETE, such as the invasion of nerves, blood vessels, or trachea. Patients with bilateral PTC underwent total thyroidectomy and bilateral central lymph node dissections. Postoperative treatment included thyroid- stimulating hormone inhibition and RAI therapies. This study was approved by the Ethics Committee of the First Hospital of China Medical University.  
+
+After the surgery, the patients were followed up every 3- 6 months with cervical ultrasounds and thyroid functional examinations. The re- examination interval was then prolonged for patients with negative ultrasounds or CT, low serum thyroglobulin level, or no persistent disease. Disease remission was defined as two consecutive negative whole- body scans and ultrasounds, with thyroglobulin and anti- thyroglobulin antibodies in the ideal range. On the other hand, two types of recurrence were considered: structural recurrence was determined as a new disease in the
+
+<--- Page Split --->
+
+thyroid bed or lymph nodes proven by cytology or histopathology, and/or confirmed by ultrasounds or CT scans, or distant metastases detected by whole- body scan; biochemical recurrence was based on one of the following criteria: (1) thyroglobulin \(>1 \mathrm{ng / mL}\) ; (2) increasing thyroglobulin antibody levels, or (3) thyrotropin- stimulated thyroglobulin \(>10 \mathrm{ng / mL}\) .  
+
+## Recurrence risk factors among clinical features  
+
+To identify recurrence risk factors among the clinical features of the PTC pediatric patients, we conducted univariate and multivariate analyses using ten clinical features. In particular, we used the Cox proportional hazard (CoxPH) model and combined prognosis information: recurrence events, the time interval between surgery and recurrence, or between surgery and the last follow- up. The ten clinical characteristics were: age, gender, maximum nodule size, multifocality, ETE, total lymph node metastasis rate (TLNR), lateral lymph node metastasis rate (LLNR), total lymph node metastasis number (TLNN), lateral lymph node metastasis number (LLNN), and Hashimoto thyroiditis (HT).  
+
+We built a univariate CoxPH model for each clinical feature, identified the factors whose \(P\) values were less than 0.05, split the PM patients into two groups ( \(< \mathrm{or} \geq\) the median value) based on each significant factor, and compared the Kaplan- Meier survival curves of the two groups. Next, we transformed the age factor from a continuous non- negative integer to a categorical variable (0 representing ages below the median value (16- year- old), 1 otherwise) and performed the same analysis. Lastly, the ten clinical features were input into the multivariate CoxPH model two times: using the continuous non- negative integer age or the categorical age. We then compared the global \(P\) value (log- rank), Akaike information criterion (AIC), and Concordance
+
+<--- Page Split --->
+
+Index (C- Index) of the two cases to determine which data format was more suitable for the age variable.  
+
+## Sample collection  
+
+We collected 240 thyroid nodules (87 PM, 85 PB, 68 AM) FFPE slides (10 \(\mu \mathrm{m}\) thick) from 234 patients (85 PM, 83 PB, 66 AM). Two samples from each group were randomly selected as technical replicates. Each slide was stained with hematoxylin and eosin and reviewed by at least two experienced histopathologists. The tumor ratio of each slide was approximately more than \(80\%\) .  
+
+## Sample preparation  
+
+FFPE slides were prepared by pressure cycling technology (PCT) as in our previous study<sup>42,43</sup>. Briefly, the slides were dewaxed, rehydrated, and de- crosslinked using heptane, three different concentrations of ethanol (100%, 90%, and 75%), and 100 mM tris- base solution (pH=10), respectively. Next, the samples were lysed using PCT with a buffer containing 6 M urea, 2 M thiourea, 10 mM tris (2- carboxymethyl) phosphine, and 40 mM iodoacetamide. Then, the samples were digested using trypsin and lysC. Finally, the digested peptides were desalted by C18 (SOLAμ columns, Thermo Fisher Scientific, USA). The chemicals were bought from Sigma- Aldrich (USA), and the enzymes were obtained from Hualishi Scientific (Beijing, China).  
+
+Cleaned peptides were labeled using TMTpro 16- plex reagents (Thermo Fisher Scientific, USA). Each batch comprised 15 samples and one pooled sample, which were separated into 30 fractions within a 60 min gradient on Ultimate Dinex 3000 (Thermo Fisher Scientific, USA) equipped with a C18 column (300 Å, \(5 \mu \mathrm{m} \times 4.6 \mathrm{mm} \times 250 \mathrm{mm}\) , XBridge Peptide BEH, Waters, USA).
+
+<--- Page Split --->
+
+## Proteomics data acquisition  
+
+Each fraction was analyzed using liquid chromatography- mass spectrometry (nanoflow DIONEX UltiMate 3000 RSLCnano System and Orbitrap Exploris 480 with FAIMS Pro™, Thermo Fisher Scientific, USA). In each acquisition, peptides were separated using a 60 min gradient (from 3% to 28% buffer B (98% acetonitrile and 0.1% formic acid)) at a 300 nL/min flowrate on an analytical column (1.9 μm 100 Å C18- Aqua, 150 mm × 75 μm). Buffer A was composed of 2% ACN, 98% H₂O, and 0.1% formic acid. All reagents were mass spectrometry- grade. The mass- to- charge (m/z) range of the MS1 was 375- 1,800 Th with a resolution of 60,000 full widths at half maximum (FWHM); the MS2 resolution was 30,000 FWHM. The turbo- TMT and advanced peak determination were enabled.  
+
+## Proteomics data analysis  
+
+Proteomic raw files were searched using Proteome Discoverer (v2.4.1.15) against a FASTA file containing 20,368 entries (human Swiss- Prot database). Channel TMT- 126 was set as the reference for each batch. The quantified proteins were filtered for a 1% false discovery rate. Other settings were left to their default values.  
+
+## Proteomics data quality control and preprocessing  
+
+The data quality was assessed by evaluating the coefficient of variation (CV) across the pooled samples and the technical replicates. When calculating CVs, the missing values were omitted, and log2- transformed protein abundance was used.
+
+<--- Page Split --->
+
+The R package NAguideR was used for missing value imputation, and the imspecprob method (for robust sequential imputation) was used. Next, the batch effects correction of the resulting protein matrix was performed using Combat, an empirical Bayes framework from the R package sva<sup>44</sup>. For the matrices after imputation and correction, the non- positive values in the matrix were replaced by half the minimum value of the positive abundances of the corresponding protein. The differentially expressed proteins (DEPs) were identified with fold change (FC) values to be greater than 1.2 or 1.5 (for different purposes), with an adjusted Welch's \(t\) - test \(P < 0.05\) . Each pair of technical replicates were then combined to one sample by calculating the mean protein abundance.  
+
+## Tumor immune microenvironment analysis  
+
+CIBERSORTx<sup>15</sup> (https://cibersortx.stanford.edu/) was utilized to profile the proportions of 7 immune cell types in our proteomic data. The software required a feature matrix which contained the gene expression profiles of each cell type of interest. We used a custom signature matrix generated from published thyroid cancer single- cell RNA data<sup>45</sup>.  
+
+## The analysis of immune checkpoints  
+
+Thirty- one immune checkpoint proteins were quantified in our data. We conducted Student's \(t\) - test to compare the abundance of these immune checkpoint proteins between PB and PM using R (v4.1.1). Samples with extreme values defined by Tukey's fences were removed before plotting the boxplots.  
+
+## Multiplex immunohistochemistry staining
+
+<--- Page Split --->
+
+Multiplex immunohistochemistry (mIHC) staining was performed using methods and reagents following the TSA Opal mIHC protocols (Akoya Biosciences/PerkinElmer). Briefly, \(5 \mu \mathrm{m}\) thickness FFPE tumor sections were stained with DAPI and antibodies against the following markers: CD3 (cat# ab135372, dilution 1:500, Abcam), CD4 (cat# ab288724, dilution 1:1000, Abcam), and CD8 (cat# ab17147, dilution 1:500, Abcam). All markers were sequentially applied and stained using their respective fluorophores in the Opal 7 kit (catalogue #NEL797001KT; Akoya Biosciences/PerkinElmer). Stained slides were scanned using the multispectral microscope, Vectra v3.0.3 imaging system (Akoya Biosciences/PerkinElmer), under fluorescence and low magnification at \(10 \times 40\) . Following scanning, around four regions of interest (each region of interest (ROI) \(0.6522 \mathrm{mm}^2\) ) were selected per sample using the phenochart viewer v1.0.9 (Akoya Biosciences/PerkinElmer). ROIs were analyzed by the image analysis software, InForm v2.8.2 (Akoya Biosciences/PerkinElmer).  
+
+## Predicting PPTC recurrence risk using clinical or/and protein features  
+
+To build models for predicting PPTC recurrence risk, the PM samples were randomly divided into a training set (n=50, \(\sim 60\%\) ) and an independent test set (n=35, \(\sim 40\%\) ). The training set was used for building prognostic prediction models, including hyperparameter tuning, feature selection, and model training. The independent test set was used to evaluate our models' generalization capability.  
+
+We built five models using the R package mlr3. Specifically, we generated a Cox proportional hazard model based on clinical features (CliCox), a random survival forest based on clinical features (CliRsf), a Cox proportional hazard model based on protein features (ProtCox), a
+
+<--- Page Split --->
+
+random survival forest based on protein features (ProtRsf), and a random survival forest based on clinical and protein features (CliProtRsf). For each model, we tuned the hyperparameters using grid search strategy and 3- fold cross- validation, selected the features, and trained the model using the training set. Lastly, we compared the training, cross- validation, and test C- Indexes of the five models. The models based on the ten clinical features did not conduct the feature selection step due to the small number of clinical features, and for the ProtCox model, we used the Least Absolute Shrinkage and Selection Operator (LASSO) for selecting the protein features. As for ProtRsf and CliProtRsf, we made the feature selection as next described. Based on the models after tuning and using 1,548 DEPs (PB versus PM; FC > 1.2, adjusted \(P < 0.05\) ), which were accompanied by clinical features in the case of CliProtRsf, the models were trained for 100 times with different initial states. In each training, we ranked the features according to permutation importance and selected the 50 most important features. Finally, we recorded the selected numbers of each protein and chose the features selected no less than 50 times as the final feature set.  
+
+## Prognostic stratification of the PM patients  
+
+Using our previously developed ProtRsf model, we predicted the prognostic survival curve of each PM patient. Then, according to the prognostic curves, the expectations corresponding to these curves were calculated and used to compute the recurrence risk score, noted as Continuous risk ranking (Crank), which is proportional to the recurrence risk. Next, using the training set, we chose the stratification threshold using the Crank scores from the recurrence and the non- recurrence groups. Specifically, the threshold was calculated by averaging the mean Cranks of two groups. We then classified the PM samples as high or low- risk using this threshold. Finally,
+
+<--- Page Split --->
+
+we validated our stratification threshold using the independent test cohort, which allowed us to evaluate the generalization ability of the final model.  
+
+## Bioinformatics and statistical analyses  
+
+Statistical analysis was conducted using R (v4.1.1). The Uniform Manifold Approximation and Projection (UMAP) visualization was performed using the R package UMAP. The heatmap was generated using the R package pheatmap, with protein- level normalization and hierarchical clustering (Euclidean distance and complete option were used). The \(P\) values of the DEPs in the volcano plots were derived from a two- sided unpaired Welch's \(t\) - test and adjusted using the Benjamini- Hochberg method. Pathways and networks were analyzed using the Ingenuity Pathway Analysis (IPA) and visualized with Cytoscape (v3.8.2). Log- rank test was used for comparing Kaplan- Meier curves in two sample groups. For the tables of clinical characteristics, continuous variables were reported as mean \(\pm\) standard deviation (SD), and categorical variables as frequency and proportion. Student's \(t\) - test (for continuous variables), chi- squared test (for categorical variables with three or more categories) and Fisher's exact test (for categorical variables with two categories) were used for comparison.  
+
+The Pearson correlations between the fractions of immune cells and the abundance of the selected 19 proteomic features were calculated using the R package Hmisc.  
+
+## DATA AVAILABILITY  
+
+The mass spectrometry proteomic raw data have been deposited to the iProX database with the identifier IPX0006407000.
+
+<--- Page Split --->
+
+## AUTHOR INFORMATION  
+
+## Corresponding Author  
+
+Yaoting Sun https://orcid.org/0000- 0001- 7613- 648X Email: sunyaoting@westlake.edu.cn  
+
+Hao Zhang https://orcid.org/0000- 0002- 9938- 8433 Email: haozhang@cmu.edu.cn  
+
+Hongkun Wu https://orcid.org/0000- 0002- 9459- 6583 Email: wuhongkun@zju.edu.cn  
+
+## Author Contributions  
+
+T.G., H.Z., Z.W., and Y.S. designed the study. C.W. prepared samples for proteomics data. T.H. and L.T. conducted database searching. H.Wang conducted data analysis. Y.Zhou conducted tumor immune microenvironment and checkpoints analysis. H.Wu processed the multiplexed immunofluorescent staining and data interpretation. Y.Zhu, M.L. and L.L. helped improve the manuscripts.  
+
+## Conflict of Interest  
+
+T.G. and Y.Zhu are shareholders of Westlake Omics Inc. C.W. and L.T. are employees of Westlake Omics Inc. Z.W., Y.S., H.Wang, H.Z. and T.G. have applied for a patent on this project. The other authors declare no competing interests in this paper.  
+
+## ACKNOWLEDGMENT  
+
+We thank Shanjun Chen from Westlake Omics Inc. for his help with improving the text. We also thank for the assistance in data storage, computation, and peptide fractionation by the Westlake University Supercomputer Center and the Mass Spectrometry & Metabolomics Core Facility at
+
+<--- Page Split --->
+
+the Center for Biomedical Research Core Facilities of Westlake University. Figures 1A was created through Biorender.com.  
+
+This work was supported by the National Key R&D Program of China (No. 2021YFA1301602,  
+
+2021YFA1301601, 2020YFE0202200 to Tiannan Guo), China Postdoctoral Science Foundation  
+
+(2022M722841 to Yaoting Sun), and National Key R&D Program of China (No.  
+
+2022YFF0608403 to Yi Zhu). This work was further supported by the Science and Technology  
+
+Project of Shenyang City (21- 173- 9- 31) and Applied Basic Research Program of Liaoning  
+
+Province (2022020225- JH2/1013 to Hao Zhang).  
+
+## ABBREVIATIONS  
+
+PTC, Papillary thyroid carcinoma; PPTC, pediatric papillary thyroid carcinoma; PB, pediatric benign nodules; PM, pediatric malignant nodules; AM, adult malignant nodules; FFPE, formalin- fixed paraffin- embedded; FNA, fine- needle aspiration; CT, computed tomography; RAI, radioactive iodine; CoxPH, Cox proportional hazard; ETE, extrathyroidal extension; TLNR, total lymph node metastasis rate; LLNR, lateral lymph node metastasis rate; TLNN, total lymph node metastasis number; LLNN, lateral lymph node metastasis number; HT, Hashimoto thyroiditis; HR, hazard ratio; AIC, Akaike information criterion; C- index, Concordance Index; TMT, tandem mass tag; NA, missing value; UMAP, Uniform Manifold Approximation and Projection; DEP, differentially expressed proteins; FC, fold change; NK, natural killer; LGALS3, galectin- 3; CHGA, chromogranin- A; COL6A3, collagen alpha- 3(VI) chain; COL23A1, collagen alpha- 1(XXIII) chain; ITGA4, integrin alpha- 4; MYOC, myocilin; RFS, recurrence- free survival; SD, standard deviation.
+
+<--- Page Split --->
+
+564 1. Bernier MO, Withrow DR, Berrington de Gonzalez A, et al. Trends in pediatric thyroid 565 cancer incidence in the United States, 1998- 2013. Cancer. Jul 15 2019;125(14):2497- 2505. 566 doi:10.1002/cncr.32125 567 2. Francis GL, Waguespack SG, Bauer AJ, et al. Management Guidelines for Children with 568 Thyroid Nodules and Differentiated Thyroid Cancer. Thyroid. Jul 2015;25(7):716- 59. 569 doi:10.1089/thy.2014.0460 570 3. Cherella CE, Wassner AJ. Pediatric thyroid cancer: Recent developments. Best Pract Res 571 Clin Endocrinol Metab. Jan 2023;37(1):101715. doi:10.1016/j.beem.2022.101715 572 4. Gallant JN, Chen SC, Ortega CA, et al. Evaluation of the Molecular Landscape of Pediatric 573 Thyroid Nodules and Use of a Multigene Genomic Classifier in Children. JAMA Oncol. Sep 1 574 2022;8(9):1323- 1327. doi:10.1001/jamaoncol.2022.1655 575 5. Hess J, Newbern DK, Beebe KL, Walsh AM, Schafernak KT. High prevalence of gene 576 fusions and copy number alterations in pediatric radiation therapy- induced papillary and follicular 577 thyroid carcinomas. Thyroid. Dec 16 2021;doi:10.1089/thy.2021.0217 578 6. Pekova B, Sykorova V, Dvorakova S, et al. RET, NTRK, ALK, BRAF, and MET Fusions 579 in a Large Cohort of Pediatric Papillary Thyroid Carcinomas. Thyroid. Dec 2020;30(12):1771- 580 1780. doi:10.1089/thy.2019.0802 581 7. Kotanidou EP, Giza S, Tsinopoulou VR, et al. The Prognostic Significance of BRAF Gene 582 Analysis in Children and Adolescents with Papillary Thyroid Carcinoma: A Systematic Review 583 and Meta- Analysis. Diagnostics (Basel). Mar 21 2023;13(6)doi:10.3390/diagnostics13061187 584 8. Nikiforov YE, Rowland JM, Bove KE, Monforte- Munoz H, Fagin JA. Distinct pattern of 585 ret oncogene rearrangements in morphological variants of radiation- induced and sporadic thyroid 586 papillary carcinomas in children. Cancer Res. May 1 1997;57(9):1690- 4. 587 9. Penko K, Livezey J, Fenton C, et al. BRAF mutations are uncommon in papillary thyroid 588 cancer of young patients. Thyroid. Apr 2005;15(4):320- 5. doi:10.1089/thy.2005.15.320 589 10. Fenton CL, Lukes Y, Nicholson D, Dinauer CA, Francis GL, Tuttle RM. The ret/PTC 590 mutations are common in sporadic papillary thyroid carcinoma of children and young adults. J 591 Clin Endocrinol Metab. Mar 2000;85(3):1170- 5. doi:10.1210/jcem.85.3.6472 592 11. Zhu Y, Aebersold R, Mann M, Guo T. SnapShot: Clinical proteomics. Cell. Sep 2 593 2021;184(18):4840- 4840 e1. doi:10.1016/j.cell.2021.08.015 594 12. Sun Y, Selvarajan S, Zang Z, et al. Artificial intelligence defines protein- based 595 classification of thyroid nodules. Cell Discov. Sep 6 2022;8(1):85. doi:10.1038/s41421- 022- 596 00442- x 597 13. Sun Y, Li L, Zhou Y, et al. Stratification of follicular thyroid tumours using data- 598 independent acquisition proteomics and a comprehensive thyroid tissue spectral library. Mol Oncol. 599 Apr 2022;16(8):1611- 1624. doi:10.1002/1878- 0261.13198 600 14. Cai X, Xue Z, Wu C, et al. High- throughput proteomic sample preparation using pressure 601 cycling technology. Nat Protoc. Oct 2022;17(10):2307- 2325. doi:10.1038/s41596- 022- 00727- 1 602 15. Newman AM, Steen CB, Liu CL, et al. Determining cell type abundance and expression 603 from bulk tissues with digital cytometry. Nat Biotechnol. Jul 2019;37(7):773- 782. 604 doi:10.1038/s41587- 019- 0114- 2 605 16. Bartolazzi A, Sciacchitano S, D'Alessandria C. Galectin- 3: The Impact on the Clinical 606 Management of Patients with Thyroid Nodules and Future Perspectives. Int J Mol Sci. Feb 2 607 2018;19(2)doi:10.3390/ijms19020445
+
+<--- Page Split --->
+
+17. Portel-Gomes GM, Grimelius L, Johansson H, Wilander E, Stridsberg M. Chromogranin A in human neuroendocrine tumors: an immunohistochemical study with region-specific antibodies. Am J Surg Pathol. Oct 2001;25(10):1261-7. doi:10.1097/00000478-200110000-00006
+18. Shibata M, Inaishi T, Ichikawa T, et al. Identifying the tumor-progressive gene expression profile in high-risk papillary thyroid cancer. Surg Today. Oct 2021;51(10):1703-1712. doi:10.1007/s00595-021-02262-0
+19. Pan Y, Wu L, He S, Wu J, Wang T, Zang H. Identification of hub genes in thyroid carcinoma to predict prognosis by integrated bioinformatics analysis. Bioengineered. Dec 2021;12(1):2928-2940. doi:10.1080/21655979.2021.1940615
+20. Song J, Wu S, Xia X, Wang Y, Fan Y, Yang Z. Cell adhesion-related gene somatic mutations are enriched in aggressive papillary thyroid microcarcinomas. J Transl Med. Oct 1 2018;16(1):269. doi:10.1186/s12967-018-1642-0
+21. Taguchi M, Kanno H, Kubota R, Miwa S, Shishiba Y, Ozawa Y. Molecular cloning and expression profile of rat myocilin. Mol Genet Metab. May 2000;70(1):75-80. doi:10.1006/mgme.2000.2986
+22. Lin P, Yao Z, Sun Y, et al. Deciphering novel biomarkers of lymph node metastasis of thyroid papillary microcarcinoma using proteomic analysis of ultrasound-guided fine-needle aspiration biopsy samples. J Proteomics. Jul 30 2019;204:103414. doi:10.1016/j.jprot.2019.103414
+23. Shi X, Sun Y, Shen C, et al. Integrated proteogenomic characterization of medullary thyroid carcinoma. Cell Discov. Nov 8 2022;8(1):120. doi:10.1038/s41421-022-00479-y
+24. Jeon MJ, Kim YN, Sung TY, et al. Practical Initial Risk Stratification Based on Lymph Node Metastases in Pediatric and Adolescent Differentiated Thyroid Cancer. Thyroid. Feb 2018;28(2):193-200. doi:10.1089/thy.2017.0214
+25. Xu Y, Wang Y, Zhang X, Huang R, Tian R, Liu B. Prognostic value of lymph node ratio in children and adolescents with papillary thyroid cancer. Clin Endocrinol (Oxf). Oct 2021;95(4):649-656. doi:10.1111/cen.14491
+26. Sugino K, Nagahama M, Kitagawa W, et al. Cutoff Age Between Pediatric and Adult Thyroid Differentiated Cancer: Is 18 Years Old Appropriate? Thyroid. Feb 2022;32(2):145-152. doi:10.1089/thy.2021.0255
+27. Zeng X, Wang Z, Gui Z, et al. High Incidence of Distant Metastasis Is Associated With Histopathological Subtype of Pediatric Papillary Thyroid Cancer - a Retrospective Analysis Based on SEER. Front Endocrinol (Lausanne). 2021;12:760901. doi:10.3389/fendo.2021.760901
+28. Baloch ZW, Asa SL, Barletta JA, et al. Overview of the 2022 WHO Classification of Thyroid Neoplasms. Endocr Pathol. Mar 2022;33(1):27-63. doi:10.1007/s12022-022-09707-3
+29. Mostoufi-Moab S, Labourier E, Sullivan L, et al. Molecular Testing for Oncogenic Gene Alterations in Pediatric Thyroid Lesions. Thyroid. Jan 2018;28(1):60-67. doi:10.1089/thy.2017.0059
+30. Dai J, Yu X, Han Y, et al. TMT-labeling Proteomics of Papillary Thyroid Carcinoma Reveal Invasive Biomarkers. J Cancer. 2020;11(20):6122-6132. doi:10.7150/jca.47290
+31. Lee JJ, Hsu YC, Li YS, Cheng SP. Galectin-3 Inhibitors Suppress Anoikis Resistance and Invasive Capacity in Thyroid Cancer Cells. Int J Endocrinol. 2021;2021:5583491. doi:10.1155/2021/5583491
+32. Li PH, Kong XY, He YZ, et al. Recent developments in application of single-cell RNA sequencing in the tumour immune microenvironment and cancer therapy. Mil Med Res. Sep 26 2022;9(1):52. doi:10.1186/s40779-022-00414-y
+33. Ferrari SM, Fallahi P, Galdiéro MR, et al. Immune and Inflammatory Cells in Thyroid Cancer Microenvironment. Int J Mol Sci. Sep 7 2019;20(18)doi:10.3390/ijms20184413
+
+<--- Page Split --->
+
+34. Liotti F, Kumar N, Prevete N, et al. PD-1 blockade delays tumor growth by inhibiting an intrinsic SHP2/Ras/MAPK signalling in thyroid cancer cells. J Exp Clin Cancer Res. Jan 7 2021;40(1):22. doi:10.1186/s13046-020-01818-1
+35. Liu WL, Zhang YQ, Li LT, et al. Application of molecular imaging in immune checkpoints therapy: From response assessment to prognosis prediction. Crit Rev Oncol Hematol. Aug 2022;176:103746. doi:10.1016/j.critrevonc.2022.103746
+36. D'Andrea G, Lassalle S, Guevara N, Mograbi B, Hofman P. From biomarkers to therapeutic targets: the promise of PD-L1 in thyroid autoimmunity and cancer. Theranostics. 2021;11(3):1310-1325. doi:10.7150/thno.50333
+37. Chera A, Stancu AL, Bucur O. Thyroid-related adverse events induced by immune checkpoint inhibitors. Front Endocrinol (Lausanne). 2022;13:1010279. doi:10.3389/fendo.2022.1010279
+38. Wang W, Green M, Choi JE, et al. CD8(+) T cells regulate tumour ferroptosis during cancer immunotherapy. Nature. May 2019;569(7755):270-274. doi:10.1038/s41586-019-1170-y
+39. Chen Z, Guo ML, Li YY, et al. Immune profiling identifies CD8(+) T-cell subset signatures as prognostic markers for recurrence in papillary thyroid cancer. Front Immunol. 2022;13:894919. doi:10.3389/fimmu.2022.894919
+40. Cunha LL, Morari EC, Guihen AC, et al. Infiltration of a mixture of immune cells may be related to good prognosis in patients with differentiated thyroid carcinoma. Clin Endocrinol (Oxf). Dec 2012;77(6):918-25. doi:10.1111/j.1365-2265.2012.04482.x
+41. Cunha LL, Marcello MA, Nonogaki S, et al. CD8+ tumour-infiltrating lymphocytes and COX2 expression may predict relapse in differentiated thyroid cancer. Clin Endocrinol (Oxf). Aug 2015;83(2):246-53. doi:10.1111/cen.12586
+42. Zhu Y, Weiss T, Zhang Q, et al. High-throughput proteomic analysis of FFPE tissue samples facilitates tumor stratification. Mol Oncol. Sep 8 2019;13(11):2305-2328. doi:10.1002/1878-0261.12570
+43. Nie X, Qian L, Sun R, et al. Multi-organ proteomic landscape of COVID-19 autopsies. Cell. Feb 4 2021;184(3):775-791 e14. doi:10.1016/j.cell.2021.01.004
+44. Leek JT, Johnson WE, Parker HS, Jaffe AE, Storey JD. The sva package for removing batch effects and other unwanted variation in high-throughput experiments. Bioinformatics. Mar 15 2012;28(6):882-3. doi:10.1093/bioinformatics/bts034
+45. Pu W, Shi X, Yu P, et al. Single-cell transcriptomic analysis of the tumor ecosystems underlying initiation and progression of papillary thyroid carcinoma. Nature Communications 2021 12:1: Nature Publishing Group; 2021. p. 1-18.
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1. Study overview. (A) Study design of analyzed cohort and experiment workflow. </center>  
+
+Created through Biorender.com. (B) Enrollment and exclusion criteria for pediatric PTC, benign and adult PTC patients.
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure 2. Analysis of the clinical recurrence risk factors. (A) Kaplan-Meier survival curves of </center>  
+
+two groups (red: patients below the median; blue: otherwise) for three significant factors: age,
+
+<--- Page Split --->
+
+701 total lymph node metastasis (TLNN) and lateral lymph node metastasis number (LLNN). (B-C)  
+
+702 Forest plots for two multivariate CoxPH models using (B) continuous non- negative integer age  
+
+703 and (C) categorical age, respectively.  
+
+704
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Figure 3. Functional analyses of the dysregulated proteins. (A and B) DEPs are shown in the </center>  
+
+volcano plots: (A) PM vs. PB and (B) PM vs. AM. The cutoff is defined by requiring the FC to
+
+<--- Page Split --->
+
+be greater than 1.2 or 1.5, with adjusted \(P< 0.05\) . The names of the up/down-regulated proteins with the top five largest FC are reported in the plots. (C) The scatter diagram shows the FC distribution of the dysregulated proteins in two pairwise comparisons: PM vs.PB and PM vs. AM. The overlapping significantly co- dysregulated proteins are colored in red. The proteins significantly dysregulated in PM/PB are colored in orange, and those dysregulated in PM/AM are colored in blue. (D) The heatmap shows 37 proteins: they are co- upregulated/co- downregulated proteins and the five most up-/down- regulated ones from the volcano plots (A- B). Proteins were clustered using hierarchical clustering. (E) Pathway enrichment of the 243 DEPs from the volcano plot with the PM/PB comparison (FC \(>1.5\) ). The red and blue bars represent the active and inhibited pathways, respectively. (F) Results of the gene ontology enrichment of the biological processes using the DEPs in PM/PB.
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Figure 4. In-silico immune infiltration analysis and expression levels of immune check </center>  
+
+points. (A) Relative proportions of seven types of immune cells in PB and PM samples imputed by CIBERSORTx. The asterisks marked significant difference between PB and PM by Student's \(t\) - test ( \(*P < 0.05\) , \(**P < 0.01\) , \(***P < 0.001\) ). (B) Representative multiplex immunohistochemistry staining in PB (N=10) and three PM samples (N=10). The scale bar represents 50 \(\mu \mathrm{m}\) . (C) The protein expression abundances of PVR and IL10RB in PB, PM-NR (non-recurrence) and PM-R (recurrence) groups.
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Figure 5. PPTC prognostic prediction. (A) The C-indexes of our five models were calculated </center>  
+
+on training, 3- fold cross- validation, and test sets. (B) Density curves of the training Crank scores of two groups (recurrence (1) or no recurrence (0)). The Fisher decision boundary was used to differentiate the low- from the high- risk groups. (C) The Kaplan- Meier survival curves of the low- and high- risk groups, calculated on the training and test sets, show significant differences. (D) Permutation importance of the 19 proteins from the ProtRsf model. (E) Network showing the 19 features of the ProtRsf model with the connected proteins enriched using the IPA software. (F) The relative protein abundances of LGALS3, the hub protein of the network (E), in the four
+
+<--- Page Split --->
+
+groups. The mild outliers were removed, and a two- sided unpaired Wilcoxon rank- sum test was  
+
+used, without continuity correction, to calculate the \(P\) values.
+
+<--- Page Split --->
+![](images/Figure_6.jpg)
+
+<center>Figure 6. Risk stratification. (A) Predicted risk stratification for PPTC patients. The sample </center>  
+
+indexes of false positives (N=8) and the false negatives (N=2) are labeled. (B and C) The predicted survival curves of the two false negatives (B) and eight false positives (C) with their Crank scores, sample indexes and recurrence or latest follow-up times (shown by the vertical lines).
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+- PPTCsupp20230602.docx
+
+<--- Page Split --->

preprint/preprint__274e95da440ff6392b39f82cfd8c0d9037b52e5365e600c5936d7b07e8956ae1/images_list.json

@@ -0,0 +1,40 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2. Biochemical and kinetical characterization of PSI assembly intermediate.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 3
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3. Superposition and comparison of pre-PSI-1 with mature PSI.",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        350,
+        886,
+        735
+      ]
+    ],
+    "page_idx": 4
+  },
+  {
+    "type": "image",
+    "img_path": "images/Supplementary_Figure_1.jpg",
+    "caption": "Supplementary Fig. 1 Data processing, Fourier Shell Correlations (FSCs), angular",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        88,
+        880,
+        475
+      ]
+    ],
+    "page_idx": 5
+  }
+]

preprint/preprint__274e95da440ff6392b39f82cfd8c0d9037b52e5365e600c5936d7b07e8956ae1/preprint__274e95da440ff6392b39f82cfd8c0d9037b52e5365e600c5936d7b07e8956ae1.mmd

@@ -0,0 +1,225 @@
+
+# Structure of plant Photosystem I in a native assembly state  
+
+Andreas Naschberger  Stockholm University  
+
+Maria Fadeeva  Tel Aviv University  
+
+Daniel Klaiman  Tel Aviv University  https://orcid.org/0000- 0002- 3807- 6591  
+
+Anna Borovikova- Sheinker  Tel Aviv University  
+
+Ido Caspy  Tel Aviv University  https://orcid.org/0000- 0003- 1981- 1017  
+
+Nathan Nelson  Tel Aviv University  https://orcid.org/0000- 0003- 3588- 7265  
+
+Alexey Amunts  
+
+amunts@scilifelab.se  
+
+Stockholm University  https://orcid.org/0000- 0002- 5302- 1740  
+
+Brief Communication  
+
+Keywords:  
+
+Posted Date: December 28th, 2022  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 2406494/v1  
+
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Plants on May 30th, 2024. See the published version at https://doi.org/10.1038/s41477- 024- 01699- 8.
+
+<--- Page Split --->
+
+## Structure of plant Photosystem I in a native assembly state  
+
+3 Andreas Naschberger \(^{1, \dagger}\) , Mariia Fadeeva \(^{2, \dagger}\) , Daniel Klaiman \(^{2}\) , Anna Borovikova- Sheinker \(^{2}\) , Ido Caspy \(^{2}\) , Nathan Nelson \(^{2, *}\) , Alexey Amunts \(^{1, *}\)  
+
+6 'Science for Life Laboratory, Department of Biochemistry and Biophysics, Stockholm University, 7 17165 Solna, Sweden  
+
+8 'The George S. Wise Faculty of Life Sciences, Department of Biochemistry and Molecular  
+
+9 Biology, Tel Aviv University, 69978 Tel Aviv, Israel  
+
+11 \*Correspondence to: nelson@tauex.tau.ac.il; amunts@scilifelab.se  
+
+12 \*These authors contributed equally.  
+
+13  
+
+14  
+
+## 15 Abstract  
+
+16 Plant Photosystem I (PSI) consists of at least 13 nuclear- encoded and four chloroplast- encoded protein subunits, acting as a sunlight- driven oxidoreductase. Here, we report the first structure of PSI assembly intermediate isolated from greening oat seedlings. The assembly state displays absence of at least eight subunits and lacks the photoreduction activity, which is attributed to an elimination of the plastocyanin (Pc) binding site, due to a missing subunit PsaF. The data shows that PsaF is a regulatory checkpoint that promotes the assembly, effectively coupling biogenesis to function.
+
+<--- Page Split --->
+
+PSI is a large protein- pigment complex embedded in photosynthetic membranes of chloroplast and cyanobacteria. It primarily functions as a transmembrane electron conductor from the luminal reduced carrier Pc to the stromal oxidised acceptor ferredoxin, providing the reducing power needed for carbon fixation. Chloroplast PSI shares a common ancestor with cyanobacterial counterpart and have additional subunits and light- harvesting proteins that can be remodelled to regulate dimerization \(^{1,2}\) . In plants, the core subunits are encoded by the chloroplast genome, whereas at least 13 tightly bound subunits are imported from the cytoplasm. Subunit PsaF is unique with respect to its location on the PSI and the assembly path. Its position in the complex is stabilised by a single- transmembrane subunit PsaJ in a way that orients the N- terminal domain towards the lumen, where it forms a docking site for Pc \(^{3,5}\) . The assembly path of PsaF involves import into the thylakoid lumen \(^{6}\) . This way, a co- localisation of the positively charged N- terminal region with the Pc is achieved, thus supporting their functional recognition for electron transfer that is based on electrostatic interactions \(^{2,7,8}\) .  
+
+The complex architecture and the requirement to coordinate two genetic systems and distinct assembly paths generate additional complexity for the biogenesis of PSI. A stepwise process has been suggested \(^{9}\) , and biochemical studies proposed that trans- acting factors might contribute to biogenesis \(^{10}\) . However, despite the central role of PSI in the light reaction of photosynthesis, the data are limited to biochemical characterisations of subcomplexes in young leaves of N. tabacum \(^{11}\) , and no structural data exist on any of the assembly steps of PSI.  
+
+To identify a stable PSI assembly intermediate, we took advantage of the etioplast to chloroplast transition system. A high chlorophyll \(a / b\) ratio during that process \(^{12}\) , as well as the presence of PsaA- PsaB \(^{9}\) indicate a preference for PSI biogenesis. Therefore, we grew A. sativa seedlings in the darkness to accumulate photosynthetically inactive organelles, and then initiated greening by irradiating the seedlings as they come out of soil. Upon ten hours of irradiation, we harvested the plants, isolated thylakoids and solubilised photosynthetic complexes. The material that migrated as a green band was detected on a sucrose gradient, eluted, and subjected to the cryo- EM analysis. We collected 3,324 images and generated a reconstruction at 2.1 Å resolution (Supplementary Fig. 1 and Supplementary Table 1). The map dimensions of \(140 \times 110\) Å compared to a typical size of \(170 \times 150\) Å for a mature PSI suggested a different protein composition with missing subunits (Fig. 1a). The well- defined density allowed us to build an atomic model that corresponds to a
+
+<--- Page Split --->
+
+60 subcomplex consisting of eight subunits: PsaA, PsaB, PsaC, PsaD, PsaE, PsaH, PsaI, PsaL, which 61 we name pre- PSI- 1 (Fig. 1b).  
+
+![](images/Figure_2.jpg)
+
+
+<--- Page Split --->
+
+## 63 Fig. 1. Structural characterization of PSI assembly intermediate.  
+
+64 a, Coomassie stained SDS- PAGE of pre- PSI- 1 and mature PSI. Local resolution of 3D 65 reconstructed map of pre- PSI- 1 and mature PSI. b, Model of pre- PSI- 1 with missing PsaF (white 66 in close- up view, Pc- binding residues are red) and mature PSI with newly identified chlorophyll 67 CLA867 (red in close- up view).  
+
+![](images/Figure_3.jpg)
+
+<center>Fig. 2. Biochemical and kinetical characterization of PSI assembly intermediate. </center>  
+
+71 a, Light- induced P700 oxidation in mature PSI preparations from \(P\) . sativum and \(A\) . sativa. b, Chlorophyll to P700 ratio of the \(A\) . sativa pre- PSI- 1 that contains 64 chlorophylls compared with mature PSI from \(P\) . sativum. c, Light- induced P700 photo- oxidation and P700+ reduction of mature PSI from \(A\) . sativa with different Pc concentrations: Black - \(4.8\mu \mathrm{g}\) , gray - \(14.3\mu \mathrm{g}\) , blue - \(47.6\mu \mathrm{g}\) and green - \(47.6\mu \mathrm{g}\) of \(\mathrm{Pc} + 200\mathrm{mMNaCl}\) and of pre- PSI- 1. d, Light- induced P700 photo- oxidation and P700+ reduction of pre- PSI- 1. Pc amounts and labelling like in (b) except of orange - \(47.6\mu \mathrm{g} + 400\mathrm{mMNaCl}\) .  
+
+79 To confirm that the pre- PSI- 1 is not a degradation product, we purified PSI from mature green 80 leaves and determined its cryo- EM structure. The refined structure reached 2.2 Å resolution (Fig. 81 1a), which represents the highest reported resolution for a plant PSI. Mature PSI from \(A\) . sativa 82 shows a similar P700 oxidation profile compared to \(P\) . sativum (Fig. 2a) indicating a fully 83 functional protein complex. The data is also consistent in the terms of chlorophyll to P700 ratio 84 (Fig. 2b). Comparison between pre- PSI- 1 and mature PSI revealed a series of structural alterations 85 with respect to cofactor and protein conformation close to the PSI- LHCI interface (Fig. 3). First,
+
+<--- Page Split --->
+
+86 the tail of chlorophyll CLA815 in pre- PSI- 1 is ordered and occupies the space where the head 87 group of beta carotene BCR856 resides in the mature structure. Second, the position of lutein 858 88 is partially preoccupied by a detergent molecule that likely represents a lipid as structural holder 89 in pre- PSI- 1 until PsaF is incorporated into the complex. Third, phylloquinone PQN844 (PsaA) 90 shows an altered conformation of the isoprenoid chain. Finally, the N- terminus of PsaA is largely 91 unstructured in pre- PSI- 1 and forms a short two- turn helix (residues 31- 40). In the mature PSI, it 92 is replaced by a loop motif, followed by an ordered N- terminus stabilized mainly by interactions 93 with Lhca3 (Fig. 3). Taken together, the structural alterations between the core of PSI and LHC1, 94 suggest for a fine- tuned mechanism involving conformational changes in PSI assembly triggered 95 by the presence or absence of PsaF and LHC1.  
+
+![](images/Supplementary_Figure_1.jpg)
+
+<center>Fig. 3. Superposition and comparison of pre-PSI-1 with mature PSI. </center>  
+
+The complete structure (top, left) and pigments (top, right) are shown for pre- PSI- 1 (coloured) and mature PSI (transparent). Close- up views of structural alterations are shown in the bottom panels (mature PSI, white), and the coloured boxes indicate the position of the respective view in the structure.  
+
+Since unlike previously reported mini- PSI \(^{13}\) , pre- PSI- 1 is lacking the functionally critical subunit PsaF (Fig. 1b), we compared its kinetic properties to the mature PSI. While the PSI exhibited
+
+<--- Page Split --->
+
+relatively high nicotinamide adenine dinucleotide phosphate (NADP) photoreduction activity of \(633\pm 36\) (umole NADPH/mg chl/h), the pre- PSI- 1 showed no activity. We next inspected light- induced P700 photo- oxidation and \(\mathrm{P700 + }\) reduction by plastocyanin. The assay is based on different Pc concentrations, where a micromolar concertation in the presence of ascorbate results in fully oxidised P700 upon illumination that is re- reduced in the dark, whereas excess of Pc doesn't lead to P700 oxidation (Fig. 2b). An addition of NaCl salt then slows down the electron transfer, resulting in a light- dependent accumulation of oxidised P700 in the mature PSI. In contrast, for the pre- PSI- 1 the electron transfers markedly slowed down (Fig. 2d), enabling accumulation of oxidized P700 independently of plastocyanin concentration. Increasing NaCl has no effect on the electron transfer but on acceleration of \(\mathrm{P700 + }\) re- reduction rates, attributed to the hydrophobic interactions between Pc and pre- PSI- 1 in the absence of PsaF.  
+
+Finally, we built a canonical mature complex with a minimal clashscore of 3.4 (Supplementary Table 1), which represents a new reference model of a plant PSI. Compared to the previous models \(^{14,15}\) , we identified a new gap chlorophyll CLA867 that is found between subunits PsaA and Lhca3 (Fig. 1b). CLA867 in our model is situated between CLA617 (3011) of Lhca3 and CLA812 (1109) of PsaA, within 10.6 Å and 13.3 Å, respectively (Fig. 1b). Since the range is favorable for fast excitation energy transfer, our structure suggests that the CLA867 position rationalises a previously undetected excitation energy path LHC1 to PSI. Importantly, CLA617 is situated between protein moieties with no direct coordination.  
+
+Overall, by using a natural system of organelle maturation, we visualized an intermediate in the biogenesis pathway of the plant PSI that contains less than half of its subunits. Our study structurally characterised a native complex, which is the first structure of a PSI assembly intermediate. It reveals that a stable binding of the conserved subunit PsaF is required for photoreduction and represents a critical step of PSI biogenesis. Complemented by the unique feature of PsaF being inserted into PSI from the thylakoid lumen through the same pathway as Pc, the structural data suggests that its attachment is mechanistically regulated. Therefore, PsaF might serve as a regulatory checkpoint that promotes the assembly, effectively coupling it to function. A previous study in Synechocystis sp. PCC6803 showed that PsaF- depleted mutant forms a stable PSI sub- complex \(^{16}\) .  
+
+Collectively, our data confirm that PsaF is intimately linked to the photosynthetic functionality and assembly in a way that the pathway is dependent on the subunit levels accumulated in the
+
+<--- Page Split --->
+
+thylakoid lumen during seedling greening. The exact mechanism of how PSI is modulated throughout the dynamic assembly to establish the catalytic complex remains to be explored, and our study opens the door for future work on more specific roles of other factors and their regulation. In addition, we define the most accurate available model of a plant PSI, including the complete set of pigments at their correct orientations that will provide a reference plant PSI model for structural and molecular sciences.  
+
+## Methods  
+
+Purification of PSI  
+
+Pre- PSI- 1 was prepared from A. sativa (var. Saja 6) grown in dark at \(25^{\circ}\mathrm{C}\) for 5 days. Following the dark period, the plants were irradiated with cool- white, fluorescent light at a photon flux density of \(50\mu \mathrm{E}\mathrm{m}^{- 2}\mathrm{s}^{- 1}\) for 10 hours. \(\sim 470\mathrm{g}\) of leaves were harvested and and ground in a blender with \(900~\mathrm{ml}\) of buffer containing \(0.4\mathrm{M}\) sucrose, Tricine- NaOH (pH 8), \(15\mathrm{mMNaCl}\) , \(30\mathrm{mM}\) , \(2\mathrm{mM}\) ascorbic acid, \(1\mathrm{mM}\) PMSF, \(1\mu \mathrm{M}\) pepstatin A. After filtration through cheesecloth, the suspension was centrifuged at \(6,000\mathrm{g}\) for \(10\mathrm{min}\) , and again at \(180,000\mathrm{g}\) for \(20\mathrm{min}\) . The pellet was resuspended in \(200\mathrm{ml}\) buffer containing \(10\mathrm{mM}\) Tricine- NaOH (pH 8), \(150\mathrm{mMNaCl}\) and pelleted through centrifugation of \(180,000\mathrm{g}\) for \(20\mathrm{min}\) . The pellet was resuspended in \(45\mathrm{ml}\) of buffer containing \(10\mathrm{mM}\) Tricine- NaOH (pH 8), \(0.4\mathrm{M}\) sucrose, to a concentration of \(1\mathrm{mg / ml}\) chlorophyll and solubilized with \(1.5\%\) n- dodecyl- \(\beta\) - D- maltoside (DDM). Following \(30\mathrm{min}\) incubation on ice, the material was centrifuged at \(176,000\) for \(20\mathrm{min}\) and applied on DEAE- cellulose column \((2.5\times 13\mathrm{cm})\) pre- equilibrated with \(20\mathrm{mM}\) Tricine- Tris (pH 8.0), \(0.2\%\) DDM. The material was eluted with \(300\mathrm{mMNaCl}\) , concentrated twice with PEG 6000 precipitation and centrifuged through a sucrose gradient of \(10\) to \(40\%\) in SW40 rotor (Beckman, Inc) at \(170,000\mathrm{g}\) for \(16\) hours. The green band containing PSI was collected, subjected to FPLC chromatography, and then applied on to a \(10 - 35\%\) sucrose gradient and centrifuged at \(336,000\mathrm{g}\) for \(4\) hours in SW- 60 rotor (Beckman, Inc). The green band was collected, and sucrose was removed by buffer exchange.  
+
+Mature PSI was prepared from 7- day old A. sativa (var. Saja 6) grown for 7 days in \(16 / 8\) light/dark cycle at a photon flux density \(50\mu \mathrm{E}\mathrm{m}^{- 2}\mathrm{s}^{- 1}\) and followed the protocol published before \(^{8}\) .
+
+<--- Page Split --->
+
+167 Kinetic measurements, P700 reduction, and NADP+ photo- reduction activity assay  
+
+168 PC was heterologously expressed, and codon- optimized; the following experiments were performed generally as previously described 17. P700 reduction was measured in a Quartz cuvette, containing: 1 ml reaction mix (20 mM Tricine- NaOH pH 8, 5 mM \(\mathrm{MgCl}_2\) and \(0.05\%\) n- 171 Dodecyl \(\beta\) - maltoside), 10 \(\mu\) mol Ascorbate, 100 nmol Methyl viologen, 16 \(\mu\) g chlorophyll PSI, 172 and 50 pmol Pc of Synechocystis PCC SP 6803 or 50 pmol Cyt \(\mathrm{c}_6\) (Cyt C533). P700 photo- 173 oxidation and re- reduction by Pc and Cyt \(\mathrm{c}_6\) was measured by JTS- 10 spectrophotometer by 174 illuminating the sample with red light (705 nm) for 5 s. Changes in absorbance were measured 175 by 2 ms of LED light flashes (700 nm passed through a 705 nm interference filter).  
+
+177 NADP+ photo- reduction activity assay was measured in a Quartz cuvette, 1 ml reaction mix (20 mM NaCl, 10 mM Tricine NaOH pH 8, 0.5 mM \(\mathrm{MgCl}_2\) ) was supplemented with 20 \(\mu\) mol 178 ascorbate, 125 \(\mu\) g Ferredoxin, 8.8 \(\mu\) g FNR, 1 \(\mu\) mol NADP+ (Roche Diagnostics), 14 nmol Pc, 179 and PSI (14.4 \(\mu\) g chlorophyll). NADPH accumulation was measured at 340 nm ( \(\epsilon = 6220 \mathrm{M}^{- 1}\) 180 \(\mathrm{cm}^{- 1}\) ) by Cary 60 spectrophotometer (Agilent technologies) under continuous illumination with 181 660 nm LED light (600 \(\mu\) E). The activity was calculated as \(\mu\) mol NADPH/(mg chl- Hr).  
+
+183 Cryo- EM data collection, processing, and model building  
+
+184 3 \(\mu\) l of pre- PSI- 1 and mature PSI at 2 mg/ml chlorophyll were applied on glow- discharged holey 185 carbon grids and vitrified for cryo- EM structural determination using Leica- EM- GP (3 sec blot at 186 \(20^{\circ}\mathrm{C}\) and \(90\%\) humidity). Data were collected on a 300- kV Titan Krios G3 Microscope (Thermo 187 Fisher Scientific) equipped with a Gatan Bioquantum energy filter and a K3 Summit direct electron 188 detector (Ametek). Movies were recorded using counting mode at a magnification of \(\times 105,000\) 189 corresponding to a calibrated pixel size of 0.84- 0.85 Å. 3,324 micrographs at a total dose of 45 190 \(\mathrm{e} / \mathrm{\AA}^2\) and 24,930 micrographs at 51 \(\mathrm{e} / \mathrm{\AA}^2\) were collected for pre- PSI- 1 and mature PSI, respectively, 191 with a defocus range from - 0.5 to - 1.9 \(\mu\) m. Movies were imported into cryoSPARC 3.1 18, motion 192 correction, CTF estimation, picking, and 2D classification was performed on the fly during data 193 collection using cryoSPARC live (with blob picker and template picker). Ab initio models were 194 generated with a subset of particles. Heterogenous and homogeneous refinement was performed 195 for pre- PSI- 1 and mature PSI respectively. Particles (383,325 for mature PSI and 546,410 for pre- 196 PSI- 1) were converted into a Star file format and were imported into RELION 3.1.1 19. Particles 197 were re- extracted (un- binned) and processed in RELION using a box size of 400 pixels for pre
+
+<--- Page Split --->
+
+PS- 1 and 500 pixels for mature PSI. For pre- PSI- 1, 3D classification with 2 classes was performed. One class with 169,213 particles of high- quality particles was selected and subjected to 3D refinement which resulted in an overall resolution of 3.1 Å. CTF refinement, 3D refinement and Bayesian polishing followed by another round of CTF refinement was performed for pre- PSI- 1 and mature PSI. Another 3D refinement resulted in an overall resolution of 2.1 Å for the pre- PSI- 1 and mature PSI. Since the Lhca2- 3 heterodimer in the mature PSI appeared to be loosely bound, we used focused classification (3 classes) with signal subtraction with a mask around the Lhca2- 3 region to improve the local density. One subclass showed a better ordered Lhca2- 3 region. The particles of this subclass (96,997) were selected, and the signal was reverted. A final 3D refinement of mature PSI resulted in an overall resolution of 2.2 Å with an improved density for the Lhca2- 3 region.  
+
+Model building and real- space refinement of pre- PSI- 1 was then carried out using Coot 9.1.4 20. The completed model was then fitted into the mature PSI map, and the remaining protein chains were built using rigid body fitted PSI (PDB ID: 6YAC) as a starting model. All protein residues as well as pigments were fitted using Coot with locally optimized map weights. For the entire modelling in Coot restraint files for pigments and ligands were used which were generated using the Grade server (http://grade.globalphasing.org), as previously described 21. Models were refined using Real- Space- Refine from the PHENIX suite 22. The refinement protocol was optimized by adjusting for optimal refinement weight parameters. Iterations of validation, model building, and refinement were carried out using MolProbity 4.2 23, Coot 9.1.4 20 and PHENIX 22.  
+
+## Data availability  
+
+Atomic coordinates and structure factors of pre- PSI- 1 have been deposited in the Protein Data Bank under accession code 8BCW. Atomic coordinates and structure factors of mature PSI have been deposited in the Protein Data Bank under accession code 8BCW. The cryo- EM map of pre- PSI- 1 has been deposited in the Electron Microscopy Data Bank under accession code EMD- 15970. The cryo- EM map of mature PSI has been deposited in the Electron Microscopy Data Bank under accession code EMD- 15969. Other atomic coordinates that were used in this study: 6YAC [https://www.rcsb.org/structure/6YAC] (PSI- ferredoxin).
+
+<--- Page Split --->
+
+## Acknowledgments  
+
+This work was supported by the Swedish Foundation for Strategic Research (FFL15:0325, ARC19- 0051), European Research Council (ERC- 2018- StG- 805230), Knut and Alice Wallenberg Foundation (2018.0080), EMBO Young Investigator Program. The cryo- EM facility is funded by the Knut and Alice Wallenberg, Family Erling Persson, and Kempe foundations.  
+
+## Author Contributions Statement  
+
+N.N. and A.A. designed the project. M.F., D.K. and A.B-S. prepared the sample for cryo- EM. A.N. collected and processed the cryo- EM data and built the model. M.F., D.K., A.B-S. and I.C. performed biochemical and kinetic analysis. A.N., N.N. and A.A. analysed the structure and wrote the manuscript with contributions from M.F. All authors contributed to the analysis and the final version of the manuscript.  
+
+## Competing Interests Statement  
+
+The authors declare no competing interests.
+
+<--- Page Split --->
+![PLACEHOLDER_11_0]
+
+<center>Supplementary Fig. 1 Data processing, Fourier Shell Correlations (FSCs), angular </center>  
+
+distributions. a, Data processing scheme for mature PSI on the left and pre- PSI- 1 on the right.  
+
+b, Angular distribution and FSC for resolution estimation for pre- PSI- 1 (top) and of mature PSI  
+
+(bottom).
+
+<--- Page Split --->
+
+
+262 Supplementary Table 1 263 Cryo-EM data collection, refinement and validation statistics of PSI and pre-PSI-1 assembly 264 intermediate of Avena sativa.   
+
+<table><tr><td></td><td>PSI<br>(EMDB-15969)<br>(PDB 8BCV)</td><td>pre-PSI-1<br>(EMDB-15970)<br>(PDB 8BCW)</td></tr><tr><td>Data collection and processing</td><td></td><td></td></tr><tr><td>Magnification</td><td>105,000</td><td>105,000</td></tr><tr><td>Voltage (kV)</td><td>300</td><td>300</td></tr><tr><td>Electron exposure (e-/Å2)</td><td>45.00</td><td>51.35</td></tr><tr><td>Defocus range (μm)</td><td>-0.5-2.1</td><td>-0.5-2.1</td></tr><tr><td>Pixel size (Å)</td><td>0.85</td><td>0.84</td></tr><tr><td>Initial particle images (no.)</td><td>24,930</td><td>3,324</td></tr><tr><td>Symmetry imposed</td><td>C1</td><td>C1</td></tr><tr><td>Final particle images (no.)</td><td>96,997</td><td>169,213</td></tr><tr><td>Map resolution (Å)</td><td rowspan="2">2.20</td><td rowspan="2">2.11</td></tr><tr><td>FSC threshold 0.143</td></tr><tr><td>Refinement</td><td></td><td></td></tr><tr><td>Initial model used (PDB code)</td><td>6YAC</td><td>6YAC</td></tr><tr><td>Map sharpening B factor (Å2)</td><td>-17.41</td><td>-21.56</td></tr><tr><td>Model composition</td><td></td><td></td></tr><tr><td>Non-hydrogen atoms</td><td>37,485</td><td>22,313</td></tr><tr><td>Protein residues</td><td>3,258</td><td>2,029</td></tr><tr><td>Ligands</td><td>224</td><td>114</td></tr><tr><td>Waters</td><td>703</td><td>517</td></tr><tr><td>B factors (Å2)</td><td></td><td></td></tr><tr><td>(min/max/mean)</td><td></td><td></td></tr><tr><td>Protein</td><td>8.32/93.7/39.3</td><td>14.1/76.9/34.1</td></tr><tr><td>Ligand</td><td>10.3/87.8/38.4</td><td>15.1/64.3/32.6</td></tr><tr><td>Waters</td><td>9.8/80.1/28.7</td><td>12.0/55.0/32.7</td></tr><tr><td>R.m.s. deviations</td><td></td><td></td></tr><tr><td>Bond lengths (Å)</td><td>0.008</td><td>0.008</td></tr><tr><td>Bond angles (°)</td><td>1.059</td><td>1.032</td></tr><tr><td>Validation</td><td></td><td></td></tr><tr><td>MolProbity score</td><td>1.13</td><td>1.04</td></tr><tr><td>Clashscore</td><td>3.43</td><td>2.56</td></tr><tr><td>Poor rotamers (%)</td><td>0.64</td><td>0.54</td></tr></table>
+
+<--- Page Split --->
+
+<table><tr><td>Ramachandran plot</td><td></td><td></td></tr><tr><td>Favored (%)</td><td>98.67</td><td>98.11</td></tr><tr><td>Allowed (%)</td><td>1.33</td><td>1.89</td></tr><tr><td>Disallowed (%)</td><td>0.00</td><td>0.00</td></tr></table>
+
+265 
+
+266 
+
+267 
+
+268 
+
+269 1. M. Chen et al., Distinct structural modulation of photosystem I and lipid environment stabilizes its tetrameric assembly. Nat Plants 6, 314-320 (2020). 
+
+270 2. A. Naschberger et al., Algal photosystem I dimer and high-resolution model of PSI-plastocyanin complex. Nat Plants 8, 1191-1201 (2022). 
+
+273 3. A. Amunts, O. Drory, N. Nelson, The structure of a plant photosystem I supercomplex at 3.4 A resolution. Nature 447, 58-63 (2007). 
+
+275 4. A. Amunts, N. Nelson, Plant photosystem I design in the light of evolution. Structure 17, 637-650 (2009). 
+
+276 5. A. Amunts, H. Toporik, A. Borovikova, N. Nelson, Structure determination and improved model of plant photosystem I. J Biol Chem 285, 3478-3486 (2010). 
+
+279 6. I. Karnauchov, D. Cai, I. Schmidt, R. G. Herrmann, R. B. Klosgen, The thylakoid translocation of subunit 3 of photosystem I, the psaF gene product, depends on a bipartite transit peptide and proceeds along an azide-sensitive pathway. J Biol Chem 269, 32871-32878 (1994). 
+
+283 7. M. Hippler, F. Drepper, W. Haehnel, J. D. Rochaix, The N-terminal domain of PsaF: precise recognition site for binding and fast electron transfer from cytochrome c6 and plastocyanin to photosystem I of Chlamydomonas reinhardtii. Proc Natl Acad Sci U S A 95, 7339-7344 (1998). 
+
+287 8. I. Caspy, A. Borovikova-Sheinker, D. Klaiman, Y. Shkolnisky, N. Nelson, The structure of a triple complex of plant photosystem I with ferredoxin and plastocyanin. Nat Plants 6, 1300-1305 (2020). 
+
+290 9. R. Nechushtai, N. Nelson, Biogenesis of photosystem I reaction center during greening of oat, bean and spinach leaves. Plant Mol Biol 4, 377-384 (1985). 
+
+292 10. S. Nellaepalli, S. I. Ozawa, H. Kuroda, Y. Takahashi, The photosystem I assembly apparatus consisting of Ycf3-Y3IP1 and Ycf4 modules. Nat Commun 9, 2439 (2018). 
+
+294 11. G. Wittenberg et al., Identification and characterization of a stable intermediate in photosystem I assembly in tobacco. Plant J 90, 478-490 (2017). 
+
+296 12. W. Rudiger, J. Benz, C. Guthoff, Detection and partial characterization of activity of chlorophyll synthetase in etioplast membranes. Eur J Biochem 109, 193-200 (1980). 
+
+298 13. A. Perez-Boerema et al., Structure of a minimal photosystem I from the green alga Dunaliella salina. Nat Plants 6, 321-327 (2020). 
+
+300 14. J. Wang et al., Structure of plant photosystem I-light harvesting complex I supercomplex at 2.4 A resolution. J Integr Plant Biol 63, 1367-1381 (2021). 
+
+302 15. Y. Mazor, A. Borovikova, I. Caspy, N. Nelson, Structure of the plant photosystem I Supercomplex at 2.6 Å Resolution. Nature Plants 3, 17014 (2017). 
+
+304 16. T. Malavath, I. Caspy, SY. Netzer-El, D. Klaiman, N. Nelson, Structure and function of wild-type and subunit-depleted photosystem I in synechocystis. Biochimica et Biophysica Acta (BBA) - Bioenergetics 1859, 645-654 (2018).
+
+<--- Page Split --->
+
+307 17. SY. Netzer- El, I. Caspy, N. Nelson N, Crystal structure of photosystem I monomer from Synechocystis PCC 6803. Front Plant Sci 9, 1- 7 (2018).308 18. A. Punjani, J. Rubinstein, D. Fleet, M. Brubaker, CryoSPARC: algorithms for rapid unsupervised cryo- EM structure determination. Nat. Methods 14, 290- 296 (2017).309 19. J. Zivanov, T. Nakane, S.H. Scheres, Estimation of high- order aberrations and anisotropic magnification from cryo- EM data sets in RELION- 3.1. IUcrJ 7, 253- 267 (2020).310 20. P. Emsley, B. Lohkamp, W.G. Scott, K. Cowtan. Features and development of Coot. Acta Crystallogr. D 66, 486- 501 (2010).311 21. A. Amunts, The revolution evolution. Nat Plants 8, 14- 17 (2022).312 22. P. Afonine et al., Real- space refinement in phenix for cryo- EM and crystallography. Acta Crystallogr D Struct Biol. 74, 531- 544 (2018).313 23. C. Williams et al., Molprobity: more and better reference data for improved all- atom structure validation. Protein Sci. 27, 293- 315 (2017).
+
+<--- Page Split --->

preprint/preprint__274e95da440ff6392b39f82cfd8c0d9037b52e5365e600c5936d7b07e8956ae1/preprint__274e95da440ff6392b39f82cfd8c0d9037b52e5365e600c5936d7b07e8956ae1_det.mmd

@@ -0,0 +1,299 @@
+<|ref|>title<|/ref|><|det|>[[42, 108, 797, 175]]<|/det|>
+# Structure of plant Photosystem I in a native assembly state  
+
+<|ref|>text<|/ref|><|det|>[[42, 195, 240, 235]]<|/det|>
+Andreas Naschberger  Stockholm University  
+
+<|ref|>text<|/ref|><|det|>[[42, 242, 216, 282]]<|/det|>
+Maria Fadeeva  Tel Aviv University  
+
+<|ref|>text<|/ref|><|det|>[[42, 289, 572, 330]]<|/det|>
+Daniel Klaiman  Tel Aviv University  https://orcid.org/0000- 0002- 3807- 6591  
+
+<|ref|>text<|/ref|><|det|>[[42, 335, 272, 375]]<|/det|>
+Anna Borovikova- Sheinker  Tel Aviv University  
+
+<|ref|>text<|/ref|><|det|>[[42, 382, 575, 423]]<|/det|>
+Ido Caspy  Tel Aviv University  https://orcid.org/0000- 0003- 1981- 1017  
+
+<|ref|>text<|/ref|><|det|>[[42, 428, 575, 469]]<|/det|>
+Nathan Nelson  Tel Aviv University  https://orcid.org/0000- 0003- 3588- 7265  
+
+<|ref|>text<|/ref|><|det|>[[42, 475, 176, 492]]<|/det|>
+Alexey Amunts  
+
+<|ref|>text<|/ref|><|det|>[[52, 502, 279, 519]]<|/det|>
+amunts@scilifelab.se  
+
+<|ref|>text<|/ref|><|det|>[[52, 547, 600, 567]]<|/det|>
+Stockholm University  https://orcid.org/0000- 0002- 5302- 1740  
+
+<|ref|>text<|/ref|><|det|>[[42, 606, 230, 625]]<|/det|>
+Brief Communication  
+
+<|ref|>text<|/ref|><|det|>[[42, 645, 136, 664]]<|/det|>
+Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[42, 683, 349, 702]]<|/det|>
+Posted Date: December 28th, 2022  
+
+<|ref|>text<|/ref|><|det|>[[42, 721, 475, 741]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 2406494/v1  
+
+<|ref|>text<|/ref|><|det|>[[42, 758, 914, 801]]<|/det|>
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 819, 535, 839]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 875, 940, 918]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Plants on May 30th, 2024. See the published version at https://doi.org/10.1038/s41477- 024- 01699- 8.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[210, 91, 787, 113]]<|/det|>
+## Structure of plant Photosystem I in a native assembly state  
+
+<|ref|>text<|/ref|><|det|>[[115, 145, 886, 193]]<|/det|>
+3 Andreas Naschberger \(^{1, \dagger}\) , Mariia Fadeeva \(^{2, \dagger}\) , Daniel Klaiman \(^{2}\) , Anna Borovikova- Sheinker \(^{2}\) , Ido Caspy \(^{2}\) , Nathan Nelson \(^{2, *}\) , Alexey Amunts \(^{1, *}\)  
+
+<|ref|>text<|/ref|><|det|>[[115, 226, 886, 272]]<|/det|>
+6 'Science for Life Laboratory, Department of Biochemistry and Biophysics, Stockholm University, 7 17165 Solna, Sweden  
+
+<|ref|>text<|/ref|><|det|>[[115, 277, 835, 297]]<|/det|>
+8 'The George S. Wise Faculty of Life Sciences, Department of Biochemistry and Molecular  
+
+<|ref|>text<|/ref|><|det|>[[115, 303, 538, 322]]<|/det|>
+9 Biology, Tel Aviv University, 69978 Tel Aviv, Israel  
+
+<|ref|>text<|/ref|><|det|>[[115, 357, 641, 376]]<|/det|>
+11 \*Correspondence to: nelson@tauex.tau.ac.il; amunts@scilifelab.se  
+
+<|ref|>text<|/ref|><|det|>[[115, 383, 396, 402]]<|/det|>
+12 \*These authors contributed equally.  
+
+<|ref|>text<|/ref|><|det|>[[115, 410, 896, 428]]<|/det|>
+13  
+
+<|ref|>text<|/ref|><|det|>[[115, 448, 896, 466]]<|/det|>
+14  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 479, 193, 496]]<|/det|>
+## 15 Abstract  
+
+<|ref|>text<|/ref|><|det|>[[112, 512, 886, 688]]<|/det|>
+16 Plant Photosystem I (PSI) consists of at least 13 nuclear- encoded and four chloroplast- encoded protein subunits, acting as a sunlight- driven oxidoreductase. Here, we report the first structure of PSI assembly intermediate isolated from greening oat seedlings. The assembly state displays absence of at least eight subunits and lacks the photoreduction activity, which is attributed to an elimination of the plastocyanin (Pc) binding site, due to a missing subunit PsaF. The data shows that PsaF is a regulatory checkpoint that promotes the assembly, effectively coupling biogenesis to function.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 886, 425]]<|/det|>
+PSI is a large protein- pigment complex embedded in photosynthetic membranes of chloroplast and cyanobacteria. It primarily functions as a transmembrane electron conductor from the luminal reduced carrier Pc to the stromal oxidised acceptor ferredoxin, providing the reducing power needed for carbon fixation. Chloroplast PSI shares a common ancestor with cyanobacterial counterpart and have additional subunits and light- harvesting proteins that can be remodelled to regulate dimerization \(^{1,2}\) . In plants, the core subunits are encoded by the chloroplast genome, whereas at least 13 tightly bound subunits are imported from the cytoplasm. Subunit PsaF is unique with respect to its location on the PSI and the assembly path. Its position in the complex is stabilised by a single- transmembrane subunit PsaJ in a way that orients the N- terminal domain towards the lumen, where it forms a docking site for Pc \(^{3,5}\) . The assembly path of PsaF involves import into the thylakoid lumen \(^{6}\) . This way, a co- localisation of the positively charged N- terminal region with the Pc is achieved, thus supporting their functional recognition for electron transfer that is based on electrostatic interactions \(^{2,7,8}\) .  
+
+<|ref|>text<|/ref|><|det|>[[111, 428, 888, 580]]<|/det|>
+The complex architecture and the requirement to coordinate two genetic systems and distinct assembly paths generate additional complexity for the biogenesis of PSI. A stepwise process has been suggested \(^{9}\) , and biochemical studies proposed that trans- acting factors might contribute to biogenesis \(^{10}\) . However, despite the central role of PSI in the light reaction of photosynthesis, the data are limited to biochemical characterisations of subcomplexes in young leaves of N. tabacum \(^{11}\) , and no structural data exist on any of the assembly steps of PSI.  
+
+<|ref|>text<|/ref|><|det|>[[111, 585, 888, 867]]<|/det|>
+To identify a stable PSI assembly intermediate, we took advantage of the etioplast to chloroplast transition system. A high chlorophyll \(a / b\) ratio during that process \(^{12}\) , as well as the presence of PsaA- PsaB \(^{9}\) indicate a preference for PSI biogenesis. Therefore, we grew A. sativa seedlings in the darkness to accumulate photosynthetically inactive organelles, and then initiated greening by irradiating the seedlings as they come out of soil. Upon ten hours of irradiation, we harvested the plants, isolated thylakoids and solubilised photosynthetic complexes. The material that migrated as a green band was detected on a sucrose gradient, eluted, and subjected to the cryo- EM analysis. We collected 3,324 images and generated a reconstruction at 2.1 Å resolution (Supplementary Fig. 1 and Supplementary Table 1). The map dimensions of \(140 \times 110\) Å compared to a typical size of \(170 \times 150\) Å for a mature PSI suggested a different protein composition with missing subunits (Fig. 1a). The well- defined density allowed us to build an atomic model that corresponds to a
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 91, 884, 137]]<|/det|>
+60 subcomplex consisting of eight subunits: PsaA, PsaB, PsaC, PsaD, PsaE, PsaH, PsaI, PsaL, which 61 we name pre- PSI- 1 (Fig. 1b).  
+
+<|ref|>image<|/ref|><|det|>[[125, 145, 857, 920]]<|/det|>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[112, 91, 658, 109]]<|/det|>
+## 63 Fig. 1. Structural characterization of PSI assembly intermediate.  
+
+<|ref|>text<|/ref|><|det|>[[112, 111, 884, 190]]<|/det|>
+64 a, Coomassie stained SDS- PAGE of pre- PSI- 1 and mature PSI. Local resolution of 3D 65 reconstructed map of pre- PSI- 1 and mature PSI. b, Model of pre- PSI- 1 with missing PsaF (white 66 in close- up view, Pc- binding residues are red) and mature PSI with newly identified chlorophyll 67 CLA867 (red in close- up view).  
+
+<|ref|>image<|/ref|><|det|>[[112, 210, 863, 540]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[112, 545, 787, 565]]<|/det|>
+<center>Fig. 2. Biochemical and kinetical characterization of PSI assembly intermediate. </center>  
+
+<|ref|>text<|/ref|><|det|>[[112, 565, 884, 705]]<|/det|>
+71 a, Light- induced P700 oxidation in mature PSI preparations from \(P\) . sativum and \(A\) . sativa. b, Chlorophyll to P700 ratio of the \(A\) . sativa pre- PSI- 1 that contains 64 chlorophylls compared with mature PSI from \(P\) . sativum. c, Light- induced P700 photo- oxidation and P700+ reduction of mature PSI from \(A\) . sativa with different Pc concentrations: Black - \(4.8\mu \mathrm{g}\) , gray - \(14.3\mu \mathrm{g}\) , blue - \(47.6\mu \mathrm{g}\) and green - \(47.6\mu \mathrm{g}\) of \(\mathrm{Pc} + 200\mathrm{mMNaCl}\) and of pre- PSI- 1. d, Light- induced P700 photo- oxidation and P700+ reduction of pre- PSI- 1. Pc amounts and labelling like in (b) except of orange - \(47.6\mu \mathrm{g} + 400\mathrm{mMNaCl}\) .  
+
+<|ref|>text<|/ref|><|det|>[[112, 725, 884, 901]]<|/det|>
+79 To confirm that the pre- PSI- 1 is not a degradation product, we purified PSI from mature green 80 leaves and determined its cryo- EM structure. The refined structure reached 2.2 Å resolution (Fig. 81 1a), which represents the highest reported resolution for a plant PSI. Mature PSI from \(A\) . sativa 82 shows a similar P700 oxidation profile compared to \(P\) . sativum (Fig. 2a) indicating a fully 83 functional protein complex. The data is also consistent in the terms of chlorophyll to P700 ratio 84 (Fig. 2b). Comparison between pre- PSI- 1 and mature PSI revealed a series of structural alterations 85 with respect to cofactor and protein conformation close to the PSI- LHCI interface (Fig. 3). First,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[68, 88, 888, 348]]<|/det|>
+86 the tail of chlorophyll CLA815 in pre- PSI- 1 is ordered and occupies the space where the head 87 group of beta carotene BCR856 resides in the mature structure. Second, the position of lutein 858 88 is partially preoccupied by a detergent molecule that likely represents a lipid as structural holder 89 in pre- PSI- 1 until PsaF is incorporated into the complex. Third, phylloquinone PQN844 (PsaA) 90 shows an altered conformation of the isoprenoid chain. Finally, the N- terminus of PsaA is largely 91 unstructured in pre- PSI- 1 and forms a short two- turn helix (residues 31- 40). In the mature PSI, it 92 is replaced by a loop motif, followed by an ordered N- terminus stabilized mainly by interactions 93 with Lhca3 (Fig. 3). Taken together, the structural alterations between the core of PSI and LHC1, 94 suggest for a fine- tuned mechanism involving conformational changes in PSI assembly triggered 95 by the presence or absence of PsaF and LHC1.  
+
+<|ref|>image<|/ref|><|det|>[[112, 350, 886, 735]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 740, 688, 759]]<|/det|>
+<center>Fig. 3. Superposition and comparison of pre-PSI-1 with mature PSI. </center>  
+
+<|ref|>text<|/ref|><|det|>[[115, 760, 886, 840]]<|/det|>
+The complete structure (top, left) and pigments (top, right) are shown for pre- PSI- 1 (coloured) and mature PSI (transparent). Close- up views of structural alterations are shown in the bottom panels (mature PSI, white), and the coloured boxes indicate the position of the respective view in the structure.  
+
+<|ref|>text<|/ref|><|det|>[[115, 860, 886, 905]]<|/det|>
+Since unlike previously reported mini- PSI \(^{13}\) , pre- PSI- 1 is lacking the functionally critical subunit PsaF (Fig. 1b), we compared its kinetic properties to the mature PSI. While the PSI exhibited
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 88, 886, 372]]<|/det|>
+relatively high nicotinamide adenine dinucleotide phosphate (NADP) photoreduction activity of \(633\pm 36\) (umole NADPH/mg chl/h), the pre- PSI- 1 showed no activity. We next inspected light- induced P700 photo- oxidation and \(\mathrm{P700 + }\) reduction by plastocyanin. The assay is based on different Pc concentrations, where a micromolar concertation in the presence of ascorbate results in fully oxidised P700 upon illumination that is re- reduced in the dark, whereas excess of Pc doesn't lead to P700 oxidation (Fig. 2b). An addition of NaCl salt then slows down the electron transfer, resulting in a light- dependent accumulation of oxidised P700 in the mature PSI. In contrast, for the pre- PSI- 1 the electron transfers markedly slowed down (Fig. 2d), enabling accumulation of oxidized P700 independently of plastocyanin concentration. Increasing NaCl has no effect on the electron transfer but on acceleration of \(\mathrm{P700 + }\) re- reduction rates, attributed to the hydrophobic interactions between Pc and pre- PSI- 1 in the absence of PsaF.  
+
+<|ref|>text<|/ref|><|det|>[[111, 377, 886, 581]]<|/det|>
+Finally, we built a canonical mature complex with a minimal clashscore of 3.4 (Supplementary Table 1), which represents a new reference model of a plant PSI. Compared to the previous models \(^{14,15}\) , we identified a new gap chlorophyll CLA867 that is found between subunits PsaA and Lhca3 (Fig. 1b). CLA867 in our model is situated between CLA617 (3011) of Lhca3 and CLA812 (1109) of PsaA, within 10.6 Å and 13.3 Å, respectively (Fig. 1b). Since the range is favorable for fast excitation energy transfer, our structure suggests that the CLA867 position rationalises a previously undetected excitation energy path LHC1 to PSI. Importantly, CLA617 is situated between protein moieties with no direct coordination.  
+
+<|ref|>text<|/ref|><|det|>[[111, 587, 886, 842]]<|/det|>
+Overall, by using a natural system of organelle maturation, we visualized an intermediate in the biogenesis pathway of the plant PSI that contains less than half of its subunits. Our study structurally characterised a native complex, which is the first structure of a PSI assembly intermediate. It reveals that a stable binding of the conserved subunit PsaF is required for photoreduction and represents a critical step of PSI biogenesis. Complemented by the unique feature of PsaF being inserted into PSI from the thylakoid lumen through the same pathway as Pc, the structural data suggests that its attachment is mechanistically regulated. Therefore, PsaF might serve as a regulatory checkpoint that promotes the assembly, effectively coupling it to function. A previous study in Synechocystis sp. PCC6803 showed that PsaF- depleted mutant forms a stable PSI sub- complex \(^{16}\) .  
+
+<|ref|>text<|/ref|><|det|>[[111, 849, 885, 895]]<|/det|>
+Collectively, our data confirm that PsaF is intimately linked to the photosynthetic functionality and assembly in a way that the pathway is dependent on the subunit levels accumulated in the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[112, 89, 886, 240]]<|/det|>
+thylakoid lumen during seedling greening. The exact mechanism of how PSI is modulated throughout the dynamic assembly to establish the catalytic complex remains to be explored, and our study opens the door for future work on more specific roles of other factors and their regulation. In addition, we define the most accurate available model of a plant PSI, including the complete set of pigments at their correct orientations that will provide a reference plant PSI model for structural and molecular sciences.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 276, 205, 296]]<|/det|>
+## Methods  
+
+<|ref|>text<|/ref|><|det|>[[115, 303, 266, 322]]<|/det|>
+Purification of PSI  
+
+<|ref|>text<|/ref|><|det|>[[111, 327, 886, 796]]<|/det|>
+Pre- PSI- 1 was prepared from A. sativa (var. Saja 6) grown in dark at \(25^{\circ}\mathrm{C}\) for 5 days. Following the dark period, the plants were irradiated with cool- white, fluorescent light at a photon flux density of \(50\mu \mathrm{E}\mathrm{m}^{- 2}\mathrm{s}^{- 1}\) for 10 hours. \(\sim 470\mathrm{g}\) of leaves were harvested and and ground in a blender with \(900~\mathrm{ml}\) of buffer containing \(0.4\mathrm{M}\) sucrose, Tricine- NaOH (pH 8), \(15\mathrm{mMNaCl}\) , \(30\mathrm{mM}\) , \(2\mathrm{mM}\) ascorbic acid, \(1\mathrm{mM}\) PMSF, \(1\mu \mathrm{M}\) pepstatin A. After filtration through cheesecloth, the suspension was centrifuged at \(6,000\mathrm{g}\) for \(10\mathrm{min}\) , and again at \(180,000\mathrm{g}\) for \(20\mathrm{min}\) . The pellet was resuspended in \(200\mathrm{ml}\) buffer containing \(10\mathrm{mM}\) Tricine- NaOH (pH 8), \(150\mathrm{mMNaCl}\) and pelleted through centrifugation of \(180,000\mathrm{g}\) for \(20\mathrm{min}\) . The pellet was resuspended in \(45\mathrm{ml}\) of buffer containing \(10\mathrm{mM}\) Tricine- NaOH (pH 8), \(0.4\mathrm{M}\) sucrose, to a concentration of \(1\mathrm{mg / ml}\) chlorophyll and solubilized with \(1.5\%\) n- dodecyl- \(\beta\) - D- maltoside (DDM). Following \(30\mathrm{min}\) incubation on ice, the material was centrifuged at \(176,000\) for \(20\mathrm{min}\) and applied on DEAE- cellulose column \((2.5\times 13\mathrm{cm})\) pre- equilibrated with \(20\mathrm{mM}\) Tricine- Tris (pH 8.0), \(0.2\%\) DDM. The material was eluted with \(300\mathrm{mMNaCl}\) , concentrated twice with PEG 6000 precipitation and centrifuged through a sucrose gradient of \(10\) to \(40\%\) in SW40 rotor (Beckman, Inc) at \(170,000\mathrm{g}\) for \(16\) hours. The green band containing PSI was collected, subjected to FPLC chromatography, and then applied on to a \(10 - 35\%\) sucrose gradient and centrifuged at \(336,000\mathrm{g}\) for \(4\) hours in SW- 60 rotor (Beckman, Inc). The green band was collected, and sucrose was removed by buffer exchange.  
+
+<|ref|>text<|/ref|><|det|>[[113, 800, 850, 870]]<|/det|>
+Mature PSI was prepared from 7- day old A. sativa (var. Saja 6) grown for 7 days in \(16 / 8\) light/dark cycle at a photon flux density \(50\mu \mathrm{E}\mathrm{m}^{- 2}\mathrm{s}^{- 1}\) and followed the protocol published before \(^{8}\) .
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 90, 770, 110]]<|/det|>
+167 Kinetic measurements, P700 reduction, and NADP+ photo- reduction activity assay  
+
+<|ref|>text<|/ref|><|det|>[[111, 115, 880, 362]]<|/det|>
+168 PC was heterologously expressed, and codon- optimized; the following experiments were performed generally as previously described 17. P700 reduction was measured in a Quartz cuvette, containing: 1 ml reaction mix (20 mM Tricine- NaOH pH 8, 5 mM \(\mathrm{MgCl}_2\) and \(0.05\%\) n- 171 Dodecyl \(\beta\) - maltoside), 10 \(\mu\) mol Ascorbate, 100 nmol Methyl viologen, 16 \(\mu\) g chlorophyll PSI, 172 and 50 pmol Pc of Synechocystis PCC SP 6803 or 50 pmol Cyt \(\mathrm{c}_6\) (Cyt C533). P700 photo- 173 oxidation and re- reduction by Pc and Cyt \(\mathrm{c}_6\) was measured by JTS- 10 spectrophotometer by 174 illuminating the sample with red light (705 nm) for 5 s. Changes in absorbance were measured 175 by 2 ms of LED light flashes (700 nm passed through a 705 nm interference filter).  
+
+<|ref|>text<|/ref|><|det|>[[111, 366, 870, 475]]<|/det|>
+177 NADP+ photo- reduction activity assay was measured in a Quartz cuvette, 1 ml reaction mix (20 mM NaCl, 10 mM Tricine NaOH pH 8, 0.5 mM \(\mathrm{MgCl}_2\) ) was supplemented with 20 \(\mu\) mol 178 ascorbate, 125 \(\mu\) g Ferredoxin, 8.8 \(\mu\) g FNR, 1 \(\mu\) mol NADP+ (Roche Diagnostics), 14 nmol Pc, 179 and PSI (14.4 \(\mu\) g chlorophyll). NADPH accumulation was measured at 340 nm ( \(\epsilon = 6220 \mathrm{M}^{- 1}\) 180 \(\mathrm{cm}^{- 1}\) ) by Cary 60 spectrophotometer (Agilent technologies) under continuous illumination with 181 660 nm LED light (600 \(\mu\) E). The activity was calculated as \(\mu\) mol NADPH/(mg chl- Hr).  
+
+<|ref|>text<|/ref|><|det|>[[112, 510, 571, 529]]<|/det|>
+183 Cryo- EM data collection, processing, and model building  
+
+<|ref|>text<|/ref|><|det|>[[111, 534, 888, 896]]<|/det|>
+184 3 \(\mu\) l of pre- PSI- 1 and mature PSI at 2 mg/ml chlorophyll were applied on glow- discharged holey 185 carbon grids and vitrified for cryo- EM structural determination using Leica- EM- GP (3 sec blot at 186 \(20^{\circ}\mathrm{C}\) and \(90\%\) humidity). Data were collected on a 300- kV Titan Krios G3 Microscope (Thermo 187 Fisher Scientific) equipped with a Gatan Bioquantum energy filter and a K3 Summit direct electron 188 detector (Ametek). Movies were recorded using counting mode at a magnification of \(\times 105,000\) 189 corresponding to a calibrated pixel size of 0.84- 0.85 Å. 3,324 micrographs at a total dose of 45 190 \(\mathrm{e} / \mathrm{\AA}^2\) and 24,930 micrographs at 51 \(\mathrm{e} / \mathrm{\AA}^2\) were collected for pre- PSI- 1 and mature PSI, respectively, 191 with a defocus range from - 0.5 to - 1.9 \(\mu\) m. Movies were imported into cryoSPARC 3.1 18, motion 192 correction, CTF estimation, picking, and 2D classification was performed on the fly during data 193 collection using cryoSPARC live (with blob picker and template picker). Ab initio models were 194 generated with a subset of particles. Heterogenous and homogeneous refinement was performed 195 for pre- PSI- 1 and mature PSI respectively. Particles (383,325 for mature PSI and 546,410 for pre- 196 PSI- 1) were converted into a Star file format and were imported into RELION 3.1.1 19. Particles 197 were re- extracted (un- binned) and processed in RELION using a box size of 400 pixels for pre
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 88, 886, 372]]<|/det|>
+PS- 1 and 500 pixels for mature PSI. For pre- PSI- 1, 3D classification with 2 classes was performed. One class with 169,213 particles of high- quality particles was selected and subjected to 3D refinement which resulted in an overall resolution of 3.1 Å. CTF refinement, 3D refinement and Bayesian polishing followed by another round of CTF refinement was performed for pre- PSI- 1 and mature PSI. Another 3D refinement resulted in an overall resolution of 2.1 Å for the pre- PSI- 1 and mature PSI. Since the Lhca2- 3 heterodimer in the mature PSI appeared to be loosely bound, we used focused classification (3 classes) with signal subtraction with a mask around the Lhca2- 3 region to improve the local density. One subclass showed a better ordered Lhca2- 3 region. The particles of this subclass (96,997) were selected, and the signal was reverted. A final 3D refinement of mature PSI resulted in an overall resolution of 2.2 Å with an improved density for the Lhca2- 3 region.  
+
+<|ref|>text<|/ref|><|det|>[[111, 376, 886, 607]]<|/det|>
+Model building and real- space refinement of pre- PSI- 1 was then carried out using Coot 9.1.4 20. The completed model was then fitted into the mature PSI map, and the remaining protein chains were built using rigid body fitted PSI (PDB ID: 6YAC) as a starting model. All protein residues as well as pigments were fitted using Coot with locally optimized map weights. For the entire modelling in Coot restraint files for pigments and ligands were used which were generated using the Grade server (http://grade.globalphasing.org), as previously described 21. Models were refined using Real- Space- Refine from the PHENIX suite 22. The refinement protocol was optimized by adjusting for optimal refinement weight parameters. Iterations of validation, model building, and refinement were carried out using MolProbity 4.2 23, Coot 9.1.4 20 and PHENIX 22.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 642, 258, 660]]<|/det|>
+## Data availability  
+
+<|ref|>text<|/ref|><|det|>[[111, 664, 886, 840]]<|/det|>
+Atomic coordinates and structure factors of pre- PSI- 1 have been deposited in the Protein Data Bank under accession code 8BCW. Atomic coordinates and structure factors of mature PSI have been deposited in the Protein Data Bank under accession code 8BCW. The cryo- EM map of pre- PSI- 1 has been deposited in the Electron Microscopy Data Bank under accession code EMD- 15970. The cryo- EM map of mature PSI has been deposited in the Electron Microscopy Data Bank under accession code EMD- 15969. Other atomic coordinates that were used in this study: 6YAC [https://www.rcsb.org/structure/6YAC] (PSI- ferredoxin).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[67, 91, 273, 109]]<|/det|>
+## Acknowledgments  
+
+<|ref|>text<|/ref|><|det|>[[66, 115, 884, 220]]<|/det|>
+This work was supported by the Swedish Foundation for Strategic Research (FFL15:0325, ARC19- 0051), European Research Council (ERC- 2018- StG- 805230), Knut and Alice Wallenberg Foundation (2018.0080), EMBO Young Investigator Program. The cryo- EM facility is funded by the Knut and Alice Wallenberg, Family Erling Persson, and Kempe foundations.  
+
+<|ref|>sub_title<|/ref|><|det|>[[67, 252, 393, 271]]<|/det|>
+## Author Contributions Statement  
+
+<|ref|>text<|/ref|><|det|>[[66, 280, 886, 405]]<|/det|>
+N.N. and A.A. designed the project. M.F., D.K. and A.B-S. prepared the sample for cryo- EM. A.N. collected and processed the cryo- EM data and built the model. M.F., D.K., A.B-S. and I.C. performed biochemical and kinetic analysis. A.N., N.N. and A.A. analysed the structure and wrote the manuscript with contributions from M.F. All authors contributed to the analysis and the final version of the manuscript.  
+
+<|ref|>sub_title<|/ref|><|det|>[[67, 448, 380, 466]]<|/det|>
+## Competing Interests Statement  
+
+<|ref|>text<|/ref|><|det|>[[67, 478, 460, 496]]<|/det|>
+The authors declare no competing interests.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 88, 880, 475]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 475, 808, 496]]<|/det|>
+<center>Supplementary Fig. 1 Data processing, Fourier Shell Correlations (FSCs), angular </center>  
+
+<|ref|>text<|/ref|><|det|>[[111, 500, 870, 520]]<|/det|>
+distributions. a, Data processing scheme for mature PSI on the left and pre- PSI- 1 on the right.  
+
+<|ref|>text<|/ref|><|det|>[[111, 526, 870, 547]]<|/det|>
+b, Angular distribution and FSC for resolution estimation for pre- PSI- 1 (top) and of mature PSI  
+
+<|ref|>text<|/ref|><|det|>[[111, 553, 193, 570]]<|/det|>
+(bottom).
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[113, 152, 902, 899]]<|/det|>
+<|ref|>table_caption<|/ref|><|det|>[[63, 92, 884, 151]]<|/det|>
+262 Supplementary Table 1 263 Cryo-EM data collection, refinement and validation statistics of PSI and pre-PSI-1 assembly 264 intermediate of Avena sativa.   
+
+<table><tr><td></td><td>PSI<br>(EMDB-15969)<br>(PDB 8BCV)</td><td>pre-PSI-1<br>(EMDB-15970)<br>(PDB 8BCW)</td></tr><tr><td>Data collection and processing</td><td></td><td></td></tr><tr><td>Magnification</td><td>105,000</td><td>105,000</td></tr><tr><td>Voltage (kV)</td><td>300</td><td>300</td></tr><tr><td>Electron exposure (e-/Å2)</td><td>45.00</td><td>51.35</td></tr><tr><td>Defocus range (μm)</td><td>-0.5-2.1</td><td>-0.5-2.1</td></tr><tr><td>Pixel size (Å)</td><td>0.85</td><td>0.84</td></tr><tr><td>Initial particle images (no.)</td><td>24,930</td><td>3,324</td></tr><tr><td>Symmetry imposed</td><td>C1</td><td>C1</td></tr><tr><td>Final particle images (no.)</td><td>96,997</td><td>169,213</td></tr><tr><td>Map resolution (Å)</td><td rowspan="2">2.20</td><td rowspan="2">2.11</td></tr><tr><td>FSC threshold 0.143</td></tr><tr><td>Refinement</td><td></td><td></td></tr><tr><td>Initial model used (PDB code)</td><td>6YAC</td><td>6YAC</td></tr><tr><td>Map sharpening B factor (Å2)</td><td>-17.41</td><td>-21.56</td></tr><tr><td>Model composition</td><td></td><td></td></tr><tr><td>Non-hydrogen atoms</td><td>37,485</td><td>22,313</td></tr><tr><td>Protein residues</td><td>3,258</td><td>2,029</td></tr><tr><td>Ligands</td><td>224</td><td>114</td></tr><tr><td>Waters</td><td>703</td><td>517</td></tr><tr><td>B factors (Å2)</td><td></td><td></td></tr><tr><td>(min/max/mean)</td><td></td><td></td></tr><tr><td>Protein</td><td>8.32/93.7/39.3</td><td>14.1/76.9/34.1</td></tr><tr><td>Ligand</td><td>10.3/87.8/38.4</td><td>15.1/64.3/32.6</td></tr><tr><td>Waters</td><td>9.8/80.1/28.7</td><td>12.0/55.0/32.7</td></tr><tr><td>R.m.s. deviations</td><td></td><td></td></tr><tr><td>Bond lengths (Å)</td><td>0.008</td><td>0.008</td></tr><tr><td>Bond angles (°)</td><td>1.059</td><td>1.032</td></tr><tr><td>Validation</td><td></td><td></td></tr><tr><td>MolProbity score</td><td>1.13</td><td>1.04</td></tr><tr><td>Clashscore</td><td>3.43</td><td>2.56</td></tr><tr><td>Poor rotamers (%)</td><td>0.64</td><td>0.54</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[112, 90, 914, 168]]<|/det|>
+<table><tr><td>Ramachandran plot</td><td></td><td></td></tr><tr><td>Favored (%)</td><td>98.67</td><td>98.11</td></tr><tr><td>Allowed (%)</td><td>1.33</td><td>1.89</td></tr><tr><td>Disallowed (%)</td><td>0.00</td><td>0.00</td></tr></table>
+
+<|ref|>title<|/ref|><|det|>[[115, 207, 213, 223]]<|/det|>
+265 
+
+<|ref|>title<|/ref|><|det|>[[115, 209, 213, 223]]<|/det|>
+266 
+
+<|ref|>title<|/ref|><|det|>[[115, 210, 213, 223]]<|/det|>
+267 
+
+<|ref|>title<|/ref|><|det|>[[115, 228, 213, 243]]<|/det|>
+268 
+
+<|ref|>text<|/ref|><|det|>[[115, 245, 848, 279]]<|/det|>
+269 1. M. Chen et al., Distinct structural modulation of photosystem I and lipid environment stabilizes its tetrameric assembly. Nat Plants 6, 314-320 (2020). 
+
+<|ref|>text<|/ref|><|det|>[[115, 280, 839, 313]]<|/det|>
+270 2. A. Naschberger et al., Algal photosystem I dimer and high-resolution model of PSI-plastocyanin complex. Nat Plants 8, 1191-1201 (2022). 
+
+<|ref|>text<|/ref|><|det|>[[115, 313, 880, 347]]<|/det|>
+273 3. A. Amunts, O. Drory, N. Nelson, The structure of a plant photosystem I supercomplex at 3.4 A resolution. Nature 447, 58-63 (2007). 
+
+<|ref|>text<|/ref|><|det|>[[115, 348, 880, 382]]<|/det|>
+275 4. A. Amunts, N. Nelson, Plant photosystem I design in the light of evolution. Structure 17, 637-650 (2009). 
+
+<|ref|>text<|/ref|><|det|>[[115, 383, 808, 416]]<|/det|>
+276 5. A. Amunts, H. Toporik, A. Borovikova, N. Nelson, Structure determination and improved model of plant photosystem I. J Biol Chem 285, 3478-3486 (2010). 
+
+<|ref|>text<|/ref|><|det|>[[115, 417, 880, 485]]<|/det|>
+279 6. I. Karnauchov, D. Cai, I. Schmidt, R. G. Herrmann, R. B. Klosgen, The thylakoid translocation of subunit 3 of photosystem I, the psaF gene product, depends on a bipartite transit peptide and proceeds along an azide-sensitive pathway. J Biol Chem 269, 32871-32878 (1994). 
+
+<|ref|>text<|/ref|><|det|>[[115, 486, 870, 555]]<|/det|>
+283 7. M. Hippler, F. Drepper, W. Haehnel, J. D. Rochaix, The N-terminal domain of PsaF: precise recognition site for binding and fast electron transfer from cytochrome c6 and plastocyanin to photosystem I of Chlamydomonas reinhardtii. Proc Natl Acad Sci U S A 95, 7339-7344 (1998). 
+
+<|ref|>text<|/ref|><|det|>[[115, 556, 884, 606]]<|/det|>
+287 8. I. Caspy, A. Borovikova-Sheinker, D. Klaiman, Y. Shkolnisky, N. Nelson, The structure of a triple complex of plant photosystem I with ferredoxin and plastocyanin. Nat Plants 6, 1300-1305 (2020). 
+
+<|ref|>text<|/ref|><|det|>[[115, 607, 884, 641]]<|/det|>
+290 9. R. Nechushtai, N. Nelson, Biogenesis of photosystem I reaction center during greening of oat, bean and spinach leaves. Plant Mol Biol 4, 377-384 (1985). 
+
+<|ref|>text<|/ref|><|det|>[[115, 642, 848, 676]]<|/det|>
+292 10. S. Nellaepalli, S. I. Ozawa, H. Kuroda, Y. Takahashi, The photosystem I assembly apparatus consisting of Ycf3-Y3IP1 and Ycf4 modules. Nat Commun 9, 2439 (2018). 
+
+<|ref|>text<|/ref|><|det|>[[115, 677, 800, 710]]<|/det|>
+294 11. G. Wittenberg et al., Identification and characterization of a stable intermediate in photosystem I assembly in tobacco. Plant J 90, 478-490 (2017). 
+
+<|ref|>text<|/ref|><|det|>[[115, 711, 845, 745]]<|/det|>
+296 12. W. Rudiger, J. Benz, C. Guthoff, Detection and partial characterization of activity of chlorophyll synthetase in etioplast membranes. Eur J Biochem 109, 193-200 (1980). 
+
+<|ref|>text<|/ref|><|det|>[[115, 746, 830, 780]]<|/det|>
+298 13. A. Perez-Boerema et al., Structure of a minimal photosystem I from the green alga Dunaliella salina. Nat Plants 6, 321-327 (2020). 
+
+<|ref|>text<|/ref|><|det|>[[115, 781, 877, 815]]<|/det|>
+300 14. J. Wang et al., Structure of plant photosystem I-light harvesting complex I supercomplex at 2.4 A resolution. J Integr Plant Biol 63, 1367-1381 (2021). 
+
+<|ref|>text<|/ref|><|det|>[[115, 816, 839, 850]]<|/det|>
+302 15. Y. Mazor, A. Borovikova, I. Caspy, N. Nelson, Structure of the plant photosystem I Supercomplex at 2.6 Å Resolution. Nature Plants 3, 17014 (2017). 
+
+<|ref|>text<|/ref|><|det|>[[115, 851, 884, 902]]<|/det|>
+304 16. T. Malavath, I. Caspy, SY. Netzer-El, D. Klaiman, N. Nelson, Structure and function of wild-type and subunit-depleted photosystem I in synechocystis. Biochimica et Biophysica Acta (BBA) - Bioenergetics 1859, 645-654 (2018).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[61, 90, 883, 336]]<|/det|>
+307 17. SY. Netzer- El, I. Caspy, N. Nelson N, Crystal structure of photosystem I monomer from Synechocystis PCC 6803. Front Plant Sci 9, 1- 7 (2018).308 18. A. Punjani, J. Rubinstein, D. Fleet, M. Brubaker, CryoSPARC: algorithms for rapid unsupervised cryo- EM structure determination. Nat. Methods 14, 290- 296 (2017).309 19. J. Zivanov, T. Nakane, S.H. Scheres, Estimation of high- order aberrations and anisotropic magnification from cryo- EM data sets in RELION- 3.1. IUcrJ 7, 253- 267 (2020).310 20. P. Emsley, B. Lohkamp, W.G. Scott, K. Cowtan. Features and development of Coot. Acta Crystallogr. D 66, 486- 501 (2010).311 21. A. Amunts, The revolution evolution. Nat Plants 8, 14- 17 (2022).312 22. P. Afonine et al., Real- space refinement in phenix for cryo- EM and crystallography. Acta Crystallogr D Struct Biol. 74, 531- 544 (2018).313 23. C. Williams et al., Molprobity: more and better reference data for improved all- atom structure validation. Protein Sci. 27, 293- 315 (2017).
+
+<--- Page Split --->

preprint/preprint__2aba0cad8c9e21f18a62e920b616ed2495a0f2c186dc941277f572951cc0528c/preprint__2aba0cad8c9e21f18a62e920b616ed2495a0f2c186dc941277f572951cc0528c_det.mmd

@@ -0,0 +1,500 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 881, 209]]<|/det|>
+# Sequence-specific capture and concentration of viral RNA by type III CRISPR system enhances diagnostic  
+
+<|ref|>text<|/ref|><|det|>[[42, 225, 640, 952]]<|/det|>
+Anna Nemudraia  Montana State University  Artem Nemudryi  Montana State University  Murat Buyukyoruk  Montana State University  Andrew Scherffius  Montana State University  Trevor Zahl  Montana State University  Tanner Wiegand  Montana State University  https://orcid.org/0000- 0002- 0528- 268X  Shishir Pandey  Montana State University  Joseph Nichols  Montana State University  Laina Hall  Montana State University  Aidan McVey  Montana State University  Helen Lee  Montana State University  Royce Wilkinson  Montana State University  https://orcid.org/0000- 0001- 8831- 2081  Laura Snyder  University of Michigan  Joshua Jones  University of Michigan  Kristin Koutmou  https://orcid.org/0000- 0002- 7763- 9262  Andrew Santiago- Frangos
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[45, 46, 636, 111]]<|/det|>
+Montana State University https://orcid.org/0000- 0001- 9615- 065XBlake Wiedenheft ( \(\boxed{ \begin{array}{r l} \end{array} }\) bwiedenheft@gmail.com)Montana State University https://orcid.org/0000- 0001- 9297- 5304  
+
+<|ref|>sub_title<|/ref|><|det|>[[45, 153, 102, 170]]<|/det|>
+## Article  
+
+<|ref|>title<|/ref|><|det|>[[45, 191, 135, 209]]<|/det|>
+# Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[45, 228, 299, 247]]<|/det|>
+Posted Date: April 19th, 2022  
+
+<|ref|>text<|/ref|><|det|>[[45, 266, 474, 285]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 1466718/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 303, 910, 346]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[72, 88, 872, 140]]<|/det|>
+# Sequence-specific capture and concentration of viral RNA by type III CRISPR system enhances diagnostic  
+
+<|ref|>text<|/ref|><|det|>[[112, 163, 877, 255]]<|/det|>
+Anna Nemudraia<sup>1,2</sup>, Artem Nemudryi<sup>1,2</sup>, Murat Buyukyoruk<sup>1,3</sup>, Andrew M. Scherffius<sup>1,3</sup>, Trevor Zahl<sup>1</sup>, Tanner Wiegand<sup>1</sup>, Shishir Pandey<sup>1</sup>, Joseph E. Nichols<sup>1</sup>, Laina Hall<sup>1</sup>, Aidan McVey<sup>1</sup>, Helen H Lee<sup>1</sup>, Royce A. Wilkinson<sup>1</sup>, Laura R. Snyder<sup>4</sup>, Joshua D. Jones<sup>4</sup>, Kristin S. Koutmou<sup>4</sup>, Andrew Santiago- Frangos<sup>1\*</sup>, and Blake Wiedenheft<sup>1,5\*</sup>  
+
+<|ref|>text<|/ref|><|det|>[[112, 283, 875, 320]]<|/det|>
+<sup>1</sup>Department of Microbiology and Cell Biology, Montana State University, Bozeman, MT 59717, USA  
+
+<|ref|>text<|/ref|><|det|>[[112, 330, 421, 349]]<|/det|>
+<sup>2</sup>These authors contributed equally  
+
+<|ref|>text<|/ref|><|det|>[[112, 359, 421, 377]]<|/det|>
+<sup>3</sup>These authors contributed equally  
+
+<|ref|>text<|/ref|><|det|>[[112, 388, 789, 408]]<|/det|>
+<sup>4</sup>Department of Chemistry, University of Michigan, Ann Arbor, MI 48105, USA  
+
+<|ref|>text<|/ref|><|det|>[[112, 417, 238, 435]]<|/det|>
+<sup>5</sup>Lead contact  
+
+<|ref|>text<|/ref|><|det|>[[112, 446, 864, 465]]<|/det|>
+<sup>5</sup>Correspondence: andrew.santiagofrangos@gmail.com and bwiedenheft@gmail.com.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[60, 90, 880, 444]]<|/det|>
+17 Abstract18 Type- III CRISPR- Cas systems have recently been adopted for sequence- specific detection of SARS- CoV- 2. Here, we make two major advances that simultaneously limit sample handling and significantly enhance the sensitivity of SARS- CoV- 221 RNA detection directly from patient samples. First, we repurpose the type III- A CRISPR complex from Thermus thermophilus (TtCsm) for programmable capture and concentration of specific RNAs from complex mixtures. The target bound TtCsm complex primarily generates two cyclic oligoadenylates (i.e., cA3 and cA4) that allosterically activate ancillary nucleases. To improve sensitivity of the diagnostic, we identify and test several ancillary nucleases (i.e., Can1, Can2, and NucC). We show that Can1 and Can2 are activated by both cA3 and cA4, and that different activators trigger changes in the substrate specificity of these nucleases. Finally, we integrate the type III- A CRISPR RNA- guided capture technique with the Can2 nuclease for 90 fM (5x10⁴ copies/ul) detection of SARS- CoV- 2 RNA directly from nasopharyngeal swab samples.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[113, 90, 232, 108]]<|/det|>
+## Introduction  
+
+<|ref|>text<|/ref|><|det|>[[111, 123, 882, 348]]<|/det|>
+Although qPCR (quantitative polymerase chain reaction) remains the "gold standard" for nucleic acid detection, it requires sophisticated equipment, trained personnel, efficient specimen transport to high- complexity labs, and reliable reporting systems<sup>1</sup>. While the complexity and turnaround times necessary for qPCR are acceptable for many diagnostic applications, the SARS- CoV- 2 (Severe Acute Respiratory Syndrome Coronavirus 2) pandemic reveals an urgent need for diagnostics that are easy to distribute, simple to perform, and fast enough to stop transmission of a contagious disease<sup>1</sup>. Although rapid antigen tests and isothermal amplification methods have helped address this need, these and other emerging methods have limitations related to sensitivity, versatility, or specificity<sup>2,3</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[111, 364, 880, 771]]<|/det|>
+CRISPR RNA- guided diagnostics (CRISPR- dx) are a diverse group of nascent technologies that aim to address current limitations by providing a versatile and programmable platform that is sufficiently sensitive for clinical applications and stable enough for distribution<sup>4,5</sup>. The first CRISPR- based viral diagnostic came from Collins and colleagues in 2016, when they demonstrated that Cas9 could be used to discriminate between different variants of the Zika virus<sup>6</sup>. This approach relies on converting viral RNA to DNA using reverse transcriptase, followed by isothermal DNA amplification prior to sequence- based discrimination by Cas9. The exclusive recognition of double- stranded DNA (dsDNA) by Cas9 seemed to be an intrinsic limitation for diagnostic applications that require RNA detection. However, Beisel and colleagues recently developed a creative method that uses the trans- acting CRISPR- RNA (tracrRNA) to capture complementary RNA guides derived from RNA viruses<sup>7</sup>. In this system, the engineered tracrRNA- crRNA hybrid guides Cas9 to a complementary dsDNA reporter. While this approach enables RNA detection, Cas9 is a single turn- over enzyme, which may limit its sensitivity. In contrast to Cas9, target recognition by type V (Cas12- DETECTR) and type VI (Cas13- SHERLOCK) CRISPR- systems activates a multi- turnover non- sequence- specific "collateral nuclease" activity that amplifies the signal by cleaving thousands of reporter molecules for every target bound<sup>8,9</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[111, 788, 882, 898]]<|/det|>
+Like type VI, type III systems also recognize complementary RNA. However, unlike any other CRISPR system, target recognition by type III complexes simultaneously activates polymerase and HD- nuclease domains in the Cas10 subunit<sup>10–12</sup>. The polymerase domain has been estimated to generate \(\sim 1000\) cyclic oligoadenylates per bound RNA<sup>13</sup>, which trans- activate and allosterically regulate diverse multi- turnover ancillary
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 879, 313]]<|/det|>
+nucleases that provide defense from invading genetic parasites<sup>14,15</sup>. This biochemical cascade exponentially amplifies the signal when a type III complex detects target RNA, suggesting that these systems have the potential to enhance the sensitivity of CRISPR- based diagnostics. However, initial efforts to implement this approach failed to be sufficiently sensitive for clinical applications without prior amplification of the target RNA<sup>16-18</sup>. The sensitivity of this first- generation diagnostic was in part limited by the use of Csm6 ancillary nucleases that also degrade the cyclic nucleotide activator<sup>19-23</sup>. Recently, Malcolm White's lab demonstrated that alternative ancillary nucleases, which efficiently cleave reporters but do not cleave the signaling molecule, can be used to enhance the sensitivity of type III- based diagnostics<sup>24</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[110, 330, 884, 757]]<|/det|>
+Despite innovations leading to new and improved CRISPR- based diagnostics, point- of- care testing requires new strategies that simplify the workflow and increase the sensitivity without prior RNA purification or amplification (e.g., PCR, LAMP, NASB, RPA, etc.). Here, we bring CRISPR- dx closer to a deployable diagnostic by developing a type III CRISPR- based method for sequence- specific capture and concentration of RNA from heterogeneous samples. To improve the sensitivity, we purify several different ancillary nucleases (i.e., Can1, Can2, and NucC), systemically test nuclease activation using a series of purified cyclic oligoadenylate standards (i.e., cA3- cA6), test for ring nuclease activity and determine how cyclic oligoadenylates, as well as metal- preferences impact substrate cleavage activities. We show that the Can1 nuclease from T. thermophilus (TtCan1) and the Can2 ortholog from Archaeoglobi archaeon JdFR- 42 (AaCan2) are activated by more than one cyclic nucleotide species (i.e., cA3 and cA4) and that substrate specificity of these nucleases changes according to the bound activator. This observation helps to explain how diverse cyclic nucleotides (i.e., cA3- cA6) produced by a single type III surveillance complex integrate distinct activities from a single effector. Finally, we demonstrate how the type III complex can be used to bypass RNA extraction methods, and that coupling type III- based RNA capture with the AaCan2 nuclease further increases the sensitivity of SARS- CoV- 2 RNA detection in patient swabs to \(5 \times 10^{4}\) copies/ul.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[115, 90, 189, 108]]<|/det|>
+## Results  
+
+<|ref|>text<|/ref|><|det|>[[115, 128, 605, 149]]<|/det|>
+Type III- mediated sequence- specific enrichment of RNA  
+
+<|ref|>text<|/ref|><|det|>[[110, 160, 880, 712]]<|/det|>
+Type III CRISPR RNA- guided complexes (i.e., Csm and Cmr) bind and cleave complementary single- stranded RNA (ssRNA) targets<sup>25</sup>. Complementary RNA is cleaved in six- nucleotide increments by metal- dependent nucleases (Csm3 or Cmr4) that form the oligomeric "backbone" of the complex<sup>26</sup>. Type III complexes release fragments of the cleaved target, which inactivates ATP polymerization by the Cas10 subunit<sup>26</sup>. Previously, we mutated residues in the Csm3 subunit responsible for target RNA cleavage (D34A), purified the RNase- dead complex (TtCsm<sup>Csm3-D34A</sup>), and showed that the mutant complex provides more sensitive detection of viral RNA than the wild- type complex<sup>16</sup>. To further increase the sensitivity, we set out to determine if TtCsm<sup>Csm3-D34A</sup> could be used to concentrate sequence- specific RNAs. To test this approach, we mixed <sup>32</sup>P- labeled target or non- target RNAs with TtCsm<sup>Csm3-D34A</sup>, incubated for 20 minutes, and concentrated the His- tagged complex using nickel- derivatized magnetic beads (Fig. 1a, Supplementary Fig. 1a). The beads were concentrated using a magnet, and RNAs were extracted from the bound and unbound fractions. The type III complex captured most of the radiolabeled target RNA (76±5.8%), while non- target RNA primarily remains in the supernatant (Fig. 1b, Supplementary Fig. 1b, c). To determine if type III CRISPR- based RNA capture and concentration results in the synthesis of more cyclic nucleotides, we mixed Csm- beads with 120 μL of a sample containing SARS- CoV- 2 RNA and total RNA extracted from HEK 293T cells (Fig. 1c, see Methods). After concentrating the beads with a magnet, we resuspended the pellet in a buffer containing α- <sup>32</sup>P- ATP, allowed the cyclic polymerization to proceed, and analyzed the reactions using thin- layer chromatography (TLC). The type III CRISPR- based concentration increases the amount of cA<sub>3</sub> and cA<sub>4</sub>, as compared to the reaction performed without RNA concentration (Fig. 1c, d, Supplementary Fig. 1d).  
+
+<|ref|>text<|/ref|><|det|>[[111, 725, 881, 905]]<|/det|>
+Previously, we repurposed TtCsm6, a cA4- activated ribonuclease, to generate a real- time fluorescent readout for Csm- based RNA detection<sup>16</sup> (Fig. 1e, top). We reasoned that increased cA4 levels after RNA enrichment will boost the nuclease activity of TtCsm6 and therefore increase the sensitivity of the RNA detection. To test this hypothesis, we titrated 10<sup>8</sup> to 10<sup>5</sup> copies/μL of SARS- CoV- 2 N- gene RNA into total RNA extracted from HEK 293T cells, concentrated the target RNA using TtCsm<sup>Csm3-D34A</sup>, resuspended the beads in a buffer containing ATP, and then transferred the polymerization products to a reaction containing TtCsm6 and a fluorescent RNA
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[112, 88, 870, 200]]<|/det|>
+reporter (i.e., FAM- RNA- lowa Black FQ). Csm- based RNA enrichment increased the sensitivity of the assay 100- fold compared to the assay without the pull- down (Fig. 1e). Taken together, these results demonstrate how type III- A CRISPR- complexes can be used to capture sequence- specify RNAs, resulting in a higher concentration of cyclic nucleotides, which improves the sensitivity of sequence- specific RNA detection.  
+
+<|ref|>text<|/ref|><|det|>[[112, 216, 825, 238]]<|/det|>
+CARF- nucleases Can1 and Can2 exhibit \(cA_{3}\) - and \(cA_{4}\) - specific nuclease activities  
+
+<|ref|>text<|/ref|><|det|>[[112, 254, 875, 435]]<|/det|>
+Csm6 proteins contain an amino- terminal CARF (CRISPR- associated Rossman Fold) and a carboxy- terminal HEPN (Higher Eukaryotes and Prokaryotes Nucleotide- binding) domains \(^{10,12}\) . Csm6 family proteins form homodimers, and the two CARF- domains bind \(cA^{23,27}\) or \(cA^{22}\) , which activate the C- terminal HEPN nuclease domain. However, the CARF domain of some Csm6 proteins also degrades the cyclic nucleotide, which inactivates the nuclease and may limit the sensitivity of Csm6- based assays \(^{28}\) . To improve the sensitivity, we sought to identify and incorporate a CARF- nuclease that is activated by but does not degrade \(cA_{4}\) .  
+
+<|ref|>text<|/ref|><|det|>[[112, 450, 875, 630]]<|/det|>
+CRISPR ancillary nucleases (Can) are another family of recently identified proteins that are activated by cyclic oligoadenylates and lack ring nuclease activity \(^{29 - 31}\) . Like Csm6 proteins, Can proteins also contain amino- terminal CARF domains, but the carboxy- terminal nucleases are distinct. The Can1 protein from Thermus thermophilus (TtCan1) has a unique monomeric architecture with two non- identical CARF domains, one nuclease- like domain (NLD) and one restriction endonuclease domain (PD- (D/E)XK) \(^{31}\) , while Can2 nucleases contain a single CARF domain and form symmetrical homodimers \(^{29,30}\) (Fig 2a).  
+
+<|ref|>text<|/ref|><|det|>[[112, 646, 881, 895]]<|/det|>
+To identify Can1 and Can2 orthologs compatible with the TtCsm complex, we generated profile Hidden Markov models (HMMs) to query publicly available microbial genomes and metagenomes from NCBI and JGI. This analysis identified 204 Can1 and 3,121 Can2 proteins. Based on this analysis, we selected TtCan1 and three Can2 orthologs from thermophilic organisms for cloning and expression (Fig. 2b). While previous research demonstrated that metal- dependent nicking of supercoiled DNA by TtCan1 is dependent on activation by \(cA_{4}^{31}\) , the impact of other cyclic oligoadenylates on TtCan1 activity has not been reported. We purified TtCan1 and tested nuclease activity against plasmid DNA in the presence of five different cyclic oligoadenylates (cA2- cA6) (Supplementary Fig. 2a- c). To our surprise, TtCan1 robustly degrades plasmid DNA to \(\sim 100\) bp fragments in the presence of \(cA_{3}\) and Mn \(^{2 + }\) , while cleavage with \(cA_{4}\) is
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 876, 335]]<|/det|>
+comparable to the background activity in the absence of an activator (Fig. 2c, left; Supplementary Fig. 2d). To determine if the TtCan1 nuclease has any sequence preference, we deep- sequenced the cleavage fragments, aligned the reads, and identified cut sites. This analysis failed to identify common sequence motifs that define the cleavage site, suggesting that TtCan1 is a non- sequence specific DNase (Supplementary Fig. 2e). Based on the unexpected activation of TtCan1 with cA3, we tested several other substrates and discovered that TtCan1 is a cA4- dependent single- stranded RNase (ssRNA) but does not cleave ssDNA (Fig. 2c, Supplementary Fig. 2f, g). Taken together, our in vitro assays show that TtCan1 is a non- sequence specific double- stranded DNase when activated with cA3 and a single- stranded RNase when activated with cA4.  
+
+<|ref|>text<|/ref|><|det|>[[111, 352, 884, 759]]<|/det|>
+Can2 genes from Clostridium thermobutyricum (CthCan2), Thermus thermophilus (TtCan2), and Archaeoglobi archaeon JdFR- 42 (AaCan2) were cloned and expressed in E. coli (Fig. 2b). However, only AaCan2 purified in quantities sufficient for biochemical assays (Supplementary Fig. 3a, b). We systematically tested the activities of AaCan2 against different substrates with a range of cyclic oligoadenylates (Supplementary Fig. 3c, d). Like TtCan1, AaCan2 is also a Mn2+- and cA3- dependent dsDNase (Fig. 2d, left gel; Supplementary Fig. 3c), or a ssRNase when activated with cA4. The ssRNase activity of AaCan2 is supported by either Mn2+ or Mg2+ (Fig. 2d, Supplementary Fig. 3d). Reproducible cleavage of ssDNA is also detectable for AaCan2, but the activity is Mn2+- specific, and robust cleavage requires a higher concentration of cA4 (i.e., 45 nM) (Fig. 2d, Supplementary Fig. 3d). Cleavage of ssDNA produces a discrete band suggesting that the enzyme processes ssDNA to a minimal cleavage product or that the activity is sequence- specific (Fig. 2d; Supplementary Fig. 3d). While cA4- dependent activities of AaCan2 are consistent with activities previously reported for the Can2 protein from Treponema succinifaciens29 (i.e., TresuCard1; Fig. 2b), cA3- dependent dsDNA cleavage has not been previously reported. Collectively, our results demonstrate that Can1 and Can2 function as either dsDNA- or ssRNA- specific nucleases, depending on the cyclic nucleotide activator (i.e., cA3 or cA4).  
+
+<|ref|>text<|/ref|><|det|>[[115, 775, 725, 796]]<|/det|>
+Can2 ancillary nuclease provides sensitive Csm- based RNA detection  
+
+<|ref|>text<|/ref|><|det|>[[112, 812, 884, 902]]<|/det|>
+To determine if incorporating TtCan1 or AaCan2 improves sensitivity of the Csm- based RNA detection assay, we screened a library of synthetic RNA reporters designed to identify sequences that might be preferred by these nucleases (Supplementary Table 1, Supplementary Fig. 4). Consistent with our gel- based assays, cA4- activated AaCan2
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 87, 876, 201]]<|/det|>
+cleaves RNA reporters in the presence of either \(\mathrm{Mg^{2 + }}\) or \(\mathrm{Mn^{2 + }}\) , but reactions with \(\mathrm{Mn^{2 + }}\) consistently result in higher fluorescent signal (Supplementary Fig. 4a, b). While TtCan1 cleaves the same RNA reporters as AaCan2, cleavage by TtCan1 requires higher concentrations of \(\mathrm{cA_4}\) and produces less fluorescent signal (Supplementary Fig. 5).  
+
+<|ref|>text<|/ref|><|det|>[[111, 216, 880, 465]]<|/det|>
+Having established that AaCan2 is more active than TtCan1, we set out to compare AaCan2 to the sensitivity of TtCsm6, which we used previously<sup>16</sup>. This comparison was performed by measuring \(\mathrm{cA_4}\) concentration- dependent activity for AaCan2 and TtCsm6 using the preferred RNA reporter for each of the respective enzymes (Supplementary Fig. 4). AaCan2 produces a similar fluorescent signal to TtCsm6 when activated with 20- fold less \(\mathrm{cA_4}\) (0.5 nM versus 10 nM) (Fig. 2e). Moreover, AaCan2 exhibits an incremental decrease in cleavage rates with decreasing \(\mathrm{cA_4}\) , while TtCsm6 exhibits a dramatic (non- linear) drop in the activity. The distinction in activity between these enzymes is consistent with the ring- nuclease activity of TtCsm6 rapidly degrading its activator, while AaCan2 binds and preserves the cyclic nucleotide (Supplementary Fig. 3e).  
+
+<|ref|>text<|/ref|><|det|>[[111, 480, 870, 750]]<|/det|>
+Finally, we incorporated AaCan2 into the type III- based detection assay and benchmarked this combination against TtCsm6- based detection (Fig. 2f, g). The TtCsm6- based assay reliably detects \(10^6\) copies/μL of target RNA (Fig. 2f), while AaCan2- based reactions are more sensitive ( \(10^5\) copies/μL) (Fig. 2g). While coupling TtCsm- detection to AaCan2 results in significantly higher sensitivity, it also results in higher background, but this background is only evident in the presence of the TtCsm- complex (Fig. 2f, g), whereas AaCan2 alone demonstrates very little non- specific cleavage (Fig. 2e). This disparity suggests that non- sequence specific activation of the Cas10 polymerase may generate low levels of \(\mathrm{cA_4}\) , which stably activates AaCan2, whereas the ring- nuclease of TtCsm6 rapidly degrades \(\mathrm{cA_4}\) limiting the background signal. Collectively, these results demonstrate that coupling AaCan2 with TtCsm<sup>Csm3- D34A</sup> provides more sensitive RNA detection.  
+
+<|ref|>text<|/ref|><|det|>[[112, 767, 872, 810]]<|/det|>
+Incorporating \(\mathrm{cA_3}\) - dependent nuclease activity does not provide additional sensitivity of RNA detection  
+
+<|ref|>text<|/ref|><|det|>[[112, 828, 835, 895]]<|/det|>
+While our assay uses \(\mathrm{cA_4}\) - activated collateral cleavage of ssRNA reporters, the TtCsm<sup>Csm3- D34A</sup>- complex also produces \(\mathrm{cA_3}\) (Fig. 1d, Supplementary Fig. 1d). We hypothesized that combining \(\mathrm{cA_3}\) - and \(\mathrm{cA_4}\) - sensing nucleases might enhance the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 880, 313]]<|/det|>
+sensitivity of TtCsm- based detection (Fig. 3a). NucC (Nuclease, CD- NTase associated) endonucleases adopt homotrimeric structures forming a 3- fold symmetric pocket for cA3 binding \(^{24,32,33}\) . Binding cA3 triggers dimerization of NucC homotrimers juxtaposing pair of active sites to cleave DNA \(^{32,33}\) . We purified three thermophilic NucC orthologs and tested cA3- dependent dsDNA cleavage (Supplementary Fig. 6). The NucC from Clostridium tepidum (CtNucC) has the highest dsDNase activity and digests plasmid DNA into 300- 400 bp fragments in the presence of cA3 (Fig. 3b, left; Supplementary Fig. 7a). Deep sequencing of cleavage fragments determined that all purified NucC nucleases have a preference for 5'- ANNT- 3' sequence motif, which is consistent with previously published work \(^{33}\) (Fig. 3b, right; Supplementary Fig. 7b- e).  
+
+<|ref|>text<|/ref|><|det|>[[111, 330, 881, 622]]<|/det|>
+Next, we set out to determine if CtNucC and AaCan2 could be combined into a single reaction to improve the sensitivity of RNA detection with TtCsm \(^{Csm3 - D34A}\) . To perform fluorescent assays with CtNucC, we designed a 31- bp dsDNA reporter comprising six repeats of the optimal cleavage site (Supplementary Table 1). The lowest concentration of cA3 detected by CtNucC is 0.5 nM, which is 10- fold more sensitive than TtCan1 and 100- fold more sensitive than AaCan2 (Fig. 3c). However, TtCsm \(^{Csm3 - D34A}\) coupled with CtNucC and dsDNA reporter only detects high concentrations of target RNA (i.e., \(10^{7}\) copies/μL; Fig. 3d). Further, combining CtNucC with AaCan2 and matching fluorescent probes (i.e., dsDNA and ssRNA, respectively) (Fig. 3a) into a single reaction does not improve the sensitivity compared to detection with AaCan2 alone (Fig. 3d, Supplementary Fig. 8a). While CtNucC is sensitive to cA3 activation, the TtCsm- complex may not produce sufficient concentrations of this cyclic nucleotide to increase sensitivity over AaCan2 detection alone.  
+
+<|ref|>text<|/ref|><|det|>[[115, 640, 750, 660]]<|/det|>
+Type III CRISPR based RNA capture and detection from patient samples  
+
+<|ref|>text<|/ref|><|det|>[[111, 678, 881, 903]]<|/det|>
+RNA extracted from nasopharyngeal swabs of COVID- 19 patients are complex mixtures of nucleic acids derived from the host, the virus, and microbial communities residing in the upper respiratory tract. To determine if TtCsm complex can capture SARS- CoV- 2 RNA in such mixtures, we extracted total RNA from nasopharyngeal swabs of 17 positive and 6 negative patients diagnosed by RT- qPCR (Supplementary Fig. 9a). We used 3 μL of each RNA sample to perform the TtCsm- AaCan2 reaction and 120 μL as input for Csm- based RNA capture followed by a polymerization reaction and fluorometric detection with AaCan2. Only samples with the highest viral RNA concentration (Ct <17) tested positive in the TtCsm- AaCan2 reactions. However, adding the Csm- based RNA capture method increases the sensitivity ~100- fold and reliably
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 88, 872, 133]]<|/det|>
+detects SARS- CoV- 2 RNA in patient samples with Ct values \(\leq 23.2\) , which corresponds to \(\sim 10^{4}\) copies/μL of viral RNA (Fig. 4a, b and Supplementary Fig. 9b, c).  
+
+<|ref|>text<|/ref|><|det|>[[110, 149, 881, 488]]<|/det|>
+RNA extraction kits are expensive, time- consuming, and require specialized equipment. To eliminate this step, we tested if the TtCsm complex can capture and concentrate target RNA directly from a nasopharyngeal swab sample without prior RNA extraction. To identify lysis conditions that do not inhibit activity of the TtCsm- complex, we tested 10 lysis buffer compositions with varying concentrations of detergents (i.e., Triton X- 100 or NP- 40) and chelators (i.e., EDTA or EGTA) (Supplementary Fig. 9d). We mixed Csm- beads with a mock sample made by spiking SARS- CoV- 2 RNA fragment into SARS- CoV- 2 negative nasopharyngeal swab, added lysis buffer, and incubated for 20 min at \(65^{\circ}C\) . This heat treatment inactivates SARS- CoV- 2, promotes lysis, and allows RNA binding by TtCsm- complex and its downstream activities<sup>34,35</sup>. After pulling down Csm- beads with a magnet, we discarded the supernatant and performed polymerization reactions followed by a TtCsm6- based fluorescent readout. The TtCsm complex detects spiked RNA in the samples treated with Triton X- 100 ( \(0.025 - 0.1\%\) ) and EGTA (1 mM), while other buffers significantly inhibited Csm- based detection (Supplementary Fig. 9d).  
+
+<|ref|>text<|/ref|><|det|>[[110, 504, 881, 840]]<|/det|>
+Finally, to assess the sensitivity of direct SARS- CoV- 2 RNA detection in swab samples using type III capture and AaCan2- based fluorescent detection (Fig. 4c), we used a SARS- CoV- 2 positive patient sample (Ct \(\sim 13.6\) ) that was 10- fold serially diluted in a negative patient swab sample (Fig. 4d). In this assay, we used lysis buffer supplemented with \(0.05\%\) Triton X- 100 and 1 mM EGTA. Csm- based RNA capture assay detects SARS- CoV- 2 RNA in unprocessed samples (i.e., no RNA purification) with Ct \(< 21.2\) (Fig. 4d, Supplementary Fig. 9f), which corresponds to \(5 \times 10^{4}\) copies/μL and \(\sim 5\) - fold less sensitive compared to detection performed using purified RNA (Fig. 4b, Supplementary Fig. 9e). To compare the efficiency of direct detection from lysed nasopharyngeal swab relative to detection from extracted RNA, we used three nasopharyngeal swab samples that previously tested positive for SARS- CoV- 2 using RT- qPCR (Supplementary Fig. 9f). All three samples tested positive using direct detection from nasal swabs, however direct detection from patient samples resulted in a higher signal- to- noise ratio. This difference suggests that further optimization of the lysis conditions may lead to higher sensitivity (Supplementary Fig. 9f).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[115, 90, 222, 108]]<|/det|>
+## Discussion  
+
+<|ref|>text<|/ref|><|det|>[[112, 128, 881, 262]]<|/det|>
+CRISPR- based diagnostics have been progressing at a remarkable pace<sup>5</sup>. Development efforts have primarily focused on type V (Cas12) and type VI (Cas13) CRISPR- systems, and the sensitivity of these techniques have improved from picomolar<sup>36</sup> to attomolar concentrations<sup>28</sup>. However, most CRISPR- based viral diagnostics described to date still require nucleic acid extraction and pre- amplification to reach clinically relevant sensitivities<sup>4</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[112, 280, 884, 550]]<|/det|>
+In 2021, the first attempts to repurpose type III CRISPR systems for SARS- CoV- 2 diagnostics achieved \(0.1 - 1\) nM sensitivity of RNA detection without preamplification<sup>16,17</sup>. More recent improvements using different type III complexes or different ancillary nucleases have been used to detect SARS- CoV- 2 RNA in purified RNA samples with \(\sim 2 - 4\) fM sensitivity<sup>18,24</sup>. Here, we contribute to the ongoing development of type III systems by developing methods for sequence- specific capture and concentration of target RNAs directly from unprocessed patient samples. This approach enables direct detection of \(5 \times 10^{4}\) copies of SARS- VoV- 2 RNA per \(\mu \mathrm{L}\) (\~90 fM) in clinical samples without laboratory- based RNA extraction or pre- amplification. While the sensitivity of the approach is still inferior to RT- qPCR, it is sufficient to identify infected individuals capable of spreading SARS- CoV- 2<sup>37</sup> and is comparable to rapid antigen tests<sup>2</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[111, 567, 881, 905]]<|/det|>
+Like Cas13, type III systems also recognize RNA, and the most sensitive detection methods developed to date for either approach rely on collateral nuclease activity to release a fluorescent signal<sup>4</sup>. While Cas13- based methods are currently more sensitive (\~50 aM), the intrinsic amplification of RNA recognition by type III system may ultimately improve sensitivity. Type III systems uniquely amplify RNA recognition in two sequential steps: first, through Cas10- mediated polymerization of cOAs and second, through cOA- mediated activation of multi- turnover effectors (e.g., Csm6). In addition to the advantages that might come from consecutive stages of signal amplification, the separation of target recognition by the type III surveillance complex (i.e., Csm or Cmr) from collateral cleavage by ancillary effectors also enables programmable RNA capture. Unlike Cas13, which relies on the same active site for target and non- target collateral cleavage<sup>38</sup>, the RNase- dead TtCsm complex (TtCsm<sup>Csm3- D33A</sup>) can be used to capture and maintain target RNA from a larger volume and concentrate these RNAs for various downstream applications. Incorporating RNA capture increases the sensitivity of type III CRISPR- based diagnostic and allows direct detection in clinical samples without RNA
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[112, 88, 879, 222]]<|/det|>
+extraction, a prerequisite for most current platforms. We anticipate that further incorporation of type III- based RNA pull- down techniques to bypass RNA extraction, optimization of lysis conditions, and next generation of readouts (e.g., real- time sequencing, digital enzymology, amperometry, etc.) will further boost the sensitivity and minimize time- to- result, bringing type III CRISPR diagnostic to current standards of rapid molecular testing.  
+
+<|ref|>text<|/ref|><|det|>[[111, 238, 883, 488]]<|/det|>
+Our work to improve type III diagnostics has also uncovered ancillary nuclease activities that are valuable for understanding the basic biology and augmenting applications for biotechnology. Both \(\mathrm{cA}_3\) and \(\mathrm{cA}_4\) , but none of the other tested cyclic oligoadenylates (i.e., \(\mathrm{cA}_2\) , \(\mathrm{cA}_5\) , \(\mathrm{cA}_6\) ), activate TtCan1 and AaCan2 to cleave specific substrates. TtCan1 is primarily a \(\mathrm{cA}_3\) - dependent dsDNase, while AaCan2 is a \(\mathrm{cA}_4\) - dependent ssRNase. Can1 nucleases may have emerged from duplication and fusion of ancestral Can2 genes \(^{30,31}\) , and we hypothesize that this fusion may enable the evolution of mechanisms for recognizing diverse (e.g., non- symmetrical) ligands that activate the effector. Similarly, SAVED (SMODS- Associated and fused to Various Effector Domains) domains appear to be derived from the fusion of two ancient CARF- like domains and are activated by cyclic trinucleotides \(^{39}\) .  
+
+<|ref|>text<|/ref|><|det|>[[111, 504, 880, 867]]<|/det|>
+Target RNA binding by type III Csm- or Cmr- complexes triggers synthesis of several cyclic oligoadenylate species in varying ratios \(^{19,24}\) . We showed that the TtCsm complex predominantly generates \(\mathrm{cA}_4\) , while \(\mathrm{cA}_3\) is produced at a lower level. We hypothesize that cOA ratios generated by type III complexes have evolved as a fine- tuned immunomodulatory mechanism that regulates ancillary nuclease activities and infection outcomes. In fact, the genome of T. thermophilus (HB8 and HB27 strains) encodes both a \(\mathrm{cA}_4\) - activated Csm6 RNase and Can1 CARF- nuclease \(^{31}\) that is activated by \(\mathrm{cA}_4\) (RNase) and \(\mathrm{cA}_3\) (DNase). \(\mathrm{cA}_4\) is the primary signal generated by target- bound TtCsm, and RNA cleavage by \(\mathrm{cA}_4\) - activated Csm6 nucleases results in growth arrest and facilitates clearance of invading genetic parasites \(^{15}\) . However, failure to clear the infection through \(\mathrm{cA}_4\) - dependent RNase activity by Csm6 would result in continuous polymerization by Cas10 and accumulation of \(\mathrm{cA}_3\) , which will activate the TtCan1 DNase. The lack of sequence preference suggests that TtCan1 might degrade the host genome and induce abortive infection and cell death. More work is necessary to understand the diversity of nucleoside- based signal generators and the diversity of signal integrators.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[115, 90, 291, 109]]<|/det|>
+## Acknowledgments  
+
+<|ref|>text<|/ref|><|det|>[[112, 128, 881, 314]]<|/det|>
+We are grateful to members of Bozeman Health who provided deidentified patient samples. A.S- F. is a postdoctoral fellow of the Life Science Research Foundation that is supported by the Simons Foundation. A.S- F. is supported by the Postdoctoral Enrichment Program Award from the Burroughs Wellcome Fund. Research in the Wiedenheft lab is supported by the NIH (R35GM134867), the M.J. Murdock Charitable Trust, a young investigator award from Amgen, a generous gift from the Rosolowsky family, and the Montana State University Agricultural Experimental Station (USDA NIFA). The Koutmou lab's contributions to this work were supported by the NIH (R35GM128836). Funders had no role in designing, performing, interpreting, or submitting the work. Figures were created using BioRender.com.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 333, 312, 352]]<|/det|>
+## Author contributions  
+
+<|ref|>text<|/ref|><|det|>[[112, 372, 881, 600]]<|/det|>
+B.W., A. Nemudraia, A. Nemudryi, and A.S.- F. conceived the experimental plans. A. Nemudraia, A. Nemudryi and R.W. developed and performed Type III Csm- based RNA concentration method. A.M.S., T.Z., R.W., M.B. and A.S.- F. purified the proteins. A. Nemudraia, A.S.- F., S.P., J.N., and R.W. performed biochemical characterization of the ancillary nucleases. A. Nemudraia performed RNA reporter's screen. A. Nemudryi performed statistical analyses and analyzed sequencing data. L.R., J.J., and K.K. contributed to the initial design of TLC assays. L.H. and A. Nemudryi performed TLC; M.B., S.P., and T.W. performed the bioinformatic analyses and phylogenetics. H.L. and A.M. performed RNA extractions and RT- qPCR of patient nasopharyngeal swab samples. A. Nemudraia and A. Nemudryi performed RT- qPCR and Csm- based detection assay. A. Nemudraia, A. Nemudryi, and B.W. wrote the manuscript. All authors edited and approved the manuscript.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 617, 335, 636]]<|/det|>
+## Declaration of interests  
+
+<|ref|>text<|/ref|><|det|>[[115, 656, 866, 717]]<|/det|>
+B.W. is the founder of SurGene LLC, and VIRIS Detection Systems Inc. B.W., A. Nemudryi, A. Nemudraia, and A.S.- F. are inventors on patents related to CRISPR- Cas systems and applications thereof.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 737, 198, 755]]<|/det|>
+## Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 772, 579, 792]]<|/det|>
+## Human clinical sample collection and preparation  
+
+<|ref|>text<|/ref|><|det|>[[115, 810, 883, 874]]<|/det|>
+Clinical samples were obtained with local IRB approval (protocol #DB033020) and informed consent from patients undergoing testing for SARS- CoV- 2 at Bozeman Health Deaconess Hospital. Nasopharyngeal swabs from patients that either tested negative or
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[60, 88, 883, 131]]<|/det|>
+positive for SARS- CoV- 2 were collected in viral transport media. RNA was extracted from all patient samples using the QIAamp Viral RNA Mini Kit (QIAGEN).  
+
+<|ref|>sub_title<|/ref|><|det|>[[63, 150, 243, 169]]<|/det|>
+## Nucleic acids  
+
+<|ref|>text<|/ref|><|det|>[[60, 187, 884, 366]]<|/det|>
+Sodium salts of cyclic di- , tri- , tetra- , penta- and hexa- adenosine monophosphates (cA2- 6) were purchased from Biolog Life Science Institute. Fluorescent reporters (RNA and DNA) were purchased from IDT (Supplementary Table 1). The dsDNA reporter was ordered as a duplex from IDT. Target and non- target RNAs of SARS- CoV- 2 N- gene were in vitro transcribed with MEGAscript T7 (Thermo Fisher Scientific) from PCR products generated from pairs of synthesized overlapping DNA oligos (Supplementary Table 1) (Eurofins). Transcribed RNAs were purified by denaturing PAGE. Total RNA from HEK 293T cells was extracted using TRIzol reagent.  
+
+<|ref|>sub_title<|/ref|><|det|>[[63, 383, 378, 403]]<|/det|>
+## Non-targeting control (NTC)  
+
+<|ref|>text<|/ref|><|det|>[[60, 420, 884, 555]]<|/det|>
+Total RNA extracted from SARS- CoV- 2 negative nasopharyngeal swabs or total RNA extracted from HEK 293T cells were used as negative controls. RNA extracted from HEK 293T cells was diluted to match the average Ct level ( \(\sim 27\) ) obtained for RNaseP mRNA in RNA samples extracted from nasopharyngeal swabs (Supplementary Table 2). The RT- qPCR for RNase P mRNA was performed using CDC RP primers and probe (2019- nCoV CDC EUA Kit, IDT#10006606).  
+
+<|ref|>sub_title<|/ref|><|det|>[[63, 572, 203, 591]]<|/det|>
+## Plasmids  
+
+<|ref|>text<|/ref|><|det|>[[60, 609, 884, 880]]<|/det|>
+Plasmids encoding the type III- A Csm complex from Thermus thermophilus (pCDF- 5xT7- TtCsm; Addgene #128572 and pACYC- TtCas6- 4xcrRNA4.5; Addgene #127764), were a gift from Jennifer Doudna. Vector pCDF- 5xT7- TtCsm was used as a template for site- directed mutagenesis to mutate the D33 residue in Csm3 to alanine (D33A) and inactivate Csm3- mediated cleavage of target RNA (pCDF- 5xT7- TtCsmCsm3- D34A) \(^{35}\) . The CRISPR array in pACYC- TtCas6- 4xcrRNA4.5 was replaced with a synthetic CRISPR array (GeneArt) containing five repeats and four identical spacers, designed to target the N- gene of SARS- CoV- 2 (i.e., pACYC- TtCas6- 4xgCoV2N1) \(^{16}\) . TtCas6 was PCR was PCR- amplified from the pACYC- TtCas6- 4xcrRNA4.5 plasmid and cloned between the Ncol and Xhol sites in the pRSF- 1b backbone (Millipore Sigma) (pRSF- TtCas6). Expression vector encoding TtCsm6 nuclease, pC0075 TtCsm6 His6- TwinStrep- SUMO- Bsal, was a gift from Feng Zhang (Addgene plasmid #115270) \(^{40}\) .
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[60, 88, 884, 358]]<|/det|>
+Gene fragments encoding for Can1 from Thermus thermophilus (TtCan1; NCBI accession=WP_011229147.1), Can2 from Archaeoglobi archaeon JdFR- 42 (AaCan2; (JGI) IMG gene accession=2730024700), Clostridium thermobutyricum (CtCan2; NCBI accession=WP_195972101.1), and Thermus thermophilus (TtCan2; NCBI accession= WP_143585921.1), were codon optimized for expression in E. coli, synthesized by GenScript, and cloned into pC0075 vector (Addgene #115270) in frame with the N- terminal His6- TwinStrep- SUMO tag using Ncol and Xhol restriction sites to replace the TtCsm6 gene. NucC from Clostridium tepidum BSD2780120874b_170522_A10 (CtNucC; NCBI accession= WP_195923598.1), Elioraea sp. Yellowstone (EsNucC; NCBI accession= WP_141855040.1) and Acidimicrobiales bacterium mtb01 (Amtb01NucC; NCBI accession= TEX45487.1), were cloned into pC0075 backbone using the same restriction sites as for Can1 and Can2 genes.  
+
+<|ref|>sub_title<|/ref|><|det|>[[64, 376, 447, 396]]<|/det|>
+## Protein expression and purification  
+
+<|ref|>text<|/ref|><|det|>[[110, 408, 884, 883]]<|/det|>
+Expression and purification of the TtCsmCsm3- D34A complex and TtCsm6 were performed as previously described<sup>16</sup>. TtCan1, AaCan2, CtCan2, TtCan2, CtNucC, EsNucC, and Amtb01NucC) were purified according to the following protocol. Each expression vector was transformed into Escherichia coli BL21(DE3) cells and grown in LB Broth (Lennox) (Thermo Fisher Scientific) at \(37^{\circ}\mathrm{C}\) to an OD600 of 0.5. Cultures were then incubated on ice for 1 hour, and then induced with 0.5 mM IPTG for overnight expression at \(16^{\circ}\mathrm{C}\) . Cells were lysed with sonication in Lysis buffer (20 mM Tris- HCl pH 8, 500 mM NaCl, 1 mM TCEP) and lysate was clarified by centrifugation at 10,000 \(\mathrm{xg}\) for 25 mins, \(4^{\circ}\mathrm{C}\) . The lysate was heat- treated at \(55^{\circ}\mathrm{C}\) for 45 minutes and clarified by centrifugation at 10,000 \(g\) for 25 mins at \(4^{\circ}\mathrm{C}\) . His<sub>6</sub>- TwinStrep- tagged protein was bound to a StrepTrap HP column (Cytiva) and washed with Lysis buffer. The protein was eluted with Lysis buffer supplemented with 2.5 mM desthiobiotin and concentrated (10k MWCO Corning Spin- X concentrators) at \(4^{\circ}\mathrm{C}\) . Affinity tags were removed from the protein using His- tagged SUMO protease (100 \(\mu \mathrm{L}\) of 2.5 mg/mL protease per 20 mg of protein) during dialysis against SUMO digest buffer (30 mM Tris- HCl pH 8, 500 mM NaCl, 1 mM dithiothreitol (DTT), 0.15% Igepal) at \(4^{\circ}\mathrm{C}\) overnight. The tag and the protease were applied to HisTrap HP column (Cytiva), and the flow- through was concentrated using Corning Spin- X concentrators at \(4^{\circ}\mathrm{C}\) . Finally, the protein was purified using a HiLoad Superdex 200 26/600 size- exclusion column (Cytiva) in storage buffer (20 mM Tris- HCl pH 7.5, 1 mM DTT, 400 mM
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 883, 133]]<|/det|>
+monopotassium glutamate, \(5\%\) glycerol). Fractions containing the target protein were pooled, concentrated, aliquoted, flash-frozen in liquid nitrogen, and stored at \(- 80^{\circ}C\) .  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 150, 361, 171]]<|/det|>
+## \(^{32}\mathrm{P}\) -labeling of RNA oligos  
+
+<|ref|>text<|/ref|><|det|>[[111, 187, 884, 344]]<|/det|>
+Target (SARS- CoV- 2 N1) and non- target RNAs were transcribed from PCR extended duplex oligos using home- made T7 RNA polymerase (Supplementary Table 3) (Eurofins). The IVT RNAs were gel purified and dephosphorylated with Quick CIP (NEB) for 20 min at \(37^{\circ}C\) in 1X CutSmart Buffer (NEB). The phosphatase was inactivated by heating at \(80^{\circ}C\) for 5 min before \(5'\) end- labeling the RNAs with T4 polynucleotide kinase (NEB) and \([\gamma - ^{32}\mathrm{P}]\cdot\) ATP (PerkinElmer) for 30 min at \(37^{\circ}C\) . The kinase was heat inactivated by heating at \(65^{\circ}C\) for 20 min.  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 361, 575, 382]]<|/det|>
+## Binding and pull-down of RNA oligos with TtCsM  
+
+<|ref|>text<|/ref|><|det|>[[111, 398, 884, 648]]<|/det|>
+For the experiments shown in Fig. 1b and Supplementary Fig. 1b,c, \(^{32}\mathrm{P}\) - labeled RNA (25 nM) was incubated with \(\mathrm{TtCsM^{Csm3 - D34A}}\) (160 nM) targeting SARS- CoV- 2 N- gene in 1X Binding Buffer (25 mM HEPES, pH 7.5, 150 mM NaCl, 1 mM TCEP) for 20 min at \(65^{\circ}C\) . The reaction mixtures were added to \(10~\mu \mathrm{L}\) of HisPur Ni- NTA Magnetic beads (ThermoFisher) equilibrated in Binding Buffer and incubated on ice 30 min with vortexing every 10 min. The beads were separated from the supernatant using a magnet and washed with \(50~\mu \mathrm{L}\) 1X binding buffer. The RNA was extracted from supernatant (unbound fraction) and beads (bound fraction) using Acid Phenol: chloroform (Ambion). Extracted RNA was resolved using UREA- PAGE, exposed to a phosphor screen, and imaged on a Typhoon 5 imager (Amersham). Bands corresponding to the IVT RNAs were quantified using ImageJ and the percent bound calculated [bound/(bound + free) \(^{*}100\%\) ].  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 664, 519, 685]]<|/det|>
+## Complexing of TtCsM with magnetic beads  
+
+<|ref|>text<|/ref|><|det|>[[112, 701, 884, 812]]<|/det|>
+The HisPur Ni- NTA Magnetic beads (ThermoFisher) were washed two times with a 1X Binding Buffer (25 mM HEPES, pH 7.5, 150 mM NaCl, 1 mM TCEP). For one reaction, 5 \(\mu \mathrm{L}\) of equilibrated beads were mixed with TtCsM<sup>dead</sup> complex (25 nM) in 1X Binding Buffer ( \(V = 50~\mu \mathrm{L}\) ) and incubated for 30 min on ice. The beads with the complex (CsM- beads) were concentrated with a magnet and resuspended in 5 \(\mu \mathrm{L}\) of 1x Binding Buffer.  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 829, 428, 849]]<|/det|>
+## Thin-layer chromatography (TLC)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 884, 358]]<|/det|>
+For the experiments shown in Fig. 1c, \(3 \mu \text{L}\) of positive sample (target RNA diluted in NTC, \(10^{10}\) copies/μL) or \(3 \mu \text{L}\) of NTC were mixed with TtCsm \(Csm3 - D34A\) complex (25 nM) and \(250 \mu \text{M ATP}\) supplemented with \([\alpha - ^{32}\text{P}]\) - ATP (PerkinElmer) in the reaction buffer (20 mM Tris- HCl pH 7.8, 250 mM monopotassium glutamate, 10 mM ammonium sulfate, 1 mM TCEP (tris(2- carboxymethyl)phosphine)), 5 mM magnesium sulfate). The reaction was incubated at \(60^{\circ}\text{C}\) for 1h. For the pull- down reactions, \(120 \mu \text{L}\) of positive or negative samples were mixed with \(5 \mu \text{L}\) of Csm- beads in Binding Buffer (25 mM HEPES, pH 7.5, 150 mM NaCl, 1 mM TCEP) for 10 min at \(60^{\circ}\text{C}\) . The Csm- beads were concentrated with a magnet and the supernatant was discarded. The Csm pellets were resuspended in 30 μL of the reaction buffer and \(250 \mu \text{M ATP}\) supplemented with \([\alpha - 32\text{P}]\) - ATP (PerkinElmer). Reaction products were phenol- chloroform extracted and resolved on silica TLC plates (Millipore).  
+
+<|ref|>text<|/ref|><|det|>[[112, 374, 884, 533]]<|/det|>
+Samples (1 μL) were mixed with 100 mM sodium acetate, pH 5.2 (2 μL) and spotted 1.5 cm above the bottom of the TLC plate. The plate was placed inside a 2 L beaker filled to \(\sim 0.5 \text{cm}\) with developing solvent (0.2 M ammonium bicarbonate pH 9.3, 70% ethanol and 30% water) and capped with aluminum foil. The plate was run for 2 h at room temperature and dried. TLC plate was exposed to a phosphor screen and imaged with Typhoon phosphor imager. Chemically synthesized standards (2μM) were resolved on the same TLC plate and visualized using UV shadowing.  
+
+<|ref|>text<|/ref|><|det|>[[112, 550, 884, 637]]<|/det|>
+To test \(\text{Ca}_3\) and \(\text{Ca}_4\) hydrolysis in the presence of ancillary nuclease, radiolabeled \(\text{Ca}_3\) and \(\text{Ca}_4\) produced above were mixed with nuclease (500 nM) in the reaction buffer and incubated for 1 hour at \(55^{\circ}\text{C}\) . Reaction products were phenol- chloroform extracted and resolved using thin- layer chromatography (TLC) for 45 min as described above.  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 655, 390, 675]]<|/det|>
+## Type III-based RNA detection  
+
+<|ref|>text<|/ref|><|det|>[[112, 692, 885, 895]]<|/det|>
+3 μL of RNA sample was mixed with 250 μM ATP, 25 nM TtCsm \(^{\text{dead}}\) complex, 300 nM of nuclease (TtCsm6, AaCan2, or CtNucC) with corresponding reporter in a reaction buffer (20 mM Tris- HCl pH 7.8, 250 mM monopotassium glutamate, 10 mM ammonium sulfate, 1 mM TCEP (tris(2- carboxymethyl)phosphine)), 5 mM magnesium sulfate (for TtCsm6 and CtNucC) or 5 mM manganese(II) chloride (for AaCan2) in a 30 μL reaction. The reporter B8 (300 nM) was used for the reaction with TtCsm6, D7 (300 nM) – with AaCan2, and dsDNA probe (300 nM) – with CtNucC. Reactions were incubated at \(55^{\circ}\text{C}\) . Cleavage of fluorescent reporters was detected by measuring fluorescence every 10 sec in a real- time PCR instrument QuantStudio 3 (Applied Biosystems).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[113, 89, 528, 110]]<|/det|>
+## Type III-based RNA pull-down and detection  
+
+<|ref|>text<|/ref|><|det|>[[113, 127, 884, 261]]<|/det|>
+To bind TtCsm<sup>dead</sup> complex with the magnetic beads, the HisPur Ni- NTA Magnetic beads (ThermoFisher) were washed two times with a 1X Binding Buffer (25 mM HEPES, pH 7.5, 150 mM NaCl, 1 mM TCEP). For one reaction, 5 μL of equilibrated beads were mixed with TtCsm<sup>dead</sup> complex (30 nM) in 1X Binding Buffer (V = 50 μL) and incubated for 30 min on ice. The beads with the complex (Csm- beads) were concentrated with a magnet and resuspended in 5 μL of 1x Binding Buffer.  
+
+<|ref|>text<|/ref|><|det|>[[112, 277, 884, 548]]<|/det|>
+Pull- down and detection from RNA sample: 120 μL of sample was mixed with 5 μL of Csm- beads in 1x Binding Buffer for 10 min at \(60^{\circ}\mathrm{C}\) . The Csm- beads were concentrated with a magnet and the supernatant was discarded. The Csm- beads pellet was resuspended in 20 μL of the 1X reaction buffer (20 mM Tris- HCl pH 7.8, 250 mM monopotassium glutamate, 10 mM ammonium sulfate, 1 mM TCEP (tris(2- carboxymethyl)phosphine)), 5 mM magnesium sulfate / manganese(II) chloride) containing ATP (250 μM). The reaction was incubated 10 min at \(60^{\circ}\mathrm{C}\) , the Csm- beads were pelleted, and the supernatant (10 μL) was transferred to a new reaction with TtCsm6 (300 nM) and B8 RNA Reporter (300 nM) or AaCan2 (300 nM) and D7 RNA Reporter (300 nM) in 1X reaction buffer (V = 30 μL) (Supplementary Table 1). Reactions were incubated at \(55^{\circ}\mathrm{C}\) . Cleavage of the fluorescent RNA reporter was detected by measuring fluorescence every 10 sec in a real- time PCR instrument QuantStudio 3.  
+
+<|ref|>text<|/ref|><|det|>[[112, 564, 884, 903]]<|/det|>
+Pull- down and detection from nasopharyngeal swab: 120 μL of a nasopharyngeal swab was mixed with 5 μL of Csm- beads in 1X Lysis Buffer and incubated for 20 min at \(65^{\circ}\mathrm{C}\) . Ten lysis buffers compositions were tested. All buffers contained 25 mM HEPES, pH 7.5, 150 mM NaCl, 1 mM TCEP and were supplemented with (A) 0.025% Triton X- 100, (B) 0.025% Triton X- 100 and 1 mM EDTA, (C1) 0.025% Triton X- 100 and 1 mM EGTA, (C2) 0.05% Triton X- 100 and 1 mM EGTA, (C3) 0.1% Triton X- 100 and 1 mM EGTA, (I) 0.025% NP- 40, (J) 0.025% NP- 40 and 1 mM EDTA, (K1) 0.025% NP- 40 and 1 mM EGTA, (K2) 0.05% NP- 40 and 1 mM EGTA, or (K3) 0.1% NP- 40 and 1 mM EGTA. Each of the supplements are lettered according to the results presented in Supplementary Fig. 9d. The Csm- beads were concentrated with a magnet and the supernatant was discarded. The Csm- beads pellet was resuspended in 20 μL of the 1x reaction buffer (20 mM Tris- HCl pH 7.8, 250 mM monopotassium glutamate, 10 mM ammonium sulfate, 1 mM TCEP (tris(2- carboxymethyl)phosphine)), 5 mM magnesium sulfate or manganese(II) chloride) containing ATP (250 μM). The reaction was incubated 10 min at \(65^{\circ}\mathrm{C}\) , the Csm- beads were pelleted, and the supernatant (10 μL) was transferred to a new reaction with TtCsm6
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 883, 177]]<|/det|>
+(300 nM) and B8 RNA Reporter (300 nM) or AaCan2 (300 nM) and D7 RNA Reporter (300 nM) in 1 x reaction buffer (the final volume of a reaction 30 μL). Reactions were incubated at 55°C. Cleavage of fluorescent RNA reporter was detected by measuring fluorescence every 10 sec in a real-time PCR instrument QuantStudio 3.  
+
+<|ref|>sub_title<|/ref|><|det|>[[111, 195, 204, 214]]<|/det|>
+## RT-qPCR  
+
+<|ref|>text<|/ref|><|det|>[[110, 231, 884, 547]]<|/det|>
+RT- qPCR was performed using N1 and RP CDC primers (2019- nCoV CDC EUA Kit, IDT#10006606). RNA was extracted from patient samples with QIAamp Viral RNA Mini Kit (QIAGEN, # 52906) and used for one- step RT- qPCR in ABI 7500 Fast Real- Time PCR System according to CDC protocols (https://www.fda.gov/media/134922/download). In brief, 20 μL reaction included 8.5 μL of Nuclease- free Water, 1.5 μL of Primer and Probe mix (IDT, 10006713), 5 μL of TaqPath 1- Step RT- qPCR Master Mix (ThermoFisher, A15299) and 5 μL of the RNA. Nuclease- free water was used as negative template control (NTC). Amplification was performed as follows: 25°C for 2 min, 50°C for 15 min, 95°C for 2 min followed by 45 cycles of 95°C for 3 s and 55°C for 30 s. To quantify viral RNA in the samples, standard curve for N1 primers was generated using a dilution series of a SARS- CoV- 2 synthetic RNA fragment (RTGM 10169, NIST) spanning N gene with concentrations ranging from 10 to 10⁶ copies per μL. Three technical replicates were performed at each dilution. The NTC showed no amplification throughout the 45 cycles of qPCR.  
+
+<|ref|>sub_title<|/ref|><|det|>[[112, 564, 584, 585]]<|/det|>
+## Nanopore sequencing of DNA cleavage fragments  
+
+<|ref|>text<|/ref|><|det|>[[110, 601, 884, 895]]<|/det|>
+DNA cleavage fragments were sequenced using Oxford Nanopore with Ligation Sequencing Kit (SQK- LSK109). After incubation with TtCan1 or NucC nucleases, cleavage fragments were column- purified using DNA Clean & Concentrator- 5 kit (Zymo Research, D4004) as instructed. Next, for each sample 50 ng of purified DNA was used to prepare sequencing libraries with NEBNext® Ultra™ II DNA Library Prep Kit (NEB, E7645S). Briefly, DNA was end- repaired with NEBNext Ultra II End Prep Enzyme Mix, which fills 5'- and removes 3'- overhangs. Next, end- repaired fragments were barcoded with Native Barcoding Expansion kit (ONT, EXP- NBD104) using Ultra II Ligation Master Mix (NEB). Barcoded DNA fragments were pooled together and purified with magnetic beads (Omega Bio- tek, M1378- 01). Freshly mixed 80% ethanol was used to wash magnetic bead pellet. Sequencing adapters (AMII) were ligated to barcoded DNA using NEBNext® Quick Ligation Module (NEB, E6056S). Ligation reactions were purified with magnetic beads. SFB buffer (ONT, EXP- SFB001) was used for washes. Resulting DNA
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[112, 88, 884, 221]]<|/det|>
+library was eluted from the beads in \(20 \mu \mathrm{L}\) of EB buffer (QIAGEN, #19086). DNA concentration was measured with Qubit dsDNA HS Assay (ThermoFisher, Q32851), and \(20 \mathrm{ng}\) was loaded on the Nanopore MinION (R9.4.1 flow cell). The flow cell was primed, and library was loaded according to Oxford Nanopore protocol (SQK- LSK109 kit). The sequencing run was performed in the high- accuracy base calling mode in the MinKNOW software.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 240, 358, 260]]<|/det|>
+## Sequencing data analysis  
+
+<|ref|>text<|/ref|><|det|>[[111, 277, 884, 592]]<|/det|>
+Sequenced reads were demultiplexed using guppy- barcoder (ONT) and aligned with minimap2 v2.17- r954- dirty (ax map- ont mode) to the reference plasmid sequence that was modified by adding 1000 bp overlaps at the \(5'\) - and \(3'\) - ends. Overlapping regions were introduced to account for circular nature of the plasmid. Resulting alignments (BAM files) were sorted and indexed using samtools v1.13. Next, bamtoed function in bedtools package was used to generate BED files and read coordinates were extracted. Read end coordinates were used to calculate cleavage fragment length distributions and map frequencies of cuts at specific locations (Supplementary Fig. 7). To analyze the sequence preferences of each nuclease, 14 bp windows surrounding read ends were extracted with getfasta function from bedtools package. Resulting fasta files were used to calculate position weigh matrices (PWMs) with getPwmFromFastaFile() function in DiffLogo R package. Finally, PWMs were plotted as sequence logos using ggseqlogo R package. Sequencing depth around the most frequent cut site for each nuclease was calculated with samtools depth function and plotted with ggplot2 package in RStudio.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 610, 323, 629]]<|/det|>
+## RNA reporter's library  
+
+<|ref|>text<|/ref|><|det|>[[111, 646, 884, 894]]<|/det|>
+To determine the optimal RNA reporter for each cOA- activated nuclease, we constructed a library of variable RNA sequences tethering a FAM fluorophore to an Iowa Black quencher. These reporters were designed as single- stranded RNA molecules (i.e., \(5'\) - FAM- AUNNNNNNAU- IABkFQ- 3'; variable region underlined) or to produce a structured RNA (e.g., \(5'\) - FAM- CGCGNNNNNNNCCGC- IABkFQ- 3'; variable region underlined). The Biostrings package in R was used to construct a library of reporter sequences containing each of the 64 unique trinucleotide combinations possible. Since multiple unique trinucleotides could be included in a single reporter (e.g. \(5'\) - FAM- AUAGAAGAAU- IABkFQ- 3' contains AGA, GAA and AAG), we narrowed our initial library of 64 reporters to remove redundant sequences. This resulted in a library of 24 unique reporter sequences, each of which were integrated into both a single- stranded RNA reporter and
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 883, 132]]<|/det|>
+a structured RNA reporter (Supplementary Table 1). The R- script used to design these reporters is accessible on GitHub (WiedenheftLab/RNA_reporter_design).  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 149, 479, 169]]<|/det|>
+## In vitro DNA and RNA cleavage assays  
+
+<|ref|>text<|/ref|><|det|>[[111, 186, 885, 502]]<|/det|>
+All reactions were performed in a buffer containing 20 mM Tris- HCl pH 7.8, 250 mM monopotassium glutamate, 10 mM ammonium sulfate, 1 mM TCEP, 5 mM magnesium sulfate or 5 mM manganese chloride. Plasmid DNA cleavage assays were performed by incubating 1 \(\mu \mathrm{g}\) of Lenti- luciferase- P2A- Neo (Addgene #105621) plasmid with TtCan1, AaCan2 or CtNucC (15- 200 nM) in the presence of cOAv (15- 45 nM) in 10 \(\mu \mathrm{L}\) reaction. After 5- 15 min incubation at \(60^{\circ}\mathrm{C}\) for TtCan1 and \(55^{\circ}\mathrm{C}\) for both AaCan2 and CtNucC, Gel Loading Dye, Purple (6X) (NEB) was added and 4 \(\mu \mathrm{L}\) was loaded on \(1\%\) agarose gel. For ssDNA and ssRNA cleavage assays, 0.425 \(\mu \mathrm{M}\) of 71 nt DNA oligo (CGTcGTACCGgTTAGAGGATGGTGCAAGCGTAATCTGGAAACATCGTTATGGGTATG CCCACGGTGTCCACGGGCG, Eurofins) or 0.425 \(\mu \mathrm{M}\) of 74 nt IVT RNA SARS- CoV- 2 N- gene (Supplementary Table 3) were incubated with TtCan1 (200 nM) or AaCan2 (200 nM) in the presence of cOAv (20- 45 nM) in 10 \(\mu \mathrm{L}\) . After 5- 15 min incubation at \(60^{\circ}\mathrm{C}\) for TtCan1 and at \(55^{\circ}\mathrm{C}\) for AaCan2, 2X RNA Loading Dye (NEB) was added and 10 \(\mu \mathrm{L}\) was loaded on \(12\%\) UREA PAGE.  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 518, 576, 539]]<|/det|>
+## Phylogenetic analysis of Can1 and Can2 proteins  
+
+<|ref|>text<|/ref|><|det|>[[111, 556, 884, 895]]<|/det|>
+A DELTA- BLAST was initiated, using previously described Can1 and Can2 proteins as queries \(^{29 - 31}\) to generate individual lists of closely related proteins with an e- value cutoff of \(10^{- 4}\) and \(50\%\) query coverage. The resulting sequences were then used as queries to initiate a PSI- BLAST search with an E- value cutoff of \(10^{- 4}\) and \(50\%\) query coverage. This step was repeated until convergence and redundant sequences were removed with CD- HIT v4. \(^{74}\) . In case of Can1, sequences from a previously published dataset \(^{14}\) that contain two CARF domains and a nuclease domain were used to generate multiple sequence alignment of Can1- related proteins. In total, 29 sequences of Can1- related proteins and 2,531 sequences of Can2- related proteins were used separately to generate multiple sequence alignment with a local version of MAFFT v7.429 \(^{42}\) (--localpair --maxiterate 1000). The generated alignments for Can1 and Can2 were curated with MaxAlign v1. \(^{143}\) to remove misaligned or non- homologous sequences. The resulting dataset—comprised of 29 Can1- like and 1,283 Can2- like proteins, respectively—were then individually realigned with MAFFT and HMMbuild \(^{44}\) (HMMER v3.2.1) was used to generate HMM profiles from each alignment. The resulting profiles were used to search a local database
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 87, 884, 359]]<|/det|>
+of prokaryotic genomes from NCBI (downloaded on June 11, 2021) and list of sequences identified in BLAST search from previous steps. An initial search performed with these HMM profiles identified 1,442 Can1 and 5,431 Can2 homologs, which were manually filtered according to the presence of domains that define each protein, as well as the presence of conserved residues found in CARF and nuclease domains. The resulting set of 204 Can1 and 3,121 Can2 proteins were merged into a single file and aligned in MAFFT (LINSI option) for downstream phylogenetic analyses. Next, Trimal v1.4<sup>45</sup> was used to remove columns in the alignment comprised of \(\geq 70\%\) gaps. Thermostable homologs of Can1 and Can2 were annotated according to organisms that they are originated. ProtTest v3.4.2<sup>46</sup> was used to select an evolutionary model, and a phylogenetic tree was constructed in IQ- TREE v1.6.1<sup>47</sup> using the recommended model (i.e., LG+G+F). The phylogenetic tree was plotted using the ggTree package in R<sup>48</sup>.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 375, 403, 396]]<|/det|>
+## Phylogenetic analysis of NucC  
+
+<|ref|>text<|/ref|><|det|>[[111, 411, 884, 775]]<|/det|>
+A phylogenetic tree of NucC proteins was generated using the same methods as described above for Can1/Can2 proteins. Briefly, DELTA- BLAST and PSI- BLAST searches with previously identified NucC proteins<sup>32</sup> generated a list of closely related proteins (e- value cutoff of \(10^{- 4}\) and minimum \(50\%\) query coverage). The resulting dataset was filtered with CD- HIT v4.7 to remove redundant sequences. The resulting 1,230 NucC sequences were aligned with MAFFT (- - localpair - - maxiterate 1000), and poorly aligned and highly gapped sequences were removed with MaxAlign. The resulting set of 896 NucC sequences were re- aligned with MAFFT as previously described, and the resulting alignment was used to generate a NucC HMM profile which we used to search within prokaryotic genomes from NCBI. This search identified 1,774 hits, which were filtered according to the presence of restriction endonuclease- like domain (i.e., ID<sub>30</sub>EAK- motif containing), gate- loop and cA<sub>3</sub> binding domains and were aligned with MAFFT. The remaining NucC homologs were curated according to organisms they are originated from to identify thermostable NucC homologs. The resulting alignment of 1,510 NucC proteins with 21 thermostable homologs was used to generate a phylogenetic tree with FastTree v2.1.10<sup>49</sup> and was plotted using the ggTree package in R.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 790, 571, 811]]<|/det|>
+## QUANTIFICATION AND STATISTICAL ANALYSIS  
+
+<|ref|>text<|/ref|><|det|>[[115, 824, 847, 889]]<|/det|>
+All statistical analyses were performed in RStudio. Analysis of Variance Models (ANOVA) were calculated with aov() function in the stats R package. Multiple comparisons between positive samples and negative controls were performed using
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 879, 382]]<|/det|>
+Dunnett's test with multcomp R package. Reaction slopes were determined by extracting coefficients from linear models fitted to fluorescence data with \(lm()\) function in R. The linear regions of the fluorescence curves were identified using rolling regression with auto_rate() function in respR package. Patient samples \((n = 17)\) for viral detection assays were randomly selected from a sample database \((n = 858)\) with base R function sample(). Statistical threshold for detecting SARS- CoV- 2 in patient samples with Csm- based assay was set as mean of negative control \(\pm 2.33\) S.D., which captures \(98\%\) of variation in negative samples ( \(2\%\) false positive). Samples with z- score \(>2.33\) were considered positive for SARS- CoV- 2. Z- scores were calculated in R using following formula: \(\mathrm{Z} = (\mathrm{F}_{\mathrm{sample}} - \mu_{\mathrm{neg}}) / \sigma_{\mathrm{neg}}\) , where \(\mathrm{F}_{\mathrm{sample}}\) is fluorescence measured in a sample, \(\mu_{\mathrm{neg}}\) is mean of the negative control, \(\sigma_{\mathrm{neg}}\) is standard deviation of the negative control. Statistical significance levels used in the figures are \*\*\* \(p < 0.001\) , \*\* \(p < 0.01\) , and \* \(p < 0.05\) .
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[55, 90, 880, 907]]<|/det|>
+705 References706 1. Drain, P. K. Rapid Diagnostic Testing for SARS-CoV-2.707 https://doi.org/10.1056/NEJMcp2117115 (2022) doi:10.1056/NEJMCp2117115.708 2. Allan-Blitz, L. T. & Klausner, J. D. A Real- World Comparison of SARS- CoV- 2 Rapid709 Antigen Testing versus PCR Testing in Florida. Journal of Clinical Microbiology 59,710 (2021).711 3. Jessica L. Prince- Guerra, P. O. A. et al. Evaluation of Abbott BinaxNOW Rapid Antigen712 Test for SARS- CoV- 2 Infection at Two Community- Based Testing Sites — Pima County,713 Arizona, November 3–17, 2020.714 https://www.cdc.gov/mmwr/volumes/70/wr/pdfs/mm7003e3- H.pdf.715 4. Kaminski, M. M., Abudayyeh, O. O., Gootenberg, J. S., Zhang, F. & Collins, J. J.716 CRISPR- based diagnostics. Nature Biomedical Engineering 2021 5:7 5, 643–656 (2021).717 5. Abudayyeh, O. O. & Gootenberg, J. S. CRISPR diagnostics. Science 372, 914–915718 (2021).719 6. Pardee, K. et al. Rapid, Low- Cost Detection of Zika Virus Using Programmable720 Biomolecular Components. Cell 165, 1255–1266 (2016).721 7. Jiao, C. et al. Noncanonical crRNAs derived from host transcripts enable multiplexable722 RNA detection by Cas9. Science 372, 941–948 (2021).723 8. Gootenberg, J. S. et al. Nucleic acid detection with CRISPR- Cas13a/C2c2. Science 356,724 438–442 (2017).725 9. Chen, J. S. et al. CRISPR- Cas12a target binding unleashes indiscriminate single-726 stranded DNase activity. Science (New York, N.Y.) 360, 436 (2018).727 10. Kazlauskiene, M., Kostiuk, G., Venclovas, Č., Tamulaitis, G. & Siksnys, V. A cyclic728 oligonucleotide signaling pathway in type III CRISPR- Cas systems. Science 357, 605–729 609 (2017).730 11. Kazlauskiene, M., Tamulaitis, G., Kostiuk, G., Venclovas, Č. & Siksnys, V.731 Spatiotemporal Control of Type III- A CRISPR- Cas Immunity: Coupling DNA Degradation732 with the Target RNA Recognition. Molecular cell 62, 295–306 (2016).733 12. Niewoehner, O. et al. Type III CRISPR–Cas systems produce cyclic oligoadenylate734 second messengers. Nature 2017 548:7669 548, 543–548 (2017).735 13. Athukoralage, J. S. et al. The dynamic interplay of host and viral enzymes in type iii736 crisp- mediated cyclic nucleotide signalling. eLife 9, (2020).737 14. Makarova, K. S. et al. Evolutionary and functional classification of the CARF domain738 superfamily, key sensors in prokaryotic antivirus defense. Nucleic Acids Research 48,739 8828–8847 (2020).740 15. Rostol, J. T. & Marraffini, L. A. Non- specific degradation of transcripts promotes plasmid741 clearance during type III- A CRISPR–Cas immunity. Nature Microbiology 2019 4:4 4, 656–742 662 (2019).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[55, 88, 879, 145]]<|/det|>
+743 16. Santiago- Frangos, A. et al. Intrinsic signal amplification by type III CRISPR- Cas systems provides a sequence- specific SARS- CoV- 2 diagnostic. Cell Reports Medicine 2, 100319 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[55, 150, 880, 204]]<|/det|>
+746 17. Steens, J. A. et al. SCOPE enables type III CRISPR- Cas diagnostics using flexible targeting and stringent CARF ribonuclease activation. Nature Communications 2021 12:1 12, 1- 12 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[55, 210, 830, 248]]<|/det|>
+749 18. Sridhara, S., Goswami, H. N., Whyms, C., Dennis, J. H. & Li, H. Virus detection via programmable Type III- A CRISPR- Cas systems. doi:10.1038/s41467- 021- 25977- 7.  
+
+<|ref|>text<|/ref|><|det|>[[55, 254, 866, 309]]<|/det|>
+751 19. Smalakyte, D. et al. Type III- A CRISPR- associated protein Csm6 degrades cyclic hexa- adenylate activator using both CARF and HEPN domains. Nucleic Acids Research 48, 9204- 9217 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[55, 315, 876, 370]]<|/det|>
+754 20. Athukoralage, J. S., Graham, S., Gruschow, S., Rouillon, C. & White, M. F. A Type III CRISPR Ancillary Ribonuclease Degrades Its Cyclic Oligoadenylate Activator. Journal of Molecular Biology 431, 2894- 2899 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[55, 376, 870, 431]]<|/det|>
+757 21. Athukoralage, J. S., Rouillon, C., Graham, S., Gruschow, S. & White, M. F. Ring nucleases deactivate type III CRISPR ribonucleases by degrading cyclic oligoadenylate. Nature 2018 562:7726 562, 277- 280 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[55, 437, 863, 493]]<|/det|>
+760 22. Garcia- Doval, C. et al. Activation and self- inactivation mechanisms of the cyclic oligoadenylate- dependent CRISPR ribonuclease Csm6. Nature Communications 2020 11:1 11, 1- 9 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[55, 499, 861, 555]]<|/det|>
+763 23. Jia, N., Jones, R., Yang, G., Ouerfelli, O. & Patel, D. J. CRISPR- Cas III- A Csm6 CARF Domain Is a Ring Nuclease Triggering Stepwise cA4 Cleavage with ApA>p Formation Terminating RNase Activity. Molecular Cell 75, 944- 956. e6 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[55, 560, 876, 617]]<|/det|>
+766 24. Gruschow, S., Gruschow, G., Adamson, C. S. & White, M. F. Specificity and sensitivity of an RNA targeting type III CRISPR complex coupled with a NucC endonuclease effector. Nucleic Acids Research (2021) doi:10.1093/NAR/GKAB1190.  
+
+<|ref|>text<|/ref|><|det|>[[55, 623, 870, 661]]<|/det|>
+769 25. Hale, C. R. et al. RNA- Guided RNA Cleavage by a CRISPR RNA- Cas Protein Complex. Cell 139, 945- 956 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[55, 667, 850, 705]]<|/det|>
+771 26. Rouillon, C., Athukoralage, J. S., Graham, S., Gruschow, S. & White, M. F. Control of cyclic oligoadenylate synthesis in a type III CRISPR system. eLife 7, (2018).  
+
+<|ref|>text<|/ref|><|det|>[[55, 711, 880, 748]]<|/det|>
+773 27. Molina, R. et al. Structure of Csex1- cOA4 complex reveals the basis of RNA decay in Type III- B CRISPR- Cas. Nature Communications 2019 10:1 10, 1- 14 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[55, 754, 846, 792]]<|/det|>
+775 28. Liu, T. Y. et al. Accelerated RNA detection using tandem CRISPR nucleases. Nature Chemical Biology 2021 17:9 17, 982- 988 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[55, 798, 808, 835]]<|/det|>
+777 29. Rostol, J. T. et al. The Card1 nuclease provides defence during type III CRISPR immunity. Nature 2021 590:7847 590, 624- 629 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[55, 841, 850, 880]]<|/det|>
+779 30. Zhu, W. et al. The CRISPR ancillary effector Can2 is a dual- specificity nuclease potentiating type III CRISPR defence. Nucleic Acids Research 49, 2777- 2789 (2021).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[55, 90, 878, 145]]<|/det|>
+781 31. McMahon, S. A. et al. Structure and mechanism of a Type III CRISPR defence DNA 782 nuclease activated by cyclic oligoadenylate. Nature Communications 2020 11:1 11, 1–11 783 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[55, 150, 876, 204]]<|/det|>
+784 32. Ye, Q. et al. HORMA Domain Proteins and a Trip13-like ATPase Regulate Bacterial 785 cGAS-like Enzymes to Mediate Bacteriophage Immunity. Molecular Cell 77, 709-722.e7 786 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[55, 210, 878, 247]]<|/det|>
+787 33. Lau, R. K. et al. Structure and Mechanism of a Cyclic Trinucleotide-Activated Bacterial 788 Endonuclease Mediating Bacteriophage Immunity. Molecular Cell 77, 723-733.e6 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[55, 253, 870, 308]]<|/det|>
+789 34. Batéjat, C., Grassin, Q., Manuguerra, J.-C. & Leclercq, I. Heat inactivation of the severe 790 acute respiratory syndrome coronavirus 2. Journal of Biosafety and Biosecurity 3, 1 791 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[55, 314, 870, 369]]<|/det|>
+792 35. Liu, T. Y., Iavarone, A. T. & Doudna, J. A. RNA and DNA Targeting by a Reconstituted 793 Thermus thermophilus Type III-A CRISPR-Cas System. PLOS ONE 12, e0170552 794 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[55, 376, 878, 414]]<|/det|>
+795 36. East-Seletsky, A. et al. Two distinct RNase activities of CRISPR-C2c2 enable guide-RNA 796 processing and RNA detection. Nature 2016 538:7624 538, 270–273 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[55, 420, 875, 458]]<|/det|>
+797 37. Larremore, D. B. et al. Test sensitivity is secondary to frequency and turnaround time for 798 COVID-19 screening. Science Advances 7, (2021).  
+
+<|ref|>text<|/ref|><|det|>[[55, 464, 875, 501]]<|/det|>
+799 38. Liu, L. et al. The Molecular Architecture for RNA-Guided RNA Cleavage by Cas13a. Cell 800 170, 714-726.e10 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[55, 508, 875, 565]]<|/det|>
+801 39. Lowey, B. et al. CBASS Immunity Uses CARF-Related Effectors to Sense 3 0-5 0-and 2 802 0-5 0-Linked Cyclic Oligonucleotide Signals and Protect Bacteria from Phage Infection In 803 Brief. Cell 182, 38–49 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[55, 571, 840, 609]]<|/det|>
+804 40. Gootenberg, J. S. et al. Multiplexed and portable nucleic acid detection platform with 805 Cas13, Cas12a and Csm6. Science 360, 439–444 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[55, 615, 840, 653]]<|/det|>
+806 41. Li, W. & Godzik, A. Cd-hit: a fast program for clustering and comparing large sets of 807 protein or nucleotide sequences. Bioinformatics 22, 1658–1659 (2006).  
+
+<|ref|>text<|/ref|><|det|>[[55, 659, 861, 715]]<|/det|>
+808 42. Katoh, K. & Standley, D. M. MAFFT Multiple Sequence Alignment Software Version 7: 809 Improvements in Performance and Usability. Molecular Biology and Evolution 30, 772– 810 780 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[55, 721, 880, 760]]<|/det|>
+811 43. Gouveia-Oliveira, R., Sackett, P. W. & Pedersen, A. G. MaxAlign: Maximizing usable data 812 in an alignment. BMC Bioinformatics 8, 1–8 (2007).  
+
+<|ref|>text<|/ref|><|det|>[[55, 766, 820, 803]]<|/det|>
+813 44. Finn, R. D., Clements, J. & Eddy, S. R. HMMER web server: interactive sequence 814 similarity searching. Nucleic Acids Research 39, W29–W37 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[55, 810, 866, 865]]<|/det|>
+815 45. Capella-Gutierrez, S., Silla-Martinez, J. M. & Gabaldón, T. trimAl: a tool for automated 816 alignment trimming in large-scale phylogenetic analyses. Bioinformatics 25, 1972–1973 817 (2009).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[52, 88, 880, 120]]<|/det|>
+818 46. Darriba, D., Taboada, G. L., Doallo, R. & Posada, D. ProtTest 3: fast selection of best-fit models of protein evolution. Bioinformatics 27, 1164–1165 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[55, 130, 861, 167]]<|/det|>
+820 47. Minh, B. Q. et al. IQ-TREE 2: New Models and Efficient Methods for Phylogenetic Inference in the Genomic Era. Molecular Biology and Evolution 37, 1530–1534 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[56, 177, 879, 231]]<|/det|>
+822 48. Yu, G., Lam, T. T. Y., Zhu, H. & Guan, Y. Two Methods for Mapping and Visualizing Associated Data on Phylogeny Using Ggtree. Molecular Biology and Evolution 35, 3041–3043 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[56, 240, 880, 277]]<|/det|>
+825 49. Price, M. N., Dehal, P. S. & Arkin, A. P. FastTree 2 – Approximately Maximum-Likelihood Trees for Large Alignments. PLOS ONE 5, e9490 (2010).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[123, 133, 839, 352]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[126, 360, 868, 811]]<|/det|>
+<center>Fig.1: Type III CRISPR-based RNA concentration enhances detection. a Schematic of Type III CRISPR-based RNA concentration. RNase-dead Type III CRISPR complex from Thermus thermophilus (e.g., TtCsmCsm3-D34A) is added to a sample to bind complementary "Target" RNA. The His-tagged complex is concentrated using nickel-derivatized magnetic beads and a magnet. b Sequence-specific RNA enrichment with TtCsmCsm3-D34A complex was tested using \(25~\mathrm{nM}^{32}\mathrm{P}5'\) -end labeled RNA. Target and non-target RNA fragments were mixed with \(125~\mathrm{nM}\) TtCsmCsm3-D34A complex, incubated at \(65^{\circ}\mathrm{C}\) for 20 min prior to concentration of the His-tagged complex with nickel-derivatized magnetic beads. After the pull-down, phenol-chloroform extracted RNAs from the supernatants and the Csm-beads were resolved using UREA-PAGE. c Csm-based direct RNA detection using \(3\mu \mathrm{L}\) of sample is compared to an assay with an additional RNA capture and concentration step. Magnetic beads decorated with TtCsmCsm3-D34A are added to the sample. After concentrating beads with a magnet, the supernatant is decanted. The pellet is then resuspended in a small volume of the reaction buffer containing ATP to activate polymerase activity of Cas10. Polymerization products (e.g., \(cA_{3}\) and \(cA_{4}\) ) are used for the downstream detection assays. d TtCsmCsm3-D34A polymerization reactions were performed with \(\alpha -32\mathrm{P}\) -ATP as shown in c and products were resolved using thin-layer chromatography (TLC). Black arrow shows migration of solvent in the TLC plate. Bands were annotated using chemically synthesized standards (Supplementary Fig. 1d). \(3\mu \mathrm{L}\) (-RNA capture) or \(120\mu \mathrm{L}\) (+RNA capture) of SARS-CoV-2 N-gene RNA \((10^{10}\) copies/μL) diluted in total human RNA (293T cells) were used for reactions. e TtCsm6-based fluorescent readout (top panel) is used for detection of \(cA_{4}\) generated by TtCsmCsm3-D34A with (red bars) or without RNA capture step (blue bars) as shown in panel c. SARS-CoV-2 N-gene RNA diluted in total human RNA (HEK 293T cells) was used as a target. Fluorescence was measured with qPCR instrument and normalized to the no target control (NTC, HEK 293T RNA only, dashed line). In each assay, means (n=3) were compared with one-way ANOVA. Pairwise comparisons between target RNA dilutions and NTC were performed using post hoc Dunnett's test. Data are shown as mean \(\pm\) SD. \(*p<\) 0.05; \(**p<\) 0.005; \(***p<\) 0.001; \(****p<\) 0.0001. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[125, 90, 850, 545]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[127, 550, 872, 965]]<|/det|>
+<center>Fig.2: Can1 and Can2 ancillary nucleases cleave RNA or DNA in an activator-dependent manner. a Domain organization of Can1 and Can2 proteins. Can2 proteins have two domains – CARF and PD-(D/E)XK superfamily nuclease domain. Can1 is predicted to be derived from Can2 by gene duplication<sup>30</sup>. NLD – nuclease-like domain. b Maximum-likelihood phylogeny of 204 Can1 (CARF2 and PD-(D/E)XK nuclease domain) and 3,121 Can2 proteins. Previously studied effectors are underlined on the tree. *, effectors chosen for purification and in vitro experiments. c Plasmid (15 nM), ssRNA (425 nM), and ssDNA (425 nM) cleavage assay with TtCan1 (200 nM) in the presence of \(\mathrm{cA_3}\) or \(\mathrm{cA_4}\) (20 nM). The reactions were incubated 15 min at \(60^{\circ}\mathrm{C}\) . d Cleavage assays with AaCan2 (200 nM) in in the presence \(\mathrm{cA_4}\) or \(\mathrm{cA_3}\) (20 nM). Assays were performed with 15 nM plasmid DNA (left), 425 nM ssRNA or ssDNA (right) for 15 min at \(55^{\circ}\mathrm{C}\) . e TtCsm6 (300 nM) and AaCan2 (300 nM) cleavage assays with fluorescent ssRNA reporter (top) in the presence of varying \(\mathrm{cA_4}\) activator concentrations (shown with colors). Data is shown as the mean (center line) of three replicates ± S.D. (ribbon). The optimal fluorescent reporter (top) was determined using RNA library screen in Supplementary Fig. 4. f, g TtCsm RNA detection assays coupled with TtCsm6- (f) and AaCan2-based (g) readouts were performed using samples with target RNA concentrations ranging from \(10^{7}\) to \(10^{2}\) copies/μL. Samples were prepared by spiking IVT fragments of SARS-CoV-2 N gene into total RNA extracted from nasopharyngeal swab patient sample negative for SARS-CoV-2. Cleavage of fluorescent RNA reporter was detected by measuring fluorescence every 10 sec in a real-time PCR instrument (left). Data were plotted as mean of 4 replicates. Simple linear regression was used to calculate slopes for linear regions of the curves. Bars show mean values \((n = 4) \pm \mathrm{S.E.M}\) . (right). Data was analyzed with one-way ANOVA followed by multiple comparisons to NTC sample using one-tailed post-hoc Dunnett's test. *** \(p < 0.001\) ; ** \(p < 0.01\) ; * \(p < 0.05\) . </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[182, 95, 828, 537]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[125, 546, 863, 927]]<|/det|>
+<center>Fig.3: Incorporation of \(\mathbf{cA}_3\) -activated nucleases into Csm-based RNA detection assay. a The target bound TtCsm complex primarily generates \(\mathrm{cA_4}\) and \(\mathrm{cA_3}\) . Schematics summarizes \(\mathrm{cA_4}\) - and \(\mathrm{cA_3}\) -dependent activities of nucleases biochemically tested. N/D - not detected; Asterisk (\*) indicates nucleases that have sequences preferences (Supplementary Fig. 4). b Left panel: CtNucC (15 nM) is activated by \(\mathrm{cA_3}\) (20 nM) and cleaves plasmid DNA into short fragments in 15 min. Right panel: The deep sequencing of DNA fragments generated after 5 min of incubation with CtNucC revealed the preferential cleavage sites (ANNT). The reduced sequencing depth at the cut site is consistent with a cleavage mechanism producing \(3^{\prime}\) -overhangs that are removed by T4 DNA polymerase when sequencing library is prepared. c CtNucC (300 nM), TtCan1 (300 nM) and AaCan2 (300 nM) cleavage assays with fluorescent dsDNA reporter across eight concentrations of \(\mathrm{cA_3}\) (shown with colors). Data is shown as mean (center line) of three replicates \(\pm\) S.E.M. (ribbon). d TtCsm RNA detection assays coupled with AaCan2 (ssRNA reporter), CtNucC (dsDNA reporter) and combination of AaCan2 and CtNucC (both reporters). Reactions were performed using samples with target RNA concentrations ranging from \(10^{7}\) to \(10^{2}\) copies/μL. Samples were prepared by spiking IVT fragment of SARS-CoV-2 N gene in total RNA of SARS-CoV-2 negative nasal swab. Cleavage of the fluorescent reporter was detected by measuring fluorescence every 10 sec in a real-time PCR instrument. Simple linear regression was used to determine slopes for 3 replicates. See Supplementary Fig. 8 for fluorescent curves used in the analysis. Data were plotted as mean \((n = 3)\pm \mathrm{S.D}\) and analyzed with one-way ANOVA. All samples were compared to the non-target RNA control (NTC) using one-tailed post-hoc Dunnett's test. \(***p< 0.001\) ; \(**p< 0.01\) ; \(*p< 0.05\) </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[123, 118, 875, 512]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[125, 520, 864, 800]]<|/det|>
+<center>Fig. 4: TtCsm-based RNA capture directly detects SARS-CoV-2 in clinical samples. a Seventeen SARS-CoV-2 positive (red lines) and six negative (blue lines) RNA samples were tested with TtCsm-AaCan2 detection assay with and without upstream RNA capture. Dots show timepoints that were used to analyze type III detection results. Error bars show mean fluorescence in negative samples \((n = 6)\pm 2.33\) S.D. Reactions that generated signal higher than upper bound of this interval were considered positive for SARS-CoV-2 RNA. b Scatter plot showing distribution of Ct values (N1 CDC primers) of RNA samples tested in a. Red dots show samples that tested positive in type III detection, blue shows samples that tested negative. c Schematic of TtCsm-based RNA capture assay from nasopharyngeal swab coupled with AaCan2-based fluorescent detection. d Nasopharyngeal swab sample positive for SARS-CoV-2 (RT-qPCR Ct = 13.6) was used to make 10-fold serial dilutions in a negative nasopharyngeal swab (Ct > 40). Total of \(120\mu \mathrm{L}\) of the sample was used for direct detection with TtCsm-based RNA capture assay depicted in c. Bars show mean values \((n =\) 3) \(\pm\) S.E.M. of the reaction slopes calculated using simple linear regression (Supplementary Fig. 9c). All slopes were compared to the negative control (NTC) with one-way ANOVA and post-hoc one-tailed Dunnett's test. \\*\\*\\* \(p< 0.001\) ; \\*\\* \(p< 0.01\) ; \\* \(p< 0.05\) </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[44, 93, 765, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 572, 150]]<|/det|>
+- Nemudraiaetal.2022Supplementarywithinlenumbers.pdf
+
+<--- Page Split --->

preprint/preprint__2ad30f94c3b90c96475bc36c1aff17d563b878206dfefe71283989f76bd09092/images_list.json

@@ -0,0 +1,107 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        150,
+        517,
+        620
+      ]
+    ],
+    "page_idx": 31
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        260,
+        636,
+        864
+      ]
+    ],
+    "page_idx": 32
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        106,
+        592,
+        435
+      ]
+    ],
+    "page_idx": 34
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        105,
+        639,
+        490
+      ]
+    ],
+    "page_idx": 35
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        240,
+        639,
+        660
+      ]
+    ],
+    "page_idx": 36
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6.",
+    "footnote": [],
+    "bbox": [
+      [
+        118,
+        98,
+        639,
+        720
+      ]
+    ],
+    "page_idx": 38
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_7.jpg",
+    "caption": "Figure 7.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        100,
+        636,
+        450
+      ]
+    ],
+    "page_idx": 40
+  }
+]

preprint/preprint__2ae36470298dc21812c2daf83a1347cd8d7520177c40d85a6ff8aa616b82b269/images_list.json

@@ -0,0 +1,152 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "FIG. 1. The theorem of universal skin effect. (a) represents the Brillouin zone. (b)(d) shows that when the spectral area of \\(\\mathcal{H}(\\mathbf{k})\\) is nonzero, the system on generic geometries must have universal skin effect. (c)(e) shows that when the spectral area of \\(\\mathcal{H}(\\mathbf{k})\\) is zero, or forming one or several arcs on the complex plane, there is no skin effect under any geometry.",
+    "footnote": [],
+    "bbox": [
+      [
+        85,
+        65,
+        914,
+        285
+      ]
+    ],
+    "page_idx": 2
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "FIG. 2. Two manifestations of skin effect. One is the CSE (a)-(d), the other is the GDSE (e)-(h). In in (a)(b)(e)(f), the light blue regions represent the spectrum under periodic boundary, and the red points represent the eigenvalues under different open-boundary geometries. The spatial distributions of eigenstates \\(W(\\pmb {x})\\) are plotted in (c)(d)(g)(h). The GDSE disappears under square geometry (open boundary 1) in (g), and reappears under triangle geometry (open boundary 2) in (h).",
+    "footnote": [],
+    "bbox": [
+      [
+        94,
+        80,
+        912,
+        355
+      ]
+    ],
+    "page_idx": 3
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "FIG. 3. Two Dirac points (a) of a two-dimensional photonic crystal model are split into four exceptional points (b) upon adding non-Hermitian term, such as radiational loss. Correspondingly, the evolution of Gaussian wave packet with initial velocity at the center of a diamond geometry for each ten time intervals is shown in (c) with \\(\\gamma = 0\\) (Hermitian) and (d) with \\(\\gamma = 1 / 4\\) (GDSE).",
+    "footnote": [],
+    "bbox": [
+      [
+        88,
+        68,
+        916,
+        270
+      ]
+    ],
+    "page_idx": 4
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "FIG. S1. Some numerical examples of the theorem. (a-d) show an example having skin effect, and (e-h) show an example without skin effect. (a) and (e) show the corresponding hoping parameters of the hamiltonian shown in Eq. (S1). (b) and (f) show the periodic boundary spectrum. (c-d) and (g-h) show the distribution of \\(W(x)\\) in Eq. (S2) under different open boundary geometries.",
+    "footnote": [],
+    "bbox": [
+      [
+        128,
+        537,
+        875,
+        833
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "FIG. S2. (a) shows the equivalence relation between spectral area, spectral winding and the universal skin effect. Each equivalence relation is sufficient and necessary. (b) illustrates the spectral winding for hamiltonian Eq. S13. Here the light blue region represents the periodic boundary spectrum, and the paths on BZ corresponds to the spectral loops (or arcs) on the complex plane with the same color.",
+    "footnote": [],
+    "bbox": [
+      [
+        128,
+        66,
+        872,
+        214
+      ]
+    ],
+    "page_idx": 7
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "FIG. S3. (a) shows the periodic-boundary spectrum of Eq. (S25) with gray color, and the pre-images of \\(E_{0} = 1 + i\\) (red point in (a)) are the four red points in (b). The periodic-boundary spectrum of Eq. (S26) is the gray line in (c), and \\(\\mathbf{k}(E_{0} = 1 + i)\\) is plotted by the red lines in (d).",
+    "footnote": [],
+    "bbox": [
+      [
+        105,
+        63,
+        900,
+        208
+      ]
+    ],
+    "page_idx": 10
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_3.jpg",
+    "caption": "FIG. S4. The distribution of \\(W(x)\\) for Hamiltonian Eq. (S27) with different parameters and on different geometries. The system size is \\(L_{x} = L_{y} = 25\\) . The probability density is proportional to the opacity of the red color. (a) shows the corner-skin effect with \\(t_{1} = t_{2} = 1\\) , \\(w = 0\\) ; (b) shows line skin with \\(t_{1} = 1\\) , \\(t_{2} = w = 0\\) ; (c) has no skin effect with \\(t_{1} = t_{2} = 0\\) , \\(w = 1\\) , and geometry-dependent-skin effect appears in (d) under triangle and diamond geometries.",
+    "footnote": [],
+    "bbox": [
+      [
+        125,
+        61,
+        875,
+        228
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_4.jpg",
+    "caption": "FIG. S5. The norm squared of all wave functions of the Hamiltonian Eq. (S13) on square lattice (a) and parallelogram lattice (b) is plotted. The system size is chosen as \\(L_{x} = L_{y} = 25\\) . The volume law is shown in (c), in which blue line represents \\(N_{skin} \\propto V\\) , gray line \\(N_{skin} \\propto \\sqrt{V}\\) and black line \\(N_{skin} \\propto 1\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        128,
+        66,
+        876,
+        240
+      ]
+    ],
+    "page_idx": 15
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_5.jpg",
+    "caption": "FIG. S6. The periodic-boundary spectrum of the photonic crystal model is shown in (a)(c) with light blue color. Under the square geometry with the systems size \\(L_{x} = L_{y} = 31\\) , the eigenvalues (red points) and the norm squared of all wave functions are shown in (a) and (b), respectively. Under the diamond geometry with the systems size \\(L_{x} = L_{y} = 45\\) , the eigenvalues (red points) and the norm squared of all wave functions are plotted in (c) and (d), respectively.",
+    "footnote": [],
+    "bbox": [
+      [
+        231,
+        63,
+        777,
+        375
+      ]
+    ],
+    "page_idx": 16
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_6.jpg",
+    "caption": "FIG. S7. Two Weyl points (a) of a three-dimensional Weyl semimetal are expanded into two exceptional rings (b) after the addition of non-Hermitian perturbations. The spatial distribution of eigenstate is plotted in (c). The modulus square of the propagator from \\(i\\) to \\(o P_{i o}(\\omega)\\) and that from \\(o\\) to \\(i P_{o i}(\\omega)\\) , as functions of \\(\\omega\\) , are plotted with red color and dark cyan color in (d), respectively.",
+    "footnote": [],
+    "bbox": [
+      [
+        151,
+        68,
+        855,
+        248
+      ]
+    ],
+    "page_idx": 18
+  }
+]

preprint/preprint__2ae36470298dc21812c2daf83a1347cd8d7520177c40d85a6ff8aa616b82b269/preprint__2ae36470298dc21812c2daf83a1347cd8d7520177c40d85a6ff8aa616b82b269.mmd

@@ -0,0 +1,591 @@
+
+# Universal non-Hermitian skin effect in two and higher dimensions  
+
+Kai Zhang  Institute of Physics, Chinese Academy of Sciences https://orcid.org/0000- 0002- 9684- 7016  Zhensen Yang  Kavli Institute for Theoretical Sciences  Chen Fang (cfang@iphys.ac.cn)  Institute of Physics, Chinese Academy of Sciences https://orcid.org/0000- 0002- 9150- 8023  
+
+## Article  
+
+Keywords: skin effect, dimensions, non- Hermitian hamiltonian  
+
+Posted Date: August 18th, 2021  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 757556/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License.  Read Full License  
+
+Version of Record: A version of this preprint was published at Nature Communications on May 6th, 2022. See the published version at https://doi.org/10.1038/s41467- 022- 30161- 6.
+
+<--- Page Split --->
+
+# Universal non-Hermitian skin effect in two and higher dimensions  
+
+Kai Zhang, \(^{1,2}\) Zhesen Yang, \(^{3,*}\) and Chen Fang \(^{1,3,4,\dagger}\) \(^{1}\) Beijing National Laboratory for Condensed Matter Physics and Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China \(^{2}\) University of Chinese Academy of Sciences, Beijing 100049, China \(^{3}\) Kavli Institute for Theoretical Sciences, Chinese Academy of Sciences, Beijing 100190, China \(^{4}\) Songshan Lake Materials Laboratory, Dongguan, Guangdong 523808, China  
+
+Skin effect, experimentally discovered in one dimension, describes the physical phenomenon that on an open chain, an extensive number of eigenstates of a non- Hermitian hamiltonian are localized at the end(s) of the chain. Here in two and higher dimensions, we establish a theorem that the skin effect exists, if and only if periodic- boundary spectrum of the hamiltonian covers a finite area on the complex plane. This theorem establishes the universality of the effect, because the above condition is satisfied in almost every generic non- Hermitian hamiltonian, and, unlike in one dimension, is compatible with all spatial symmetries. We propose two new types of skin effect in two and higher dimensions: the corner- skin effect where all eigenstates are localized at one corner of the system, and the geometry- dependent- skin effect where skin modes disappear for systems of a particular shape, but appear on generic polygons. An immediate corollary of our theorem is that any non- Hermitian system having exceptional points (lines) in two (three) dimensions exhibits skin effect, making this phenomenon accessible to experiments in photonic crystals, Weyl semimetals, and Kondo insulators.  
+
+## INTRODUCTION  
+
+The study of non- Hermitian hamiltonians, which can be regarded as the effective description of dissipative processes, can be traced back to the investigation of alpha decay, where real and imaginary parts of the complex energy are related to the experimentally observed energy level and decay rate [1]. When a lattice system is coupled with environments and has dissipations, e.g. photonic crystals having radiational loss [2- 4] and electronic systems having finite quasiparticle lifetime [5, 6], the non- Hermitian band theory becomes a conceptually simple and efficient approach [7- 12].  
+
+Skin effect [13- 21], a phenomenon unique to the nonHermitian band theory, refers to the localization of eigenstates at the boundary, the number of which scales with the volume of the system. For example, in one dimension, all eigenstates of a non- Hermitian hamiltonian can be localized at the ends of a chain [13]. This suggests the failure of Bloch's theorem [22, 23], which states that eigenstates in the bulk are modulated plane waves. As Bloch's theorem plays a fundamental role in the development of condensed- matter physics [24], the emergence of skin effect indicates a new and possibly revolutionary direction. Especially, the skin effect has been experimentally observed in one- dimensional classical systems [25- 27], inspiring further studies on their higher dimensional generalizations [14, 28- 35]. However, a general theory for the higher- dimensional skin effect has not been established.  
+
+Apart form the skin effect, another focus topic in non- Hermitian band systems is the exceptional point (or line) [36- 45] that refers to stable point- type (or line- type) non- Hermitian band degeneracy in the Brillouin zone. At the exceptional point, not only eigenvalues but also eigenstates of the Bloch hamiltonian coalesce [37]. Many a novel phenomenon related to exceptional points has been predicted and observed [45- 50], such as the emergence of bulk- Fermi arc terminated at the exceptional points [5, 43]. Since the bulk- boundary correspondence plays a central role in the development of topological phases [51], it is natural to ask if there exists a nonHermitian bulk- boundary correspondence in bands having exceptional points, analogous to the surface Fermi arc in the Weyl semimetals in the Hermitian counterpart [52].   
+
+In this paper, we establish a theorem that reveals a universal bulk- boundary correspondence in two and higher dimensional non- Hermitian bands, as shown in Fig. 1. The "bulk" refers to the area of the spectrum of the hamiltonian on the complex plane with periodic boundary condition, and "boundary" the presence (absence) of the skin effect for open- boundary system of an arbitrary shape. The theorem states that the skin effect appears if and only if the spectral area is nonzero. This skin effect is "universal" for three reasons: (i) a randomly generated local non- Hermitian hamiltonian has the skin effect with probability one; (ii) the skin effect is, unlike in one dimension, compatible with all spatial symmetries and time- reversal symmetry; and (iii) it does not require any special geometry of the open- boundary system. We also propose two manifestations, restricted to two and higher dimensions, of the universal skin effect, i.e., the cornerskin effect and the geometry- dependent skin effect.  
+
+A surprising corollary of our theorem is that all stable exceptional points [8, 39, 41] imply the presence of skin effect. Because exceptional points have been either observed or proposed in meta- materials as well as in condensed matter, this corollary makes skin effect observable in known systems. We predict the geometry- dependent
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>FIG. 1. The theorem of universal skin effect. (a) represents the Brillouin zone. (b)(d) shows that when the spectral area of \(\mathcal{H}(\mathbf{k})\) is nonzero, the system on generic geometries must have universal skin effect. (c)(e) shows that when the spectral area of \(\mathcal{H}(\mathbf{k})\) is zero, or forming one or several arcs on the complex plane, there is no skin effect under any geometry. </center>  
+
+skin effect in the two- dimensional photonic crystal studied in Ref. [43], and propose to observe this effect in the anomalous dynamics of wave packets.  
+
+## THEOREM: AN EQUIVALENCE BETWEEN SPECTRAL AREA AND SKIN EFFECT  
+
+For generic one- dimensional non- Hermitian systems, the correspondence between the spectral shape and the skin effect has been derived [17, 18], i.e., when the Bloch spectrum is a loop- type (an arc- type), the skin effect appears (disappears).  
+
+Generalizing the correspondence to two dimensions, we note two main differences. One difference is in the periodic- boundary spectrum, \(E_{i}(\mathbf{k})\) , where \(i\) is the band index and \(\mathbf{k}\) the crystal momentum in the first Brillouin zone (BZ). Generally speaking, \(E_{i}(\mathbf{k})\) is a mapping from the \(d\) - dimensional torus to the complex plane, \(\mathbb{C}\) . When \(d = 1\) , the image of \(E_{i}(k)\) forms a loop; but when \(d > 1\) , the image is generically a continuum on \(\mathbb{C}\) , denoted by \(E_{i}(\mathrm{BZ})\) . The area covered by \(E_{i}(\mathrm{BZ})\) on the complex plane is called the spectral area, denoted by \(A_{i}\) . Another difference is in the variety of open- boundary condition. There is only one geometry for an open system in one dimension, i.e., an open chain; but there are an infinite number of geometries in two dimensions such as triangle, rectangle and pentagon.  
+
+Now we are ready to state the theorem of universal skin effect: in the thermodynamic limit, the skin effect is present in a hamiltonian having open boundary of arbitrary geometry, if the spectral area is nonzero ( \(A_{i} \neq 0\) ); vice versa, the skin effect is absent for all possible geometries, if the spectral area is zero ( \(A_{i} = 0\) ). As the periodic- boundary hamiltonian describes the dynamics in the bulk, the theorem relates a bulk property (spectral area) to a boundary one (existence of skin modes).  
+
+Fig. 1 shows some schematic examples. In the Supplementary Section I, a complete proof of the theorem has been provided.  
+
+The above theorem has implied the universality of skin effect in two and higher dimensions. As \(E_{i}(\mathrm{BZ})\) is the image of the \(d \geq 2\) - dimensional torus on the complex plane, it takes fine tuning of parameters to make \(A_{i} = 0\) for every \(i\) . In fact, for single- band hamiltonian, we can prove that \(A = 0\) if and only if \(\mathcal{H}(\mathbf{k}) = P[h(\mathbf{k})]\) , where \(h(\mathbf{k})\) is a Hermitian hamiltonian and \(P\) is a polynomial. In other words, a randomly generated non- Hermitian hamiltonian \(\mathcal{H}(\mathbf{k})\) has skin effect: the first meaning of universality. In previous studies, other types of skin effect, such as the line- skin and the high- order- skin effect, in two and higher dimensions have been proposed [28, 32]. These types all require the open- boundary system take a special geometry (usually a rectangle) and are hence considered special and non- generic. The skin effect when \(A_{i} \neq 0\) assumes a completely generic geometry of boundary: the second meaning of universality. The third meaning of universality lies in the fact that, unlike in one dimension, higher- dimensional skin effect is compatible with all spatial symmetries. A standing wave explanation for the above theorem is provided in the Supplementary Section II.  
+
+## THE CORNER-SKIN AND THE GEOMETRY-DEPENDENT-SKIN EFFECT  
+
+While the theorem shows that the skin effect is universal, it does not specify what skin modes look like in higher dimensions. Here, we report two types of the universal skin effect, the corner- skin effect (CSE) and the geometry- dependent skin effect (GDSE). Note that the CSE we mentioned in this paper has the nature of non- reciprocity, similar to the one- dimensional skin effect.
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>FIG. 2. Two manifestations of skin effect. One is the CSE (a)-(d), the other is the GDSE (e)-(h). In in (a)(b)(e)(f), the light blue regions represent the spectrum under periodic boundary, and the red points represent the eigenvalues under different open-boundary geometries. The spatial distributions of eigenstates \(W(\pmb {x})\) are plotted in (c)(d)(g)(h). The GDSE disappears under square geometry (open boundary 1) in (g), and reappears under triangle geometry (open boundary 2) in (h). </center>  
+
+The hamiltonian of the example for CSE is  
+
+\[\begin{array}{r}\mathcal{H}(\mathbf{k}) = [5(\cos k_x + \cos 2k_x) - i(\sin k_x + 3\sin 2k_x)\\ +5\cos k_y + i\sin k_y] / 2, \end{array} \quad (1)\]  
+
+The hamiltonian of the example for GDSE reads  
+
+\[J_{\alpha}[n] = \sum_{i}\oint_{\mathrm{BZ}}dk^{d}n(E_{i},E_{i}^{*})\partial_{k_{\alpha}}E_{i}(\mathbf{k}) \quad (2)\]  
+
+under the periodic- boundary condition, where \(n(E,E^{*})\) is any smooth function [17]. The current functional is shown to vanish in two and three dimensions under point groups \(C_i\) , \(D_{2,3,4,6}\) , \(C_{2h,3h,4h,6h}\) , \(D_{2d,3d,2h,3h,4h,6h}\) , \(T\) , \(T_{d,h}\) , \(O\) and \(O_h\) . Therefore, the CSE is only compatible with point groups \(C_m\) and \(C_{2,3,4,6,2v,3v,4v,6v}\) (see details in the Supplementary Section III).  
+
+The hamiltonian of the example for GDSE reads  
+
+\[\mathcal{H}(\mathbf{k}) = 2\cos k_x + 2i\cos k_y. \quad (3)\]  
+
+Since the spectral area is nonzero, our theorem tells us that the system must have skin effect for generic geometry, such as a random polygon. However, an interesting phenomenon in this example is that the skin effect disappears under the square geometry due to the existence of two mirror symmetries shown in Fig. 2 (g). Once we choose other types of boundaries where mirror symmetries are broken, the skin effect reappears as shown in Fig. 2 (h). Since the emergence of the skin effect and the position of localization depend on the geometry, it is called the GDSE. Besides the distribution of the eigenstates, another feature of the GDSE is that area of the open- boundary spectrum seems to be the same as \(A_i\) . However, the corresponding density of states is dependent by the choice of geometry as shown in Fig. 2 (e)(f). We conjecture this is a universal phenomenon for the GDSE. In the Supplementary Section III, we have provided some additional examples to illustrate this new type of skin effect, and shown that the increase in the number of skin modes is proportional to the increase in the system volume. For GDSE, there is at least one spatial geometry such that skin modes vanish, and as such is mutually exclusive with CSE. Additionally, GDSE is compatible with all point groups, in contrast to CSE.
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>FIG. 3. Two Dirac points (a) of a two-dimensional photonic crystal model are split into four exceptional points (b) upon adding non-Hermitian term, such as radiational loss. Correspondingly, the evolution of Gaussian wave packet with initial velocity at the center of a diamond geometry for each ten time intervals is shown in (c) with \(\gamma = 0\) (Hermitian) and (d) with \(\gamma = 1 / 4\) (GDSE). </center>  
+
+## COROLLARY: SKIN EFFECT FROM EXCEPTIONAL POINTS  
+
+An immediate corollary of our theorem is that all lattice hamiltonians having stable exceptional points have universal skin effect, connecting two unique phenomena in the non- Hermitian band theory. Consider a stable exceptional point \(\mathbf{k}_0\) in two dimensions. Due to the branch point structure of exceptional point, the dispersion around \(\mathbf{k}_0\) can be expressed as \(E_{\pm}(\mathbf{k}) = \pm c_0\sqrt{q_x + c_1q_y} + O(|\mathbf{k} - \mathbf{k}_0|)\) , where \(q_{i = x,y}\) denotes a small derivation from exceptional point in \(x\) or \(y\) direction, that is, \(q_{i} = k_{i} - k_{0i}\) . Here \(c_0, c_1\) are nonzero complex numbers and \(c_1 \notin \mathbb{R}\) . Suppose the range of the expansion is \(r_0\) , then it is clear that \(A_{\pm} \geq |c_0|\pi r_0^2 /2 \neq 0\) . By the theorem, the system must have skin effect (see more details in the Supplementary Section IV).  
+
+Now we use the photonic crystal model that has been experimentally realized in Ref. [43] to demonstrate our corollary. The tight- binding model hamiltonian with periodic boundary can be written as  
+
+\[H(k) = \pmb {d}(k)\cdot \pmb {\sigma} - i\gamma /2(\sigma_0 - \sigma_z), \quad (4)\]  
+
+where \(\pmb {\sigma} = (\sigma_0,\sigma_x,\sigma_y,\sigma_z)\) is a vector of the Pauli matrices and \(\pmb {d}(k)\) is a vector with four components, that is, \(\pmb {d}(k) = \{\mu_0 - (t_2 + t_3)(\cos k_x + \cos k_y),t_1[1 - \cos k_x - \cos k_y + \cos (k_x - k_y)],t_1[\sin k_x - \sin k_y - \sin (k_x - k_y)],\mu_z + (t_2 - t_3)(\cos k_x - \cos k_y)\}\) . The parameters are chosen as follows, \((t_1,t_2,t_3,\mu_0,\mu_z) = (0.4, - 0.1,0.5,1.35, - 0.02)\) . As shown in Fig. 3 (a), in the Hermitian limit, i.e. \(\gamma = 0\) , the system has two Dirac points along the \(x\) - axis. When external dissipation or radiational loss is added, i.e., \(\gamma \neq 0\) , each Dirac point splits into two exceptional points shown in Fig. 3 (b), connected by the bulk Fermi arc. According to our theorem, the system must have the universal skin effect, more precisely, the GDSE, proved in the Supplementary  
+
+Section IV. The skin effect disappears under square geometry but reappears under diamond geometry shown in Fig. 3 (c)(d). In this case, the majority of the eigenstates are localized on the four edges.  
+
+As mentioned in the previous discussion, the appearance of the skin effect can be reflected in the dynamical properties. In order to show this, we simulate the time evolution of the wave packet starting at the center of the diamond geometry with an initial velocity perpendicular to one edge. Here the initial state is chosen to be Gaussian form \(|\psi_0\rangle = \mathcal{N}\exp [-(x - x_0)^2 /10 - (y - y_0)^2 /10 - i2x - i2y](1,1)^T\) , where \(\mathcal{N}\) is the normalization factor and \(x_0 = y_0 = 21\) is the center coordinate of the diamond geometry. We plot the corresponding normalized final states \(|\psi (t_f)\rangle = \mathcal{N}(t_f)e^{- i\mathcal{H}_{\mathrm{0BC}}t_f}|\psi_0\rangle\) for every ten time intervals, where \(\mathcal{H}_{\mathrm{0BC}}\) represents the open- boundary hamiltonian on the diamond geometry. As shown in Fig. 3 (c), in the Hermitian case, the center of the wave packets obeys the simple law of reflection: the center of the wave packet just bounces between the two edges while slowly dispersing with time. However, in the non- Hermitian case \((\gamma = 1 / 4)\) with GDSE, after several oscillations between two edges, the wave packet makes a side jump into the upper left corner as shown in Fig. 3 (d). This anomalous dynamical behavior is an experimental signature of GDSE.  
+
+We also propose the realization for CSE in a three- dimensional system with exceptional lines in the Supplementary Section IV. Experimentally, the non- reciprocity of the CSE can be detected in the two- point Green's function.  
+
+## DISCUSSION  
+
+Our work has built a bridge between two distinct phenomena that only exist in non- Hermitian systems, i.e.,
+
+<--- Page Split --->
+
+the exceptional points (lines) and the non- Hermitian skin effect, by establishing the correspondence between bulk (spectral area) and boundary (universal skin effect). We prove that the skin effect be universal and compatible with all spatial symmetries and reciprocity in two and higher dimensions. Due to the universality, it is expected that the skin effect is observable in a wide range of platforms, such as photonic crystals with natural radiational loss, acoustic meta- materials and circuit networks with lossy components such as resistors. Beyond these classical systems, the skin effect can also be realized in condensed matter, e.g., the heavy- fermion material with finite quasiparticle lifetime and the Weyl- exceptional- ring semimetal. The latter is realizable in Weyl semimetals made from inverting bands that have disparate effective masses, such as d- and f- bands.  
+
+One should be reminded, however, that the results in this paper assume the coherent dynamics of the constituent degrees of freedom, which is unlikely the case in macroscopic condensed- matter systems where the coherence length is shorter than the system size. On the contrary, for the systems where the system size and the coherent length are comparable, as in mesoscopic systems, we believe that the universal skin effect has a significant contribution to the transport properties, a subject for future exploration.
+
+<--- Page Split --->
+
+## Supplementary Materials  
+
+### I. THE PROOF OF THE THEOREM  
+
+In this section, we will prove the following theorem appeared in the main text:  
+
+Theorem: In the thermodynamic limit, the skin effect is present in a hamiltonian having open boundary of arbitrary shape, if the periodic- boundary spectral area of the same hamiltonian is nonzero; vice versa, the skin effect is absent for all possible shapes of open boundary, if the spectral area is zero.  
+
+Here the spectral area refers to the area of the region covered by the periodic boundary spectrum on the complex plane. In the following contents, we will first show some numerical verifications of the theorem, and then prove the theorem in two dimensional systems, and finally extend the proof to three- dimensional cases.  
+
+### A. Some numerical examples of the theorem  
+
+In this subsection, we provide some numerical examples of the theorem. In order to simplify the discussion, we consider the following single- band model  
+
+\[\mathcal{H}(\mathbf{k}) = \sum_{i,j}t_{ij}\beta_{x}^{-i}\beta_{y}^{-j},\qquad \beta_{x / y} = e^{ik_{x / y}}, \quad (S1)\]  
+
+where \(t_{ij}\) represents the hoping parameter that an electron or particle hopes from site \((m,n)\) to site \((m,n) + (i,j)\)  
+
+The two rows in Fig. S1 represent two different models. In the first one, the hoping parameters are shown in Fig. S1 (a), and the periodic boundary spectrum, i.e. \(E(\mathbf{k}) = \mathcal{H}(\mathbf{k})\) with \(\mathbf{k}\) belonging to the Brillouin zone (BZ), is shown in Fig. S1 (b). Here the color strength represents the cover times of \(E_{0}\in E(\mathbf{k})\) when \(\mathbf{k}\) sweeps over the whole BZ. One can notice that the spectral area of the first model is nonzero. As a result, the open boundary eigenstates show the localization behaviors, namely, the emergence of skin effect, as shown in Fig. S1 (c- d). In order to illustrate the  
+
+![](images/Figure_unknown_0.jpg)
+
+<center>FIG. S1. Some numerical examples of the theorem. (a-d) show an example having skin effect, and (e-h) show an example without skin effect. (a) and (e) show the corresponding hoping parameters of the hamiltonian shown in Eq. (S1). (b) and (f) show the periodic boundary spectrum. (c-d) and (g-h) show the distribution of \(W(x)\) in Eq. (S2) under different open boundary geometries. </center>
+
+<--- Page Split --->
+![](images/Figure_unknown_1.jpg)
+
+<center>FIG. S2. (a) shows the equivalence relation between spectral area, spectral winding and the universal skin effect. Each equivalence relation is sufficient and necessary. (b) illustrates the spectral winding for hamiltonian Eq. S13. Here the light blue region represents the periodic boundary spectrum, and the paths on BZ corresponds to the spectral loops (or arcs) on the complex plane with the same color. </center>  
+
+localization properties,  
+
+\[W(x) = \frac{1}{N}\sum_{n}|\psi_{n}(x)|^{2} \quad (S2)\]  
+
+is plotted in Fig. S1 (c- d), , where \(\psi_{n}(x)\) is the \(n\) - th normalized eigenstate of the open boundary hamiltonian, and \(N\) is the number of open boundary eigenstates. For the second example, since the spectral area is zero, as shown in Fig. S1 (f), there is no skin effect. Indeed, \(W(x)\) shown in Fig. S1 (g- h) are extended.  
+
+In the following subsection, we will prove the theorem in two- dimensional systems. Our strategy of the proof is illustrated in Fig. S2(a). The equivalence relation between "spectral area" and "the universal skin effect" is linked by "spectral winding" (see the following discussion).  
+
+### B. The proof of the theorem in two-dimensions  
+
+## 1. Spectral area and spectral winding  
+
+We begin with a two- dimensional single- band non- Hermitian model with periodic boundary in both \(x\) and \(y\) directions  
+
+\[\mathcal{H}(\mathbf{k}) = u(\mathbf{k}) + iv(\mathbf{k}), \quad (S3)\]  
+
+where \(u\) and \(v\) are real functions about \(\mathbf{k} = (k_{x},k_{y})\) . For any \(\mathbf{k}_{r}\in \mathrm{BZ}\) , one can define the following winding number  
+
+\[\nu (\mathbf{k}_{r}) = \oint_{\Gamma_{\mathbf{k}_{r}}}\frac{d\mathbf{k}}{2\pi i}\cdot \nabla_{\mathbf{k}}\log \operatorname *{det}[\mathcal{H}(\mathbf{k}) - E_{r}],\quad \mathbf{k}_{r}\in \mathrm{BZ}, \quad (S4)\]  
+
+where \(\Gamma_{\mathbf{k}_{r}}\) represents the infinitesimal counterclockwise loop enclosing \(\mathbf{k}_{r}\) . Here \(E_{r} = \mathcal{H}(\mathbf{k}_{r})\) represents the reference energy, which is shown in Fig. S2 (b) with red point. We note that for different \(\mathbf{k}_{r}\) , the reference energy is different. This topological invariant describes the spectral winding on the complex plane. As shown in Fig. S2 (b), if \(\nu (\mathbf{k}_{r})\) is nonzero, the image of \(\Gamma_{\mathbf{k}_{r}}\) , i.e. \(\mathcal{H}(\Gamma_{\mathbf{k}_{r}})\) , forms a closed loop that encloses \(E_{r}\) .  
+
+First we prove the following two statements,  
+
+1. if there are some \(\mathbf{k}_{r}\) points in the BZ with nonzero topological charge, the spectral area must be nonzero;  
+
+2. if all the \(\mathbf{k}_{r}\) points on the BZ have zero topological charge, the spectral area must be zero.  
+
+The above two statement can be represented by the equivalence relation between "spectral winding" and "spectral area" shown in Fig. S2 (a).  
+
+Based on the definition of winding number, the statement 1 is obvious. Therefore, we only need to prove the statement 2. In order to show this, we expand the hamiltonian at the point \(\mathbf{k}_{r}\equiv (k_{x}^{r},k_{y}^{r})\) as follows  
+
+\[\mathcal{H}(\mathbf{k}) - \mathcal{H}(\mathbf{k}_{r})\approx \partial_{x}\mathcal{H}(\mathbf{k}_{r})q_{x} + \partial_{y}\mathcal{H}(\mathbf{k}_{r})q_{y}, \quad (S5)\]
+
+<--- Page Split --->
+
+where \(q_{x},q_{y}\) are the displacements \(\mathbf{k} - \mathbf{k}_{r}\) in \(x,y\) directions respectively. For a generic \(\mathbf{k}_{r}\) point, the first derivative of \(\mathcal{H}\) does not vanish, i.e., the coefficients of \(q_{x}\) and \(q_{y}\) in Eq. (S5) cannot be zero at the same time. The reason is that in order to make \(\partial_{x}\mathcal{H}(\mathbf{k}_{r}) = \partial_{y}\mathcal{H}(\mathbf{k}_{r}) = 0\) , four independent real conditions, \(\mathrm{Re}[\partial_{x}\mathcal{H}(\mathbf{k}_{r})] = \mathrm{Im}[\partial_{x}\mathcal{H}(\mathbf{k}_{r})] =\) \(\mathrm{Re}[\partial_{y}\mathcal{H}(\mathbf{k}_{r})] = \mathrm{Im}[\partial_{y}\mathcal{H}(\mathbf{k}_{r})] = 0\) , need to be satisfied. However, in two dimensions, there are only two free parameters \((k_{x},k_{y})\) , which cannot satisfy the above four equations generally.  
+
+In order to calculate the topological charge of a generic \(\mathbf{k}_{r}\) point, according to Eq. (S3), one can define  
+
+\[C(\mathbf{k}_{r}) = \left( \begin{array}{cc}\partial_{x}u(\mathbf{k}_{r}) & \partial_{y}u(\mathbf{k}_{r})\\ \partial_{x}v(\mathbf{k}_{r}) & \partial_{y}v(\mathbf{k}_{r}) \end{array} \right), \quad (S6)\]  
+
+where the notation \(\partial_{x / y}\) refers to \(\partial /\partial_{k_{x / y}}\) . When \(\operatorname *{det}[C(\mathbf{k}_{r})] \neq 0\) , the topological charge of \(\mathbf{k}_{r}\) is the sign of the determinant of \(C(\mathbf{k}_{r})\) , expressed as  
+
+\[\nu (\mathbf{k}_{r}) = \operatorname {sgn}[\operatorname *{det}[C(\mathbf{k}_{r})]]. \quad (S7)\]  
+
+Therefore, a sufficient and necessary condition for the zero charge for each \(\mathbf{k} \in \mathrm{BZ}\) is  
+
+\[\operatorname *{det}[C(\mathbf{k})] = \partial_{x}u(\mathbf{k})\partial_{y}v(\mathbf{k}) - \partial_{y}u(\mathbf{k})\partial_{x}v(\mathbf{k}) = 0. \quad (S8)\]  
+
+A theorem (the corollary of theorem 13.2) in Ref. [53] tells us that if \(C(\mathbf{k}) \neq 0\) and \(\operatorname *{det}[C(\mathbf{k})] = 0\) , for an open set \(S\) , then, \(u(\mathbf{k})\) and \(v(\mathbf{k})\) have a functional dependent relation with \(\mathbf{k} \in S \subset \mathrm{BZ}\) . Applying the theorem to the entire BZ (except for some isolated points at which the first derivative vanishes), one can reexpress the single- band hamiltonian that satisfies Eq. (S8) as  
+
+\[\mathcal{H}(\mathbf{k}) = P[h(\mathbf{k})], \quad (S9)\]  
+
+where \(h(\mathbf{k})\) is a real and periodic function of \(\mathbf{k}\) , and \(P\) is a complex polynomial of \(h\) . Since \(h(\mathbf{k})\) is a real periodic function, its image must be an arc on the real axis, e.g. \(h(\mathbf{k}) \in [h_{1}, h_{2}]\) , where \(h_{1 / 2}\) are real numbers. Therefore, the image of \(P[h(\mathbf{k})]\) is also an arc on the complex plane. This completes the proof of the statement 2 for single- band models.  
+
+Generalizing the above discussion to the multi- band case, for each \(\mathbf{k}_{r} \in \mathrm{BZ}\) , the topological charge defined for the \(m\) - th band is  
+
+\[\begin{array}{l}{\nu_{m}(\mathbf{k}_{r}) = \oint_{\Gamma_{\mathbf{k}_{r}}}\frac{d\mathbf{k}}{2\pi i}\cdot \nabla_{\mathbf{k}}\log \operatorname *{det}[\mathcal{H}(\mathbf{k}) - E_{m}(\mathbf{k}_{r})]}\\ {= \sum_{n}\oint_{\Gamma_{\mathbf{k}_{r}}}\frac{d\mathbf{k}}{2\pi i}\cdot \nabla_{\mathbf{k}}\log [E_{n}(\mathbf{k}) - E_{m}(\mathbf{k}_{r})],} \end{array} \quad (S10)\]  
+
+where \(E_{m}(\mathbf{k}_{r})\) is the energy of the \(m\) - th band with the momentum \(\mathbf{k}_{r}\) . For the second equal sign in Eq. (S10), we have used \(\operatorname *{det}[\mathcal{H}(\mathbf{k}) - E_{m}(\mathbf{k}_{r})] = \prod_{n}[E_{n}(\mathbf{k}) - E_{m}(\mathbf{k}_{r})]\) . For Eq. (S10), if \(\mathbf{k}_{r}\) is not the degeneracy point, only \(n = m\) term in the summation has contributions to the topological charge. Therefore, the Eq. (S10) further becomes  
+
+\[\nu_{m}(\mathbf{k}_{r}) = \oint_{\Gamma_{\mathbf{k}_{r}}}\frac{d\mathbf{k}}{2\pi i}\cdot \nabla_{\mathbf{k}}\log [E_{m}(\mathbf{k}) - E_{m}(\mathbf{k}_{r})]. \quad (S11)\]  
+
+Using the similar approaches in the single- band case, one can conclude that, the the real and imaginary parts of \(E_{m}(\mathbf{k})\) are locally functional dependent on the neighbourhood of \(\mathbf{k} \in \mathrm{BZ}\) . As a result, the spectrum of \(E_{m}(\mathbf{k})\) must be an arc. The above conclusion applies for each band. So far, we have proved that if each \(\mathbf{k}\) point on the BZ has zero topological charge, the spectral area must be zero.  
+
+## 2. Spectral winding and the universal skin effect  
+
+In the above contents, we have proved the equivalence relation between "spectral winding" and "spectral area". Here, we will prove the equivalence condition between "spectral winding" and "universal skin effect", as shown in Fig. S2 (a).  
+
+We first notice that the BZ in two- dimensional systems can be covered by a set of straight lines of any slope, labeled as \(\{L_{s}\}\) . Here, the subscript \(s\) indicates the slope of the set \(\{L_{s}\}\) , and \(L_{s}\) represents a generic straight line belonging to \(\{L_{s}\}\) . For example, if we fix \(k_{y}(k_{x})\) and change \(k_{x}(k_{y})\) from 0 to \(2\pi\) , we get a horizontal (vertical) straight line
+
+<--- Page Split --->
+
+with the slope being \(0\) ( \(\infty\) ) on BZ, and the set of all horizontal or vertical straight lines ( \(\{L_{0}\}\) or \(\{L_{\infty}\}\) ) covers the entire BZ. Particularly, an inclined straight line goes out from one side of BZ and again enters from another side as shown in Fig. S2(b). Since the straight lines on the BZ are periodic, one can define the spectral winding number for each straight lines with respect to the prescribed reference energy \(E_{r}\) .  
+
+\[\nu (L_{s},E_{r}) = \oint_{L_{s}}\frac{d\mathbf{k}}{2\pi i}\cdot \nabla_{\mathbf{k}}\log \operatorname *{det}[\mathcal{H}(\mathbf{k}) - E_{r}]. \quad (S12)\]  
+
+Obviously, if all the \(\mathbf{k}\) points on the BZ have zero topological charge, \(\nu (L_{s},E_{r})\) must be zero for arbitrary \(L_{s}\) and \(E_{r}\) . Otherwise, one can always find some \(L_{s}\) , such that \(\nu (L_{s},E_{r})\) is nonzero. Next we briefly prove the latter statement. Assuming that each straight line in \(\{L_{0}\}\) and \(\{L_{\infty}\}\) has zero spectral winding, and there is a \(k_{r}\) point carrying nonzero topological charge on the BZ. For a generic inclined straight line, one can always find corresponding horizontal and vertical straight lines, such that together with the inclined straight line to form a closed path enclosing \(\mathbf{k}_{r}\) on BZ. Therefore, the closed path has nonzero winding number with respect to \(E_{r}\) . Due to the zero spectral winding of the horizontal and vertical straight lines as we assumed, hence a generic inclined straight line must have nonzero spectral winding number.  
+
+Let's use an example to show this. Consider the following hamiltonian  
+
+\[\mathcal{H}(k_{x},k_{y}) = 2\cos k_{x} + 2i\sin k_{y}. \quad (S13)\]  
+
+As shown in Fig. S2, the spectrum of the hamiltonian along each horizontal or vertical straight line (gray lines) on BZ has zero winding number with respect to any reference energy on the spectral area (lightblue square region). However, once we choose the straight line with darker cyan color, these three straight lines (two gray lines with the darker cyan line) together form a closed path that encloses \(\mathbf{k}_{r}\) . Therefore, the closed path has nonzero spectral winding number with respect to \(E_{r}\) . Due to the zero spectral winding of two gray lines, the darker cyan straight line must have nonzero winding number for \(E_{r}\) as illustrated in Fig. S2(b).  
+
+If the system has no skin effect under a specific parallelepiped open boundary, the winding number \(\nu_{m}(L_{m},E_{r})\) along each straight lines that are perpendicular to the boundary cut directions should be zero. (This is a conjecture that we cannot exactly proved currently. However, we believe this statement is true as all the numerical results we obtained obey this conclusion. Furthermore, it is also a natural generalization of the one- dimensional results [17, 18] and has been mentioned or applied in some recent works [16, 54]) More generally, if the system has no skin effect under any parallelepiped open boundary, then the spectral windings of all straight lines on BZ are required to be zero, which is satisfied when the spectral area of the system is zero. Therefore, nonzero spectral area means that there must be skin effect under certain open boundaries, namely, the existence of the universal skin effect.  
+
+### C. The proof of the theorem in three-dimensions  
+
+In this section, we extend the above proof of two- dimensional systems into three dimensions. We obtain the similar conclusion that nonzero spectral area is equivalent to the existence of the universal skin effect.  
+
+Consider a general three- dimensional single- band tight- binding hamiltonian, which consists of real- and imaginary- part functions  
+
+\[\mathcal{H}(k_{x},k_{y},k_{z}) = u(k_{x},k_{y},k_{z}) + i\nu (k_{x},k_{y},k_{z}). \quad (S14)\]  
+
+We choose a generic \(\mathbf{k}_{r}\) point and use its energy \(\mathcal{H}(\mathbf{k}_{r})\) as the reference energy. For a given reference energy \(E_{r}\) , we can obtain a one- dimensional curve in the three- dimensional BZ by solving the following two real equations,  
+
+\[\begin{array}{r l} & {u(k_{x},k_{y},k_{z}) = \mathrm{Re}(E_{r});}\\ & {v(k_{x},k_{y},k_{z}) = \mathrm{Im}(E_{r}).} \end{array} \quad (S15)\]  
+
+Each equation determines a surface, and the intersection of two surfaces is one- dimensional curve in three- dimensional BZ. The tangent direction of the curve at \(\mathbf{k}_{r}\) is perpendicular to the normal vector of the two surfaces at this \(\mathbf{k}_{r}\) point. The tangent vector at \(\mathbf{k}_{r}\) is expressed as  
+
+\[T_{\mathbf{k}_{r}} = \nabla u(\mathbf{k}_{r})\times \nabla v(\mathbf{k}_{r}), \quad (S16)\]  
+
+where \(\nabla u(\mathbf{k}_{r})\) represents the gradient of \(u\) . We choose the local coordinate system \((R^{3}\) space) with \(\mathbf{k}_{r}\) as the origin, and the gradient is reexpressed as  
+
+\[\nabla u(\mathbf{k}_{r}) = \partial_{x}u(\mathbf{k}_{r})q_{x} + \partial_{y}u(\mathbf{k}_{r})q_{y} + \partial_{z}u(\mathbf{k}_{r})q_{z}, \quad (S17)\]
+
+<--- Page Split --->
+![](images/Figure_unknown_2.jpg)
+
+<center>FIG. S3. (a) shows the periodic-boundary spectrum of Eq. (S25) with gray color, and the pre-images of \(E_{0} = 1 + i\) (red point in (a)) are the four red points in (b). The periodic-boundary spectrum of Eq. (S26) is the gray line in (c), and \(\mathbf{k}(E_{0} = 1 + i)\) is plotted by the red lines in (d). </center>  
+
+where \(q_{i} \equiv (\frac{\mathbf{k} - \mathbf{k}_{0}}{|\mathbf{k} - \mathbf{k}_{0}|})_{i}\) , \(\mathbf{k}\) and \(\mathbf{k}_{0}\) represent two vectors in the global coordinate system.  
+
+Next we expand the hamiltonian into Taylor series around the origin of the local coordinate system,  
+
+\[\mathcal{H}(\mathbf{k}) - \mathcal{H}(\mathbf{k}_{r}) = \sum_{i = x,y,z}\partial_{i}\mathcal{H}(\mathbf{k}_{r})q_{i} + o(|\mathbf{q}|), \quad (S18)\]  
+
+where the subscription \(i\) represents the partial differential to \(x,y,z\) . And \(q_{i}\) represents the deviation of \(k_{i}\) from \(k_{r,i}\) and the last term is the infinitesimal of higher order of \(|\mathbf{q}|\) . Obvious, the zero winding condition requires  
+
+\[\begin{array}{rl} & {\partial_{x}u(\mathbf{k}_{r})\partial_{y}v(\mathbf{k}_{r}) - \partial_{x}v(\mathbf{k}_{r})\partial_{y}u(\mathbf{k}_{r}) = 0;}\\ & {\partial_{x}v(\mathbf{k}_{r})\partial_{z}u(\mathbf{k}_{r}) - \partial_{x}u(\mathbf{k}_{r})\partial_{z}v(\mathbf{k}_{r}) = 0;}\\ & {\partial_{y}u(\mathbf{k}_{r})\partial_{z}v(\mathbf{k}_{r}) - \partial_{y}v(\mathbf{k}_{r})\partial_{z}u(\mathbf{k}_{r}) = 0,} \end{array} \quad (S19)\]  
+
+or equivalently,  
+
+\[T_{k_{r}} = 0. \quad (S20)\]  
+
+Next, we prove that if all \(\mathbf{k}\) points in three- dimensional BZ satisfy \(T_{k} = \mathbf{0}\) , then the entire 3D periodic- boundary spectrum must be an arc in the complex plane. We define a two- tuple function \(W(\mathbf{k}) := [u(\mathbf{k}) \quad v(\mathbf{k})]^{t}\) with three variables, the exterior derivative of the vector- valued function is expressed as  
+
+\[dW = \left( \begin{array}{ll}\partial_{x}u(\mathbf{k}) & \partial_{y}u(\mathbf{k}) & \partial_{z}u(\mathbf{k})\\ \partial_{x}v(\mathbf{k}) & \partial_{y}v(\mathbf{k}) & \partial_{z}v(\mathbf{k}) \end{array} \right). \quad (S21)\]  
+
+Eq. (S20) implies that the rank of \(dW\) less than 2 (the number of components of \(W\) ). To be precise, there are the following cases. (i.) Both the gradients of \(u\) and \(v\) are not zero vector, and they are linearly dependent on each other. (ii.) One of the gradients of \(u\) and \(v\) is zero vector. (iii.) Both the gradients of \(u\) and \(v\) are zero vector. In all these cases, we can obtain the final conclusion that \(u\) and \(v\) are linearly functional dependent on each other. Therefore, the spectrum must be arcs on the complex plane.  
+
+A 3D BZ can be divided into a series of plane systems, each plane corresponds to a two- dimensional subsystem. If the spectral area of a three- dimensional system is nonzero, then for each reference energy on the spectral area, its preimage (1D ring) has nonzero topological charge. Equivalently, the two- dimensional subsystem, of which the BZ (2D plane) has intersections with the ring, also has nonzero topological charge for the intersecting \(k\) points. Hence, the 2D subsystem has nonzero spectral area, and has the universal skin effect. Correspondingly, we come to the same conclusion in 3D systems that nonzero spectral area signifies the existence of the universal skin effect.  
+
+## II. A PHYSICAL EXPLANATION FOR THE THEOREM  
+
+Here, we use some examples to illustrate the intuition that motivates the theorem. Consider the following one- dimensional model  
+
+\[\mathcal{H}_{0}(k) = 2\cos k \quad (S22)\]
+
+<--- Page Split --->
+
+placed on a chain of length \(L\) . Under the periodic boundary condition, the two Bloch waves \(e^{ik_0x}\) and \(e^{- ik_0x}\) have the same energy \(E(k_0) = 2\cos k_0\) . When the system has open boundary condition, the Bloch wave \(e^{ik_0x}\) will be reflected to \(e^{- ik_0x}\) with a \(\pi\) - phase shift. Their linear superposition \(e^{ik_0x} - e^{- ik_0x}\) is an eigenstate with energy \(2\cos k_0\) that satisfies the zero boundary condition at \(x = 0, L\) , thus being an open- boundary eigenstate. When the system is added a momentum- dependent dissipation,  
+
+\[\mathcal{H}(k) = 2\cos k + i\sin k, \quad (S23)\]  
+
+the spectrum \(E(k)\) becomes complex and forms an ellipse in the complex plane. In this case, the degeneracy is broken, e.g. \(E(k) \neq E(- k)\) , which implies the open- boundary eigenstates are no longer the linear superposition of the extended Bloch waves. This implies the emergence of skin effect.  
+
+Extend the above arguments to two dimensions, and we can provide a physical explanation for the theorem proved in the section I.  
+
+Formally, we consider a single- band model  
+
+\[\mathcal{H}(\mathbf{k}) = \mathcal{H}_0(\mathbf{k}) + i\Gamma (\mathbf{k}). \quad (S24)\]  
+
+When the real and imaginary parts of which are functionally independent, the hamiltonian will have a non- zero spectral area. For a given eigenvalue \(E_0\) of the Bloch hamiltonian, by solving \(\mathcal{H}_0(\mathbf{k}) = \mathrm{Re}E_0\) and \(\Gamma (\mathbf{k}) = \mathrm{Im}E_0\) , one can obtain a finite set of pre- images of \(E_0\) , i.e, \(\mathbf{K}(E_0) = \{\mathbf{k}_1, \ldots , \mathbf{k}_m\}\) , which includes all Bloch waves having energy \(E_0\) . Now suppose that one of the Bloch waves \(\mathbf{k}_i \in \mathbf{K}(E_0)\) is incident on the boundary, depending on the normal direction of the boundary, the corresponding momentum of the reflected wave can be arbitrary. However, the number of elements of \(\mathbf{K}(E_0)\) is finite, and as such cannot support so many reflection channels. This failure of reflection mechanism at a generic boundary means the failure in forming an open boundary eigenstate from Bloch waves, which implies the emergence of skin effect under a generic open- boundary geometry. However, the spectrum collapses into an arc (zero spectral area) if the real and imaginary parts of the hamiltonian are functionally dependent, and the number of the corresponding solutions of \(\mathcal{H}(\mathbf{k}) = E_0\) is infinite. It means that there are infinite reflection channels to satisfy the open boundary of any shape, and an open boundary eigenstate can be formed from superimposing all Bloch- wave channels.  
+
+Concretely, we choose two examples to demonstrate the above arguments. The first example is  
+
+\[\mathcal{H}(\mathbf{k}) = 2\cos k_x + 2i\cos k_y, \quad (S25)\]  
+
+of which the spectral area is nonzero shown in Fig. S3(a). For a given eigenvalue \(E_0 = 1 + i\) , by solving \(2\cos k_x = 1\) and \(2\cos k_y = 1\) , we can obtain a finite set of pre- images of \(E_0\) , that is, \(\mathbf{K}(E_0) = \{k_1, k_2, k_3, k_4\}\) [red points in Fig. S3(b)]. The finite solutions of \(\mathcal{H}(\mathbf{k}) = E_0\) cannot support so many reflection channels, that is to say, cannot form an open- boundary eigenstate on a generic geometry by superimposing these Bloch waves specified by \(k_{i = 1,2,3,4}\) . Therefore, the hamiltonian Eq. (S25) has skin effect under open- boundary geometry of a generic shape. The second example reads  
+
+\[\mathcal{H}(\mathbf{k}) = 2\cos (k_x + k_y) + 2i\cos (k_x + k_y), \quad (S26)\]  
+
+the periodic- boundary spectrum of which is an arc [the gray line in Fig. S3(c)]. The set of pre- images of \(E_0 = 1 + i\) has infinite elements [the red lines in Fig. S3(d)], which means that there are infinite ways of superimposing these Bloch waves to satisfy the open boundary condition of any shape. Therefore, the hamiltonian Eq. (S26) has no skin effect under any open- boundary geometry.  
+
+## III. CORNER-SKIN EFFECT AND GEOMETRY-DEPENDENT-SKIN EFFECT  
+
+In this section, we will provide some examples to demonstrate the characteristics of the two manifestations of the universal skin effect. We discuss the role of symmetry on the universal skin effect. We define the current functional to explain the appearance of corner- skin effect. In addition, we numerically verified that geometry- dependent- skin effect obeys the volume law, which is a significant feature to distinguish skin modes form conventional boundary states.  
+
+### A. Symmetry and the universal skin effect  
+
+We have proved that if the system has nonzero spectral area, the universal skin effect will occur under a general open boundary condition. According to the symmetry restriction, the universal skin effect has two manifestations,
+
+<--- Page Split --->
+![](images/Figure_unknown_3.jpg)
+
+<center>FIG. S4. The distribution of \(W(x)\) for Hamiltonian Eq. (S27) with different parameters and on different geometries. The system size is \(L_{x} = L_{y} = 25\) . The probability density is proportional to the opacity of the red color. (a) shows the corner-skin effect with \(t_{1} = t_{2} = 1\) , \(w = 0\) ; (b) shows line skin with \(t_{1} = 1\) , \(t_{2} = w = 0\) ; (c) has no skin effect with \(t_{1} = t_{2} = 0\) , \(w = 1\) , and geometry-dependent-skin effect appears in (d) under triangle and diamond geometries. </center>  
+
+that is, corner- skin effect and geometry- dependent- skin effect. Consider a system with nonzero spectral area, if its Hamiltonian has no any symmetry, the skin modes are localized at one or several vertices on an open geometry of any shape, and the number of modes are proportional to the volume of the system. If the Hamiltonian has certain spatial symmetries, such as mirror symmetry, the corner- skin effect will be forbidden. The reason is that the corner- skin effect has the nature of non- reciprocity, which is incompatible with mirror symmetry. However, if we change the open boundary geometry such that the symmetry is broken on the boundary, then geometry- dependent- skin modes will appear, which is a unique but universal phenomenon in higher- dimensional systems. We take a concrete example to demonstrate the role of symmetry on the skin effect. Consider a tight- binding model, of which the periodic- boundary Hamiltonian reads  
+
+\[\begin{array}{r l} & {H(k_{x},k_{y}) = h_{0}(k_{x},k_{y}) + i h_{1}(k_{x},k_{y})}\\ & {\qquad = \sin k_{x}\sigma_{x} + \sin k_{y}\sigma_{y} + (2 - \cos k_{x} - \cos k_{y})\sigma_{z} + i[t_{1}\sin k_{x} + t_{2}\sin k_{y} + w(\cos k_{x} - \cos k_{y})]\sigma_{z},} \end{array} \quad (S27)\]  
+
+where \(h_{0}\) and \(h_{1}\) are the Hermitian and non- Hermitian parts, respectively. The Hermitian part is in a gapless phase with a Dirac point at \(k_{x} = k_{y} = 0\) , and has inversion symmetry \(\sigma_{z}h_{0}(k_{x},k_{y})\sigma_{z} = h_{0}(- k_{x}, - k_{y})\) .  
+
+If we only add the \(w\) non- Hermitian term \((t_{1} = t_{2} = 0\) ; \(w = 1\) ), the combined mirror and non- Hermitian time- reversal symmetry \(M_{x}T = A_{t}\) and \(M_{y}T = \sigma_{z}A_{t}\) are preserved ( \(A_{t}\) representing transpose operator),  
+
+\[\begin{array}{r l} & {(M_{x}T)H(k_{x},k_{y})(M_{x}T)^{-1} = H(k_{x}, - k_{y});}\\ & {(M_{y}T)H(k_{x},k_{y})(M_{y}T)^{-1} = H(- k_{x},k_{y}).} \end{array} \quad (S28)\]  
+
+There is no skin effect under open boundary with square geometry as shown in Fig. S4(c). However, if we cut the square lattice into triangles and diamond lattices, the skin effect will retrieve on the boundaries that destroy the two symmetries, which is shown in Fig. S4(d).  
+
+If we only turn on \(t_{1}\) non- Hermitian term \((t_{1} = 1;t_{2} = w = 0)\) , the Hamiltonian preserves \(M_{x}T\) symmetry but destroys \(M_{y}T\) symmetry. In this case, the system has skin modes along \(x\) direction as shown in Fig. S4(b), the number of the skin modes are proportional to the volume of the system. If we add \(t_{1}\) and \(t_{2}\) non- Hermitian terms \((t_{1} = t_{2} = 1;w = 0)\) , the Hamiltonian destroys the two symmetries, and skin modes will be concentrated on the corner as shown in Fig. S4(a).  
+
+### B. Current functional  
+
+In this section, we first simply introduce the concept of current functional, then discuss the restrictions of point groups in two and three dimensions on the current functional, a quantity faithfully depicting the appearance of corner- skin effect, which demonstrates that corner- skin effect is only compatible with point groups \(C_{m}\) and \(C_{2,3,4,6,2v,3v,4v,6v}\) .
+
+<--- Page Split --->
+
+We define the current functional to depict the corner- skin effect in \(d\) dimensions.  
+
+\[J_{\alpha}[n] = \sum_{i}J_{i,\alpha}[n] = \sum_{i}\oint_{\mathrm{BZ}}dk^{d}n(E_{i},E_{i}^{*})\partial_{k_{\alpha}}E_{i}(\mathbf{k}), \quad (S29)\]  
+
+where \(i\) is band index, and \(k_{\alpha}\) is a vector, expressed as \(k_{\alpha} = \sum_{i}k_{i}\mathbf{e}_{i}\) in \(d\) - dimensional momentum space with unit vector basis \(\mathbf{e}_{i}\) . Here \(n(E,E^{*})\) is a distribution function depending on \(E\) and \(E^{*}\) , but does not depend on \(k\) explicitly, such as the Bose distribution \(n(E,E^{*}) = (e^{\mathrm{Re}E(k) / k_{B}T} - 1)^{- 1}\) . If there exists a \(n(E,E^{*})\) such that the current functional is nonzero for any \(\alpha\) , then the system must have the corner- skin effect. If for any possible \(n(E,E^{*})\) and \(\alpha\) , the current functional is zero, then the system has no corner- skin effect.  
+
+For example, we take a single- band tight- binding model [Eq. (1) in the main text] as  
+
+\[\mathcal{H}(\mathbf{k}) = [5(\cos k_{x} + \cos 2k_{x}) - i(\sin k_{x} + 3\sin 2k_{x}) + 5\cos k_{y} + i\sin k_{y}] / 2, \quad (S30)\]  
+
+and \(n(E,E^{*})\) as \(\operatorname {Im}[\mathcal{H}(\mathbf{k})]\) . In this case, \(J_{x}\) is equal to \(25\pi^{2} / 2\) and \(J_{y}\) is equal to \(- 5\pi^{2} / 2\) . Hence, the system has corner- skin effect. Another example of tight- binding model [Eq. (3) in the main text] reads  
+
+\[\mathcal{H}(\mathbf{k}) = 2\cos k_{x} + 2i\cos k_{y}, \quad (S31)\]  
+
+of which the current functional \(J_{x} = J_{y} = 0\) regardless of the choose of \(n(E,E^{*})\) . Therefore, the hamiltonian has no corner- skin effect (although hosts the geometry- dependent- skin effect due to the existence of two mirror symmetries). Meanwhile, the example also reminds us that certain symmetries may prohibit the corner- skin effect. Next, we will systematically analyze the interplay of point- group symmetries and the corner- skin effect.  
+
+## 2. Corner-skin effect under point groups  
+
+Here we consider the hamiltonian with only spatial symmetries (without any anti- unitary symmetry such as non- Hermitian time- reversal symmetry), and investigate the restrictions of the point- group symmetries on the current functional, and further conclude that corner- skin effect is only compatible with point groups \(C_{m}\) and \(C_{2,3,4,6,2v,3v,4v,6v}\) .  
+
+Inversion: Consider a system that only has inversion symmetry \(I\) , and each band satisfies \(E_{i}(\mathbf{k}) = E_{i}(I\mathbf{k}) = E_{i}(-\mathbf{k})\) . The current functional for \(i\) - th band can be expressed as  
+
+\[J_{i,\alpha}[n] = \oint_{\mathrm{BZ}}dk^{d}n(E_{i},E_{i}^{*})\partial_{k_{\alpha}}E_{i}(\mathbf{k}), \quad (S32)\]  
+
+which is invariant by replacing \(\mathbf{k}\) with \(\mathbf{k}^{\prime}\equiv - \mathbf{k}\) . After the transformation, Eq. (S32) becomes  
+
+\[\begin{array}{l}{{J_{i,\alpha}[n]=\oint_{B Z}(-1)^{d}d k^{\prime d}n(E_{i},E_{i}^{*})\partial_{-k_{\alpha}^{\prime}}E_{i}(-\mathbf{k}^{\prime})=(-1)^{d}\oint_{B Z}(-1)^{d}d k^{\prime d}n(E_{i},E_{i}^{*})\partial_{-k_{\alpha}^{\prime}}E_{i}(-\mathbf{k}^{\prime})}}\\ {{=-\oint_{B Z}d k^{\prime d}n(E_{i},E_{i}^{*})\partial_{k_{\alpha}^{\prime}}E_{i}(\mathbf{k}^{\prime})=-\oint_{B Z}d k^{d}n(E_{i},E_{i}^{*})\partial_{k_{\alpha}}E_{i}(\mathbf{k})=-J_{i,\alpha}[n]=0.}}\end{array} \quad (S33)\]  
+
+It means that if the hamiltonian has only inversion symmetry, the current functional for each band must be zero regardless of the choose of \(n(E,E^{*})\) . Equivalently, the corner- skin effect must vanish in the system with inversion symmetry, and is incompatible with the point groups including inversion symmetry, such as \(C_{i,3i,2h,4h,6h}\) , \(D_{3d,2h,4h,6h}\) , \(T_{h}\) and \(O_{h}\) .  
+
+Rotation: Consider a system that is invariant under a point group including rotation operator \(R\) , then \(E_{i}(\mathbf{k}) = E_{i}(R\mathbf{k})\) . The Eq. (S32) is also invariant under the transformation from \(\mathbf{k}\) to \(\mathbf{k}^{\prime}\equiv R^{- 1}\mathbf{k} = \sum_{ij}\mathbf{e}_{i}c_{ij}\mathbf{k}_{j}\) , where \(k_{j}\) is the \(j\) - th component (along \(\mathbf{e}_{j}\) ) of \(\mathbf{k}\) . After the transformation, the current functional becomes  
+
+\[\begin{array}{l}{{J_{i,\alpha}[n]=\oint_{\mathrm{BZ}^{\prime}}\mathrm{det}\left[J_{\mathbf{k},\mathbf{k}^{\prime}}\right]d k^{\prime d}n(E_{i},E_{i}^{*})\partial_{R k_{\alpha}^{\prime}}E_{i}(R\mathbf{k}^{\prime})=\oint_{\mathrm{BZ}}d k^{\prime d}n(E,E^{*})\partial_{R k_{\alpha}^{\prime}}E_{i}(\mathbf{k}^{\prime})}}\\ {{=\oint_{\mathrm{BZ}}d k^{d}n(E,E^{*})\partial_{R k_{\alpha}}E_{i}(\mathbf{k})=\oint_{\mathrm{BZ}}d k^{d}n(E,E^{*})\partial_{k_{\alpha}}E_{i}(\mathbf{k}),}}\end{array} \quad (S34)\]  
+
+where \(\operatorname *{det}\left[J_{\mathbf{k},\mathbf{k}^{\prime}}\right]\) in the second term is the determinant of the Jacobian \(J_{\mathbf{k},\mathbf{k}^{\prime}}\) that measures the change of differential volume element under different representations and the sign of \(\operatorname *{det}\left[J_{\mathbf{k},\mathbf{k}^{\prime}}\right]\) is positive because the rotational operator
+
+<--- Page Split --->
+
+preserves orientation. One can always choose an appropriate basis transformation such that \(\operatorname *{det}\left[J_{\mathbf{k},\mathbf{k}^{\prime}}\right] = 1\) . In addition, since the Brillouin zone has the same symmetry group as the hamiltonian and rotational operator \(R\) does not change the orientation, the integral region BZ is invariant under the point group. For example, consider a rotation \(R\) that rotates \(\pi /4\) along \(\mathbf{e}_{z}\) axis, then \(k_{x} = - k_{y}^{\prime}\) and \(k_{y} = k_{x}^{\prime}\) . The Jacobian matrix can be written as  
+
+\[J_{\mathbf{k},\mathbf{k}^{\prime}} = \left( \begin{array}{ll}\partial_{x^{\prime}}k_{x} & \partial_{y^{\prime}}k_{x}\\ \partial_{x^{\prime}}k_{y} & \partial_{y^{\prime}}k_{y^{\prime}} \end{array} \right) \quad (S35)\]  
+
+In this case, the determinant of Jacobian \(\operatorname *{det}\left[J_{\mathbf{k},\mathbf{k}^{\prime}}\right]\) is 1.  
+
+The last equation of Eq. (S34) requires  
+
+\[R k_{\alpha} = k_{\alpha}, \quad (S36)\]  
+
+which means that the direction of \(k_{\alpha}\) is parallel to the rotational axis of \(R\) , equivalently, the component of \(k_{\alpha}\) perpendicular to the rotational axis must be zero. For example, if \(R\) is a rotation that rotates \(\theta\) along \(\mathbf{e}_{z}\) axis, that is,  
+
+\[R = \left( \begin{array}{ccc}\cos \theta & -\sin \theta & 0\\ \sin \theta & \cos \theta & 0\\ 0 & 0 & 1 \end{array} \right), \quad (S37)\]  
+
+then Eq. (S36) restricts \(k_{\alpha}\) as \(0\mathbf{e}_{x} + 0\mathbf{e}_{y} + k_{z}\mathbf{e}_{z}\) , along \(\mathbf{e}_{z}\) axis.  
+
+As a consequence, if a point group contains two or more rotations with non- parallel rotational axes, the current functional for each band must be zero. If the point group contains only one rotation, a nonzero current functional for each band is allowed, thus corner- skin effect is compatible with the point groups including only one rotation, such as \(C_{2,3,4,6}\) .  
+
+Mirror: In a similar way, a mirror symmetry requires  
+
+\[M k_{\alpha} = k_{\alpha} \quad (S38)\]  
+
+if nonzero current functional for each band is allowed, which means the corner- skin effect is compatible with the group containing only one mirror symmetry, that is, \(C_{m}\) .  
+
+It is notable that an additional rotation symmetry is allowed when the rotational axis lies in the mirror plane, and \(k_{\alpha}\) is restricted as  
+
+\[R M k_{\alpha} = k_{\alpha}. \quad (S39)\]  
+
+Therefore, the corner- skin effect is also compatible with such point groups, i.e., \(C_{2v,3v,4v,6v}\) .  
+
+In summary, we have explicitly shown that corner- skin effect does not exist under most point groups because the current functional is restricted to zero under such point groups, regardless of the choice of \(n(E, E^{*})\) , but is allowed to appear under these point groups, i.e.,  
+
+\[\{C_{m},C_{2},C_{3},C_{4},C_{6},C_{2v},C_{3v},C_{4v},C_{6v}\} . \quad (S40)\]  
+
+## C. Volume Law  
+
+We numerically show that the geometry- dependent- skin effect satisfies the volume law, that is, the increase in the number of skin modes is proportional to the increase in volume of the system,  
+
+\[\delta N_{skin}\propto \delta V. \quad (S41)\]  
+
+For example, If the shape of the open boundary is a parallelogram whose side- lengths are \(L_{x}\) and \(L_{y}\) as shown in Fig. S5, then, \(V = L_{x}L_{y}\) . Our criterion for judging a mode as a skin mode is to check whether ninety percent of the probability density of this mode lies within the boundary we appointed.  
+
+Consider a tight- binding model with periodic- boundary Hamiltonian \(\mathcal{H}(k_{x},k_{y}) = 2\cos k_{x} + 2i\sin k_{y}\) , there is no skin effect under square geometry Fig. S5(a), but the skin effect appears under parallelogram geometry Fig. S5(b) due to the spectral area being nonzero, which is geometry- dependent- skin effect. The distributions of \(W(\mathbf{x})\) under different open boundaries are plotted. For parallelogram geometry, we specify the thickness of the boundary to be the width of three unit cells, and use black dashed lines to distinguish the boundary from the bulk. If ninety percent of the probability density of a mode lies in the boundary, we count it as a skin mode. We count the number of skin modes for different volumes ( \(\mathcal{L}_{x}\mathcal{L}_{y}\) ), and the fitting curve (blue curve in Fig. S5(c)) shows that the two are in a linear relation, that is, \(\delta N_{skin} = 0.52\delta V\) . The volume law of geometry- dependent- skin effect has been verified numerically.
+
+<--- Page Split --->
+![](images/Figure_unknown_4.jpg)
+
+<center>FIG. S5. The norm squared of all wave functions of the Hamiltonian Eq. (S13) on square lattice (a) and parallelogram lattice (b) is plotted. The system size is chosen as \(L_{x} = L_{y} = 25\) . The volume law is shown in (c), in which blue line represents \(N_{skin} \propto V\) , gray line \(N_{skin} \propto \sqrt{V}\) and black line \(N_{skin} \propto 1\) . </center>  
+
+## IV. EXCEPTIONAL SEMIMETALS  
+
+In this section, we will prove the corollary of our theorem, that is all stable exceptional semimetals imply the universal skin effect. We first review the topological charge of non- Hermitian band degeneracies.  
+
+### A. Non-Hermitian band degeneracy  
+
+Consider a general \(m\) - band non- Hermitian Bloch Hamiltonian (with periodic boundary condition),  
+
+\[\mathcal{H}(\mathbf{k}) = \sum_{s = 1}^{m^{2} - 1}\left[h_{s}^{s}(\mathbf{k}) + ih_{s}^{i}(\mathbf{k})\right]\Gamma_{s}, \quad (S42)\]  
+
+where \(\Gamma_{s}\) are the generators of Lei algebra \(\mathfrak{su}(m)\) and \(h_{s}^{i}(\mathbf{k})\) and \(h_{s}^{i}(\mathbf{k})\) are real functions of \(\mathbf{k}\) . When \(m = 2,3,4\) , \(\Gamma_{s}\) are the Pauli, GellMann, and \(\gamma\) matrices, respectively. The eigenvalues of \(\mathcal{H}(\mathbf{k})\) can be obtained by solving the following characteristic polynomial  
+
+\[f_{E}(\mathbf{k}) = \operatorname *{det}[E - \mathcal{H}(\mathbf{k})] = \prod_{i = 1}^{m}[E - E_{i}(\mathbf{k})], \quad (S43)\]  
+
+where \(E_{i}(\mathbf{k})\) is the \(i\) th eigenvalue of the non- Hermitian Hamiltonian \(\mathcal{H}(\mathbf{k})\) . At the degeneracy point \(\mathbf{k}_{D}\) , two bands must have the same energy, i.e.  
+
+\[E_{i}(\mathbf{k}_{D}) = E_{j}(\mathbf{k}_{D}) \quad (S44)\]  
+
+for some \(i \neq j\) . In Ref. [40, 41], the authors have shown that the above condition is equivalent to the vanishing of the discriminant of \(f_{E}(\mathbf{k})\) , i.e.  
+
+\[\mathrm{Disc}_{E}[\mathcal{H}]({\bf k}_{D}) = 0, \quad (S45)\]  
+
+where  
+
+\[\mathrm{Disc}_{E}[\mathcal{H}]({\bf k}) = \prod_{i< j}[E_{i}({\bf k}) - E_{j}({\bf k})]^{2} \quad (S46)\]  
+
+is the discriminant of \(f_{E}(\mathbf{k})\) . Although the discriminant is defined by the roots of \(f_{E}(\mathbf{k}) = 0\) , it can be computed directly from the determinant of the Sylvester matrix of \(f_{E}(\mathbf{k})\) and \(\partial_{E}f_{E}(\mathbf{k})\) , which can be expressed by the coefficients of \(f_{E}(\mathbf{k})\) . Now we show a concrete example of the discriminant method.  
+
+Example: Consider a generic two- band model  
+
+\[\mathcal{H}(\mathbf{k}) = h_{0}(\mathbf{k})\sigma_{0} + h_{x}(\mathbf{k})\sigma_{x} + h_{y}(\mathbf{k})\sigma_{y} + h_{z}(\mathbf{k})\sigma_{z}, \quad (S47)\]
+
+<--- Page Split --->
+![](images/Figure_unknown_5.jpg)
+
+<center>FIG. S6. The periodic-boundary spectrum of the photonic crystal model is shown in (a)(c) with light blue color. Under the square geometry with the systems size \(L_{x} = L_{y} = 31\) , the eigenvalues (red points) and the norm squared of all wave functions are shown in (a) and (b), respectively. Under the diamond geometry with the systems size \(L_{x} = L_{y} = 45\) , the eigenvalues (red points) and the norm squared of all wave functions are plotted in (c) and (d), respectively. </center>  
+
+where \(h_{\mu}(\mathbf{k}) = h_{\mu}^{\tau}(\mathbf{k}) + ih_{\mu}^{i}(\mathbf{k})\) are complex functions of \(\mathbf{k}\) . The characteristic polynomial of the two- band model can be written as  
+
+\[f_{E}(\mathbf{k}) = E^{2} + b(\mathbf{k})E + c(\mathbf{k}), \quad (S48)\]  
+
+where \(b(\mathbf{k}) = - 2h_{0}(\mathbf{k})\) and \(c(\mathbf{k}) = h_{0}^{2}(\mathbf{k}) - h_{x}^{2}(\mathbf{k}) - h_{y}^{2}(\mathbf{k}) - h_{z}^{2}(\mathbf{k})\) . Computing the discriminant of polynomial (S48) with respect to the energy \(E\) , we obtain the following condition for the existence of DPs  
+
+\[\mathrm{Disc}_{E}[\mathcal{H}](\mathbf{k}) = b^{2}(\mathbf{k}) - 4c(\mathbf{k}) = 4[h_{x}^{2}(\mathbf{k}) + h_{y}^{2}(\mathbf{k}) + h_{z}^{2}(\mathbf{k})] = 0. \quad (S49)\]  
+
+This condition can also be obtained from the energy spectrum, that is the two bands \(E_{\pm} = h_{0}(\mathbf{k})\pm ([h_{x}^{2}(\mathbf{k}) + h_{y}^{2}(\mathbf{k}) + h_{z}^{2}(\mathbf{k})]^{1 / 2}\) are degenerate, whenever the square root is vanishing.  
+
+From the above example, one can notice that the discriminant \(\mathrm{Disc}_{E}[\mathcal{H}](\mathbf{k})\) is a complex periodic function of \(\mathbf{k}\) . Its vanishing is equivalent to the vanishing of the real and imaginary parts, i.e.  
+
+\[\mathrm{Re}\mathrm{Disc}_{E}[\mathcal{H}](\mathbf{k}) = \mathrm{Im}\mathrm{Disc}_{E}[\mathcal{H}](\mathbf{k}) = 0. \quad (S50)\]  
+
+The solution of the above equation are the non- Hermitian degeneracy points in 2D and lines in 3D.  
+
+### B. Topological charge of non-Hermitian band degeneracies  
+
+In this subsection, we will review the topological charge of the non- Hermitian band degeneracies. Based on the discriminant of the characteristic polynomial, one can define the topological charge of the degeneracy point \(\mathbf{k}_{D}\) , i.e.  
+
+\[\nu (\pmb {k}_D) = \frac{1}{2\pi i}\oint_{\Gamma (\pmb {k}_D)}d\pmb {k}\cdot \nabla_{\pmb {k}}\ln \mathrm{Disc}_E[\mathcal{H}](\pmb {k}). \quad (S51)\]  
+
+where \(\Gamma (\pmb {k}_D)\) is a loop encircling the degeneracy point \(\mathbf{k}_{D}\) . Since \(\mathrm{Disc}_{E}[\mathcal{H}](\mathbf{k})\) is single valued, this invariant is quantized, which is called the discriminant number in Ref. [41]. Putting  
+
+\[\mathrm{Disc}_{E}[\mathcal{H}](\mathbf{k}) = \prod_{i< j}\left[E_{i}(\mathbf{k}) - E_{j}(\mathbf{k})\right]^{2} \quad (S52)\]
+
+<--- Page Split --->
+
+into \(\nu (\mathbf{k}_{D})\) , one can obtain  
+
+\[\begin{array}{l}{\nu (\mathbf{k}_{D}) = \frac{1}{2\pi i}\oint_{\Gamma (\mathbf{k}_{D})}d\mathbf{k}\cdot \nabla_{\mathbf{k}}\ln \prod_{1\leq i< j\leq n}[E_{i}(\mathbf{k}) - E_{j}(\mathbf{k})]^{2}}\\ {= \frac{1}{2\pi i}\sum_{i\neq j}\oint_{\Gamma (\mathbf{k}_{D})}d\mathbf{k}\cdot \nabla_{\mathbf{k}}\ln [E_{i}(\mathbf{k}) - E_{j}(\mathbf{k})]}\\ {= \frac{1}{2\pi}\sum_{i\neq j}\oint_{\Gamma (\mathbf{k}_{D})}d\mathbf{k}\cdot \nabla_{\mathbf{k}}\arg [E_{i}(\mathbf{k}) - E_{j}(\mathbf{k})].} \end{array} \quad (S53)\]  
+
+Therefore, for a two- band system,  
+
+\[\nu (\mathbf{k}_{D}) = \frac{1}{2\pi}\oint_{\Gamma (\mathbf{k}_{D})}d\mathbf{k}\cdot \nabla_{\mathbf{k}}\arg [E_{+}(\mathbf{k}) - E_{-}(\mathbf{k})] \quad (S54)\]  
+
+which describes the winding of the complex energy between two bands. Now we show a concrete example of the winding number.  
+
+Example: Consider the following low energy Hamiltonian around \(\mathbf{k}_{D}\) ,  
+
+\[\mathcal{H}_{1}(\delta \mathbf{k}) = \sigma_{+} + (\delta k_{x} + i\delta k_{y})\sigma_{-}, \quad (S55)\]  
+
+where \(\delta \mathbf{k} = \mathbf{k} - \mathbf{k}_{D}\) and \(\sigma_{\pm} = (\sigma_{x} \pm i\sigma_{y}) / 2\) . The eigenvalues of \(\mathcal{H}(\delta \mathbf{k})\) are  
+
+\[E_{\pm}(\delta \mathbf{k}) = \pm \sqrt{\delta k_{x} + i\delta k_{y}}. \quad (S56)\]  
+
+When \(\mathbf{k} = \mathbf{k}_{D}\) , which is equivalent to \(\delta k_{x} = \delta k_{y} = 0\) , one can find \(E_{+}(\delta \mathbf{k} = 0) = E_{-}(\delta \mathbf{k} = 0) = 0\) . This means \(\mathbf{k}_{D}\) is a non- Hermitian degeneracy point. Now we choose \(\Gamma (\mathbf{k}_{D}) = \mathbf{k}_{D} + \delta k_{r}(\cos \theta , \sin \theta)\) , then,  
+
+\[E_{\pm}(\delta \mathbf{k}) = \pm \delta k_{r}^{1 / 2}e^{i\theta /2},\qquad \theta \in (-\pi ,\pi ]. \quad (S57)\]  
+
+One can find that \(E_{+}(\delta \mathbf{k})\) and \(E_{- }(\delta \mathbf{k})\) forms a spectral loop that encloses \(E_{+}(\mathbf{k}_{D}) = E_{- }(\mathbf{k}_{D}) = 0\) . The winding number \(\nu (\mathbf{k}_{D})\) describes this topological properties of degeneracy points.  
+
+The topological charge \(\nu (\mathbf{k}_{D})\) can be used to classify the non- Hermitian degeneracies. However, the classification is not complete. As a comparison with \(\mathcal{H}_{1}(\delta \mathbf{k})\) , consider the following two low energy Hamiltonians,  
+
+\[\mathcal{H}_{2}(\delta \mathbf{k}) = (\delta k_{x} + i\delta k_{y})^{2}\sigma_{+} + (\delta k_{x} - i\delta k_{y})\sigma_{-}. \quad (S58)\]  
+
+Obvious, \(\delta \mathbf{k} = 0\) is a degeneracy point. One can further prove that its topological charge is \(+1\) , which is equal to the charge of \(\delta \mathbf{k} = 0\) in \(\mathcal{H}_{1}(\delta \mathbf{k})\) . However, these two degeneracy points have different properties. For example  
+
+\[\mathcal{H}_{1}(\delta \mathbf{k} = 0) = \sigma_{+},\qquad \mathcal{H}_{2}(\delta \mathbf{k} = 0) = 0. \quad (S59)\]  
+
+One can notice that \(\mathcal{H}_{1}(\delta \mathbf{k} = 0)\) is non- diagonal. This type of non- Hermitian degeneracy points are called exceptional points. In Ref. [41], the authors have shown that only the exceptional points with \(\nu (\mathbf{k}_{D}) = \pm 1\) are robust in 2D. Any other non- Hermitian band degeneracies are unstable against non- Hermitian perturbations.  
+
+Having clarifying the topological charge of non- Hermitian degeneracies, now we can prove the corollary of our theorem. Since in 2D, the topological charge of the stable exceptional points must be \(\pm 1\) , the corresponding spectrum area must be nonzero.  
+
+### C. The photonic crystal model  
+
+In this subsection, we numerically calculate the spectrum and spatial distribution of the wave function, i.e. \(W(x)\) in Eq. S2, for the photonic crystal model in the main text under different geometries.  
+
+It shows that the skin effect disappears under square geometry in Fig. S6(b), and reappears under diamond geometry in Fig. S6(d), which is a characteristic signature of geometry- dependent- skin effect. Here we take the non- Hermitian parameter \(\gamma\) as \(1 / 4\) . The spectrum under square geometry (red points in Fig. S6(a)) coincides with the spectrum of periodic boundary (light blue region in Fig. S6(a)(c)). We conjecture that the spectrum under diamond geometry (red points in Fig. S6(d)) will also coincide with the periodic- boundary spectrum as the system size increases. However, it clearly shows that the density of states under different geometries is completely different. The dependence of density of states on the choice of boundary geometry is another significant feature of geometry- dependent- skin effect.
+
+<--- Page Split --->
+![](images/Figure_unknown_6.jpg)
+
+<center>FIG. S7. Two Weyl points (a) of a three-dimensional Weyl semimetal are expanded into two exceptional rings (b) after the addition of non-Hermitian perturbations. The spatial distribution of eigenstate is plotted in (c). The modulus square of the propagator from \(i\) to \(o P_{i o}(\omega)\) and that from \(o\) to \(i P_{o i}(\omega)\) , as functions of \(\omega\) , are plotted with red color and dark cyan color in (d), respectively. </center>  
+
+### D. The corner-skin effect in a three-dimensional exceptional-line semimetal  
+
+We propose the realization for corner- skin effect in a three- dimensional system with exceptional lines. Consider a Weyl semimetal with non- Hermitian term as a perturbation, of which the periodic- boundary hamiltonian reads  
+
+\[H(k) = [d_{\mathbf{r}}(k) + i\delta d_{\mathbf{i}}(k)]\cdot \pmb {\sigma}, \quad (S60)\]  
+
+where \(d_{\mathbf{r}}(k)\) and \(d_{\mathbf{i}}(k)\) are vectors with four components, that is,  
+
+\[\begin{array}{l}{d_{\mathbf{r}}(k) = (0,\sin k_{x},\sin k_{y},2 - \cos k_{x} - \cos k_{y} + \sin k_{z}),}\\ {d_{\mathbf{i}}(k) = (-\sqrt{5},1 + \cos k_{z},1 - \cos k_{z},\cos k_{z}).} \end{array} \quad (S61)\]  
+
+The Hermitian part \(d_{\mathbf{r}}\cdot \pmb{\sigma}\) is a Weyl semimetal with two Weyl points. One Weyl point with \(+1\) topological charge [red cone in Fig S7 (a)] is at \((0,0,0)\) , and another with \(- 1\) topological charge [blue cone in Fig S7 (a)] is at \((0,0,\pi)\) . Upon turning on the non- Hermitian term ( \(\delta \neq 0\) ), the Weyl points evolve into two exceptional rings as shown in Fig. S7 (b). According to our theorem, the system with exceptional lines must have the universal skin effect. Specially, the system described in Eq. (S60) always has corner- skin effect as shown in Fig. S7 (c) with \(\delta = 1 / 6\) .  
+
+Due to the non- reciprocity of the corner- skin effect, we propose an experimental approach of two- point Green function to detect the corner- skin effect. We give a source at \(i = (1,1,1)\) position, and probe it at \(o = (16,16,16)\) position in Fig. S7 (c). The modulus square of the propagator from \(i\) to \(o\) is expressed as  
+
+\[P_{o}(\omega) = \sum_{\alpha ,\beta = 1,2}|\langle \alpha ,\beta |\frac{1}{\omega - H} |i,\alpha \rangle |^{2}, \quad (S62)\]  
+
+where \(\alpha ,\beta\) label the orbitals of the unit cell. We calculate \(P_{i o}(\omega)\) and plot it with red color in Fig. S7 (d). We do the same process but interchange \(i\) and \(o\) , and \(P_{o i}(\omega)\) is plotted with dark cyan color in Fig. S7 (d). A significant difference between \(P_{i o}(\omega)\) and \(P_{o i}(\omega)\) demonstrates the non- reciprocity of corner- skin effect.
+
+<--- Page Split --->
+
+[7] Leykam, D., Bliokh, K. Y., Huang, C., Chong, Y. D. & Nori, F. Edge Modes, Degeneracies, and Topological Numbers in Non-Hermitian Systems. Phys. Rev. Lett. 118, 040401 (2017).[8] Shen, H., Zhen, B. & Fu, L. Topological Band Theory for Non-Hermitian Hamiltonians. Phys. Rev. Lett. 120, 146402 (2018).[9] Gong, Z. et al. Topological Phases of Non-Hermitian Systems. Phys. Rev. X 8, 031079 (2018).[10] Bergholtz, E. J., Budich, J. C. & Kunst, F. K. Exceptional topology of non-Hermitian systems. Rev. Mod. Phys. 93, 015005 (2021).[11] Kawabata, K., Shiozaki, K., Ueda, M. & Sato, M. Symmetry and Topology in Non-Hermitian Physics. Phys. Rev. X 9, 041015 (2019).[12] Ashida, Y., Gong, Z. & Ueda, M. Non-Hermitian physics. Advances in Physics 69, 249- 435 (2020). https://doi.org/10.1080/00018732.2021.1876991. [13] Yao, S. & Wang, Z. Edge States and Topological Invariants of Non-Hermitian Systems. Phys. Rev. Lett. 121, 086803 (2018).[14] Yao, S., Song, F. & Wang, Z. Non-Hermitian Chern Bands. Phys. Rev. Lett. 121, 136802 (2018).[15] Kunst, F. K., Edvardsson, E., Budich, J. C. & Bergholtz, E. J. Biorthogonal Bulk-Boundary Correspondence in Non-Hermitian Systems. Phys. Rev. Lett. 121, 026808 (2018).[16] Yokomizo, K. & Murakami, S. Non-Bloch Band Theory of Non-Hermitian Systems. Phys. Rev. Lett. 123, 066404 (2019).[17] Zhang, K., Yang, Z. & Fang, C. Correspondence between Winding Numbers and Skin Modes in Non-Hermitian Systems. Phys. Rev. Lett. 125, 126402 (2020).[18] Okuma, N., Kawabata, K., Shiozaki, K. & Sato, M. Topological Origin of Non-Hermitian Skin Effects. Phys. Rev. Lett. 124, 086801 (2020).[19] Yang, Z., Zhang, K., Fang, C. & Hu, J. Non-Hermitian Bulk-Boundary Correspondence and Auxiliary Generalized Brillouin Zone Theory. Phys. Rev. Lett. 125, 226402 (2020).[20] Martinez Alvarez, V. M., Barrios Vargas, J. E. & Foa Torres, L. E. F. Non-Hermitian robust edge states in one dimension: Anomalous localization and eigenspace condensation at exceptional points. Phys. Rev. B 97, 121401(R) (2018).[21] Lee, C. H. & Thomale, R. Anatomy of skin modes and topology in non-Hermitian systems. Phys. Rev. B 99, 201103(R) (2019).[22] Xiong, Y. Why does bulk boundary correspondence fail in some non-hermitian topological models. Journal of Physics Communications 2, 035043 (2018).[23] Yang, Z. Non-perturbative Breakdown of Bloch's Theorem and Hermitian Skin Effects. arXiv:2012.03333 (2020).[24] Ashcroft, N. W. & Mermin, N. D. Solid State Physics (Holt-Saunders, 1976).[25] Xiao, L. et al. Non-Hermitian bulk-boundary correspondence in quantum dynamics. Nature Physics 16, 761- 766 (2020).[26] Ghatak, A., Brandenburger, M., van Wezel, J. & Coulais, C. Observation of non-Hermitian topology and its bulk- edge correspondence in an active mechanical metamaterial. Proceedings of the National Academy of Sciences 117, 29561- 29568 (2020).[27] Helbig, T. et al. Generalized bulk- boundary correspondence in non-Hermitian topolectrical circuits. Nature Physics 16, 747- 750 (2020).[28] Lee, C. H., Li, L. & Gong, J. Hybrid Higher- Order Skin- Topological Modes in Nonreciprocal Systems. Phys. Rev. Lett. 123, 016805 (2019).[29] Hofmann, T. et al. Reciprocal skin effect and its realization in a topolectrical circuit. Phys. Rev. Research 2, 023265 (2020).[30] Yoshida, T., Mizoguchi, T. & Hatsugai, Y. Mirror skin effect and its electric circuit simulation. Phys. Rev. Research 2, 022062(R) (2020).[31] Liu, T. et al. Second- Order Topological Phases in Non- Hermitian Systems. Phys. Rev. Lett. 122, 076801 (2019).[32] Kawabata, K., Sato, M. & Shiozaki, K. Higher- order non- Hermitian skin effect. Phys. Rev. B 102, 205118 (2020).[33] Scheibner, C., Irvine, W. T. M. & Vitelli, V. Non- Hermitian Band Topology and Skin Modes in Active Elastic Media. Phys. Rev. Lett. 125, 118001 (2020).[34] Li, L., Lee, C. H. & Gong, J. Topological Switch for Non- Hermitian Skin Effect in Cold- Atom Systems with Loss. Phys. Rev. Lett. 124, 250402 (2020).[35] Song, F., Wang, H.- Y. & Wang, Z. Non- Bloch PT symmetry breaking: Universal threshold and dimensional surprise. arXiv:2102.02230 (2021).[36] Lin, Z. et al. Unidirectional Invisibility Induced by \(\mathcal{PT}\) - Symmetric Periodic Structures. Phys. Rev. Lett. 106, 213901 (2011).[37] Miri, M.- A. & Alu, A. Exceptional points in optics and photonics. Science 363, eaar7709 (2019).[38] Özdemir, Ş. K., Rotter, S., Nori, F. & Yang, L. Parity- time symmetry and exceptional points in photonics. Nature Materials 18, 783- 798 (2019).[39] Kawabata, K., Bessho, T. & Sato, M. Classification of Exceptional Points and Non- Hermitian Topological Semimetals. Phys. Rev. Lett. 123, 066405 (2019).[40] Yang, Z., Chiu, C.- K., Fang, C. & Hu, J. Jones Polynomial and Knot Transitions in Hermitian and non- Hermitian Topological Semimetals. Phys. Rev. Lett. 124, 186402 (2020).[41] Yang, Z., Schnyder, A. P., Hu, J. & Chiu, C.- K. Fermion Doubling Theorems in Two- Dimensional Non- Hermitian Systems for Fermi Points and Exceptional Points. Phys. Rev. Lett. 126, 086401 (2021).[42] Zhen, B. et al. Spawning rings of exceptional points out of Dirac cones. Nature 525, 354- 358 (2015).
+
+<--- Page Split --->
+
+[43] Zhou, H. et al. Observation of bulk Fermi arc and polarization half charge from paired exceptional points. Science 359, 1009- 1012 (2018).  
+
+[44] Cerjan, A. et al. Experimental realization of a Weyl exceptional ring. Nature Photonics 13, 623- 628 (2019).  
+
+[45] Tang, W. et al. Exceptional nexus with a hybrid topological invariant. Science 370, 1077- 1080 (2020).  
+
+[46] Dembowski, C. et al. Experimental Observation of the Topological Structure of Exceptional Points. Phys. Rev. Lett. 86, 787- 790 (2001).  
+
+[47] Regensburger, A. et al. Parity- time synthetic photonic lattices. Nature 488, 167- 171 (2012).  
+
+[48] Gao, T. et al. Observation of non- Hermitian degeneracies in a chaotic exciton- polariton billiard. Nature 526, 554- 558 (2015).  
+
+[49] Doppler, J. et al. Dynamically encircling an exceptional point for asymmetric mode switching. Nature 537, 76- 79 (2016).  
+
+[50] Hodaei, H. et al. Enhanced sensitivity at higher- order exceptional points. Nature 548, 187- 191 (2017).  
+
+[51] Hasan, M. Z. & Kane, C. L. Colloquium: Topological insulators. Rev. Mod. Phys. 82, 3045- 3067 (2010).  
+
+[52] Armitage, N. P., Mele, E. J. & Vishwanath, A. Weyl and Dirac semimetals in three- dimensional solids. Rev. Mod. Phys. 90, 015001 (2018).  
+
+[53] Loomis, L. H. & Sternberg, S. Advanced calculus (World Scientific, 1968).  
+
+[54] Denner, M. M. et al. Exceptional Topological Insulators (2020). arXiv:2008.01090.
+
+<--- Page Split --->

preprint/preprint__2af794e8bdc6d7671634e08f6af1da53069c1fb5bf84792aa87006addcb20726/images_list.json

@@ -0,0 +1,77 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1. In-situ observation of aggregating growth between 5-FT NPs with SC NPs across diverse size ratios. (a-d) Configuration evolution after a \\(2.1\\mathrm{nm}\\) SC NP attached with a 5-FT NP (3.6 nm) with \\(R = 0.58\\) . White arrows show formation of the re-entrant surface. (e-o) Configuration evolution after a \\(4.6\\mathrm{nm}\\) SC NP aggregated with a \\(4.2\\mathrm{nm}\\) 5-FT NP \\((R = 1.10)\\) . (p) The surface outlines of the aggregated NP. The migration directions of the surface outlines with durations are denoted by red arrows. Notable, twin boundaries are highlighted by red lines. Twin units are denoted by numbers ranging from 1 to 5 in (a) and (e). Yellow arrows show migration directions of the TBs, compared with the immediate prior image. Partial dislocations are denoted by blue “L” and their slip directions are denoted by the blue arrows.",
+    "footnote": [],
+    "bbox": [
+      [
+        152,
+        100,
+        845,
+        610
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2. MD simulation of the aggregation growth of 5-FT NPs by attaching with various SC NPs. (a-e) Evolution of an aggregated 5-FT NP, formed by attaching of a \\(5.0 \\mathrm{nm}\\) symmetrical 5-FT NP (yellow, \\(\\mathrm{A}_{5\\mathrm{F}}\\) ) and a \\(2.6 \\mathrm{nm}\\) SC NP (cyan, \\(\\mathrm{B}_{35}\\) ) with \\(R = 0.52\\) . Re-entrant {111} surfaces are highlighted by gray solid lines. (f) The biggest cross-section perpendicular to the Z axis showing atomistic distribution of the initial SC NP. (g) Variation of the relative energy of the simulated system from (a) to (e). (h-r) The aggregation evolution process after a \\(5 \\mathrm{nm}\\) 5-FT Au NP ( \\(\\mathrm{A}_{5\\mathrm{F}}\\) ) attached with a \\(4.8 \\mathrm{nm}\\) SC NP ( \\(\\mathrm{B}_{35}\\) ) with \\(R = 0.96\\) . The relative slip between the 5-FT and the SC is represented by cyan arrows in (i-k) and (m). (s) Variation of the relative energy of the simulated system from (h) to (r). (t) TBs slip induced migrations of the twin pole from \\(2.5 \\mathrm{ns}\\) (i) to \\(4.8 \\mathrm{ns}\\) (o).",
+    "footnote": [],
+    "bbox": [
+      [
+        156,
+        95,
+        839,
+        650
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3. Aggregating growth between 5-FT NPs. (a-d) Atomic evolution process between a 3.7 nm 5-FT and a \\(4.6 \\mathrm{nm}\\) 5-FT with \\(R = 0.78\\) . (e-l) Atomic evolution process between a \\(3.3 \\mathrm{nm}\\) 5-FT and a \\(4.0 \\mathrm{nm}\\) 5-FT with a size ratio of 0.82. Twin units are denoted by numbers ranging from 1 to 5. TBs and unclear TBs are highlighted by red solid and dashed lines, respectively. The white arrow shows the aggregation of another NP in (d). Yellow arrows show migration directions of the TBs, compared with the immediate prior image. Partial dislocation is denoted by blue “L” and its slip direction is denoted by the blue arrow.",
+    "footnote": [],
+    "bbox": [
+      [
+        152,
+        279,
+        852,
+        686
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4. MD simulation of the aggregation growth process of S5-FT NPs by attaching with other S5-FT NPs. (a-e) and (h-m) Evolution of the aggregated 5-FT NPs, formed by attaching of a \\(5.0 \\mathrm{nm}\\) S5-FT NP (yellow, \\(\\mathrm{A_{5F}}\\) ) with \\(3.2 \\mathrm{nm}\\) ( \\(\\mathrm{B_{3F}}\\) ) and \\(5.0 \\mathrm{nm}\\) ( \\(\\mathrm{B_{5F}}\\) ) S5-FT NPs (cyan), respectively. The corresponding size ratios are 0.64 and 1.00, respectively. (f) and (n) The biggest cross-sections perpendicular to the Z axes showing atomistic distribution of the initial 5-FT NPs of \\(\\mathrm{B_{3F}}\\) and \\(\\mathrm{B_{5F}}\\) , respectively. (g) and (o) Variation of the relative energies of the corresponding simulated systems. TBs are highlighted by red lines. Twin units are denoted by numbers ranging from 1 to 5. Slip directions and layers of the TBs are denoted by yellow arrows and numbers, respectively.",
+    "footnote": [],
+    "bbox": [
+      [
+        150,
+        149,
+        847,
+        558
+      ]
+    ],
+    "page_idx": 14
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Fig. 5. Aggregation growth size effect of 5-FT NP. (a) Configuration obtained after aggregation between SC and 5-FT NPs. (b) Configuration obtained after aggregation between two 5-FT NPs.",
+    "footnote": [],
+    "bbox": [
+      [
+        225,
+        416,
+        770,
+        615
+      ]
+    ],
+    "page_idx": 16
+  }
+]

preprint/preprint__2af794e8bdc6d7671634e08f6af1da53069c1fb5bf84792aa87006addcb20726/preprint__2af794e8bdc6d7671634e08f6af1da53069c1fb5bf84792aa87006addcb20726.mmd

@@ -0,0 +1,376 @@
+
+# In-situ atomic observations unveil the aggregation growth and evolution of five-fold twin structures  
+
+Miao Song songmiao@csu.edu.cn  
+
+Central South University https://orcid.org/0000- 0002- 0483- 6580  
+
+Dingri Zhang Central South University  
+
+Dan Leng Central South University  
+
+Jaewon Lee j.lee@missouri.edu  
+
+Ziang Yang Central South University  
+
+Jiaxuan Chen Central South University  
+
+Dan Li Central South University  
+
+Lei Wang Shandong University  
+
+Gang Zhou Institute of Metal Research, Chinese Academy of Sciences  
+
+Rui Yang Institution of Metal Research, Chinese Academy of Sciences https://orcid.org/0000- 0003- 4405- 5882  
+
+Kechao Zhou Central South University  
+
+Article  
+
+Keywords:  
+
+Posted Date: April 23rd, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4283157/v1
+
+<--- Page Split --->
+
+License: © ① This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on October 25th, 2024. See the published version at https://doi.org/10.1038/s41467-024-53501-0.
+
+<--- Page Split --->
+
+# In-situ atomic observations unveil the aggregation growth and evolution of five-fold twin structures  
+
+Miao Song \(^{1,2,*}\) , Dingri Zhang \(^{1}\) , Dan Leng \(^{1}\) , Jaewon Lee \(^{3}\) , Ziang Yang \(^{1}\) , Jiaxuan Chen \(^{1}\) , Dan Li \(^{1}\) , Lei Wang \(^{2}\) , Gang Zhou \(^{4,*}\) , Rui Yang \(^{4}\) , Kechao Zhou \(^{1}\)  
+
+\(^{1}\) State Key Laboratory of Powder Metallurgy, Central South University, Changsha, Hunan 410083, China \(^{2}\) State Key Laboratory of Crystal Materials, Shandong University, Jinan, Shandong, 250100, China \(^{3}\) Department of Mechanical and Aerospace Engineering, College of Engineering, University of Missouri, Columbia, MO 65203, USA \(^{4}\) Shi-changxu Innovation Center for Advanced Materials, Institute of Metal Research, Chinese Academy of Sciences, Shenyang, Liaoning 110016, China  
+
+\*Corresponding author. Email: songmiao@csu.edu.cn; gzhou@imr.ac.cn  
+
+The unique twin boundaries and inherent lattice strain of five- fold twin (5- FT) structures offer a promising and innovative approach to tune nanocrystal configurations and properties, enriching nanomaterial performance. However, due to constraints imposed by small thermodynamically stable size and complex twin configurations, gaps persist in understanding the nonclassical growth models of 5- FT nanoparticles. Here, we in- situ investigated the mechanisms underlying size- dependent and twin configuration- related aggregation growth phenomena between 5- FT and other nanoparticles at the atomic scale. The results find that surface diffusion shapes the morphology of aggregated nanoparticles, promoting symmetrical 5- FT formation, particularly involving smaller nanoparticles. Additionally, the inherent structure of 5- FT mitigates the dominance of surface diffusion in its morphological evolution, retarding the aggregation evolution process and fostering intricate twin structures. Our findings contribute to advancing our ability to manipulate the configuration of twinned particles and achieve a more predictable synthesis of novel functional nanomaterials for engineering applications.
+
+<--- Page Split --->
+
+Uncovering the growth of nanoparticles (NPs) is significant not only for elucidating the underlying formation mechanisms of minerals and understanding the comprehensively evolution process of natural environment, but also holds scientific and engineering significance to tune size, morphology, and properties of synthetic nanomaterials. Despite numerous evidences challenge the classical interpretations of crystal growth, which emphasize surface reactions and the monomer diffusion to the surface \(^{1,2}\) , nonclassical crystal growth scenarios have only begun to be recognized gradually just over a decade ago \(^{2,3,4,5}\) . As a main nonclassical crystal growth mechanism, the particle- based aggregation \(^{3}\) , generally including orientated attachment (OA) \(^{4,6}\) , nearly OA (such as mesocrystal \(^{7,8,9}\) and dislocation induced tilt attachment \(^{4}\) ), and non- OA (such as intraparticle growth \(^{10,11}\) , aggregation and transformation of thermodynamically metastable particle to stable phases \(^{8,12}\) , and aggregation and grain boundary/surface atom migration dominated growth \(^{10,13}\) ), requires systematic and intensive investigation. These are determined by unknown fundamental and challengeable aspects, such as high spatial and temporal resolution of atomic diffusion and migration, particle relative movement, interfacial interactions, and grain boundary evolution.  
+
+The 5- FT structure has been extensively detected in diverse natural and synthetic systems \(^{14,15}\) , \(^{16}\) and exhibits unique properties attributed to its crystallographically forbidden pentagonal symmetry and inherent lattice strain, such as enhanced mechanical properties \(^{17}\) , attractive catalytic properties \(^{18}\) , and excellent optic properties \(^{19}\) . In addition to NP size, morphology, and composition, the twin plane and inherent lattice strain of 5- FT provide an appealing and novel avenue to tailor the configuration and properties of nanocrystals, thus diversifying and enhancing the performance of nanomaterials. For instance, 5- FT structure holds potential for producing hierarchical materials \(^{20,21,22}\) , which preserve the properties of the nanoscale building blocks and may exhibit novel performance characteristics \(^{22,23}\) . Nevertheless, limited by complicated twin structures (5 twin units), significantly small thermodynamic stability size (3- 14 nm) \(^{24,25,26}\) , together with movement of NPs and the only proper [110] twin pole observation direction, revealing the atomic formation and growth mechanisms of 5- FT remains seriously experimental challenging. By trigging NPs aggregation through electron- beam- induced decomposition of the organic ligands surrounding the NPs, we uncovered two underlying atomic formation mechanisms of 5- FT within Au, Pd, and Pt nanomaterials \(^{6}\) . However, nonclassical growth mechanisms of 5- FT are still mysterious and elusive, particularly potentially under the coupling effect of thermodynamic and kinetics landscapes,
+
+<--- Page Split --->
+
+involving but not limited to the atomic formation and evolution of complicated twin structure (twining and de- twinning processes), atom surface diffusion, and relative slip and configuration modulation of NPs. Thereby, clarifying the particle- based aggregation growth of 5- FT is scientifically and engineeringly significant.  
+
+In this work, in- situ high- resolution transmission electron microscopy (HRTEM) combined with molecular dynamics (MD) simulation was utilized to study the atomic aggregation growth and evolution mechanisms involving 5- FT and diverse NPs, i.e., NPs with varied size ratios and twinned configurations (single crystal- SC and 5- FT). Additionally, the impact of various thermodynamic and kinetic landscapes on the aggregation evolution was systematically investigated and comprehensively elucidated.  
+
+## Results and discussion  
+
+## In-situ HRTEM observation of aggregation evolution between 5-FT and SC NPs  
+
+Spherical Au NPs with various twin configurations ( \(\sim 4.4 \mathrm{nm}\) , Supplementary Fig. 1) embedded in 1- dodecanethiol organic matrix were drop- casted onto a TEM grid. Then, NP aggregation was induced by decomposing organics under electron- beam irradiation (see “Method” for details). As shown in Fig. 1a- b, aggregation between a 2.1 nm SC and a 3.6 nm symmetrical 5- FT (S5- FT, size ratio \(R = 0.58\) ) induces obvious growth of the twin units 1 and 5, resulting in the formation of an asymmetrical 5- FT (AS5- FT). Nevertheless, the AS5- FT evolves into a symmetrical twin structure (Fig. 1c), and transforms into a stable re- entrant {111} (highlighted by white arrows in Fig. 1d) Marks S5- FT at last. Considering none visible twin boundaries (TBs) migrations, the above evolution process is supposed to be dominated by surface diffusion from the small NP to the 5- FT, which will be further discussed in Fig. 2. Wherein, the formation of highly faceted structure (Fig. 1d) has been proved to be associated with the balance between surface and strain energy \(^{27,28}\) , i.e., the re- entrant {111} Marks 5- FT can significantly release the intrinsic lattice strain energy and slightly increases the surface energy of the 5- FT structure.
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Fig. 1. In-situ observation of aggregating growth between 5-FT NPs with SC NPs across diverse size ratios. (a-d) Configuration evolution after a \(2.1\mathrm{nm}\) SC NP attached with a 5-FT NP (3.6 nm) with \(R = 0.58\) . White arrows show formation of the re-entrant surface. (e-o) Configuration evolution after a \(4.6\mathrm{nm}\) SC NP aggregated with a \(4.2\mathrm{nm}\) 5-FT NP \((R = 1.10)\) . (p) The surface outlines of the aggregated NP. The migration directions of the surface outlines with durations are denoted by red arrows. Notable, twin boundaries are highlighted by red lines. Twin units are denoted by numbers ranging from 1 to 5 in (a) and (e). Yellow arrows show migration directions of the TBs, compared with the immediate prior image. Partial dislocations are denoted by blue “L” and their slip directions are denoted by the blue arrows. </center>  
+
+Different from the above small size ratios (Fig. 1a- b), the large SC \((R = 1.10)\) only stimulates continuous growth of the twin unit 1 with substantially de- twinning process until the final
+
+<--- Page Split --->
+
+manifestation of a new SC (Fig. 1e- o). And the detailed evolution processes are as follows: Initially, a \(4.6\mathrm{nmSC - 1}\) attaches with a \(2.1\mathrm{nmSC - 2}\) and a \(4.2\mathrm{nmS5 - FT}\) successively, resulting in attractive growth of the twin unit 1 (Fig. 1e- g). Then, a partial dislocation nucleates at the TB \(\Sigma 3_{1}\) from the periphery of the NP and slips toward the center of the 5- FT, leading to a single layer of migration of the \(\Sigma 3_{1}\) near the periphery (Fig. 1h). Subsequently, continuous nucleation and slip of partial dislocations at TBs induces de- twinning process of the 5- FT (Fig. 1i- o), giving rise to reduction of the twin units 2, 3, and 4. During this process, when the twin pole of the 5- FT is close to overlap (Fig. 1j), specifically, it involves one layer of migrations of the four TBs \((\Sigma 3_{1},\Sigma 3_{2},\Sigma 3_{3},\) and \(\Sigma 3_{4})\) . Alternatively, there is obvious split of the twin pole. At last, when the 5- FT twin pole is close to \(\sim 2 - 3\) layers of {111} planes away from) the periphery of the NP, 5- FT structure disappears within 0.8 s, resulting in the formation of the new SC NP (Fig. 1n- o). To access the influence of surface diffusion on the entire aggregation process, further analysis is conducted on the variations in surface outlines. As presented in Fig. 1g- p, distinct variations of the surface outlines are detected from 83.2 s to 179.4 s, especially at the concave surfaces and the region near the 5- FT twin pole (Fig. 1p), nevertheless, the variation can be ignored from 179.4 s to 220.8 s (Fig. 1p). Therefore, during the above aggregation induced evolution process, the surface diffusion mainly occurs at the initial stage (Fig. 1g- k), and the partial dislocations induced twin boundaries slip mainly happens at the subsequent stage (Fig. 1i- o) with an overlapping duration from 142.6 s to 179.4 s (Fig. 1i- k). Notably, surface diffusion without significant morphology variations, such as mutual diffusion, remains undetectable in Fig. 1. The relative slip between the initial 5- FT and SC NPs is also indistinguishable, which can potentially induce morphological variations. These phenomena will be comprehensively analyzed in Fig. 2 based on high time resolution data and 3- D atomic evolution tracing using MD simulations.  
+
+Therefore, as presented in Fig. 1, after attachment between a 5- FT NP with a small \((R = 0.58\) Fig. 1a- d) or a relatively large \((R = 0.73\) , Supplementary Fig. 6) SC NP, surface diffusion dominates the aggregation growth processes, resulting in formation of a S5- FT or an asymmetrical 5- FT (AS5- FT) NP, respectively. Nevertheless, during the aggregation growth process of a 5- FT with a large \((R = 1.10\) , Fig. 1e- o) SC NP, the surface diffusion exclusively dominates the initial stage of surface modulation. Subsequent to this, the activation of partial dislocations governs the de- twinning process, resulting in the manifestation of a new SC NP. Notably, based on our previous observation
+
+<--- Page Split --->
+
+6, asymmetrical 5- FTs can also evolve into symmetrical ones through NPs aggregation at small twin units. Meanwhile, with the growing of 5- FT NPs, its inherent lattice strain \(^{29}\) can be relieved through various mechanisms, including but not limited to the formation of plane defects in twin units \(^{30}\) , the migration of the 5- FT twin pole \(^{30,31}\) , and the formation of re- entrant Marks decahedral morphology \(^{32}\) . Thereby, the aforementioned AS5- FT, including various unstable intermediate 5- FT NPs, can be potentially stabilized via attachment with other NPs in conjunction with the modulation of the inherent lattice strain.  
+
+## Aggregation growth mechanisms between 5-FT and SC NPs elucidated via MD simulation  
+
+To uncover detailed evolution mechanisms of NPs' 3D configurations, the aggregation processes between S5- FT and SC NPs are further investigated by employing high temporal resolution MD simulation (see “Method”, Supplementary Fig. 2 and Table 1). As shown in Fig. 2a- c, the SC NP (B3s, \(R = 0.52\) , Fig. 2a) was initially designed to attach with the twin unit 1. However, the SC NP evolves into amorphous state and its atoms diffuse symmetrically towards the S5- FT surface quickly (Fig. 2b and also Supplementary Fig. 7), inducing significantly growth of the twin units 1 and 2, resembling the phenomena detected in Fig. 1a- b. The mutual diffusion of surface several layers' atoms, are also detected after 4.0 ns with formation of re- entrant {111} surfaces (Fig. 2c- f). Notably, significantly fewer atoms diffuse from the 5- FT to the B3s SC NP, and the atoms from the initial NP can even diffuse to the far- away twin unit 4 of the 5- FT (Fig. 2f). Correspondingly, the system experiences a substantial decrease in relative energy primarily at the initial stage (0.0- 4.0 ns, Fig. 2g), indicating significant configuration modifications. Therefore, the surface diffusion from small SC to the big 5- FT NP dominates the aggregation growth processes, consistent with the experimental results detected in Fig. 1a- d. Similar mutual diffusion is also detected at the initial stage of the simulated system with a relatively large particle size ratio (0.76, Supplementary Fig. 8). Nevertheless, the 5- FT structure swiftly transforms into a bi- crystal within 0.1 ns (Supplementary Fig. 8d- e), accompanying with significant decreasing of the relative energy (Supplementary Fig. 8g). Additionally, there is also diffusion induced growth of both twin units, i.e., the twin units 1 and 5 (Supplementary Fig. 8d). This verifies that the formation of initial asymmetrical configuration, i.e., significant growth of both twin units (Fig. 1a- b and Supplementary Figs. 6 and 8), is an inherent characteristic of surface diffusion dominated aggregation processes when \(R < - 0.76\) .
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Fig. 2. MD simulation of the aggregation growth of 5-FT NPs by attaching with various SC NPs. (a-e) Evolution of an aggregated 5-FT NP, formed by attaching of a \(5.0 \mathrm{nm}\) symmetrical 5-FT NP (yellow, \(\mathrm{A}_{5\mathrm{F}}\) ) and a \(2.6 \mathrm{nm}\) SC NP (cyan, \(\mathrm{B}_{35}\) ) with \(R = 0.52\) . Re-entrant {111} surfaces are highlighted by gray solid lines. (f) The biggest cross-section perpendicular to the Z axis showing atomistic distribution of the initial SC NP. (g) Variation of the relative energy of the simulated system from (a) to (e). (h-r) The aggregation evolution process after a \(5 \mathrm{nm}\) 5-FT Au NP ( \(\mathrm{A}_{5\mathrm{F}}\) ) attached with a \(4.8 \mathrm{nm}\) SC NP ( \(\mathrm{B}_{35}\) ) with \(R = 0.96\) . The relative slip between the 5-FT and the SC is represented by cyan arrows in (i-k) and (m). (s) Variation of the relative energy of the simulated system from (h) to (r). (t) TBs slip induced migrations of the twin pole from \(2.5 \mathrm{ns}\) (i) to \(4.8 \mathrm{ns}\) (o). </center>
+
+<--- Page Split --->
+
+(u) A double Thompson tetrahedron with an FCC crystal structure. Vertex-to-vertex (e.g., \(AB\) ), vertex-to-orthocenter (e.g., \(A\gamma\) ), and orthocenter-to-orthocenter (e.g., \(\alpha \gamma\) ) denote a perfect, partial, and stair-rod dislocations, respectively. (v) Relationship of the Thompson tetrahedron in 5-FT units verifying that dislocations within different twin units can undergo mutual transformations via various dislocation reactions. (w) Schematic illustration of twin pole migration (o to o'). TBs before and after slip are denoted by dashed yellow lines and solid red lines, respectively. Migration directions and layers of the TBs are denoted by yellow arrows and numbers, respectively. Partial dislocations are denoted by blue "L" and their slip directions are denoted by the blue arrows. Twin units are denoted by numbers ranging from 1 to 5 in (a) and (h).  
+
+Divergent from the aforementioned surface diffusion dominated NPs aggregation evolution mechanisms ( \(R< \sim 0.76\) , Fig. 2a-f and Supplementary Fig. 8), MD simulation indicates that the detwinning process induced by TBs' migrations governs the aggregation evolution as the \(R\) increases to 0.96 (Fig. 2h-r). This aligns with the experimental results delineated in Fig. 1i-o. And the aggregation evolution process can be segmented into the following three distinct stages: (1) The initial relative slip. As presented in Fig. 2h-i, the interaction between the 5-FT and the SC gives rise to the relative slip of 2 layers of {111}, and the slip direction of the 5-FT aligns parallel to the \(\Sigma 3_{5}\) (highlighted by the cyan arrow). (2) TB migration dominated de-twinning process. This process is trigged by nucleation and slip of partial dislocations (Fig. 2j-o). As shown in Fig. 2j, the nucleation and slip of a partial dislocation from the periphery of the concave surface to the twin pole instigate the migration of the TB \(\Sigma 3_{5}\) towards the initial SC. Notably, there are also relative slip between the 5-FT and the SC (Fig. 2j, k, and m). These determinations arise from the layer variation of {200} planes between the twin pole and the periphery of the initial SC (prior to the migration of the twin pole, Fig. 2h-i), alternatively, the layer variations of {111} planes from the \(\Sigma 3_{4}\) (without migration) to the {111} intersection plane in the twin unit 5 with the initial {200} interface between the 5-FT and the SC (Fig. 2j, k and m). (3) Surface diffusion dominated de-twinning process. When the twin units are as small as 3 layers of {111} planes (Fig. 2o-r), surface diffusion takes precedence as the predominant mechanism during the final de-twinning process, while no discernible TB migration can be detected. This differs from the detected TB migration dominated de-twinning process in Fig. 1n-o, and is supposed to be associated with NP's configuration. That is, when the twin pole of the AS5-FT is close to the convex surface (Fig. 1n-o and Supplementary Fig. 10f-g),
+
+<--- Page Split --->
+
+TB migration induces the fast de- twinning process. Nevertheless, when the serious AS5- FT twin structures are close to the concave surface (Fig. 2o- r), surface diffusion dominates the de- twinning process. Fig. 2s shows there is a notable increase in the system energy during the initial stage of the relative slip (Fig. 2h- k), followed by a substantial energy decrease attributed to the de- twinning process dominated by partial dislocations. This is mainly associated with reduction of TB energy and lattice strain energy \(^{31,33}\) . After that, the variation in the system energy becomes moderate. Notably, mutual diffusion is consistently detected throughout the entire evolution process. The amorphous detected in MD simulations (Fig. 2b- c and Supplementary Fig. 8), as opposed to experimental observation (Fig. 1a- c), is supposed to correlate with the relatively elevated simulation temperature employed to expedite the simulated experimental process (see "Method").  
+
+## De-twinning mechanisms  
+
+Effective TBs migration layers during the TBs migrations dominated de- twinning process (Fig. 2i- o) are illustrated in Fig. 2t. That is, TBs \(\Sigma 3_{1}\) and \(\Sigma 3_{2}\) migrate 4 layers, TBs \(\Sigma 3_{3}\) and \(\Sigma 3_{5}\) migrate 5 layers, remaining \(\Sigma 3_{4}\) are unchanged. The migration of TBs is closely associated with nucleation and slip of partial dislocations, and the detailed mechanisms are depicted in Fig. 2u- w. A double Thompson tetrahedron is generally employed to define various dislocations in face- centered cubic crystal \(^{34}\) , including perfect dislocation (vertex to vertex, such as \(AB\) and \(BD\) ), partial dislocation (vertex to orthocenter, such as \(A\delta\) and \(B\delta\) ), and stair- rod dislocation (orthocenter to orthocenter, such as \(\alpha \beta\) and \(\beta \delta\) ). Dislocations have the capability to undergo transformation into one another through reactions, guided by the principles of vector geometry addition and subtraction rules \(^{34}\) . For instance, \(DB = DA + AB\) , \(\beta \delta = 1 / 3DB\) , \(DB = D\beta +\beta \delta +\delta B = 1 / 3DB + D\beta +\delta B\) , \(DB = 3 / 2(D\beta +\delta B)\) . Here, the 5- FT NP can be regarded as a five Thompson tetrahedron (Fig. 2v), and dislocations can transform into one another by diverse dislocation reactions. Based on experimental observation (Fig. 1h) and MD simulation (Fig. 2j), partial dislocations initially nucleate at the periphery of the NPs. For instance, a partial dislocation \(C\beta '\) nucleates on \(\Sigma 3_{5}\) (Fig. 2w) and slips towards the twin pole \(O\) , inducing one layer migration of \(\Sigma 3_{5}\) . Then, \(C\beta '\) dissociates into other two partial dislocations at the twin pole, i.e., \(C\beta ' \rightarrow C\eta +\eta \beta '\) , and \(C\eta\) can slip along \(\Sigma 3_{3}\) , resulting in one layer of migration of \(\Sigma 3_{3}\) . Meanwhile, \(\eta \beta '\) can be equivalent to \(1 / 2DB\) (\(\eta \beta '\) serves as the median of the triangle FBD, Fig. 2v). And as analyzed above, \(DB = 3 / 2(D\beta +\delta B)\) , thereby, \(\eta \beta '\) can be dissociated into \(D\beta\) and \(\delta B\) , i.e.,
+
+<--- Page Split --->
+
+\(\eta \beta^{\prime} = 1 / 2DB = 3 / 4(D\beta +\delta B)\) . Correspondingly, partial dislocations \(D\beta\) and \(\delta B\) can trigger migration of TBs \(\Sigma 3_{1}\) and \(\Sigma 3_{2}\) , resulting in migration of the twin pole form \(O\) to \(O^{\prime}\) (Fig. 2w), i.e., one layer of de- twinning process of 5- FT. Based on the above dislocation reactions, one \(C\beta^{\prime}\) only generates \(3 / 4(D\beta +\delta B)\) , that is, a single layer of migration of \(\Sigma 3_{5}\) results in \(3 / 4\) layer of migrations of \(\Sigma 3_{1}\) and \(\Sigma 3_{2}\) . Nevertheless, based on the geometrical relation of 5- FT, one \(C\beta^{\prime}\) still generates \(\sim 0.18\) more \(D\beta\) and \(\delta B\) than that to maintain the perfect overlap of the twin pole. This residual mismatch can be mediated by various migration layers of TBs, such as 5 layers of \(\Sigma 3_{3}\) and \(\Sigma 3_{5}\) and 4 layers of \(\Sigma 3_{1}\) and \(\Sigma 3_{2}\) ( \(5\times 0.18 = 0.90\) layer, closed to a single layer of difference, Fig. 2i- o and t). This observation is also consistent with the experimental observation (Fig. 1h- j), i.e., the TBs \(\Sigma 3_{1}\) , \(\Sigma 3_{2}\) , \(\Sigma 3_{3}\) , and \(\Sigma 3_{4}\) migrate one layer to maintain the nearly overlapped twin pole.  
+
+## In-situ HRTEM observation of aggregation evolution between 5-FT NPs  
+
+Aggregation induced growth and TB evolution process are also investigated between 5- FT NPs with various size ratios (Fig. 3). As presented in Fig. 3a- b, aggregation between a 3.7 nm S5- FT and a 4.6 nm S5- FT induces the formation of an AS5- FT (Fig. 3d). And the AS5- FT can potentially evolve into S5- FT by aggregating with other NPs (denoted by a white arrow, Fig. 3d), as discussed above. With increasing of the size ratio between the two aggregated 5- FT NPs, intricate twin structures can be induced ( \(R = 0.82\) , Fig. 3e- l), and the detailed evolution processes are as follows. At first, the twin unit 4 of the small S5- FT NP orientationally attaches with the twin unit 2 of the big S5- FT NP (Fig. 3e- f), followed by the formation of a TBs sealed region. This is introduced by the relative slip between the two NPs (layers variation without corresponding TBs migration, highlighted by cyan arrows in Fig. 3f- g), the surface diffusion towards the concave surface (denoted by red arrows, Fig. 3f- h), and the TBs migration (highlighted by yellow arrows, Fig. 3f- i). Then, the sealed region maintains unchanged until the small 5- FT evolves into three- fold twin at 72.0 s (Fig. 3i- k). This process is dominated by surface diffusion, verified by significant morphology variation and two reserved TBs, i.e., only disappearing of the twin units 1 and 2 of the S5- FT NP. Thereby, the de- twinning process (Fig. 3j- k) is significantly different from the TB migration decided de- twinning process (Fig. 3n- o), in which the morphology variation can be ignored. Thereby, the TBs sealed region can stabilize the 5- FT structures and modify the evolution pathway of the aggregation growth mechanism between 5- FT NPs. Nevertheless, the sealed region becomes unstable after 5-
+
+<--- Page Split --->
+
+FT evolving into 3- fold twin within 1.0 s (Fig. 3k- l), resulting in the formation of a complicated twin structure, which can potentially evolve into AS5- FT by de- twinning of the TB at bottom left. Notably, when the size ratios between two aggregated 5- FT is small, such as 0.47 and 0.53 (Supplementary Fig. 9), S5- FT NPs ultimately formed. Additionally, beside the orientation attachment dominated NPs aggregation (with no detectable formation of new GBs at attachment sites, Fig. 3e- f), another scenario emerged, where a new zig- zag {200}/{111} GB was induced after aggregation ( \(R = 0.72\) , see Supplementary Fig. 10 for details).  
+
+![](images/Figure_3.jpg)
+
+<center>Fig. 3. Aggregating growth between 5-FT NPs. (a-d) Atomic evolution process between a 3.7 nm 5-FT and a \(4.6 \mathrm{nm}\) 5-FT with \(R = 0.78\) . (e-l) Atomic evolution process between a \(3.3 \mathrm{nm}\) 5-FT and a \(4.0 \mathrm{nm}\) 5-FT with a size ratio of 0.82. Twin units are denoted by numbers ranging from 1 to 5. TBs and unclear TBs are highlighted by red solid and dashed lines, respectively. The white arrow shows the aggregation of another NP in (d). Yellow arrows show migration directions of the TBs, compared with the immediate prior image. Partial dislocation is denoted by blue “L” and its slip direction is denoted by the blue arrow. </center>
+
+<--- Page Split --->
+
+## Aggregation growth mechanisms between 5-FT NPs elucidated via MD simulation  
+
+Fig. 4 shows aggregation growth between 5- FT NPs with various size ratios. As presented in Fig. 4a- b, after orientation attachment with the twin unit 1 of the \(5.0 \mathrm{nm} 5\mathrm{- FT}\) , the small 5- FT NP evolves into amorphous state within 0.1 ns (Fig. 4b). There is no detectable TB migration. Then, mutual diffusion dominates the subsequent evolution process and induces the formation of a symmetrical Marks 5- FT (Fig. 4c- e). Cross- section (Fig. 4f) shows that atoms originating from the initial small 5- FT can even diffuse over considerable distance to the twin units 3 and 4 (Fig. 4f). The system energy variation is presented in Fig. 4g. The above process analogous to that detected in Fig. 1a- g. Thereby, when the size ratio is small enough, such as 0.52 in Fig. 2a- g and 0.62 in Fig. 4a- g, the aggregation kinetics is governed by surface diffusion from small NPs to the big NPs, resulting in the formation of the S5- FT structure. With increasing of the size ratio to 0.82, apparent relative slip (Supplementary Fig. 13a- b) and TBs migration (Supplementary Fig. 13b- c) are detected, corresponding to significant decreasing of the system energy (Supplementary Fig. 13h). Although the relatively small 5- FT NP still evolves into amorphous state accompanying with surface diffusion towards the big 5- FT (Supplementary Fig. 13c- g), an AS5- FT ultimately forms.  
+
+For the large size ratio between two aggregated 5- FT NPs, MD simulation shows there is also relative slip at the initial stage (Fig. 4h- i) inducing the formation of a sealed stable region S1 (Fig. 4i). The region S1 can stabilize for a long time until the twin units 3 and 4 of the left 5- FT decrease to three and four layers, respectively, resulted by surface diffusion (Fig. 4i- l). Then, TBs migration induces \(\Sigma 3_{2}\) , \(\Sigma 3_{3}\) , \(\Sigma 3_{4}\) , and \(\Sigma 3_{5}\) TBs disappear within 0.02 ns (Fig. 4k- l), accompanying with significant decreasing of the system energy (Fig. 4o). Therefore, MD simulation results are consistent with the phenomena detected during in- situ observation (Fig. 3e- l). As for the reserved TB \(\Sigma 3_{1}\) , it disappears at last with forming of an AS5- FT (Fig. 3l- m). The system energy decreases slowly during the initial 69.50 ns, accompanying with moderate surface modulation, and decreases seriously with de- twinning of the left 5- FT NP. This means that the system energy of the aggregated 5- FT NPs is mainly introduced by the 5- FT structures. Additionally, while the surface diffusion remains a dominated factor during 5- FT NP's aggregation, de- twinning process, i.e., TB migration, is often observed across experimental (Fig. 3) and simulated (Fig. 4) conditions. This phenomenon stems from the intricate and highly strained twin configuration inherent in 5- FT NPs, as opposed to the SC NPs. Notably, due to the formation of the stable sealed region S1, a significant long
+
+<--- Page Split --->
+
+simulation duration was employed in Fig. 4 (80.00 ns) compared with that of the previous MD simulations (30.00 ns, Fig. 2 and Supplementary Figs. 8, 12, and 13).  
+
+![](images/Figure_4.jpg)
+
+<center>Fig. 4. MD simulation of the aggregation growth process of S5-FT NPs by attaching with other S5-FT NPs. (a-e) and (h-m) Evolution of the aggregated 5-FT NPs, formed by attaching of a \(5.0 \mathrm{nm}\) S5-FT NP (yellow, \(\mathrm{A_{5F}}\) ) with \(3.2 \mathrm{nm}\) ( \(\mathrm{B_{3F}}\) ) and \(5.0 \mathrm{nm}\) ( \(\mathrm{B_{5F}}\) ) S5-FT NPs (cyan), respectively. The corresponding size ratios are 0.64 and 1.00, respectively. (f) and (n) The biggest cross-sections perpendicular to the Z axes showing atomistic distribution of the initial 5-FT NPs of \(\mathrm{B_{3F}}\) and \(\mathrm{B_{5F}}\) , respectively. (g) and (o) Variation of the relative energies of the corresponding simulated systems. TBs are highlighted by red lines. Twin units are denoted by numbers ranging from 1 to 5. Slip directions and layers of the TBs are denoted by yellow arrows and numbers, respectively. </center>  
+
+## Size effect of aggregation growth between various NPs  
+
+Symmetry evolution of the aggregated 5-FT NPs is closely associated with surface diffusion and TBs migration. As observed in all experimental results (Figs. 1 and 3), during aggregation between various NPs, the surface diffusion occurs almost throughout the whole evolution processes.
+
+<--- Page Split --->
+
+And surface diffusion generally promotes the generation of S5- FT (formed based on the big 5- FT NPs), especially for small particle size ratios, by transferring atoms from small NPs to the twin units far away from the attached regions (Figs. 1a- d, Figs. 2a- f, Figs. 3a- h, Fig. 4a- f). Nevertheless, the effect of TBs migration on the symmetry of the aggregated NPs depends on the twin pole migration direction. The migration of the reserved 5- FT twin pole towards the periphery of the NPs results in a reduction in symmetry, and vice versa. Based on dozens of experimental observations, the size effect of the aggregation growth between 5- FT and various NPs (SC and 5- FT) are illustrated in Fig. 5. Correspondingly, to obtain the S5- FT, the critical ratios should be roughly smaller than 0.83 and 0.72, respectively. According to the theoretical models (Supplementary Table 1), the numbers of atoms of SC is significantly larger than that of the 5- FT with the same particle size. And even \(R = 0.76\) (model 1- 2, Supplementary Table 1, smaller than the critical \(R = 0.83\) ), the number of atoms of SC is also larger than that of the corresponding attached 5- FT. This indicates that: (1) under investigated particles size (several nanometers), S5- FT exhibits relative stability, maintaining its symmetrical structure after attaching with small SC NP, and even the SC possesses a slightly larger number of atoms than that of the 5- FT. That is, surface diffusion predominantly governs the aggregation evolution, rendering negligible the migration of TBs or twin poles. (2) Relatively small critical \(R\) (0.72, atoms significantly less than that of the attached 5- FT, model 2- 2, Supplementary Table 1) between two aggregated 5- FT also verifies the stability of the 5- FT NP, i.e., the 5- FT structure can greatly mitigate surface diffusion. (3) The mutual surface diffusion from small NP to 5- FT is markedly greater than the reverse diffusion, i.e., the surface energy plays a significant role during NP's aggregation growth.  
+
+To maintain symmetrical configuration of 5- FT NPs, the critical transfer ratio registers at \(\sim\) 0.72 for 5- FT/5- FT, which is smaller than that of SC/5- FT ( \(\sim 0.83\) ). As discussed above, the surface diffusion persists throughout the entirety evolution processes and facilitates to maintain the symmetrical configuration of the big 5- FT NPs. Thereby, for two aggregated 5- FT NPs, the formation of S5- FT is mainly associate with the de- twinning process of the small 5- FT. This can be introduced by TBs migration (Fig. 3c- d and Supplementary Figs. 12- 13) and/or surface diffusion (Fig. 4a- e and Supplementary Figs. 12- 13). Nevertheless, the S5- FT remains relatively stable and maintains its symmetry even when attaching to SC NPs with a slightly larger number of atoms compared to the 5- FT, that is, it is relative difficult to trigger TBs migration of symmetrical 5- FT
+
+<--- Page Split --->
+
+NPs. Additionally, van der Waals (vdW) force calculation (Supplementary Fig. 14) shows that the force between 5- FT and SC NPs is significantly stronger than that between two 5- FT NPs. This is introduced by the structural difference between 5- FT and SC, i.e., a greater number of atoms can exhibit a close interatomic distance in the 5- FT/SC interactions compared to the 5- FT/5- FT (see Supplementary Fig. 14 and related discussion). Thereby, the force exerted on the small S5- FT is minimal, which not only results in a diminished driving force to initiate TBs migration but also mitigates morphological modulation. Therefore, the relatively smaller critical transfer ratio for 5- FT/5- FT (0.72) than that of the SC/5- FT ( \(\sim 0.83\) ) is determined by the stable twin structure and decreases interaction force introduced by the small S5- FT. Notably, the particle size of various aggregated NPs investigated is smaller than \(10 \mathrm{nm}\) (Fig. 5). The above critical transfer ratio is supposed to vary slightly with further growth of the particle size of 5- FT, considering the size effect on 5- FT, which is stable with the size of \(3 - 14 \mathrm{nm}\) owing to thermodynamics \(^{24,25,26}\) .  
+
+![](images/Figure_5.jpg)
+
+<center>Fig. 5. Aggregation growth size effect of 5-FT NP. (a) Configuration obtained after aggregation between SC and 5-FT NPs. (b) Configuration obtained after aggregation between two 5-FT NPs. </center>  
+
+In conclusion, our findings reveal the fundamental atomic- scale mechanisms governing the aggregation growth and evolution of 5- FT NPs. We have demonstrated that when a 5- FT attaches to small NPs ( \(R< 0.83\) for SC and \(R< 0.72\) for 5- FT), surface diffusion of the small nanoparticles to the big 5- FT dominates the configuration evolution, resulting in the formation of S5- FT NPs and decreasing of the system energies. Nevertheless, when a 5- FT attaches to a larger SC NP, beside surface diffusion, TB migration induced de- twinning process contributes significantly to the evolution process. This dynamic interplay induces the formation of a SC or a simple twinned structure. Additionally, in the case of attachment to another big 5- FT NP ( \(R > 0.72\) ), this results in
+
+<--- Page Split --->
+
+the development of intricate 5- FT configurations.  
+
+Comprehending the mechanisms and dynamics underlying the process of particles aggregation growth is critical to constructing quantitative models to elucidate formation and evolution mechanisms of diverse twinned mineral materials in natural environments. Simultaneously, mastering the deterministic manipulation of NPs' growth holds the innovative synthesis pathways for fabricating twinned crystals endowed with precisely controlled morphologies and properties. Consequently, this endeavor provides guides to unlock the full potential of twinned crystal structures and promotes materials design and synthesis across a myriad of applications.  
+
+## Methods  
+
+## Sample preparation  
+
+Synthesis of gold (Au) nanoparticles (NPs): Gold(III) chloride trihydrate (HAuCl4·3H2O, \(\geq 99.9\%\) trace metals basis), tetrakis(decyl)ammonium bromide \((\geq 99.0\%)\) , TDAB), sodium borohydride (NaBH4, \(98\%\) ), 1- dodecanethiol \((\geq 98\%)\) and toluene \((\geq 99.5\%)\) were purchased from Sigma- Aldrich and used without further purification. The Au NPs were synthesized by the TDAB method \(6,35\) . 0.27 mmol of TDAB was dissolved in \(4\mathrm{ml}\) Toluene. The TDAB solution was mixed with \(6\mathrm{ml}\) of a \(0.01\mathrm{M}\) HAuCl4·3H2O aqueous solution for 1 hour. When the color of the toluene layer was changed to yellow due to the migration of AuCl4 ions, the only toluene layer was collected. Then, \(200\mu \mathrm{l}\) of a \(26\mathrm{mM}\) NaBH4 aqueous solution, which was prepared under an ice bath, was injected into the Au- TDAB toluene solution. After \(20\mathrm{min}\) , \(800\mu \mathrm{l}\) of 1- dodecanethiol was added to this mixture. The product was precipitated by excessive ethanol, and then collected by centrifuge.  
+
+After centrifuge, the sample was dispersed in \(5\mathrm{ml}\) toluene solution, drop- cast onto a TEM grid, and labeled as the as- prepared sample. As shown in Supplementary Fig. 1a- c, the as- prepared sample is SC dominated Au NPs with an average size of \(3.5\pm 0.6\mathrm{nm}\) . To tune particles' structure and size, half (2.5 ml) of the above toluene solution was sealed and kept in the dark at room temperature for \(\sim 2\) months (Supplementary Fig. 1d- f), and another half (2.5 ml) of the above toluene solution with centrifuged samples was cleaned four times with toluene to remove 1- dodecanethiol on the Au NP surface (Supplementary Fig. 1g- i). About \(1 / 3\) of NPs has evolved into 5- FT after staying for 2 months, and the particle size increases into \(4.4\pm 1.4\mathrm{nm}\) (Supplementary Fig. 1e- f). After purifying for 4 times, beside formation of some 5- FT NPs, significant aggregation of NPs was also detected
+
+<--- Page Split --->
+
+(highlighted by yellow arrows in Supplementary Fig. 1h). And the average particle size grows into \(6.2 \pm 2.6 \mathrm{nm}\) (Supplementary Fig. 1i). To investigate aggregation growth and evolution of 5- FT nanoparticles, the sample after staying for 2 months was employed for further study. Particle sizes were statistically analyzed based on more than 300 particles for each sample and the standard errors were employed (Supplementary Fig. 1).  
+
+## In-situ high-resolution TEM experiment  
+
+To investigate atomic scale 5- FT NPs growth and evolution mechanisms by particle aggregation, electron- beam induced NPs aggregation method \(^{6}\) was employed. This method has been successfully engaged to tune the formation kinetics of 5- FT Au nanoparticles, by controlling decomposition process of organic ligand coated on the Au nanoparticles under various electron beam dose rate \(^{6}\) . Here, a cold- field emission aberration- corrected TEM (Spectra 300, Thermal Fisher, USA), was employed at \(300 \mathrm{kV}\) for the above in- situ high- resolution TEM (HRTEM) imaging with an image frame rate of \(\sim 0.5 \mathrm{s}\) . An optimized dose rate of \(\sim (2 - 4) \times 10^{6} \mathrm{e} / \mathrm{nm}^{2} \cdot \mathrm{s}\) was applied. When the ratio between the longest TB and the shortest TB of 5- FT is smaller than 2.0, the 5- FT NT is referred to as a S5- FT. When the ratio is larger than 2.0, the 5- FT NP is referred to as an AS5- FT. Theoretically, the configuration of NPs evolves continuously under the employed electron- beam dose rate with time. Here, the terminal time was roughly determined according to the evolution rate of the NPs' configurations, including morphologies and twin structures. The ultimate configurations were obtained if their variation can be ignored within \(\sim 20 \mathrm{s}\) . Correspondingly, most of the in- situ experimental duration are \(\sim 150 - 200 \mathrm{s}\) under the employed electron- beam dose rate. Slip directions of partial dislocations are determined by TBs migration directions based on the immediate prior and subsequent images.  
+
+## MD simulation  
+
+MD simulations were conducted to explore the evolution processes of both the three- dimensional configuration and energy of simulated models of Au NPs (Supplementary Figs. 2- 3). Two types of models were employed in our experiment and all invariable decahedral 5- FT NPs (left side, Supplementary Figs. 2- 3) are the same, i.e., all surfaces of the big 5- FT are {111} planes with 1773 atoms. Model 1 was built with a \(5 \mathrm{nm}\) decahedral 5- FT NP attached with SC NPs in a box with
+
+<--- Page Split --->
+
+a side length of \(20\mathrm{nm}\) . The SC NPs with approximately \(2.6\mathrm{nm}\) , \(3.8\mathrm{nm}\) , and \(4.8\mathrm{nm}\) (Supplementary Fig. 2), present truncated octahedral morphologies and have corresponding numbers (Supplementary Table 1) of 811, 1862, and 4033 atoms, respectively. Model 2 was built with a \(5\mathrm{nm}\) decahedral 5- FT NP attached with decahedral 5- FT NPs in a box with a side length of \(20\mathrm{nm}\) . The size and numbers of atoms of various 5- FT NPs are \(3.2\mathrm{nm}\) , \(4.1\mathrm{nm}\) , \(5.0\mathrm{nm}\) and 569, 1061, 1773, respectively (Supplementary Fig. 3 and Table 1). The models were visualized by OVITO software<sup>36</sup>.  
+
+MD simulations were performed using the LAMMPS code<sup>37</sup>. The Au embedded atom potential developed by Ackland et al.<sup>38</sup> was employed in the simulation. This interatomic potential has been applied in previous investigations on the evolution of Au NPs during OA process successfully<sup>6</sup>. A Nose- Hoover thermostat<sup>39</sup> in the canonical ensemble was used to maintain the temperature of \(1100\mathrm{K}\) and the time step was configured to 1fs.  
+
+## Force calculation  
+
+Although the electrostatic force may argue as the source of the jump connection during the orientation attachment<sup>4</sup>, in our experiment, when the distance between the two aggregated NPs is smaller than \(\sim 2\mathrm{nm}\) , the vdW force is supposed to be the determining force<sup>35</sup>, especially the evolution process after aggregation, such as the relative slip between NPs.  
+
+vdW interactions between two NPs (NP<sub>1</sub> and NP<sub>2</sub>) were performed based on Hamaker's approach. To describe the vdW interactions of heterogeneous structures such as 5- FT structures, each NP was divided into many small cubes with length \((L) = 0.2882\mathrm{nm}\) , which is the closest Au- Au separation distance. Next, the vdW interactions between the two NPs were calculated by the summations of each pairwise interaction.  
+
+## 1) Dividing into small cubes  
+
+The center positions of the small cubes of the NP<sub>1</sub> and the NP<sub>2</sub> were defined as \((x_{1,i}, y_{1,i}, z_{1,i})\) and \((x_{2,j}, y_{2,j}, z_{2,j})\) , respectively.  
+
+## 2) vdW interaction potential  
+
+The vdW interaction potentials between the small cubes of the NP<sub>1</sub> and the small cubes of the NP<sub>2</sub> were calculated based on the formulation shown in previous work that followed Hamaker's approach with geometrical consideration.<sup>40</sup> The formulations can produce the vdW interaction energy \((V)\) between small cube- \(\alpha\) , i and cube- \(\beta\) , j with parallel configurations at arbitrary separations.  
+
+\[V_{\alpha \beta ,ij} = -\left(\frac{A}{\pi^2}\right)V_{R,\alpha \beta ,ij} \quad (Eq (1))\]  
+
+where A is the Hamaker constant \((359.579 \times 10^{- 21}\mathrm{J})\) obtained from the reference paper<sup>41</sup>, which calculated dielectric responses of Au in vacuum based on the Lifshitz theory. \(V_{R}\) is the London- vdW interaction energy between cubes,
+
+<--- Page Split --->
+
+\[V_{R,\alpha \beta ,ij} = \iint v_{\alpha ,i}v_{\beta ,j}(1 / r^6)\mathrm{d}v_{\alpha ,i}\mathrm{d}v_{\beta ,j} \quad (Eq 2)\]  
+
+where, \(r\) is the distance between an arbitrary point in the small cube- \(\alpha\) , \(i\) and an arbitrary point in the small cube- \(\beta\) , \(j\) , \(v_{\alpha ,i}\) and \(v_{\beta ,j}\) are the volume of the respective bodies. For the numerical computations, the integrations of Eq (2) were modified by  
+
+\[\begin{array}{r l} & {i f a_{m}\neq 0,b_{n}\neq 0a n d c_{p}\neq 0:}\\ & {V_{R,\alpha \beta ,i j}(c_{p},b_{n},a_{m})}\\ & {\qquad = \left[\left(\frac{c_{p}(a_{m}^{2} + b_{n}^{2})^{\frac{3}{2}}}{24a_{m}^{2}b_{n}^{2}}\right)\tan^{-1}\left(\frac{c_{p}}{\sqrt{a_{m}^{2} + b_{n}^{2}}}\right)\right.}\\ & {\qquad +\left(\frac{3}{32}\right)\left(\frac{c_{p}}{b_{n}} -\frac{b_{n}}{c_{p}}\right)\tan^{-1}\left(\frac{b_{n}}{c_{p}}\right)}\\ & {\qquad +\left(\frac{1}{24}\right)b_{n}\left(\frac{1}{a_{m}^{2}} +\frac{1}{c_{p}^{2}}\right)\sqrt{a_{m}^{2} + c_{p}^{2}}\tan^{-1}\left(\frac{b_{n}}{\sqrt{a_{m}^{2} + c_{p}^{2}}}\right)}\\ & {\qquad +\left(\frac{1}{24}\right)a_{m}\left(\frac{1}{b_{n}^{2}} +\frac{1}{c_{p}^{2}}\right)\sqrt{b_{n}^{2} + c_{p}^{2}}\tan^{-1}\left(\frac{a_{m}}{\sqrt{b_{n}^{2} + c_{p}^{2}}}\right)}\\ & {\qquad +\left(\frac{1}{32}\right)\ln \left(\frac{(b_{n}^{2} + c_{p}^{2})^{3}}{c_{p}^{2}(a_{m}^{2} + b_{n}^{2} + c_{p}^{2})^{2}}\right)\Bigg]}\\ & {\qquad i f a_{m}\neq 0,b_{n}\neq 0a n d c_{p} = 0:}\\ & {V_{x y z}(c_{p},b_{n},0) = \frac{1}{32}\left(\frac{c_{p}}{b_{n}} -\frac{b_{n}}{c_{p}}\right)\tan^{-1}\left(\frac{b_{n}}{c_{p}}\right) + \frac{1}{32}\ln \left(1 + \frac{b_{n}^{2}}{c_{p}^{2}}\right)}\\ & {\qquad i f a_{i}\neq 0,b_{j} = 0a n d c_{k}\neq 0:}\\ & {V_{x y z}(c_{p},0,a_{m}) = -\frac{1}{16}\left(\frac{c_{p}}{a_{m}} -\frac{a_{m}}{c_{p}}\right)\tan^{-1}\left(\frac{a_{m}}{c_{p}}\right) - \frac{1}{16}\ln \left(1 + \frac{a_{m}^{2}}{c_{p}^{2}}\right)} \end{array} \quad (Eq 3)\]  
+
+\[V_{x y z}(c_{p},b_{n},0) = \frac{1}{32}\left(\frac{c_{p}}{b_{n}} -\frac{b_{n}}{c_{p}}\right)\tan^{-1}\left(\frac{b_{n}}{c_{p}}\right) + \frac{1}{32}\ln \left(1 + \frac{b_{n}^{2}}{c_{p}^{2}}\right)\]  
+
+\[i f a_{i}\neq 0,b_{j} = 0a n d c_{k}\neq 0:\]  
+
+\[V_{x y z}(c_{p},0,a_{m}) = -\frac{1}{16}\left(\frac{c_{p}}{a_{m}} -\frac{a_{m}}{c_{p}}\right)\tan^{-1}\left(\frac{a_{m}}{c_{p}}\right) - \frac{1}{16}\ln \left(1 + \frac{a_{m}^{2}}{c_{p}^{2}}\right)\]  
+
+Parameter \(a_{m},b_{n}\) , and \(c_{p}\) were defined as below,  
+
+\[\begin{array}{r l} & {a_{1} = h z + L;a_{2} = h z + 2L;a_{3} = h z + L;a_{4} = h z;}\\ & {b_{1} = h y + L;b_{2} = h y + 2L;b_{3} = h y + L;b_{4} = h y;}\\ & {c_{1} = h x + L;c_{2} = h x + 2L;c_{3} = h x + L;c_{4} = h x;} \end{array} \quad (Eq 4)\]  
+
+where \(h x, h y\) , and \(h z\) were denoted, which are shown in the Supplementary Fig. 4.  
+
+Next, the vdW interaction potentials \((V)\) between \(\mathrm{NP}_{1}\) and \(\mathrm{NP}_{2}\) were calculated by  
+
+\[V = \sum_{i}\sum_{j}V_{\alpha \beta ,i j} \quad (Eq 5)\]  
+
+The interparticle distance between \(\mathrm{NP}_{1}\) and \(\mathrm{NP}_{2}\) \((h_{d})\) were defined as the distance between
+
+<--- Page Split --->
+
+closed atoms of each NC, which was shown in Supplementary Fig. 5.  
+
+3) vdW interaction force  
+
+Based on the vdW potential as a function of \(h_{d}\) , the vdW force was calculated numerically.  
+
+\[F = -\frac{V(h_{d} + \Delta d) - V(h_{d} - \Delta d)}{2\times\Delta d}, \quad (Eq (6))\]  
+
+Here, \(\Delta d = 0.05 \mathrm{nm}\) was employed.  
+
+## Data availability  
+
+The authors declare that the data supporting the findings of this study are available within the paper and its supplementary information. All raw data generated during the current study are available from the corresponding authors upon request.  
+
+## References  
+
+1. Sugimoto T. Monodispersed particles. second ed., Elsevier (2001).  
+
+2. Polte J. Fundamental growth principles of colloidal metal nanoparticles — a new perspective. CrystEngComm 17, 6809-6830 (2015).  
+
+3. De Yoreo JJ, et al. Crystallization by particle attachment in synthetic, biogenic, and geologic environments. Science 349, aaa6760 (2015).  
+
+4. Li D, Nielsen MH, Lee JRI, Frandsen C, Banfield JF, De Yoreo JJ. Direction-Specific interactions control crystal growth by oriented attachment. Science 336, 1014-1018 (2012).  
+
+5. Ye H, Zhang Z, Wang R. Nucleation and growth of nanocrystals investigated by in situ transmission electron microscopy. Small 19, 2303872 (2023).  
+
+6. Song M, et al. Oriented attachment induces fivefold twins by forming and decomposing high-energy grain boundaries. Science 367, 40 (2020).  
+
+7. Zhou L, O'Brien P. Mesocrystals — properties and applications. The Journal of Physical Chemistry Letters 3, 620-628 (2012).  
+
+8. Zhu G, et al. Self-similar mesocrystals form via interface-driven nucleation and assembly. Nature 590, 416-422 (2021).  
+
+9. Park BC, et al. Surface-ligand-induced crystallographic disorder-order transition in oriented attachment for the tuneable assembly of mesocrystals. Nature Communications 13, 1144 (2022).  
+
+10. Li J, et al. In situ atomic-scale study of particle-mediated nucleation and growth in amorphous bismuth to nanocrystal phase transformation. Advanced Science 5, 1700992 (2018).  
+
+11. Thanh NTK, Maclean N, Mahiddine S. Mechanisms of nucleation and growth of nanoparticles in
+
+<--- Page Split --->
+
+solution. Chemical Reviews 114, 7610- 7630 (2014).  
+
+12. Baumgartner J, et al. Nucleation and growth of magnetite from solution. Nature Materials 12, 310-314 (2013).  
+
+13. Huang B, et al. Identification of plasmon-driven nanoparticle-coalescence-dominated growth of gold nanoplates through nanopore sensing. Nature Communications 13, 1402 (2022).  
+
+14. Rose G. Ueber die krystallformen des goldes und des silbers. Annalen der Physik 99, 196-204 (1831).  
+
+15. Hofmeister H. Fivefold twinned nanoparticles. Encyclopedia of Nanoscience and Nanotechnology 3, 431-452 (2004).  
+
+16. Zhou S, Zhao M, Yang TH, Xia Y. Decahedral nanocrystals of noble metals: Synthesis, characterization, and applications. Materials Today 22, 108-131 (2019).  
+
+17. Wu JY, Nagao S, He JY, Zhang ZL. Role of five-fold twin boundary on the enhanced mechanical properties of fcc Fe Nanowires. Nano Letters 11, 5264-5273 (2011).  
+
+18. Wang X, et al. Pd@Pt Core-Shell Concave Decahedra: A class of catalysts for the oxygen reduction reaction with enhanced activity and durability. Journal of the American Chemical Society 137, 15036-15042 (2015).  
+
+19. Pietrobon B, Kitaev V. Photochemical synthesis of monodisperse size-controlled silver decahedral nanoparticles and their remarkable optical Properties. Chemistry of Materials 20, 5186-5190 (2008).  
+
+20. Maksimuk S, Teng X, Yang H. Roles of twin defects in the formation of platinum multipod nanocrystals. The Journal of Physical Chemistry C 111, 14312-14319 (2007).  
+
+21. Clasen P, Tiruvalam R, Ma S, Harmer MP, Kiely C. Effect of synthesis parameters on the gold nanostructures produced using an oleamine reductant. Microscopy and Microanalysis 13, 776-777 (2007).  
+
+22. Lakes R. Materials with structural hierarchy. Nature 361, 511-515 (1993).  
+
+23. Han H, et al. Multiscale hierarchical structures from a nanocluster mesophase. Nature Materials 21, 518-525 (2022).  
+
+24. Barnard AS, Young NP, Kirkland AI, van Huis MA, Xu H. Nanogold: A quantitative phase map. ACS Nano 3, 1431-1436 (2009).  
+
+25. Young NP, van Huis MA, Zandbergen HW, Xu H, Kirkland AI. Transformations of gold nanoparticles investigated using variable temperature high-resolution transmission electron microscopy. Ultramicroscopy 110, 506-516 (2010).
+
+<--- Page Split --->
+
+26. Koga K, Ikeshoji T, Sugawara K. Size- and temperature-dependent structural transitions in gold nanoparticles. Physical Review Letters 92, 115507 (2004).  
+
+27. Marks LD. Surface structure and energetics of multiply twinned particles. Philosophical Magazine A 49, 81-93 (1984).  
+
+28. Cleveland CL, Landman U, Schaaff TG, Shafigullin MN, Stephens PW, Whetten RL. Structural evolution of smaller gold nanocrystals: the truncated decahedral motif. Physical Review Letters 79, 1873-1876 (1997).  
+
+29. Johnson CL, et al. Effects of elastic anisotropy on strain distributions in decahedral gold nanoparticles. Nature Materials 7, 120 (2007).  
+
+30. Gryaznov VG, Heydenreich J, Kaprelov AM, Nepijko SA, Romanov AE, Urban J. Pentagonal symmetry and disclinations in small particles. Crystal Research and Technology 34, 1091-1119 (1999).  
+
+31. Song M, Wu Z, Lu N, Li D. Strain relaxation-induced twin interface migration and morphology evolution of silver nanoparticles. Chemistry of Materials 31, 842-850 (2019).  
+
+32. Patala S, Marks LD, Olvera de la Cruz M. Elastic strain energy effects in faceted decahedral nanoparticles. The Journal of Physical Chemistry C 117, 1485-1494 (2013).  
+
+33. Song M, et al. Uneven strain distribution induces consecutive dislocation slip, plane gliding, and subsequent detwinning of penta-twinned nanoparticles. Nano Letters 24, 1153-1159 (2024).  
+
+34. Zhu YT, Wu XL, Liao XZ, Narayan J, Kecskes LJ, Mathaudhu SN. Dislocation-twin interactions in nanocrystalline fcc metals. Acta Materialia 59, 812-821 (2011).  
+
+35. Lee J, Nakouzi E, Song M, Wang B, Chun J, Li D. Mechanistic understanding of the growth kinetics and dynamics of nanoparticle superlattices by coupling interparticle forces from real-time measurements. ACS nano 12, 12778-12787 (2018).  
+
+36. Stukowski A. Visualization and analysis of atomistic simulation data with OVITO-the open visualization tool. Modelling and Simulation in Materials Science and Engineering 18, 015012 (2010).  
+
+37. Thompson AP, et al. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Computer Physics Communications 271, 108171 (2022).  
+
+38. Ackland GJ, Tichy G, Vitek V, Finnis MW. Simple N-body potentials for the noble metals and nickel. Philosophical Magazine A 56, 735-756 (1987).  
+
+39. Nose S. A unified formulation of the constant temperature molecular dynamics methods. The Journal of Chemical Physics 81, 511-519 (1984).
+
+<--- Page Split --->
+
+40. Maeda H, Maeda Y. Orientation-dependent london-van der Waals interaction energy between macroscopic bodies. Langmuir 31, 7251-7263 (2015).  
+
+41. Pinchuk P, Jiang K. Size-dependent Hamaker constants for silver and gold nanoparticles. In: Physical Chemistry of Interfaces and Nanomaterials XIV). SPIE (2015).  
+
+## Acknowledgements  
+
+This project is supported by the National Natural Science Foundation of China (No. 52101025) and Excellent Young Scientists Fund Program (Overseas), Natural Science Foundation of Hunan Province (No. 2023JJ30684), the Changsha Municipal Natural Science Foundation (kq2202091), and State Key Laboratory of Crystal Materials, Shandong University (No. KF2206). We would also like to acknowledge the support from State Key Laboratory of Powder Metallurgy, Central South University, Changsha, China.  
+
+## Author contributions  
+
+M.S. designed the experiments, conducted experiments and data analysis, wrote the manuscript, and supervised the study. D. Z. assisted with image processing and data analysis. D. Leng revised the manuscript. J. L. synthesized Au NPs and conducted force calculation. Z. Y., J. C., D. L., Z. G., and L. W. contributed to the discussion of the results and commented on the manuscript. G. Z. conducted MD simulations and wrote the manuscript. R. Y. and K. Z. supervised the study and commented on the manuscript.  
+
+## Competing interests  
+
+The authors declare no competing interests.  
+
+## Additional information  
+
+Supplementary information The online version contains supplementary material available at XXXX.  
+
+Correspondence and requests for materials should be addressed to Miao Song.
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SupplementaryInformation.docx
+
+<--- Page Split --->

preprint/preprint__2af794e8bdc6d7671634e08f6af1da53069c1fb5bf84792aa87006addcb20726/preprint__2af794e8bdc6d7671634e08f6af1da53069c1fb5bf84792aa87006addcb20726_det.mmd

@@ -0,0 +1,505 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 914, 177]]<|/det|>
+# In-situ atomic observations unveil the aggregation growth and evolution of five-fold twin structures  
+
+<|ref|>text<|/ref|><|det|>[[44, 197, 268, 240]]<|/det|>
+Miao Song songmiao@csu.edu.cn  
+
+<|ref|>text<|/ref|><|det|>[[50, 269, 625, 288]]<|/det|>
+Central South University https://orcid.org/0000- 0002- 0483- 6580  
+
+<|ref|>text<|/ref|><|det|>[[44, 293, 268, 334]]<|/det|>
+Dingri Zhang Central South University  
+
+<|ref|>text<|/ref|><|det|>[[44, 341, 268, 381]]<|/det|>
+Dan Leng Central South University  
+
+<|ref|>text<|/ref|><|det|>[[44, 387, 225, 428]]<|/det|>
+Jaewon Lee j.lee@missouri.edu  
+
+<|ref|>text<|/ref|><|det|>[[44, 433, 268, 474]]<|/det|>
+Ziang Yang Central South University  
+
+<|ref|>text<|/ref|><|det|>[[44, 480, 268, 520]]<|/det|>
+Jiaxuan Chen Central South University  
+
+<|ref|>text<|/ref|><|det|>[[44, 526, 268, 566]]<|/det|>
+Dan Li Central South University  
+
+<|ref|>text<|/ref|><|det|>[[44, 572, 238, 612]]<|/det|>
+Lei Wang Shandong University  
+
+<|ref|>text<|/ref|><|det|>[[44, 618, 565, 659]]<|/det|>
+Gang Zhou Institute of Metal Research, Chinese Academy of Sciences  
+
+<|ref|>text<|/ref|><|det|>[[44, 665, 937, 707]]<|/det|>
+Rui Yang Institution of Metal Research, Chinese Academy of Sciences https://orcid.org/0000- 0003- 4405- 5882  
+
+<|ref|>text<|/ref|><|det|>[[44, 711, 268, 752]]<|/det|>
+Kechao Zhou Central South University  
+
+<|ref|>text<|/ref|><|det|>[[44, 792, 104, 810]]<|/det|>
+Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 830, 137, 848]]<|/det|>
+Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[44, 868, 300, 887]]<|/det|>
+Posted Date: April 23rd, 2024  
+
+<|ref|>text<|/ref|><|det|>[[44, 906, 474, 924]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 4283157/v1
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 44, 912, 87]]<|/det|>
+License: © ① This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 105, 535, 125]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 161, 941, 205]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on October 25th, 2024. See the published version at https://doi.org/10.1038/s41467-024-53501-0.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[160, 100, 845, 144]]<|/det|>
+# In-situ atomic observations unveil the aggregation growth and evolution of five-fold twin structures  
+
+<|ref|>text<|/ref|><|det|>[[148, 157, 850, 194]]<|/det|>
+Miao Song \(^{1,2,*}\) , Dingri Zhang \(^{1}\) , Dan Leng \(^{1}\) , Jaewon Lee \(^{3}\) , Ziang Yang \(^{1}\) , Jiaxuan Chen \(^{1}\) , Dan Li \(^{1}\) , Lei Wang \(^{2}\) , Gang Zhou \(^{4,*}\) , Rui Yang \(^{4}\) , Kechao Zhou \(^{1}\)  
+
+<|ref|>text<|/ref|><|det|>[[148, 208, 850, 285]]<|/det|>
+\(^{1}\) State Key Laboratory of Powder Metallurgy, Central South University, Changsha, Hunan 410083, China \(^{2}\) State Key Laboratory of Crystal Materials, Shandong University, Jinan, Shandong, 250100, China \(^{3}\) Department of Mechanical and Aerospace Engineering, College of Engineering, University of Missouri, Columbia, MO 65203, USA \(^{4}\) Shi-changxu Innovation Center for Advanced Materials, Institute of Metal Research, Chinese Academy of Sciences, Shenyang, Liaoning 110016, China  
+
+<|ref|>text<|/ref|><|det|>[[150, 293, 595, 308]]<|/det|>
+\*Corresponding author. Email: songmiao@csu.edu.cn; gzhou@imr.ac.cn  
+
+<|ref|>text<|/ref|><|det|>[[147, 349, 852, 701]]<|/det|>
+The unique twin boundaries and inherent lattice strain of five- fold twin (5- FT) structures offer a promising and innovative approach to tune nanocrystal configurations and properties, enriching nanomaterial performance. However, due to constraints imposed by small thermodynamically stable size and complex twin configurations, gaps persist in understanding the nonclassical growth models of 5- FT nanoparticles. Here, we in- situ investigated the mechanisms underlying size- dependent and twin configuration- related aggregation growth phenomena between 5- FT and other nanoparticles at the atomic scale. The results find that surface diffusion shapes the morphology of aggregated nanoparticles, promoting symmetrical 5- FT formation, particularly involving smaller nanoparticles. Additionally, the inherent structure of 5- FT mitigates the dominance of surface diffusion in its morphological evolution, retarding the aggregation evolution process and fostering intricate twin structures. Our findings contribute to advancing our ability to manipulate the configuration of twinned particles and achieve a more predictable synthesis of novel functional nanomaterials for engineering applications.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 90, 853, 470]]<|/det|>
+Uncovering the growth of nanoparticles (NPs) is significant not only for elucidating the underlying formation mechanisms of minerals and understanding the comprehensively evolution process of natural environment, but also holds scientific and engineering significance to tune size, morphology, and properties of synthetic nanomaterials. Despite numerous evidences challenge the classical interpretations of crystal growth, which emphasize surface reactions and the monomer diffusion to the surface \(^{1,2}\) , nonclassical crystal growth scenarios have only begun to be recognized gradually just over a decade ago \(^{2,3,4,5}\) . As a main nonclassical crystal growth mechanism, the particle- based aggregation \(^{3}\) , generally including orientated attachment (OA) \(^{4,6}\) , nearly OA (such as mesocrystal \(^{7,8,9}\) and dislocation induced tilt attachment \(^{4}\) ), and non- OA (such as intraparticle growth \(^{10,11}\) , aggregation and transformation of thermodynamically metastable particle to stable phases \(^{8,12}\) , and aggregation and grain boundary/surface atom migration dominated growth \(^{10,13}\) ), requires systematic and intensive investigation. These are determined by unknown fundamental and challengeable aspects, such as high spatial and temporal resolution of atomic diffusion and migration, particle relative movement, interfacial interactions, and grain boundary evolution.  
+
+<|ref|>text<|/ref|><|det|>[[147, 479, 853, 914]]<|/det|>
+The 5- FT structure has been extensively detected in diverse natural and synthetic systems \(^{14,15}\) , \(^{16}\) and exhibits unique properties attributed to its crystallographically forbidden pentagonal symmetry and inherent lattice strain, such as enhanced mechanical properties \(^{17}\) , attractive catalytic properties \(^{18}\) , and excellent optic properties \(^{19}\) . In addition to NP size, morphology, and composition, the twin plane and inherent lattice strain of 5- FT provide an appealing and novel avenue to tailor the configuration and properties of nanocrystals, thus diversifying and enhancing the performance of nanomaterials. For instance, 5- FT structure holds potential for producing hierarchical materials \(^{20,21,22}\) , which preserve the properties of the nanoscale building blocks and may exhibit novel performance characteristics \(^{22,23}\) . Nevertheless, limited by complicated twin structures (5 twin units), significantly small thermodynamic stability size (3- 14 nm) \(^{24,25,26}\) , together with movement of NPs and the only proper [110] twin pole observation direction, revealing the atomic formation and growth mechanisms of 5- FT remains seriously experimental challenging. By trigging NPs aggregation through electron- beam- induced decomposition of the organic ligands surrounding the NPs, we uncovered two underlying atomic formation mechanisms of 5- FT within Au, Pd, and Pt nanomaterials \(^{6}\) . However, nonclassical growth mechanisms of 5- FT are still mysterious and elusive, particularly potentially under the coupling effect of thermodynamic and kinetics landscapes,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 90, 852, 190]]<|/det|>
+involving but not limited to the atomic formation and evolution of complicated twin structure (twining and de- twinning processes), atom surface diffusion, and relative slip and configuration modulation of NPs. Thereby, clarifying the particle- based aggregation growth of 5- FT is scientifically and engineeringly significant.  
+
+<|ref|>text<|/ref|><|det|>[[147, 200, 852, 357]]<|/det|>
+In this work, in- situ high- resolution transmission electron microscopy (HRTEM) combined with molecular dynamics (MD) simulation was utilized to study the atomic aggregation growth and evolution mechanisms involving 5- FT and diverse NPs, i.e., NPs with varied size ratios and twinned configurations (single crystal- SC and 5- FT). Additionally, the impact of various thermodynamic and kinetic landscapes on the aggregation evolution was systematically investigated and comprehensively elucidated.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 377, 344, 394]]<|/det|>
+## Results and discussion  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 404, 752, 421]]<|/det|>
+## In-situ HRTEM observation of aggregation evolution between 5-FT and SC NPs  
+
+<|ref|>text<|/ref|><|det|>[[147, 432, 853, 784]]<|/det|>
+Spherical Au NPs with various twin configurations ( \(\sim 4.4 \mathrm{nm}\) , Supplementary Fig. 1) embedded in 1- dodecanethiol organic matrix were drop- casted onto a TEM grid. Then, NP aggregation was induced by decomposing organics under electron- beam irradiation (see “Method” for details). As shown in Fig. 1a- b, aggregation between a 2.1 nm SC and a 3.6 nm symmetrical 5- FT (S5- FT, size ratio \(R = 0.58\) ) induces obvious growth of the twin units 1 and 5, resulting in the formation of an asymmetrical 5- FT (AS5- FT). Nevertheless, the AS5- FT evolves into a symmetrical twin structure (Fig. 1c), and transforms into a stable re- entrant {111} (highlighted by white arrows in Fig. 1d) Marks S5- FT at last. Considering none visible twin boundaries (TBs) migrations, the above evolution process is supposed to be dominated by surface diffusion from the small NP to the 5- FT, which will be further discussed in Fig. 2. Wherein, the formation of highly faceted structure (Fig. 1d) has been proved to be associated with the balance between surface and strain energy \(^{27,28}\) , i.e., the re- entrant {111} Marks 5- FT can significantly release the intrinsic lattice strain energy and slightly increases the surface energy of the 5- FT structure.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[152, 100, 845, 610]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[148, 609, 852, 848]]<|/det|>
+<center>Fig. 1. In-situ observation of aggregating growth between 5-FT NPs with SC NPs across diverse size ratios. (a-d) Configuration evolution after a \(2.1\mathrm{nm}\) SC NP attached with a 5-FT NP (3.6 nm) with \(R = 0.58\) . White arrows show formation of the re-entrant surface. (e-o) Configuration evolution after a \(4.6\mathrm{nm}\) SC NP aggregated with a \(4.2\mathrm{nm}\) 5-FT NP \((R = 1.10)\) . (p) The surface outlines of the aggregated NP. The migration directions of the surface outlines with durations are denoted by red arrows. Notable, twin boundaries are highlighted by red lines. Twin units are denoted by numbers ranging from 1 to 5 in (a) and (e). Yellow arrows show migration directions of the TBs, compared with the immediate prior image. Partial dislocations are denoted by blue “L” and their slip directions are denoted by the blue arrows. </center>  
+
+<|ref|>text<|/ref|><|det|>[[148, 858, 850, 904]]<|/det|>
+Different from the above small size ratios (Fig. 1a- b), the large SC \((R = 1.10)\) only stimulates continuous growth of the twin unit 1 with substantially de- twinning process until the final
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[146, 80, 855, 723]]<|/det|>
+manifestation of a new SC (Fig. 1e- o). And the detailed evolution processes are as follows: Initially, a \(4.6\mathrm{nmSC - 1}\) attaches with a \(2.1\mathrm{nmSC - 2}\) and a \(4.2\mathrm{nmS5 - FT}\) successively, resulting in attractive growth of the twin unit 1 (Fig. 1e- g). Then, a partial dislocation nucleates at the TB \(\Sigma 3_{1}\) from the periphery of the NP and slips toward the center of the 5- FT, leading to a single layer of migration of the \(\Sigma 3_{1}\) near the periphery (Fig. 1h). Subsequently, continuous nucleation and slip of partial dislocations at TBs induces de- twinning process of the 5- FT (Fig. 1i- o), giving rise to reduction of the twin units 2, 3, and 4. During this process, when the twin pole of the 5- FT is close to overlap (Fig. 1j), specifically, it involves one layer of migrations of the four TBs \((\Sigma 3_{1},\Sigma 3_{2},\Sigma 3_{3},\) and \(\Sigma 3_{4})\) . Alternatively, there is obvious split of the twin pole. At last, when the 5- FT twin pole is close to \(\sim 2 - 3\) layers of {111} planes away from) the periphery of the NP, 5- FT structure disappears within 0.8 s, resulting in the formation of the new SC NP (Fig. 1n- o). To access the influence of surface diffusion on the entire aggregation process, further analysis is conducted on the variations in surface outlines. As presented in Fig. 1g- p, distinct variations of the surface outlines are detected from 83.2 s to 179.4 s, especially at the concave surfaces and the region near the 5- FT twin pole (Fig. 1p), nevertheless, the variation can be ignored from 179.4 s to 220.8 s (Fig. 1p). Therefore, during the above aggregation induced evolution process, the surface diffusion mainly occurs at the initial stage (Fig. 1g- k), and the partial dislocations induced twin boundaries slip mainly happens at the subsequent stage (Fig. 1i- o) with an overlapping duration from 142.6 s to 179.4 s (Fig. 1i- k). Notably, surface diffusion without significant morphology variations, such as mutual diffusion, remains undetectable in Fig. 1. The relative slip between the initial 5- FT and SC NPs is also indistinguishable, which can potentially induce morphological variations. These phenomena will be comprehensively analyzed in Fig. 2 based on high time resolution data and 3- D atomic evolution tracing using MD simulations.  
+
+<|ref|>text<|/ref|><|det|>[[147, 729, 853, 914]]<|/det|>
+Therefore, as presented in Fig. 1, after attachment between a 5- FT NP with a small \((R = 0.58\) Fig. 1a- d) or a relatively large \((R = 0.73\) , Supplementary Fig. 6) SC NP, surface diffusion dominates the aggregation growth processes, resulting in formation of a S5- FT or an asymmetrical 5- FT (AS5- FT) NP, respectively. Nevertheless, during the aggregation growth process of a 5- FT with a large \((R = 1.10\) , Fig. 1e- o) SC NP, the surface diffusion exclusively dominates the initial stage of surface modulation. Subsequent to this, the activation of partial dislocations governs the de- twinning process, resulting in the manifestation of a new SC NP. Notably, based on our previous observation
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 88, 853, 273]]<|/det|>
+6, asymmetrical 5- FTs can also evolve into symmetrical ones through NPs aggregation at small twin units. Meanwhile, with the growing of 5- FT NPs, its inherent lattice strain \(^{29}\) can be relieved through various mechanisms, including but not limited to the formation of plane defects in twin units \(^{30}\) , the migration of the 5- FT twin pole \(^{30,31}\) , and the formation of re- entrant Marks decahedral morphology \(^{32}\) . Thereby, the aforementioned AS5- FT, including various unstable intermediate 5- FT NPs, can be potentially stabilized via attachment with other NPs in conjunction with the modulation of the inherent lattice strain.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 293, 828, 310]]<|/det|>
+## Aggregation growth mechanisms between 5-FT and SC NPs elucidated via MD simulation  
+
+<|ref|>text<|/ref|><|det|>[[147, 317, 853, 896]]<|/det|>
+To uncover detailed evolution mechanisms of NPs' 3D configurations, the aggregation processes between S5- FT and SC NPs are further investigated by employing high temporal resolution MD simulation (see “Method”, Supplementary Fig. 2 and Table 1). As shown in Fig. 2a- c, the SC NP (B3s, \(R = 0.52\) , Fig. 2a) was initially designed to attach with the twin unit 1. However, the SC NP evolves into amorphous state and its atoms diffuse symmetrically towards the S5- FT surface quickly (Fig. 2b and also Supplementary Fig. 7), inducing significantly growth of the twin units 1 and 2, resembling the phenomena detected in Fig. 1a- b. The mutual diffusion of surface several layers' atoms, are also detected after 4.0 ns with formation of re- entrant {111} surfaces (Fig. 2c- f). Notably, significantly fewer atoms diffuse from the 5- FT to the B3s SC NP, and the atoms from the initial NP can even diffuse to the far- away twin unit 4 of the 5- FT (Fig. 2f). Correspondingly, the system experiences a substantial decrease in relative energy primarily at the initial stage (0.0- 4.0 ns, Fig. 2g), indicating significant configuration modifications. Therefore, the surface diffusion from small SC to the big 5- FT NP dominates the aggregation growth processes, consistent with the experimental results detected in Fig. 1a- d. Similar mutual diffusion is also detected at the initial stage of the simulated system with a relatively large particle size ratio (0.76, Supplementary Fig. 8). Nevertheless, the 5- FT structure swiftly transforms into a bi- crystal within 0.1 ns (Supplementary Fig. 8d- e), accompanying with significant decreasing of the relative energy (Supplementary Fig. 8g). Additionally, there is also diffusion induced growth of both twin units, i.e., the twin units 1 and 5 (Supplementary Fig. 8d). This verifies that the formation of initial asymmetrical configuration, i.e., significant growth of both twin units (Fig. 1a- b and Supplementary Figs. 6 and 8), is an inherent characteristic of surface diffusion dominated aggregation processes when \(R < - 0.76\) .
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[156, 95, 839, 650]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[147, 664, 852, 905]]<|/det|>
+<center>Fig. 2. MD simulation of the aggregation growth of 5-FT NPs by attaching with various SC NPs. (a-e) Evolution of an aggregated 5-FT NP, formed by attaching of a \(5.0 \mathrm{nm}\) symmetrical 5-FT NP (yellow, \(\mathrm{A}_{5\mathrm{F}}\) ) and a \(2.6 \mathrm{nm}\) SC NP (cyan, \(\mathrm{B}_{35}\) ) with \(R = 0.52\) . Re-entrant {111} surfaces are highlighted by gray solid lines. (f) The biggest cross-section perpendicular to the Z axis showing atomistic distribution of the initial SC NP. (g) Variation of the relative energy of the simulated system from (a) to (e). (h-r) The aggregation evolution process after a \(5 \mathrm{nm}\) 5-FT Au NP ( \(\mathrm{A}_{5\mathrm{F}}\) ) attached with a \(4.8 \mathrm{nm}\) SC NP ( \(\mathrm{B}_{35}\) ) with \(R = 0.96\) . The relative slip between the 5-FT and the SC is represented by cyan arrows in (i-k) and (m). (s) Variation of the relative energy of the simulated system from (h) to (r). (t) TBs slip induced migrations of the twin pole from \(2.5 \mathrm{ns}\) (i) to \(4.8 \mathrm{ns}\) (o). </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 90, 852, 330]]<|/det|>
+(u) A double Thompson tetrahedron with an FCC crystal structure. Vertex-to-vertex (e.g., \(AB\) ), vertex-to-orthocenter (e.g., \(A\gamma\) ), and orthocenter-to-orthocenter (e.g., \(\alpha \gamma\) ) denote a perfect, partial, and stair-rod dislocations, respectively. (v) Relationship of the Thompson tetrahedron in 5-FT units verifying that dislocations within different twin units can undergo mutual transformations via various dislocation reactions. (w) Schematic illustration of twin pole migration (o to o'). TBs before and after slip are denoted by dashed yellow lines and solid red lines, respectively. Migration directions and layers of the TBs are denoted by yellow arrows and numbers, respectively. Partial dislocations are denoted by blue "L" and their slip directions are denoted by the blue arrows. Twin units are denoted by numbers ranging from 1 to 5 in (a) and (h).  
+
+<|ref|>text<|/ref|><|det|>[[147, 339, 852, 916]]<|/det|>
+Divergent from the aforementioned surface diffusion dominated NPs aggregation evolution mechanisms ( \(R< \sim 0.76\) , Fig. 2a-f and Supplementary Fig. 8), MD simulation indicates that the detwinning process induced by TBs' migrations governs the aggregation evolution as the \(R\) increases to 0.96 (Fig. 2h-r). This aligns with the experimental results delineated in Fig. 1i-o. And the aggregation evolution process can be segmented into the following three distinct stages: (1) The initial relative slip. As presented in Fig. 2h-i, the interaction between the 5-FT and the SC gives rise to the relative slip of 2 layers of {111}, and the slip direction of the 5-FT aligns parallel to the \(\Sigma 3_{5}\) (highlighted by the cyan arrow). (2) TB migration dominated de-twinning process. This process is trigged by nucleation and slip of partial dislocations (Fig. 2j-o). As shown in Fig. 2j, the nucleation and slip of a partial dislocation from the periphery of the concave surface to the twin pole instigate the migration of the TB \(\Sigma 3_{5}\) towards the initial SC. Notably, there are also relative slip between the 5-FT and the SC (Fig. 2j, k, and m). These determinations arise from the layer variation of {200} planes between the twin pole and the periphery of the initial SC (prior to the migration of the twin pole, Fig. 2h-i), alternatively, the layer variations of {111} planes from the \(\Sigma 3_{4}\) (without migration) to the {111} intersection plane in the twin unit 5 with the initial {200} interface between the 5-FT and the SC (Fig. 2j, k and m). (3) Surface diffusion dominated de-twinning process. When the twin units are as small as 3 layers of {111} planes (Fig. 2o-r), surface diffusion takes precedence as the predominant mechanism during the final de-twinning process, while no discernible TB migration can be detected. This differs from the detected TB migration dominated de-twinning process in Fig. 1n-o, and is supposed to be associated with NP's configuration. That is, when the twin pole of the AS5-FT is close to the convex surface (Fig. 1n-o and Supplementary Fig. 10f-g),
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 89, 853, 357]]<|/det|>
+TB migration induces the fast de- twinning process. Nevertheless, when the serious AS5- FT twin structures are close to the concave surface (Fig. 2o- r), surface diffusion dominates the de- twinning process. Fig. 2s shows there is a notable increase in the system energy during the initial stage of the relative slip (Fig. 2h- k), followed by a substantial energy decrease attributed to the de- twinning process dominated by partial dislocations. This is mainly associated with reduction of TB energy and lattice strain energy \(^{31,33}\) . After that, the variation in the system energy becomes moderate. Notably, mutual diffusion is consistently detected throughout the entire evolution process. The amorphous detected in MD simulations (Fig. 2b- c and Supplementary Fig. 8), as opposed to experimental observation (Fig. 1a- c), is supposed to correlate with the relatively elevated simulation temperature employed to expedite the simulated experimental process (see "Method").  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 378, 341, 394]]<|/det|>
+## De-twinning mechanisms  
+
+<|ref|>text<|/ref|><|det|>[[147, 404, 853, 896]]<|/det|>
+Effective TBs migration layers during the TBs migrations dominated de- twinning process (Fig. 2i- o) are illustrated in Fig. 2t. That is, TBs \(\Sigma 3_{1}\) and \(\Sigma 3_{2}\) migrate 4 layers, TBs \(\Sigma 3_{3}\) and \(\Sigma 3_{5}\) migrate 5 layers, remaining \(\Sigma 3_{4}\) are unchanged. The migration of TBs is closely associated with nucleation and slip of partial dislocations, and the detailed mechanisms are depicted in Fig. 2u- w. A double Thompson tetrahedron is generally employed to define various dislocations in face- centered cubic crystal \(^{34}\) , including perfect dislocation (vertex to vertex, such as \(AB\) and \(BD\) ), partial dislocation (vertex to orthocenter, such as \(A\delta\) and \(B\delta\) ), and stair- rod dislocation (orthocenter to orthocenter, such as \(\alpha \beta\) and \(\beta \delta\) ). Dislocations have the capability to undergo transformation into one another through reactions, guided by the principles of vector geometry addition and subtraction rules \(^{34}\) . For instance, \(DB = DA + AB\) , \(\beta \delta = 1 / 3DB\) , \(DB = D\beta +\beta \delta +\delta B = 1 / 3DB + D\beta +\delta B\) , \(DB = 3 / 2(D\beta +\delta B)\) . Here, the 5- FT NP can be regarded as a five Thompson tetrahedron (Fig. 2v), and dislocations can transform into one another by diverse dislocation reactions. Based on experimental observation (Fig. 1h) and MD simulation (Fig. 2j), partial dislocations initially nucleate at the periphery of the NPs. For instance, a partial dislocation \(C\beta '\) nucleates on \(\Sigma 3_{5}\) (Fig. 2w) and slips towards the twin pole \(O\) , inducing one layer migration of \(\Sigma 3_{5}\) . Then, \(C\beta '\) dissociates into other two partial dislocations at the twin pole, i.e., \(C\beta ' \rightarrow C\eta +\eta \beta '\) , and \(C\eta\) can slip along \(\Sigma 3_{3}\) , resulting in one layer of migration of \(\Sigma 3_{3}\) . Meanwhile, \(\eta \beta '\) can be equivalent to \(1 / 2DB\) (\(\eta \beta '\) serves as the median of the triangle FBD, Fig. 2v). And as analyzed above, \(DB = 3 / 2(D\beta +\delta B)\) , thereby, \(\eta \beta '\) can be dissociated into \(D\beta\) and \(\delta B\) , i.e.,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 88, 854, 358]]<|/det|>
+\(\eta \beta^{\prime} = 1 / 2DB = 3 / 4(D\beta +\delta B)\) . Correspondingly, partial dislocations \(D\beta\) and \(\delta B\) can trigger migration of TBs \(\Sigma 3_{1}\) and \(\Sigma 3_{2}\) , resulting in migration of the twin pole form \(O\) to \(O^{\prime}\) (Fig. 2w), i.e., one layer of de- twinning process of 5- FT. Based on the above dislocation reactions, one \(C\beta^{\prime}\) only generates \(3 / 4(D\beta +\delta B)\) , that is, a single layer of migration of \(\Sigma 3_{5}\) results in \(3 / 4\) layer of migrations of \(\Sigma 3_{1}\) and \(\Sigma 3_{2}\) . Nevertheless, based on the geometrical relation of 5- FT, one \(C\beta^{\prime}\) still generates \(\sim 0.18\) more \(D\beta\) and \(\delta B\) than that to maintain the perfect overlap of the twin pole. This residual mismatch can be mediated by various migration layers of TBs, such as 5 layers of \(\Sigma 3_{3}\) and \(\Sigma 3_{5}\) and 4 layers of \(\Sigma 3_{1}\) and \(\Sigma 3_{2}\) ( \(5\times 0.18 = 0.90\) layer, closed to a single layer of difference, Fig. 2i- o and t). This observation is also consistent with the experimental observation (Fig. 1h- j), i.e., the TBs \(\Sigma 3_{1}\) , \(\Sigma 3_{2}\) , \(\Sigma 3_{3}\) , and \(\Sigma 3_{4}\) migrate one layer to maintain the nearly overlapped twin pole.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 377, 692, 395]]<|/det|>
+## In-situ HRTEM observation of aggregation evolution between 5-FT NPs  
+
+<|ref|>text<|/ref|><|det|>[[147, 401, 853, 896]]<|/det|>
+Aggregation induced growth and TB evolution process are also investigated between 5- FT NPs with various size ratios (Fig. 3). As presented in Fig. 3a- b, aggregation between a 3.7 nm S5- FT and a 4.6 nm S5- FT induces the formation of an AS5- FT (Fig. 3d). And the AS5- FT can potentially evolve into S5- FT by aggregating with other NPs (denoted by a white arrow, Fig. 3d), as discussed above. With increasing of the size ratio between the two aggregated 5- FT NPs, intricate twin structures can be induced ( \(R = 0.82\) , Fig. 3e- l), and the detailed evolution processes are as follows. At first, the twin unit 4 of the small S5- FT NP orientationally attaches with the twin unit 2 of the big S5- FT NP (Fig. 3e- f), followed by the formation of a TBs sealed region. This is introduced by the relative slip between the two NPs (layers variation without corresponding TBs migration, highlighted by cyan arrows in Fig. 3f- g), the surface diffusion towards the concave surface (denoted by red arrows, Fig. 3f- h), and the TBs migration (highlighted by yellow arrows, Fig. 3f- i). Then, the sealed region maintains unchanged until the small 5- FT evolves into three- fold twin at 72.0 s (Fig. 3i- k). This process is dominated by surface diffusion, verified by significant morphology variation and two reserved TBs, i.e., only disappearing of the twin units 1 and 2 of the S5- FT NP. Thereby, the de- twinning process (Fig. 3j- k) is significantly different from the TB migration decided de- twinning process (Fig. 3n- o), in which the morphology variation can be ignored. Thereby, the TBs sealed region can stabilize the 5- FT structures and modify the evolution pathway of the aggregation growth mechanism between 5- FT NPs. Nevertheless, the sealed region becomes unstable after 5-
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 88, 852, 274]]<|/det|>
+FT evolving into 3- fold twin within 1.0 s (Fig. 3k- l), resulting in the formation of a complicated twin structure, which can potentially evolve into AS5- FT by de- twinning of the TB at bottom left. Notably, when the size ratios between two aggregated 5- FT is small, such as 0.47 and 0.53 (Supplementary Fig. 9), S5- FT NPs ultimately formed. Additionally, beside the orientation attachment dominated NPs aggregation (with no detectable formation of new GBs at attachment sites, Fig. 3e- f), another scenario emerged, where a new zig- zag {200}/{111} GB was induced after aggregation ( \(R = 0.72\) , see Supplementary Fig. 10 for details).  
+
+<|ref|>image<|/ref|><|det|>[[152, 279, 852, 686]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[147, 697, 852, 882]]<|/det|>
+<center>Fig. 3. Aggregating growth between 5-FT NPs. (a-d) Atomic evolution process between a 3.7 nm 5-FT and a \(4.6 \mathrm{nm}\) 5-FT with \(R = 0.78\) . (e-l) Atomic evolution process between a \(3.3 \mathrm{nm}\) 5-FT and a \(4.0 \mathrm{nm}\) 5-FT with a size ratio of 0.82. Twin units are denoted by numbers ranging from 1 to 5. TBs and unclear TBs are highlighted by red solid and dashed lines, respectively. The white arrow shows the aggregation of another NP in (d). Yellow arrows show migration directions of the TBs, compared with the immediate prior image. Partial dislocation is denoted by blue “L” and its slip direction is denoted by the blue arrow. </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[148, 90, 767, 108]]<|/det|>
+## Aggregation growth mechanisms between 5-FT NPs elucidated via MD simulation  
+
+<|ref|>text<|/ref|><|det|>[[147, 118, 853, 499]]<|/det|>
+Fig. 4 shows aggregation growth between 5- FT NPs with various size ratios. As presented in Fig. 4a- b, after orientation attachment with the twin unit 1 of the \(5.0 \mathrm{nm} 5\mathrm{- FT}\) , the small 5- FT NP evolves into amorphous state within 0.1 ns (Fig. 4b). There is no detectable TB migration. Then, mutual diffusion dominates the subsequent evolution process and induces the formation of a symmetrical Marks 5- FT (Fig. 4c- e). Cross- section (Fig. 4f) shows that atoms originating from the initial small 5- FT can even diffuse over considerable distance to the twin units 3 and 4 (Fig. 4f). The system energy variation is presented in Fig. 4g. The above process analogous to that detected in Fig. 1a- g. Thereby, when the size ratio is small enough, such as 0.52 in Fig. 2a- g and 0.62 in Fig. 4a- g, the aggregation kinetics is governed by surface diffusion from small NPs to the big NPs, resulting in the formation of the S5- FT structure. With increasing of the size ratio to 0.82, apparent relative slip (Supplementary Fig. 13a- b) and TBs migration (Supplementary Fig. 13b- c) are detected, corresponding to significant decreasing of the system energy (Supplementary Fig. 13h). Although the relatively small 5- FT NP still evolves into amorphous state accompanying with surface diffusion towards the big 5- FT (Supplementary Fig. 13c- g), an AS5- FT ultimately forms.  
+
+<|ref|>text<|/ref|><|det|>[[147, 507, 853, 914]]<|/det|>
+For the large size ratio between two aggregated 5- FT NPs, MD simulation shows there is also relative slip at the initial stage (Fig. 4h- i) inducing the formation of a sealed stable region S1 (Fig. 4i). The region S1 can stabilize for a long time until the twin units 3 and 4 of the left 5- FT decrease to three and four layers, respectively, resulted by surface diffusion (Fig. 4i- l). Then, TBs migration induces \(\Sigma 3_{2}\) , \(\Sigma 3_{3}\) , \(\Sigma 3_{4}\) , and \(\Sigma 3_{5}\) TBs disappear within 0.02 ns (Fig. 4k- l), accompanying with significant decreasing of the system energy (Fig. 4o). Therefore, MD simulation results are consistent with the phenomena detected during in- situ observation (Fig. 3e- l). As for the reserved TB \(\Sigma 3_{1}\) , it disappears at last with forming of an AS5- FT (Fig. 3l- m). The system energy decreases slowly during the initial 69.50 ns, accompanying with moderate surface modulation, and decreases seriously with de- twinning of the left 5- FT NP. This means that the system energy of the aggregated 5- FT NPs is mainly introduced by the 5- FT structures. Additionally, while the surface diffusion remains a dominated factor during 5- FT NP's aggregation, de- twinning process, i.e., TB migration, is often observed across experimental (Fig. 3) and simulated (Fig. 4) conditions. This phenomenon stems from the intricate and highly strained twin configuration inherent in 5- FT NPs, as opposed to the SC NPs. Notably, due to the formation of the stable sealed region S1, a significant long
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 90, 850, 135]]<|/det|>
+simulation duration was employed in Fig. 4 (80.00 ns) compared with that of the previous MD simulations (30.00 ns, Fig. 2 and Supplementary Figs. 8, 12, and 13).  
+
+<|ref|>image<|/ref|><|det|>[[150, 149, 847, 558]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[147, 572, 852, 784]]<|/det|>
+<center>Fig. 4. MD simulation of the aggregation growth process of S5-FT NPs by attaching with other S5-FT NPs. (a-e) and (h-m) Evolution of the aggregated 5-FT NPs, formed by attaching of a \(5.0 \mathrm{nm}\) S5-FT NP (yellow, \(\mathrm{A_{5F}}\) ) with \(3.2 \mathrm{nm}\) ( \(\mathrm{B_{3F}}\) ) and \(5.0 \mathrm{nm}\) ( \(\mathrm{B_{5F}}\) ) S5-FT NPs (cyan), respectively. The corresponding size ratios are 0.64 and 1.00, respectively. (f) and (n) The biggest cross-sections perpendicular to the Z axes showing atomistic distribution of the initial 5-FT NPs of \(\mathrm{B_{3F}}\) and \(\mathrm{B_{5F}}\) , respectively. (g) and (o) Variation of the relative energies of the corresponding simulated systems. TBs are highlighted by red lines. Twin units are denoted by numbers ranging from 1 to 5. Slip directions and layers of the TBs are denoted by yellow arrows and numbers, respectively. </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 803, 560, 820]]<|/det|>
+## Size effect of aggregation growth between various NPs  
+
+<|ref|>text<|/ref|><|det|>[[148, 831, 851, 904]]<|/det|>
+Symmetry evolution of the aggregated 5-FT NPs is closely associated with surface diffusion and TBs migration. As observed in all experimental results (Figs. 1 and 3), during aggregation between various NPs, the surface diffusion occurs almost throughout the whole evolution processes.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[146, 88, 853, 666]]<|/det|>
+And surface diffusion generally promotes the generation of S5- FT (formed based on the big 5- FT NPs), especially for small particle size ratios, by transferring atoms from small NPs to the twin units far away from the attached regions (Figs. 1a- d, Figs. 2a- f, Figs. 3a- h, Fig. 4a- f). Nevertheless, the effect of TBs migration on the symmetry of the aggregated NPs depends on the twin pole migration direction. The migration of the reserved 5- FT twin pole towards the periphery of the NPs results in a reduction in symmetry, and vice versa. Based on dozens of experimental observations, the size effect of the aggregation growth between 5- FT and various NPs (SC and 5- FT) are illustrated in Fig. 5. Correspondingly, to obtain the S5- FT, the critical ratios should be roughly smaller than 0.83 and 0.72, respectively. According to the theoretical models (Supplementary Table 1), the numbers of atoms of SC is significantly larger than that of the 5- FT with the same particle size. And even \(R = 0.76\) (model 1- 2, Supplementary Table 1, smaller than the critical \(R = 0.83\) ), the number of atoms of SC is also larger than that of the corresponding attached 5- FT. This indicates that: (1) under investigated particles size (several nanometers), S5- FT exhibits relative stability, maintaining its symmetrical structure after attaching with small SC NP, and even the SC possesses a slightly larger number of atoms than that of the 5- FT. That is, surface diffusion predominantly governs the aggregation evolution, rendering negligible the migration of TBs or twin poles. (2) Relatively small critical \(R\) (0.72, atoms significantly less than that of the attached 5- FT, model 2- 2, Supplementary Table 1) between two aggregated 5- FT also verifies the stability of the 5- FT NP, i.e., the 5- FT structure can greatly mitigate surface diffusion. (3) The mutual surface diffusion from small NP to 5- FT is markedly greater than the reverse diffusion, i.e., the surface energy plays a significant role during NP's aggregation growth.  
+
+<|ref|>text<|/ref|><|det|>[[147, 673, 853, 914]]<|/det|>
+To maintain symmetrical configuration of 5- FT NPs, the critical transfer ratio registers at \(\sim\) 0.72 for 5- FT/5- FT, which is smaller than that of SC/5- FT ( \(\sim 0.83\) ). As discussed above, the surface diffusion persists throughout the entirety evolution processes and facilitates to maintain the symmetrical configuration of the big 5- FT NPs. Thereby, for two aggregated 5- FT NPs, the formation of S5- FT is mainly associate with the de- twinning process of the small 5- FT. This can be introduced by TBs migration (Fig. 3c- d and Supplementary Figs. 12- 13) and/or surface diffusion (Fig. 4a- e and Supplementary Figs. 12- 13). Nevertheless, the S5- FT remains relatively stable and maintains its symmetry even when attaching to SC NPs with a slightly larger number of atoms compared to the 5- FT, that is, it is relative difficult to trigger TBs migration of symmetrical 5- FT
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 90, 853, 412]]<|/det|>
+NPs. Additionally, van der Waals (vdW) force calculation (Supplementary Fig. 14) shows that the force between 5- FT and SC NPs is significantly stronger than that between two 5- FT NPs. This is introduced by the structural difference between 5- FT and SC, i.e., a greater number of atoms can exhibit a close interatomic distance in the 5- FT/SC interactions compared to the 5- FT/5- FT (see Supplementary Fig. 14 and related discussion). Thereby, the force exerted on the small S5- FT is minimal, which not only results in a diminished driving force to initiate TBs migration but also mitigates morphological modulation. Therefore, the relatively smaller critical transfer ratio for 5- FT/5- FT (0.72) than that of the SC/5- FT ( \(\sim 0.83\) ) is determined by the stable twin structure and decreases interaction force introduced by the small S5- FT. Notably, the particle size of various aggregated NPs investigated is smaller than \(10 \mathrm{nm}\) (Fig. 5). The above critical transfer ratio is supposed to vary slightly with further growth of the particle size of 5- FT, considering the size effect on 5- FT, which is stable with the size of \(3 - 14 \mathrm{nm}\) owing to thermodynamics \(^{24,25,26}\) .  
+
+<|ref|>image<|/ref|><|det|>[[225, 416, 770, 615]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[147, 628, 850, 672]]<|/det|>
+<center>Fig. 5. Aggregation growth size effect of 5-FT NP. (a) Configuration obtained after aggregation between SC and 5-FT NPs. (b) Configuration obtained after aggregation between two 5-FT NPs. </center>  
+
+<|ref|>text<|/ref|><|det|>[[147, 684, 853, 894]]<|/det|>
+In conclusion, our findings reveal the fundamental atomic- scale mechanisms governing the aggregation growth and evolution of 5- FT NPs. We have demonstrated that when a 5- FT attaches to small NPs ( \(R< 0.83\) for SC and \(R< 0.72\) for 5- FT), surface diffusion of the small nanoparticles to the big 5- FT dominates the configuration evolution, resulting in the formation of S5- FT NPs and decreasing of the system energies. Nevertheless, when a 5- FT attaches to a larger SC NP, beside surface diffusion, TB migration induced de- twinning process contributes significantly to the evolution process. This dynamic interplay induces the formation of a SC or a simple twinned structure. Additionally, in the case of attachment to another big 5- FT NP ( \(R > 0.72\) ), this results in
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 91, 500, 106]]<|/det|>
+the development of intricate 5- FT configurations.  
+
+<|ref|>text<|/ref|><|det|>[[147, 118, 853, 302]]<|/det|>
+Comprehending the mechanisms and dynamics underlying the process of particles aggregation growth is critical to constructing quantitative models to elucidate formation and evolution mechanisms of diverse twinned mineral materials in natural environments. Simultaneously, mastering the deterministic manipulation of NPs' growth holds the innovative synthesis pathways for fabricating twinned crystals endowed with precisely controlled morphologies and properties. Consequently, this endeavor provides guides to unlock the full potential of twinned crystal structures and promotes materials design and synthesis across a myriad of applications.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 331, 226, 347]]<|/det|>
+## Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 359, 301, 375]]<|/det|>
+## Sample preparation  
+
+<|ref|>text<|/ref|><|det|>[[147, 386, 854, 654]]<|/det|>
+Synthesis of gold (Au) nanoparticles (NPs): Gold(III) chloride trihydrate (HAuCl4·3H2O, \(\geq 99.9\%\) trace metals basis), tetrakis(decyl)ammonium bromide \((\geq 99.0\%)\) , TDAB), sodium borohydride (NaBH4, \(98\%\) ), 1- dodecanethiol \((\geq 98\%)\) and toluene \((\geq 99.5\%)\) were purchased from Sigma- Aldrich and used without further purification. The Au NPs were synthesized by the TDAB method \(6,35\) . 0.27 mmol of TDAB was dissolved in \(4\mathrm{ml}\) Toluene. The TDAB solution was mixed with \(6\mathrm{ml}\) of a \(0.01\mathrm{M}\) HAuCl4·3H2O aqueous solution for 1 hour. When the color of the toluene layer was changed to yellow due to the migration of AuCl4 ions, the only toluene layer was collected. Then, \(200\mu \mathrm{l}\) of a \(26\mathrm{mM}\) NaBH4 aqueous solution, which was prepared under an ice bath, was injected into the Au- TDAB toluene solution. After \(20\mathrm{min}\) , \(800\mu \mathrm{l}\) of 1- dodecanethiol was added to this mixture. The product was precipitated by excessive ethanol, and then collected by centrifuge.  
+
+<|ref|>text<|/ref|><|det|>[[147, 664, 853, 904]]<|/det|>
+After centrifuge, the sample was dispersed in \(5\mathrm{ml}\) toluene solution, drop- cast onto a TEM grid, and labeled as the as- prepared sample. As shown in Supplementary Fig. 1a- c, the as- prepared sample is SC dominated Au NPs with an average size of \(3.5\pm 0.6\mathrm{nm}\) . To tune particles' structure and size, half (2.5 ml) of the above toluene solution was sealed and kept in the dark at room temperature for \(\sim 2\) months (Supplementary Fig. 1d- f), and another half (2.5 ml) of the above toluene solution with centrifuged samples was cleaned four times with toluene to remove 1- dodecanethiol on the Au NP surface (Supplementary Fig. 1g- i). About \(1 / 3\) of NPs has evolved into 5- FT after staying for 2 months, and the particle size increases into \(4.4\pm 1.4\mathrm{nm}\) (Supplementary Fig. 1e- f). After purifying for 4 times, beside formation of some 5- FT NPs, significant aggregation of NPs was also detected
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 89, 852, 219]]<|/det|>
+(highlighted by yellow arrows in Supplementary Fig. 1h). And the average particle size grows into \(6.2 \pm 2.6 \mathrm{nm}\) (Supplementary Fig. 1i). To investigate aggregation growth and evolution of 5- FT nanoparticles, the sample after staying for 2 months was employed for further study. Particle sizes were statistically analyzed based on more than 300 particles for each sample and the standard errors were employed (Supplementary Fig. 1).  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 247, 454, 264]]<|/det|>
+## In-situ high-resolution TEM experiment  
+
+<|ref|>text<|/ref|><|det|>[[147, 275, 853, 710]]<|/det|>
+To investigate atomic scale 5- FT NPs growth and evolution mechanisms by particle aggregation, electron- beam induced NPs aggregation method \(^{6}\) was employed. This method has been successfully engaged to tune the formation kinetics of 5- FT Au nanoparticles, by controlling decomposition process of organic ligand coated on the Au nanoparticles under various electron beam dose rate \(^{6}\) . Here, a cold- field emission aberration- corrected TEM (Spectra 300, Thermal Fisher, USA), was employed at \(300 \mathrm{kV}\) for the above in- situ high- resolution TEM (HRTEM) imaging with an image frame rate of \(\sim 0.5 \mathrm{s}\) . An optimized dose rate of \(\sim (2 - 4) \times 10^{6} \mathrm{e} / \mathrm{nm}^{2} \cdot \mathrm{s}\) was applied. When the ratio between the longest TB and the shortest TB of 5- FT is smaller than 2.0, the 5- FT NT is referred to as a S5- FT. When the ratio is larger than 2.0, the 5- FT NP is referred to as an AS5- FT. Theoretically, the configuration of NPs evolves continuously under the employed electron- beam dose rate with time. Here, the terminal time was roughly determined according to the evolution rate of the NPs' configurations, including morphologies and twin structures. The ultimate configurations were obtained if their variation can be ignored within \(\sim 20 \mathrm{s}\) . Correspondingly, most of the in- situ experimental duration are \(\sim 150 - 200 \mathrm{s}\) under the employed electron- beam dose rate. Slip directions of partial dislocations are determined by TBs migration directions based on the immediate prior and subsequent images.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 739, 264, 753]]<|/det|>
+## MD simulation  
+
+<|ref|>text<|/ref|><|det|>[[147, 766, 852, 895]]<|/det|>
+MD simulations were conducted to explore the evolution processes of both the three- dimensional configuration and energy of simulated models of Au NPs (Supplementary Figs. 2- 3). Two types of models were employed in our experiment and all invariable decahedral 5- FT NPs (left side, Supplementary Figs. 2- 3) are the same, i.e., all surfaces of the big 5- FT are {111} planes with 1773 atoms. Model 1 was built with a \(5 \mathrm{nm}\) decahedral 5- FT NP attached with SC NPs in a box with
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 89, 852, 248]]<|/det|>
+a side length of \(20\mathrm{nm}\) . The SC NPs with approximately \(2.6\mathrm{nm}\) , \(3.8\mathrm{nm}\) , and \(4.8\mathrm{nm}\) (Supplementary Fig. 2), present truncated octahedral morphologies and have corresponding numbers (Supplementary Table 1) of 811, 1862, and 4033 atoms, respectively. Model 2 was built with a \(5\mathrm{nm}\) decahedral 5- FT NP attached with decahedral 5- FT NPs in a box with a side length of \(20\mathrm{nm}\) . The size and numbers of atoms of various 5- FT NPs are \(3.2\mathrm{nm}\) , \(4.1\mathrm{nm}\) , \(5.0\mathrm{nm}\) and 569, 1061, 1773, respectively (Supplementary Fig. 3 and Table 1). The models were visualized by OVITO software<sup>36</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[147, 256, 852, 386]]<|/det|>
+MD simulations were performed using the LAMMPS code<sup>37</sup>. The Au embedded atom potential developed by Ackland et al.<sup>38</sup> was employed in the simulation. This interatomic potential has been applied in previous investigations on the evolution of Au NPs during OA process successfully<sup>6</sup>. A Nose- Hoover thermostat<sup>39</sup> in the canonical ensemble was used to maintain the temperature of \(1100\mathrm{K}\) and the time step was configured to 1fs.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 415, 280, 430]]<|/det|>
+## Force calculation  
+
+<|ref|>text<|/ref|><|det|>[[147, 435, 851, 517]]<|/det|>
+Although the electrostatic force may argue as the source of the jump connection during the orientation attachment<sup>4</sup>, in our experiment, when the distance between the two aggregated NPs is smaller than \(\sim 2\mathrm{nm}\) , the vdW force is supposed to be the determining force<sup>35</sup>, especially the evolution process after aggregation, such as the relative slip between NPs.  
+
+<|ref|>text<|/ref|><|det|>[[147, 521, 851, 625]]<|/det|>
+vdW interactions between two NPs (NP<sub>1</sub> and NP<sub>2</sub>) were performed based on Hamaker's approach. To describe the vdW interactions of heterogeneous structures such as 5- FT structures, each NP was divided into many small cubes with length \((L) = 0.2882\mathrm{nm}\) , which is the closest Au- Au separation distance. Next, the vdW interactions between the two NPs were calculated by the summations of each pairwise interaction.  
+
+<|ref|>sub_title<|/ref|><|det|>[[179, 631, 403, 647]]<|/det|>
+## 1) Dividing into small cubes  
+
+<|ref|>text<|/ref|><|det|>[[147, 653, 850, 692]]<|/det|>
+The center positions of the small cubes of the NP<sub>1</sub> and the NP<sub>2</sub> were defined as \((x_{1,i}, y_{1,i}, z_{1,i})\) and \((x_{2,j}, y_{2,j}, z_{2,j})\) , respectively.  
+
+<|ref|>sub_title<|/ref|><|det|>[[178, 697, 401, 713]]<|/det|>
+## 2) vdW interaction potential  
+
+<|ref|>text<|/ref|><|det|>[[147, 719, 851, 801]]<|/det|>
+The vdW interaction potentials between the small cubes of the NP<sub>1</sub> and the small cubes of the NP<sub>2</sub> were calculated based on the formulation shown in previous work that followed Hamaker's approach with geometrical consideration.<sup>40</sup> The formulations can produce the vdW interaction energy \((V)\) between small cube- \(\alpha\) , i and cube- \(\beta\) , j with parallel configurations at arbitrary separations.  
+
+<|ref|>equation<|/ref|><|det|>[[400, 805, 835, 839]]<|/det|>
+\[V_{\alpha \beta ,ij} = -\left(\frac{A}{\pi^2}\right)V_{R,\alpha \beta ,ij} \quad (Eq (1))\]  
+
+<|ref|>text<|/ref|><|det|>[[147, 842, 851, 904]]<|/det|>
+where A is the Hamaker constant \((359.579 \times 10^{- 21}\mathrm{J})\) obtained from the reference paper<sup>41</sup>, which calculated dielectric responses of Au in vacuum based on the Lifshitz theory. \(V_{R}\) is the London- vdW interaction energy between cubes,
+
+<--- Page Split --->
+<|ref|>equation<|/ref|><|det|>[[350, 83, 835, 108]]<|/det|>
+\[V_{R,\alpha \beta ,ij} = \iint v_{\alpha ,i}v_{\beta ,j}(1 / r^6)\mathrm{d}v_{\alpha ,i}\mathrm{d}v_{\beta ,j} \quad (Eq 2)\]  
+
+<|ref|>text<|/ref|><|det|>[[147, 112, 850, 174]]<|/det|>
+where, \(r\) is the distance between an arbitrary point in the small cube- \(\alpha\) , \(i\) and an arbitrary point in the small cube- \(\beta\) , \(j\) , \(v_{\alpha ,i}\) and \(v_{\beta ,j}\) are the volume of the respective bodies. For the numerical computations, the integrations of Eq (2) were modified by  
+
+<|ref|>equation<|/ref|><|det|>[[211, 177, 840, 545]]<|/det|>
+\[\begin{array}{r l} & {i f a_{m}\neq 0,b_{n}\neq 0a n d c_{p}\neq 0:}\\ & {V_{R,\alpha \beta ,i j}(c_{p},b_{n},a_{m})}\\ & {\qquad = \left[\left(\frac{c_{p}(a_{m}^{2} + b_{n}^{2})^{\frac{3}{2}}}{24a_{m}^{2}b_{n}^{2}}\right)\tan^{-1}\left(\frac{c_{p}}{\sqrt{a_{m}^{2} + b_{n}^{2}}}\right)\right.}\\ & {\qquad +\left(\frac{3}{32}\right)\left(\frac{c_{p}}{b_{n}} -\frac{b_{n}}{c_{p}}\right)\tan^{-1}\left(\frac{b_{n}}{c_{p}}\right)}\\ & {\qquad +\left(\frac{1}{24}\right)b_{n}\left(\frac{1}{a_{m}^{2}} +\frac{1}{c_{p}^{2}}\right)\sqrt{a_{m}^{2} + c_{p}^{2}}\tan^{-1}\left(\frac{b_{n}}{\sqrt{a_{m}^{2} + c_{p}^{2}}}\right)}\\ & {\qquad +\left(\frac{1}{24}\right)a_{m}\left(\frac{1}{b_{n}^{2}} +\frac{1}{c_{p}^{2}}\right)\sqrt{b_{n}^{2} + c_{p}^{2}}\tan^{-1}\left(\frac{a_{m}}{\sqrt{b_{n}^{2} + c_{p}^{2}}}\right)}\\ & {\qquad +\left(\frac{1}{32}\right)\ln \left(\frac{(b_{n}^{2} + c_{p}^{2})^{3}}{c_{p}^{2}(a_{m}^{2} + b_{n}^{2} + c_{p}^{2})^{2}}\right)\Bigg]}\\ & {\qquad i f a_{m}\neq 0,b_{n}\neq 0a n d c_{p} = 0:}\\ & {V_{x y z}(c_{p},b_{n},0) = \frac{1}{32}\left(\frac{c_{p}}{b_{n}} -\frac{b_{n}}{c_{p}}\right)\tan^{-1}\left(\frac{b_{n}}{c_{p}}\right) + \frac{1}{32}\ln \left(1 + \frac{b_{n}^{2}}{c_{p}^{2}}\right)}\\ & {\qquad i f a_{i}\neq 0,b_{j} = 0a n d c_{k}\neq 0:}\\ & {V_{x y z}(c_{p},0,a_{m}) = -\frac{1}{16}\left(\frac{c_{p}}{a_{m}} -\frac{a_{m}}{c_{p}}\right)\tan^{-1}\left(\frac{a_{m}}{c_{p}}\right) - \frac{1}{16}\ln \left(1 + \frac{a_{m}^{2}}{c_{p}^{2}}\right)} \end{array} \quad (Eq 3)\]  
+
+<|ref|>equation<|/ref|><|det|>[[211, 572, 680, 612]]<|/det|>
+\[V_{x y z}(c_{p},b_{n},0) = \frac{1}{32}\left(\frac{c_{p}}{b_{n}} -\frac{b_{n}}{c_{p}}\right)\tan^{-1}\left(\frac{b_{n}}{c_{p}}\right) + \frac{1}{32}\ln \left(1 + \frac{b_{n}^{2}}{c_{p}^{2}}\right)\]  
+
+<|ref|>equation<|/ref|><|det|>[[211, 616, 444, 636]]<|/det|>
+\[i f a_{i}\neq 0,b_{j} = 0a n d c_{k}\neq 0:\]  
+
+<|ref|>equation<|/ref|><|det|>[[211, 640, 728, 681]]<|/det|>
+\[V_{x y z}(c_{p},0,a_{m}) = -\frac{1}{16}\left(\frac{c_{p}}{a_{m}} -\frac{a_{m}}{c_{p}}\right)\tan^{-1}\left(\frac{a_{m}}{c_{p}}\right) - \frac{1}{16}\ln \left(1 + \frac{a_{m}^{2}}{c_{p}^{2}}\right)\]  
+
+<|ref|>text<|/ref|><|det|>[[183, 708, 521, 725]]<|/det|>
+Parameter \(a_{m},b_{n}\) , and \(c_{p}\) were defined as below,  
+
+<|ref|>equation<|/ref|><|det|>[[298, 730, 838, 797]]<|/det|>
+\[\begin{array}{r l} & {a_{1} = h z + L;a_{2} = h z + 2L;a_{3} = h z + L;a_{4} = h z;}\\ & {b_{1} = h y + L;b_{2} = h y + 2L;b_{3} = h y + L;b_{4} = h y;}\\ & {c_{1} = h x + L;c_{2} = h x + 2L;c_{3} = h x + L;c_{4} = h x;} \end{array} \quad (Eq 4)\]  
+
+<|ref|>text<|/ref|><|det|>[[147, 800, 728, 817]]<|/det|>
+where \(h x, h y\) , and \(h z\) were denoted, which are shown in the Supplementary Fig. 4.  
+
+<|ref|>text<|/ref|><|det|>[[181, 821, 765, 838]]<|/det|>
+Next, the vdW interaction potentials \((V)\) between \(\mathrm{NP}_{1}\) and \(\mathrm{NP}_{2}\) were calculated by  
+
+<|ref|>equation<|/ref|><|det|>[[423, 841, 838, 881]]<|/det|>
+\[V = \sum_{i}\sum_{j}V_{\alpha \beta ,i j} \quad (Eq 5)\]  
+
+<|ref|>text<|/ref|><|det|>[[181, 886, 850, 903]]<|/det|>
+The interparticle distance between \(\mathrm{NP}_{1}\) and \(\mathrm{NP}_{2}\) \((h_{d})\) were defined as the distance between
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 84, 641, 100]]<|/det|>
+closed atoms of each NC, which was shown in Supplementary Fig. 5.  
+
+<|ref|>text<|/ref|><|det|>[[180, 106, 375, 122]]<|/det|>
+3) vdW interaction force  
+
+<|ref|>text<|/ref|><|det|>[[180, 128, 824, 144]]<|/det|>
+Based on the vdW potential as a function of \(h_{d}\) , the vdW force was calculated numerically.  
+
+<|ref|>equation<|/ref|><|det|>[[322, 148, 835, 184]]<|/det|>
+\[F = -\frac{V(h_{d} + \Delta d) - V(h_{d} - \Delta d)}{2\times\Delta d}, \quad (Eq (6))\]  
+
+<|ref|>text<|/ref|><|det|>[[147, 188, 400, 204]]<|/det|>
+Here, \(\Delta d = 0.05 \mathrm{nm}\) was employed.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 235, 293, 251]]<|/det|>
+## Data availability  
+
+<|ref|>text<|/ref|><|det|>[[148, 258, 850, 311]]<|/det|>
+The authors declare that the data supporting the findings of this study are available within the paper and its supplementary information. All raw data generated during the current study are available from the corresponding authors upon request.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 336, 245, 352]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[147, 363, 636, 380]]<|/det|>
+1. Sugimoto T. Monodispersed particles. second ed., Elsevier (2001).  
+
+<|ref|>text<|/ref|><|det|>[[147, 391, 848, 434]]<|/det|>
+2. Polte J. Fundamental growth principles of colloidal metal nanoparticles — a new perspective. CrystEngComm 17, 6809-6830 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[147, 446, 850, 490]]<|/det|>
+3. De Yoreo JJ, et al. Crystallization by particle attachment in synthetic, biogenic, and geologic environments. Science 349, aaa6760 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[147, 501, 850, 546]]<|/det|>
+4. Li D, Nielsen MH, Lee JRI, Frandsen C, Banfield JF, De Yoreo JJ. Direction-Specific interactions control crystal growth by oriented attachment. Science 336, 1014-1018 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[147, 557, 850, 601]]<|/det|>
+5. Ye H, Zhang Z, Wang R. Nucleation and growth of nanocrystals investigated by in situ transmission electron microscopy. Small 19, 2303872 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[147, 612, 850, 656]]<|/det|>
+6. Song M, et al. Oriented attachment induces fivefold twins by forming and decomposing high-energy grain boundaries. Science 367, 40 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[147, 668, 850, 712]]<|/det|>
+7. Zhou L, O'Brien P. Mesocrystals — properties and applications. The Journal of Physical Chemistry Letters 3, 620-628 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[147, 724, 850, 768]]<|/det|>
+8. Zhu G, et al. Self-similar mesocrystals form via interface-driven nucleation and assembly. Nature 590, 416-422 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[147, 780, 850, 824]]<|/det|>
+9. Park BC, et al. Surface-ligand-induced crystallographic disorder-order transition in oriented attachment for the tuneable assembly of mesocrystals. Nature Communications 13, 1144 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[147, 836, 850, 880]]<|/det|>
+10. Li J, et al. In situ atomic-scale study of particle-mediated nucleation and growth in amorphous bismuth to nanocrystal phase transformation. Advanced Science 5, 1700992 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[147, 891, 850, 907]]<|/det|>
+11. Thanh NTK, Maclean N, Mahiddine S. Mechanisms of nucleation and growth of nanoparticles in
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 90, 504, 106]]<|/det|>
+solution. Chemical Reviews 114, 7610- 7630 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[147, 118, 850, 161]]<|/det|>
+12. Baumgartner J, et al. Nucleation and growth of magnetite from solution. Nature Materials 12, 310-314 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[147, 173, 851, 218]]<|/det|>
+13. Huang B, et al. Identification of plasmon-driven nanoparticle-coalescence-dominated growth of gold nanoplates through nanopore sensing. Nature Communications 13, 1402 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[147, 229, 851, 273]]<|/det|>
+14. Rose G. Ueber die krystallformen des goldes und des silbers. Annalen der Physik 99, 196-204 (1831).  
+
+<|ref|>text<|/ref|><|det|>[[147, 284, 850, 328]]<|/det|>
+15. Hofmeister H. Fivefold twinned nanoparticles. Encyclopedia of Nanoscience and Nanotechnology 3, 431-452 (2004).  
+
+<|ref|>text<|/ref|><|det|>[[147, 340, 850, 384]]<|/det|>
+16. Zhou S, Zhao M, Yang TH, Xia Y. Decahedral nanocrystals of noble metals: Synthesis, characterization, and applications. Materials Today 22, 108-131 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[147, 395, 850, 440]]<|/det|>
+17. Wu JY, Nagao S, He JY, Zhang ZL. Role of five-fold twin boundary on the enhanced mechanical properties of fcc Fe Nanowires. Nano Letters 11, 5264-5273 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[147, 451, 851, 523]]<|/det|>
+18. Wang X, et al. Pd@Pt Core-Shell Concave Decahedra: A class of catalysts for the oxygen reduction reaction with enhanced activity and durability. Journal of the American Chemical Society 137, 15036-15042 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[147, 535, 850, 579]]<|/det|>
+19. Pietrobon B, Kitaev V. Photochemical synthesis of monodisperse size-controlled silver decahedral nanoparticles and their remarkable optical Properties. Chemistry of Materials 20, 5186-5190 (2008).  
+
+<|ref|>text<|/ref|><|det|>[[147, 590, 850, 634]]<|/det|>
+20. Maksimuk S, Teng X, Yang H. Roles of twin defects in the formation of platinum multipod nanocrystals. The Journal of Physical Chemistry C 111, 14312-14319 (2007).  
+
+<|ref|>text<|/ref|><|det|>[[147, 645, 852, 690]]<|/det|>
+21. Clasen P, Tiruvalam R, Ma S, Harmer MP, Kiely C. Effect of synthesis parameters on the gold nanostructures produced using an oleamine reductant. Microscopy and Microanalysis 13, 776-777 (2007).  
+
+<|ref|>text<|/ref|><|det|>[[147, 701, 683, 717]]<|/det|>
+22. Lakes R. Materials with structural hierarchy. Nature 361, 511-515 (1993).  
+
+<|ref|>text<|/ref|><|det|>[[147, 729, 850, 773]]<|/det|>
+23. Han H, et al. Multiscale hierarchical structures from a nanocluster mesophase. Nature Materials 21, 518-525 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[147, 785, 850, 829]]<|/det|>
+24. Barnard AS, Young NP, Kirkland AI, van Huis MA, Xu H. Nanogold: A quantitative phase map. ACS Nano 3, 1431-1436 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[147, 840, 850, 912]]<|/det|>
+25. Young NP, van Huis MA, Zandbergen HW, Xu H, Kirkland AI. Transformations of gold nanoparticles investigated using variable temperature high-resolution transmission electron microscopy. Ultramicroscopy 110, 506-516 (2010).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[144, 90, 852, 135]]<|/det|>
+26. Koga K, Ikeshoji T, Sugawara K. Size- and temperature-dependent structural transitions in gold nanoparticles. Physical Review Letters 92, 115507 (2004).  
+
+<|ref|>text<|/ref|><|det|>[[147, 145, 850, 190]]<|/det|>
+27. Marks LD. Surface structure and energetics of multiply twinned particles. Philosophical Magazine A 49, 81-93 (1984).  
+
+<|ref|>text<|/ref|><|det|>[[147, 200, 850, 273]]<|/det|>
+28. Cleveland CL, Landman U, Schaaff TG, Shafigullin MN, Stephens PW, Whetten RL. Structural evolution of smaller gold nanocrystals: the truncated decahedral motif. Physical Review Letters 79, 1873-1876 (1997).  
+
+<|ref|>text<|/ref|><|det|>[[147, 284, 850, 329]]<|/det|>
+29. Johnson CL, et al. Effects of elastic anisotropy on strain distributions in decahedral gold nanoparticles. Nature Materials 7, 120 (2007).  
+
+<|ref|>text<|/ref|><|det|>[[147, 339, 850, 384]]<|/det|>
+30. Gryaznov VG, Heydenreich J, Kaprelov AM, Nepijko SA, Romanov AE, Urban J. Pentagonal symmetry and disclinations in small particles. Crystal Research and Technology 34, 1091-1119 (1999).  
+
+<|ref|>text<|/ref|><|det|>[[147, 395, 850, 440]]<|/det|>
+31. Song M, Wu Z, Lu N, Li D. Strain relaxation-induced twin interface migration and morphology evolution of silver nanoparticles. Chemistry of Materials 31, 842-850 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[147, 450, 850, 495]]<|/det|>
+32. Patala S, Marks LD, Olvera de la Cruz M. Elastic strain energy effects in faceted decahedral nanoparticles. The Journal of Physical Chemistry C 117, 1485-1494 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[147, 506, 850, 551]]<|/det|>
+33. Song M, et al. Uneven strain distribution induces consecutive dislocation slip, plane gliding, and subsequent detwinning of penta-twinned nanoparticles. Nano Letters 24, 1153-1159 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[147, 561, 850, 606]]<|/det|>
+34. Zhu YT, Wu XL, Liao XZ, Narayan J, Kecskes LJ, Mathaudhu SN. Dislocation-twin interactions in nanocrystalline fcc metals. Acta Materialia 59, 812-821 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[147, 617, 850, 690]]<|/det|>
+35. Lee J, Nakouzi E, Song M, Wang B, Chun J, Li D. Mechanistic understanding of the growth kinetics and dynamics of nanoparticle superlattices by coupling interparticle forces from real-time measurements. ACS nano 12, 12778-12787 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[147, 701, 850, 746]]<|/det|>
+36. Stukowski A. Visualization and analysis of atomistic simulation data with OVITO-the open visualization tool. Modelling and Simulation in Materials Science and Engineering 18, 015012 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[147, 756, 850, 801]]<|/det|>
+37. Thompson AP, et al. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Computer Physics Communications 271, 108171 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[147, 812, 850, 856]]<|/det|>
+38. Ackland GJ, Tichy G, Vitek V, Finnis MW. Simple N-body potentials for the noble metals and nickel. Philosophical Magazine A 56, 735-756 (1987).  
+
+<|ref|>text<|/ref|><|det|>[[147, 867, 850, 912]]<|/det|>
+39. Nose S. A unified formulation of the constant temperature molecular dynamics methods. The Journal of Chemical Physics 81, 511-519 (1984).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 90, 850, 135]]<|/det|>
+40. Maeda H, Maeda Y. Orientation-dependent london-van der Waals interaction energy between macroscopic bodies. Langmuir 31, 7251-7263 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[148, 145, 850, 190]]<|/det|>
+41. Pinchuk P, Jiang K. Size-dependent Hamaker constants for silver and gold nanoparticles. In: Physical Chemistry of Interfaces and Nanomaterials XIV). SPIE (2015).  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 220, 317, 237]]<|/det|>
+## Acknowledgements  
+
+<|ref|>text<|/ref|><|det|>[[147, 246, 852, 405]]<|/det|>
+This project is supported by the National Natural Science Foundation of China (No. 52101025) and Excellent Young Scientists Fund Program (Overseas), Natural Science Foundation of Hunan Province (No. 2023JJ30684), the Changsha Municipal Natural Science Foundation (kq2202091), and State Key Laboratory of Crystal Materials, Shandong University (No. KF2206). We would also like to acknowledge the support from State Key Laboratory of Powder Metallurgy, Central South University, Changsha, China.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 433, 334, 450]]<|/det|>
+## Author contributions  
+
+<|ref|>text<|/ref|><|det|>[[147, 460, 852, 617]]<|/det|>
+M.S. designed the experiments, conducted experiments and data analysis, wrote the manuscript, and supervised the study. D. Z. assisted with image processing and data analysis. D. Leng revised the manuscript. J. L. synthesized Au NPs and conducted force calculation. Z. Y., J. C., D. L., Z. G., and L. W. contributed to the discussion of the results and commented on the manuscript. G. Z. conducted MD simulations and wrote the manuscript. R. Y. and K. Z. supervised the study and commented on the manuscript.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 647, 323, 664]]<|/det|>
+## Competing interests  
+
+<|ref|>text<|/ref|><|det|>[[149, 675, 459, 690]]<|/det|>
+The authors declare no competing interests.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 720, 350, 737]]<|/det|>
+## Additional information  
+
+<|ref|>text<|/ref|><|det|>[[148, 747, 850, 792]]<|/det|>
+Supplementary information The online version contains supplementary material available at XXXX.  
+
+<|ref|>text<|/ref|><|det|>[[148, 803, 712, 820]]<|/det|>
+Correspondence and requests for materials should be addressed to Miao Song.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[43, 92, 768, 112]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 366, 149]]<|/det|>
+SupplementaryInformation.docx
+
+<--- Page Split --->

preprint/preprint__2af9ac54537a84d7e65fbe5670f2af99cc7c43c1b79b4d127a1a2be44b23db02/images_list.json

@@ -0,0 +1,62 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1 | Thermal simulation and preparation of gradient nanostructured aerogel fibers (GAFs). a Coarse-Grained molecular dynamics simulations of thermal insulation for skin-core \\((\\mathbf{a}_1)\\) and gradient nanostructured \\((\\mathbf{a}_2)\\) models, with the radial temperature distribution map within the region marked by dashed box placed at the bottom, and the simulated thermal conductivities (in Lennard-Jones unit) through the gradient interface of the gradient nanostructured model and the bulk of the skin-core model on the right \\((\\mathbf{a}_3)\\) ; b Sketch map for the preparation of GAFs, including nanoexfoliation, microfluidic spinning, sol-gel translation, and supercritical drying; c, d Representative SEM images of skin-core and gradient nanostructured aerogel fibers; e, f). Scalability of GAFs and textiles.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        87,
+        880,
+        536
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2 | Forming mechanism and morphological characterization of GAFs. a Schematic illustration of nanostructure changes during GAFs preparation, including ANF dispersion in the microfluidic system, sol-gel transition in the coagulation bath, and ethanol solvent exchange; b Photograph of GAFs; c Cross-sectional SEM images of SAF and GAF at various locations during wet-spinning and microfluidic spinning, including center positions (c1, c4), edge positions (c2, c3, and c5), and gradient interface position (c6); d Schematic of Raman line imaging along the diameter (d1), with representative Raman spectra selected at equal intervals from the exterior to the interior of SAF (d2) and GAF (d3).",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        88,
+        880,
+        590
+      ]
+    ],
+    "page_idx": 7
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3 | Study on the mechanical properties and deformation mechanisms of GAF. a Stress-strain curves of SAF and GAF; b Comparison of toughness and tensile strength based on ANF aerogel materials<sup>28</sup>-<sup>32</sup>; c Photograph of a single GAF suspending a \\(50 \\mathrm{g}\\) weight; d Simulation trajectories (d1) and simulated stress-strain curves (d2) for SAF and GAF models during tensile processes; e Raman imaging technique detecting internal stress levels in poly(p-phenylene terephthalate) based on characteristic Raman shift of C=O group (e1) and the representative 2D Raman images and Raman shifts of Kevlar fiber (e2, e3), SAF (e3, e6), and GAF (e4, e7) under \\(5\\%\\) tensile strain.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        160,
+        880,
+        730
+      ]
+    ],
+    "page_idx": 9
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4 | Investigation of thermal insulation properties and mechanism of GAFs. a Heat plate experiment with equal-weight fabrics: Infrared thermal images of Kevlar, SAF, and GAF fabrics on a heat source (a1); Temperature drop between the heat plate and Kevlar, SAF, and GAF fabrics (a2); b Heat plate experiment with 0.5 mm thick fabrics: Kevlar, SAF, and GAF; c Temperature-heating time curves for Kevlar, SAF, and GAF fabrics on a heat source; d Comparison of thermal conductivity of GAF with other natural and synthetic fibers<sup>28-30,33-37</sup>; e Radar plot comparing thermal insulating materials on mechanical properties, thermal insulation, weavability, cost, and processing efficiency; f Simulated temperature profiles at t = 10000τ for a uniform density model under identical heat source and sink conditions; g Simulated temperature profiles at t = 10000τ for a gradient density model under identical heat source and sink conditions; h Energy-time curves at heat sources and sinks for uniform density and gradient density models during simulation; i Simulated thermal conductivities (in Lennard-Jones unit) of SAF and GAF.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        90,
+        880,
+        700
+      ]
+    ],
+    "page_idx": 12
+  }
+]

preprint/preprint__2af9ac54537a84d7e65fbe5670f2af99cc7c43c1b79b4d127a1a2be44b23db02/preprint__2af9ac54537a84d7e65fbe5670f2af99cc7c43c1b79b4d127a1a2be44b23db02.mmd

@@ -0,0 +1,272 @@
+
+# Gradient All-Nanostructured Aerogel Fibers for Enhanced Thermal Insulation and Mechanical Properties  
+
+Dongdong Ye ydd@whu.edu.cn  
+
+Anhui Agricultural University https://orcid.org/0000- 0002- 3377- 0656  
+
+Xiaotong Fu Anhui Agricultural University  
+
+Lianmeng Si Xi'an Jiaotong University  
+
+Zhaoxin Zhang Zhejiang University  
+
+Tingting Yang Anhui Agricultural University  
+
+Jianwei Song Xi'an Jiaotong University, Xi'an  
+
+Shuze Zhu Zhejiang University https://orcid.org/0000- 0001- 7849- 6067  
+
+## Article  
+
+# Keywords:  
+
+Posted Date: August 29th, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4912597/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on March 10th, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 57646- 4.
+
+<--- Page Split --->
+
+# Gradient All-Nanostructured Aerogel Fibers for Enhanced Thermal Insulation and Mechanical Properties  
+
+Xiaotong \(\mathrm{Fu}^{1,4}\) , Lianmeng \(\mathrm{Si}^{2,4}\) , Zhaoxin Zhang \(^{3}\) , Tingting Yang \(^{1}\) , Jianwei Song \(^{*2}\) , Shuze Zhu \(^{*3}\) , Dongdong \(\mathrm{Ye}^{*1}\)  
+
+\(^{1}\) School of Materials and Chemistry, Anhui Agricultural University, Hefei, Anhui Province 230036, China  
+
+\(^{2}\) State Key Laboratory for Strength and Vibration of Mechanical Structures, School of Aerospace Engineering, Xi'an Jiaotong University, Xi'an, 710049, China  
+
+\(^{3}\) Department of Engineering Mechanics, Institute of Applied Mechanics, Zhejiang University, Hangzhou 310000, China  
+
+\(^{4}\) These authors contributed equally to this work  
+
+Corresponding authors' emails: ydd@whu.edu.cn; shuezhu@zju.edu.cn; songjianwei@xjtue.edu.cn  
+
+Abstract: Lightweight, nanoporous aerogel fibers are crucial for personal thermal management and specialized heat protection. However, wet-spinning methods, exemplified by aramid aerogels, inevitably form a dense outer layer, significantly reducing the volume fraction of efficient thermal barrier nanovoids and limiting the development of ultimate thermal resistance in fibers. Herein, we develop a microfluidic spinning method to prepare gradient all-nanostructure aramid aerogel fibers (GAFs). Benefiting from the simultaneous shear alignment and diffusion dilution of a good solvent within the channels, the precursor gel fibers assemble into a structure with a sparse exterior and dense interior, which reverses during supercritical drying to form sheath and core layers with average pore diameters of 150 nm and 600 nm, respectively. Experiments and simulations reveal that the gradient nanostructure creates high interfacial thermal resistance at heat transfer interfaces, resulting in a GAF radial thermal conductivity as low as 0.0228 W m \(^{-1}\) K \(^{-1}\) , far below that of air and wet-spun aerogel fibers. Moreover, GAF's unique nano-entangled network efficiently dissipates stress, achieving exceptionally high tensile strength (29.5 MPa) and fracture strain (39.2%). This work establishes a correlation between multiscale nanostructures and superlative performance, thereby expanding the scope of aerogel applications in intricate environments.
+
+<--- Page Split --->
+
+## Introduction  
+
+Thermal management systems within buildings and vehicles represent a substantial portion of global energy consumption, accounting for more than one- third of the total \(^{1}\) . This significant energy expenditure notably contributes to \(\mathrm{CO_2}\) emissions, thereby intensifying global carbon footprint \(^{2,3}\) . Developing advanced insulation materials presents a viable solution to mitigate heat loss in various applications, including building walls \(^{4}\) , roofs, vehicle exteriors, and apparel \(^{5}\) , thereby enhancing user comfort and curbing carbon emissions \(^{6}\) . Actually, materials exhibiting high porosity \((>0.9)\) are particularly effective in enhancing thermal insulation by limiting collisions between air molecules, elongating heat transfer pathways, and augmenting interface thermal resistance \(^{7- 9}\) . Distinguished by their abundant nanoporous structures, aerogels can be transformed into fibrous forms to preserve superior thermal insulation properties while attaining flexibility and fabricability, making them ideal for personal thermal management textiles. Currently, inorganic, polymeric, and composite aerogel fibers, including those made from silica, polyimide, cellulose, and hybrid materials, have shown promising insulation characteristics \(^{10- 13}\) . Nonetheless, challenges remain, such as inadequate mechanical strength, limited thermal insulation performance, and elevated manufacturing costs, primarily due to the inconsistent control over micro- and nanoscale structures.  
+
+Aramid nanofibers (ANFs) are regarded as ideal nanostructured units due to their inherent high strength, environmental resilience, active sites, and excellent dispersibility, derived from polyphenylene terephthalate (PPTA) \(^{14,15}\) . Aerogels assembled by ANFs exhibit outstanding thermal insulation properties to withstand extreme temperatures through three mechanisms: 1) reduced thermal conduction via the solid skeleton \(^{16}\) , 2) restricted thermal convection via the porous structure \(^{17}\) , and 3) multiple thermal radiation \(^{18}\) . Currently, representative aerogel fiber manufacturing methods such as wet spinning \(^{19}\) , reactive spinning \(^{20}\) , and confined spinning \(^{21}\) , while
+
+<--- Page Split --->
+
+enhancing scalability, inevitably reduce the precision of nanostructure control. This results in fibers exhibiting a dense skin layer on the exterior and a porous interior<sup>22,23</sup>, along with a wide range of thermal conductivity (0.027- 0.5 W m<sup>- 1</sup> K<sup>- 1</sup>)<sup>24</sup>. Previous researches indicate that the nanopore structure of aerogels significantly affects thermal conduction; transforming the dense packing structure of the shell layer into a loose nanostructure and optimizing the radial nanostructure distribution is expected to enhance the thermal insulation capabilities of aerogel fibers.  
+
+In this study, we unveil the development of gradient nanostructured aerogel fibers (GAFs), which exhibit exceptional thermal insulation properties, as corroborated by molecular simulations. Through the design of skin- core and gradient nanostructure models, we demonstrate that the outer- layer nanostructure of GAFs manifests a significantly superior thermal storage capacity and interfacial thermal resistance compared to skin- core aerogel fibers (SAFs), resulting in a \(67\%\) reduction in simulated thermal conductivity. Moreover, we have pioneered a manufacturing technique for producing GAFs using continuous microfluidic spinning technology. Coupled with supercritical drying, we have successfully fabricated aramid nanofiber aerogel fibers by precisely controlling solvent shear, diffusion, and sol- gel transition rates within the spinning channel. Compared to wet- spun aerogel fibers, GAFs exhibit increased porosity (from \(98\%\) to \(98.6\%\) ), reduced density (from \(20.5 \mathrm{kg / m}^3\) to \(15.7 \mathrm{kg / m}^3\) , enhanced tensile strength (from \(10.9 \mathrm{MPa}\) to \(29.5 \mathrm{MPa}\) ), and significantly reduced thermal conductivity (from \(0.0327 \mathrm{W m}^{- 1} \mathrm{K}^{- 1}\) to \(0.0228 \mathrm{W m}^{- 1} \mathrm{K}^{- 1}\) ). Notably, the superior durability of aramid ensures stability under extreme cold, high temperatures, and vacuum conditions. The excellent comprehensive performance of GAF significantly broadens its potential applications in textiles, construction, firefighting equipment, and even extreme aerospace scenarios, offering significant environmental benefits.
+
+<--- Page Split --->
+
+## Results and Discussion  
+
+The thermal insulation mechanisms of aerogel materials, encompassing multiple reflected radiation, reduced conduction, and restricted convection, are intricately linked to pore size and spatial distribution. Intriguingly, natural animal hair, exemplified by polar bears<sup>25</sup>, and synthetic aerogel fibers, typically fabricated through wet spinning, exhibit dense outer layers, porous inner cores, and superior thermal insulation performance. To investigate the correlation between the nanostructure and thermal insulation, we employed coarse- grained molecular dynamics simulations under the same conditions to assess the radial heat transfer properties of two fibrous structural models with skin- core and gradient nanostructures (Fig. 1a; Supplementary Fig. S1). The results reveal that the two nanostructured fibers exhibit markedly different thermal distribution characteristics under the imposed flux from heat source with temperature 2 (Lennard- Jones unit) to heat sink with temperature 1 (Lennard- Jones unit). Unlike the skin- core fibers, whose dense outer layer nanostructure results in low interfacial thermal resistance, the gradient nanostructured fibers possess a loose nanoporous outer layer, which demonstrates high heat storage capacity and forms substantial interfacial thermal resistance at the gradient interface, thereby restricting radial heat transfer. Furthermore, quantitative simulations of the dimensionless thermal conductivity for both fiber types indicate a \(40\%\) reduction in thermal conductivity for the gradient nanostructured fibers compared to the skin- core fibers. These findings provide theoretical evidence that regulating aerogels' radial nanostructure size and spatial distribution can significantly enhance thermal insulation performance.  
+
+Guided with the simulation, we developed the gradient all- nanostructured aramid aerogel fibers (GAFs) from aramid nanofibers (ANFs) dispersion based on a spinning system integrated with nanoexfoliation, microfluidic spinning, sol- gel transition, and supercritical \(\mathrm{CO_2}\) drying
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Fig. 1 | Thermal simulation and preparation of gradient nanostructured aerogel fibers (GAFs). a Coarse-Grained molecular dynamics simulations of thermal insulation for skin-core \((\mathbf{a}_1)\) and gradient nanostructured \((\mathbf{a}_2)\) models, with the radial temperature distribution map within the region marked by dashed box placed at the bottom, and the simulated thermal conductivities (in Lennard-Jones unit) through the gradient interface of the gradient nanostructured model and the bulk of the skin-core model on the right \((\mathbf{a}_3)\) ; b Sketch map for the preparation of GAFs, including nanoexfoliation, microfluidic spinning, sol-gel translation, and supercritical drying; c, d Representative SEM images of skin-core and gradient nanostructured aerogel fibers; e, f). Scalability of GAFs and textiles. </center>  
+
+process (Fig. 1b). Contrary to the skin- core aerogel fibers (SAFs) with dense shell and porous core configuration engendered by wet spinning (Fig. 1c), GAF is devoid of a dense exterior skin layer (Fig. 1d). ANF interweaves to forge a wholly nanoporous structure, congruent with the theoretical blueprint and rendering it an optimal candidate for thermal insulation. Additionally, the spinning system demonstrates the feasibility for large- scale fiber production (Fig. 1e,
+
+<--- Page Split --->
+
+Supplementary Movie 1), which can be further woven into fabrics (Figure 1f), owing to improved mechanical properties (discussed in detail later).  
+
+The formation of GAFs undergoes three key stages (Fig. 2a): 1) Heterogeneous spatial distribution of ANF induced by complex flow in a microfluidic chip: When the sheath flow of DMSO solvent is injected at an angle into the central channel containing ANF dispersion (DMSO/KOH), it not only enhances the shear orientation of ANF but also dilutes the ANF solution from the outside in, creating a concentration gradient with lower concentration at the exterior and higher concentration at the center due to shear orientation along the channel and diffusion dilution perpendicular to the channel; 2) Sol- gel transition induced by protonation: Upon entering the acidic coagulation bath, the ANF undergoes protonation, rapidly triggering gelation and freezing the gradient structure, resulting in gel fibers with loose pores on the exterior and smaller pores on the interior (Supplementary Fig. S2); 3) Structural inversion during supercritical drying: During the solvent exchange process to form alcohol gel fibers, the large pores with high surface porosity collapse into smaller pores (Supplementary Figs. S3), leading to the formation of gradient all- nanostructured aerogel fibers after supercritical \(\mathrm{CO_2}\) drying (Fig. 2b).  
+
+We subsequently compared the structural assembly differences between SAFs produced by wet- spinning and GAFs constructed via microfluidic spinning (Fig. 2c). SAF exhibited a nanoporous core (Fig. 2c1) and a distinct dense skin layer approximately \(100 \mathrm{nm}\) thick (Fig. 2c2, 2c3), as confirmed by cross- sectional SEM images. The GAF core displays a similar nanoporous structure to SAF (Fig. 2c4), with an average pore size of about \(500 \mathrm{nm}\) (Supplementary Fig. S4), while the outer layer maintains a nanoporous structure with an average pore size of \(160 \mathrm{nm}\) (Fig. 2c5; Supplementary Fig. S5) and shows a pronounced gradient interfacial layer (Fig. 2c6). Notably, altering the flow rate of DMSO from \(1200 \mathrm{to} 200 \mu \mathrm{L} / \mathrm{min}\) enables the regulation of the gradient
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Fig. 2 | Forming mechanism and morphological characterization of GAFs. a Schematic illustration of nanostructure changes during GAFs preparation, including ANF dispersion in the microfluidic system, sol-gel transition in the coagulation bath, and ethanol solvent exchange; b Photograph of GAFs; c Cross-sectional SEM images of SAF and GAF at various locations during wet-spinning and microfluidic spinning, including center positions (c1, c4), edge positions (c2, c3, and c5), and gradient interface position (c6); d Schematic of Raman line imaging along the diameter (d1), with representative Raman spectra selected at equal intervals from the exterior to the interior of SAF (d2) and GAF (d3). </center>  
+
+nanoporous outer layer structure (from dense to loose) and thickness (from 1.5 to \(35 \mu \mathrm{m}\) ), showcasing the excellent customizable structural capabilities of the microfluidic spinning system (Supplementary Fig. S6).  
+
+Furthermore, we characterized the variation in the intensity of the characteristic peak of the aramid I band (peak position at \(1610 \mathrm{cm}^{- 1}\) ) along the diameter of the aerogel fibers (Fig. 2d1),
+
+<--- Page Split --->
+
+which visually illustrates the differences in ANF packing density between the two types of aerogel fibers. The results indicate that the Raman scattering intensity on the surface of SAF is significantly higher than that within the fiber (Fig. 2d₂), suggesting the presence of a highly compact thin layer on the surface. In contrast, the GAF shows a uniformly decreasing scattering intensity from the exterior to the interior (Fig. 2d₃), indicating that the packing density outside the fiber is slightly higher than that inside, exhibiting a gradient transition. Raman imaging further confirms the distinct assembly structures of the two fiber types.  
+
+To elucidate the impact of nanostructures on the properties of aerogel fibers, we initially assessed the mechanical properties of SAFs and GAFs. The results indicate that SAF with a dense skin layer exhibits a higher Young's modulus, reaching 112.74 MPa, a 37.7% increase over GAF (81.9 MPa) (Fig. 3a; Supplementary Fig. S7). However, SAF's tensile strength (10.9 MPa), fracture strain (17.8%), and toughness (1.06 MJ m⁻³) are significantly lower than those of GAF (29.5 MPa, 39.2%, and 5.7 MJ m⁻³), demonstrating the potential of gradient nanostructures in enhancing aerogel mechanical properties. Notably, GAFs developed in this work show superior mechanical properties compared to classical ANF- based fibers²⁹, ³⁰ and films³¹ (Fig. 3b). As a performance demonstration, a single GAF can bear a weight of 50 g (Fig. 3c). To explore the deeper mechanisms of gradient nanostructure reinforcement and toughening, we employed molecular dynamics simulations to evaluate the deformation processes of SAF and GAF (Supplementary Fig. S8). The simulations revealed distinct fracture behaviors: as the simulated strain (ε) increased from 2 to 5, due to the small thickness of the dense nanofiber layer on the surface of SAF, at microscopic large deformation the dense layer can no longer maintain structure integrity, leading to micro- crack formation that exposes the internal nanostructure (Fig. 3d₁, Supplementary Movie 2); in contrast,
+
+<--- Page Split --->
+
+171 GAF maintained structural integrity during stretching, with outer nanopores aligning and  
+
+172 deforming with the fiber, avoiding  
+
+![](images/Figure_3.jpg)
+
+<center>Fig. 3 | Study on the mechanical properties and deformation mechanisms of GAF. a Stress-strain curves of SAF and GAF; b Comparison of toughness and tensile strength based on ANF aerogel materials<sup>28</sup>-<sup>32</sup>; c Photograph of a single GAF suspending a \(50 \mathrm{g}\) weight; d Simulation trajectories (d1) and simulated stress-strain curves (d2) for SAF and GAF models during tensile processes; e Raman imaging technique detecting internal stress levels in poly(p-phenylene terephthalate) based on characteristic Raman shift of C=O group (e1) and the representative 2D Raman images and Raman shifts of Kevlar fiber (e2, e3), SAF (e3, e6), and GAF (e4, e7) under \(5\%\) tensile strain. </center>
+
+<--- Page Split --->
+
+severe damage and retaining good mechanical properties (Fig. 3d₁, Supplementary Movie 3). The simulated stress- strain curves also reflected this trend (Fig. 3d₂).  
+
+Previous researches indicate that the specific functional group deformation under external force induces a redshift (Δv) in characteristic Raman peaks \(^{26,27}\) . Thus, we employed Raman imaging to investigate the relationship between the Raman shifts of the amide I band C=O stretching vibration in 5% deformed Kevlar fiber, SAF, GAF and even isotropic ANF- assembled membrane (Supplementary Fig. S9), aiming to elucidate the correlation between nanostructure and internal stress during deformation. We first tested the Raman spectra of isotropic ANF- assembled membrane prepared by the vacuum filtration method, confirming that the characteristic Raman shift of C=O in the relaxed state of PPTA molecular chains is 1656.5 cm\(^{-1}\) (Supplementary Fig. S9a). Comparing the 2D Raman imaging of the three types of fibers, we found that the Kevlar fiber mainly exhibits a low Raman shift in the red distribution (Fig. 3e₂; Supplementary Fig. S9b), with an average Δv of 5.6 cm\(^{-1}\) (Fig. 3e₅). In contrast, SAF shows significant stress concentration only in the densely packed skin layer (Fig. 3e₃; Supplementary Fig. S9c), with a Δv of 1.2 cm\(^{-1}\) (Fig. 3e₆), leading to initial high modulus and outer layer failure due to high internal stress during stretching. Interestingly, GAF shows the smallest overall Raman red shift (Fig. 3e₄; Supplementary Fig. S9d), with a Δv of only 0.3 cm\(^{-1}\) (Fig. 3e₇), indicating that the nanostructure effectively dissipates stress during stretching, resulting in the lowest internal stress and superior mechanical properties.  
+
+Subsequently, we evaluated the thermal insulation properties of skin- core and gradient- structured aerogel fibers and fabrics. Initially, we monitored the temperature changes atop a hot stage using an infrared thermal imager for both SAF and GAF specimens (Supplementary Fig. S10). The results indicated that when the hot stage was set at 200 °C, the top temperature of the
+
+<--- Page Split --->
+
+GAF sample was only \(190.3^{\circ}\mathrm{C}\) , significantly lower than that of the SAF specimen (193.5 \(^{\circ}\mathrm{C}\) ), demonstrating superior insulation capabilities. Further, we weaved the two aerogel fibers into fabrics and compared their thermal insulation properties under the same weight or thickness conditions (Fig. 4a). The findings revealed enhanced performance across various metrics (Figs. 4b, c). Additionally, we examined the thermal insulation properties of GAF fabrics of varying thicknesses across a broad temperature range (- 30 \(^{\circ}\mathrm{C}\) to 270 \(^{\circ}\mathrm{C}\) ) (Supplementary Figs. S11a, c). The results showed that a 0.5 mm GAF fabric could achieve temperature differentials of 33.4 \(^{\circ}\mathrm{C}\) and 63 \(^{\circ}\mathrm{C}\) at cold and heat templates, respectively (Supplementary Figs. S11b, d); when the fabric thickness increased to 1mm, the insulation effect improved, yielding temperature differentials of 45.4 \(^{\circ}\mathrm{C}\) and 113 \(^{\circ}\mathrm{C}\) at cold and heat templates, respectively (Supplementary Figs. S11b, d). Notably, the insulation effect of GAF fibers based on a microfluidic system surpassed that of most previously reported thermal insulation fibers, including synthetic fibers \(^{33,34}\) (e.g., PVC, PI, porous PI, PS, PET), natural fibers \(^{7,34 - 37}\) (e.g., silk, wood, wool, cotton, cotton aerogel, TCNF), inorganic fibers \(^{38 - 40}\) (e.g., glass, rock wool, ceramic, silicon), and ANF aerogels \(^{28 - 30}\) (KNF fibers, KNF films, ANF films) (Fig. 4d). Therefore, GAF, integrating mechanical properties, weavability, time and economic costs, and thermal insulation, exhibits superior comprehensive performance compared to skin-core, nanowood, and cellulose aerogel fibers (Fig. 4e).  
+
+To investigate the relationship between GAF- enhanced thermal insulating performance and nanostructures, we established simplified models: a uniform density model and a gradient density model to examine the temperature field spatial distribution during heat transfer (Supplementary Fig. S12) at simulated time \(10000\tau\) . Under a simulated temperature gradient of \(450\mathrm{K}\) (left) to \(300\mathrm{K}\) (right), the uniform density model (representing SAF) showed a smoothly varying temperature trend (Fig. 4f). Conversely, the gradient density model (representing GAF) exhibited a noticeable
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Fig. 4 | Investigation of thermal insulation properties and mechanism of GAFs. a Heat plate experiment with equal-weight fabrics: Infrared thermal images of Kevlar, SAF, and GAF fabrics on a heat source (a1); Temperature drop between the heat plate and Kevlar, SAF, and GAF fabrics (a2); b Heat plate experiment with 0.5 mm thick fabrics: Kevlar, SAF, and GAF; c Temperature-heating time curves for Kevlar, SAF, and GAF fabrics on a heat source; d Comparison of thermal conductivity of GAF with other natural and synthetic fibers<sup>28-30,33-37</sup>; e Radar plot comparing thermal insulating materials on mechanical properties, thermal insulation, weavability, cost, and processing efficiency; f Simulated temperature profiles at t = 10000τ for a uniform density model under identical heat source and sink conditions; g Simulated temperature profiles at t = 10000τ for a gradient density model under identical heat source and sink conditions; h Energy-time curves at heat sources and sinks for uniform density and gradient density models during simulation; i Simulated thermal conductivities (in Lennard-Jones unit) of SAF and GAF. </center>
+
+<--- Page Split --->
+
+temperature drop at the density interface (Fig. 4g), indicating greater thermal resistance at this interface. We calculated the energy accumulation of heat sources and sinks for both models. Fig. 4h shows two models as a function of time. At a given time, less energy is transmitted in the gradient density model due to the interface thermal resistance. Furthermore, the simulated thermal conductivity shows the contrast in energy supplied from heat sources and energy drained to heat sinks for the (Fig. 4i) of the uniform density structure was 11.7 (Lennard- Jones unit, or 0.054 W·m<sup>-1</sup>·K<sup>-1</sup> in real unit), while that of the gradient density structure was 7.0 (Lennard- Jones unit, or 0.032 W·m<sup>-1</sup>·K<sup>-1</sup> in real unit), whose magnitudes and decreasing trend agree with experimental measurements. Both the computational insights and the experimental results confirm the feasibility and necessity of gradient nanostructures in enhancing thermal barrier performance.  
+
+## Conclusions  
+
+In summary, we develop a gradient all- nanostructure aramid aerogel fiber (GAF) facilitated by microfluidics. The concurrent shear alignment and solvent diffusion within the microfluidic channels enable the precursor gel fibers to form a structure characterized by a sparse outer layer and a dense inner core, which undergoes a reversal during supercritical drying, resulting in sheath and core layers with average pore diameters of \(150 \mathrm{nm}\) and \(600 \mathrm{nm}\) , respectively. Both experimental data and simulations indicate that this gradient nanostructure induces significant interfacial thermal resistance at heat transfer interfaces, leading to a remarkably low radial thermal conductivity of \(0.0228 \mathrm{W m^{-1} K^{-1}}\) for GAF, significantly lower than that of air and wet- spun aerogel fibers. Additionally, the distinctive nano- entangled network within GAF effectively dissipates stress, yielding extraordinary tensile strength (29.5 MPa) and fracture strain (39.2%). This research demonstrates a link between gradient nanostructures and exceptional mechanical and thermal performance, thereby broadening the potential applications of high- performance aerogels in construction, transportation, and specialized textiles.
+
+<--- Page Split --->
+
+## 263 References  
+
+1. Zhong, X. et al. Global greenhouse gas emissions from residential and commercial building materials and mitigation strategies to 2060. Nat. Commun.12, 6126-6136 (2021).  
+2. Tian, X., Stranks, S. D. & You, F. Life cycle energy use and environmental implications of high-performance perovskite tandem solar cells. Sci. Adv. 6, eabb0055 (2020).  
+3. Hill, C., Norton, A. & Dibdiakova, J. A comparison of the environmental impacts of different categories of insulation materials. Energy Build. 162, 12-20 (2018).  
+4. Hua, S., Okello, M. O. & Zhang, J. Development of a vibration-damping, sound-insulating, and heat-insulating porous sphere foam system and its application in green buildings. Sci. Rep. 14, 14277-14299 (2024).  
+5. Wu, R. et al. Spectrally engineered textile for radiative cooling against urban heat islands. Science. 384, 1203-1212 (2024).  
+6. Li, Z. et al. Solution-shearing of dielectric polymer with high thermal conductivity and electric insulation. Sci. Adv. 7, eabi7410 (2021).  
+7. Zhao, X. et al. A scalable high-porosity wood for sound absorption and thermal insulation. Nat. Sustain. 6, 306-315 (2023).  
+8. Liu, H. & Zhao, X. P. Thermal conductivity analysis of high porosity structures with open and closed pores. Int. J. Heat Mass Transfer. 183, 122089-122100 (2022).  
+9. Zhao, X. P., Brozena, A. H. & Hu, L. B. Critical roles of pores and moisture in sustainable nanocellulose-based super-thermal insulators. Matter. 4, 769-772 (2021).  
+10. Ahankari, S., Paliwal, P., Subhedar, A. & Kargarzadeh, H. Recent developments in nanocellulose-based aerogels in thermal applications: a review. ACS Nano. 15, 3849-3874 (2021).  
+11. Tafreshi, O. A. et al. A review on multifunctional aerogel fibers: processing, fabrication, functionalization, and applications. Mater. Today Chem. 23, 100736-110761 (2022).  
+12. Liu, Q. et al. Recent advances in novel aerogels through the hybrid aggregation of inorganic nanomaterials and polymeric fibers for thermal insulation. Aggregate. 2, e30 (2021).  
+13. Li, C. et al. Silica aerogels: from materials research to industrial applications. Int. Mater. Rev. 68, 862-900 (2023).  
+14. Yang, B. et al. Fabrication, Applications, and Prospects of Aramid Nanofiber. Adv. Funct. Mater. 30, 2000186-2000214 (2020).  
+15. He, H. et al. Ultrastrong and multifunctional aerogels with hyperconnective network of composite polymeric nanofibers. Nat. Commun. 13, 4242-4250 (2022).  
+16. Zhao, H. Y. et al. Bioinspired Intelligent Solar-Responsive Thermally Conductive Pyramidal Phase Change Composites with Radially Oriented Layered Structures toward Efficient Solar–Thermal–Electric Energy Conversion. Adv. Funct. Mater. 33, 2302527-2302540 (2023).
+
+<--- Page Split --->
+
+17. Guo, Y., Ruan, K. & Gu, J. Controllable thermal conductivity in composites by constructing thermal conduction networks. Mater. Today Phys. 20, 100449-100458 (2021).  
+
+18. Shoaib, M. et al. Numerical investigation for rotating flow of MHD hybrid nanofluid with thermal radiation over a stretching sheet. Sci. Rep. 10, 18533-18447 (2020).  
+
+19. Tang, Y., Chen, W. & Zhu, P. Towards superior thermal insulation: Flexible Kevlar aerogel fibers with ultrafine size. Mater. Today Sustain. 27, 100902-100911 (2024).  
+
+20. Wei, P. et al. Reactive spinning to achieve nanocomposite gel fibers: from monomer to fiber dynamically with enhanced anisotropy. Mater. Horiz. 7, 811-819 (2020).  
+
+21. Hu, P. et al. Robust and Flame-Retardant Zylon Aerogel Fibers for Wearable Thermal Insulation and Sensing in Harsh Environment. Adv. Mater. 36, e2310023 (2024).  
+
+22. Sheng, Z. et al. The Rising Aerogel Fibers: Status, Challenges, and Opportunities. Adv. Sci. 10, e2205762 (2023).  
+
+23. Bao, Y., Lyu, J., Liu, Z., Ding, Y. & Zhang, X. Bending Stiffness-Directed Fabricating of Kevlar Aerogel-Confined Organic Phase-Change Fibers. ACS Nano. 15, 15180-15190 (2021).  
+
+24. Li, M. et al. Facile Preparation of Continuous and Porous Polyimide Aerogel Fibers for Multifunctional Applications. ACS Appl. Mater. Interfaces. 13, 10416-10427 (2021).  
+
+25. Sheng, Z. & Zhang, X. Mimicking polar bear hairs in aerogel fibers. Science. 382, 1358-1359 (2023).  
+
+26. Fu, X. et al. Aesthetic Cellulose Filaments with Water-Triggered Switchable Internal Stress and Customizable Polarized Iridescence Toward Green Fashion Innovation. ACS Nano. 18, 7496-7503 (2024).  
+
+27. Lin, Z. Y. & Dongdong. Research progress in structural design of biomass materials aided by Raman imaging technology based on spectral information. Basic Sci. J. Textile Univ. 37, 1-10 (2024).  
+
+28. Hu, Y. et al. Proton Donor-Regulated Mechanically Robust Aramid Nanofiber Aerogel Membranes for High-Temperature Thermal Insulation. ACS Nano. 16, 5984-5993 (2022).  
+
+29. Li, J. et al. Highly compressible, heat-insulating and self-extinguishing cellulose nanofiber/aramid nanofiber nanocomposite foams. Carbohydr. Polym. 261, 117837-117846 (2021).  
+
+30. Li, L. et al. Folk arts-inspired twice-coagulated configuration-editable tough aerogels enabled by transformable gel precursors. Nat. Commun. 14, 8450-8461 (2023).  
+
+31. Li, X., Dong, G., Liu, Z. & Zhang, X. Polyimide Aerogel Fibers with Superior Flame Resistance, Strength, Hydrophobicity, and Flexibility Made via a Universal Sol-Gel Confined Transition Strategy. ACS Nano. 15, 4759-4768 (2021).  
+
+32. Lyu, J., Sheng, Z., Xu, Y., Liu, C. & Zhang, X. Nanoporous Kevlar Aerogel Confined Phase Change Fluids Enable Super-Flexible Thermal Diodes. Adv. Funct. Mater. 32, 2200137-2200147 (2022).  
+
+33. Dos Santos, W. N., de Sousa, J. A. & Gregorio, R. Thermal conductivity behaviour of polymers around glass transition and crystalline melting temperatures. Polymer Testing. 32, 987-994 (2013).  
+
+34. Xue, T. et al. Polyimide Aerogel Fibers with Controllable Porous Microstructure for Super-Thermal Insulation Under Extreme Environments. Adv. Fiber Mater. 4, 1118-1128 (2022).
+
+<--- Page Split --->
+
+35. Jiang, Y. et al. Contrasting soil thermal responses to fire in Alaskan tundra and boreal forest. J. Geophys. Res.: Earth Surf. 120, 363-378 (2015).  
+
+36. Xue, Y., Lofland, S. & Hu, X. Thermal Conductivity of Protein-Based Materials: A Review. Polymers. 11, 456-464 (2019).  
+
+37. Tong, T. et al. Thermal conductivity of single silk fibroin fibers measured from the 3ω method. Int. J. Therm. Sci. 185, 108057-108064 (2023).  
+
+38. Kono, J., Goto, Y., Ostermeyer, Y., Frischknecht, R. & Wallbaum, H. Factors for eco-efficiency improvement of thermal insulation materials. Key Eng. Mater. 678, 1-13 (2016).  
+
+39. Uma, S., McConnell, A. D., Asheghi, M., Kurabayashi, K. & Goodson, K. E. Conduction and radiation in thermal analysis. Int. J. Thermophys. 22, 605-616 (2001).  
+
+40. Shintani, Y., Nagaoka, T., Deguchi, Y., Ido, K. & Harada, K. Simple method to predict downward heat flux from flame to floor. Fire Sci. Technol. 33, 17-34 (2014).  
+
+41. Si, L. et al. A superstable, flexible, and scalable nanofluidic ion regulation composite membrane. Sci. Bull. 68, 2344-2353 (2023).  
+
+42. Kremer, K. & Grest, G. S. Erratum: Dynamics of entangled polymer melts: A molecular-dynamics simulation. J. Chem. Phys. 94, 4103 (1991).  
+
+43. Thompson, A. P. et al. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Comput. Phys. Commun. 271, 108171-108205 (2022).  
+
+44. Stukowski, A. Visualization and analysis of atomistic simulation data with OVITO-the Open Visualization Tool. Modelling Simul. Mater. Sci. Eng. 18, 015012 (2010).  
+
+## Materials and Methods  
+
+Materials. Kevlar was purchased from Dupont company. Dimethyl sulfoxide (DMSO), sodium hydroxide (NaOH), acetic acid, and Ethanol were all purchased from Aladdin Biochemical Technology Co., Ltd (Shanghai, China). Deionized water (3.8 μS/cm) were prepared by the Millipore Milli-Q system. All other chemicals were utilized without additional purification.  
+
+Preparation of ANF Dispersion. The synthesis of ANF involved a deprotonation method<sup>41</sup>. In this procedure, Kevlar yarn (10 g) and KOH (15 g) were added to a mixture containing DMSO and water (500:25). The mixture was vigorously stirred for a week, resulting in a dark red dispersion of ANF.  
+
+Preparation of GAF. A customized microfluidic chip with two phases and three channels, which consisted of a core flow containing ANF solution and a sheath flow of DMSO, was utilized for the production of gel fibril. The flow rates were meticulously regulated, with the core flow operating at 80 μL/min and the sheath
+
+<--- Page Split --->
+
+flow ranging from 200 to \(1200\mu \mathrm{L / min}\) , achieved by employing two sets of flow pumps. The as- prepared gel fibril was wound onto a plastic roller at a constant speed of \(25\mathrm{rpm}\) . It was then subjected to an extensive washing procedure using deionized water to effectively remove any residual DMSO, KOH, and Acetic acid. Subsequent solvent exchange involved transferring the material into a blend of Ethanol for a duration of 24 h and repeated three times. The GAF was finally prepared by supercritical drying for \(48\mathrm{h}\) .  
+
+Characterization. The morphology of GAF and SAF samples, coated with gold vapor, was characterized using a field emission scanning electron microscope (SIGMA 500, Zeiss, Germany) operated at an accelerating voltage of \(5\mathrm{kV}\) . The pore size distribution in GAF was quantified using Image J software. Raman spectroscopic analysis, including Raman spectra and spatial Raman imaging, was conducted using a Renishaw inVia Raman spectrometer equipped with a \(532\mathrm{nm}\) laser. The concentration gradient of ANF was visualized through multivariate curve resolution, facilitated by supplementary software. Mechanical properties of GAF and SAF were evaluated using a nanomechanical actuating transducer (NMAT) integrated into a Universal Testing Machine (Nano UTM T150, KLA- Tencor, USA) at a strain rate of \(0.0027\mathrm{s}^{- 1}\) . Thermal insulation properties of the fabrics were assessed using a thermal infrared camera (InfReC R550, Avio, Japan) in an environment maintained at constant relative humidity ( \(35\%\) ) and controlled specific temperatures, achieved via a heating stage (LINKAM, LTS420, UK). The camera captured thermal images at a frequency of one per second, monitoring temperature changes across the test area at a rate of \(6^{\circ}\mathrm{C}\mathrm{s}^{- 1}\) . Thermal conductivity was measured using a thermal conductivity analyzer (HS- DR- 5, HESON Technology, China), employing a transient plane sensor with a \(6.4\mathrm{mm}\) radius positioned between two identical ANF aerogel fabrics. A compression device ensured optimal thermal contact between the sensor and the fabrics. The heating power was set at \(20\mathrm{mW}\) , with each measurement lasting \(10\mathrm{s}\) .  
+
+Coarse- Grained Molecular Dynamics Simulation. The effective thermal conductivity \(\Lambda \mathrm{e}\) of the thermal conduction region was calculated using the following formula:  
+
+\[\mathrm{q} = \Lambda_{\mathrm{e}}\frac{\Delta\mathrm{T}}{\Delta\mathrm{L}} = \frac{1}{2\mathrm{A}}\left(\left|\frac{\Delta\mathrm{E}_{\mathrm{Hot}}}{\Delta\mathrm{t}}\right| + \left|\frac{\Delta\mathrm{E}_{\mathrm{Cold}}}{\Delta\mathrm{t}}\right|\right)\]
+
+<--- Page Split --->
+
+where \(\mathbf{q}\) is the heat flux, A is the cross- sectional area perpendicular to the direction of heat conduction, \(\Delta \mathrm{t}\) is the time step, \(\mathrm{E}_{\mathrm{Source}}\) and \(\mathrm{E}_{\mathrm{Sink}}\) are respectively the heat added to the source and subtracted from the sink, \(\Delta \mathrm{L}\) is the length of the thermal conduction region and \(\Delta \mathrm{T}\) is the temperature difference in that region.  
+
+We utilize coarse- grained molecular dynamics simulations to validate the impact of pores on the mechanical properties of aramid fibers. The coarse- grained bead- spring \(\mathrm{KG}^{42}\) model is employed in our simulations. Lennard- Jones units are used. Energy, length, and mass units are respectively set as \(\epsilon\) , \(\sigma\) , and \(\mathrm{m}\) . The pair- wise interaction forces between beads are represented by the Lennard- Jones potential:  
+
+\[\mathrm{E}_{\mathrm{ij}} = 4 \epsilon_{\mathrm{ij}} \left[ \left(\frac{\sigma_{\mathrm{i}}}{\mathrm{r}}\right)^{12} - \left(\frac{\sigma_{\mathrm{i}}}{\mathrm{r}}\right)^{6} \right] - 4 \epsilon_{\mathrm{ij}} \left[ \left(\frac{\sigma_{\mathrm{i}}}{\mathrm{r_{\mathrm{c}}}}\right)^{12} - \left(\frac{\sigma_{\mathrm{i}}}{\mathrm{r_{\mathrm{c}}}}\right)^6 \right] \qquad r < \mathrm{r_{\mathrm{c}}} = 2.5,\]  
+
+Where \(\mathrm{r}\) is the distance between two beads, \(\mathrm{r}_{\mathrm{c}} = 2.5\) is the cutoff of interaction. \(\epsilon_{ij}\) is the potential strength and \(\sigma_{i}\) is the size of the bead. For intrachain pair- wise interaction forces, \(\epsilon_{ij} = 1\epsilon\) and \(\sigma_{i} = 1\sigma\) . For interchain pair- wise interaction forces, different \(\epsilon_{inter}\) values are used to tune the density of fibers. Arithmetic mixing is used for pair- wise interaction between core and skin. All \(\sigma_{i}\) values are uniformly set to \(1\sigma\) . The Unbreakable Finitely Extensible Nonlinear Elastic (FENE) potential is employed to describe the chemical bond interactions between particles.  
+
+\[\mathrm{U}_{\mathrm{FENE}}(\mathrm{r}) = -0.5\mathrm{KR}_{0}^{2}\ln \left[1 - \left(\frac{\mathrm{r}}{\mathrm{R}_{0}}\right)^{2}\right];\]  
+
+With \(R_{0} = 1.5\sigma\) is the maximum extent of the bond. \(\mathrm{K}\) is the strength of bond interaction, \(\mathrm{K} = 30 \epsilon /\sigma^{2}\) . The bond bending potential is harmonic style. \(U_{angle} = \frac{K_{\theta}}{2} (\theta - \theta_{0})^{2}\) with \(\mathrm{K}_{\theta} = 10\epsilon\) per radian \(^{2}\) is the strength of the interaction and \(\theta_{0} = 180^{\circ}\) is the equilibrium bond angle. To scale the simulation units to real units, Set the LJ temperature unit \(\mathrm{T}^{*} = 1\) to correspond to the real unit of \(300\mathrm{K}\) , so that \(1\epsilon = 4.14\mathrm{E}^{- 21}\mathrm{J}\) . Set two consecutive beads to represent a \(\mathrm{C}_{14}\mathrm{H}_{10}\mathrm{O}_{2}\mathrm{N}_{2}\) monomer (the chemical formula for each bead is \(\mathrm{C}_{7}\mathrm{H}_{5}\mathrm{ON}\) ), so that \(1\sigma = 6.57\mathrm{E}^{- 10}\mathrm{m}\) , and \(1\mathrm{m} = 0.11912\mathrm{kg / NA}\) where \(\mathrm{NA}\) is the avagardo number. The simulation time unit is \(\tau = \sqrt{\mathrm{m} \frac{\sigma^{2}}{\epsilon}}\) . The simulated thermal conductivity unit is \(\epsilon /(\tau \sigma \mathrm{T}^{*})\) . All simulations are conducted using
+
+<--- Page Split --->
+
+413 the LAMMPS<sup>43</sup> simulation package and visualization is performed using the OVITO<sup>44</sup> software package.  
+
+414 Model setup and simulation details for each model can be found in Supplementary Materials.  
+
+## Acknowledgements  
+
+417 This study was supported by the National Natural Science Foundation of China (Grant number. 52103124, 418 12322205, 12002304), the Distinguished Young Scientists Fund from the Natural Science Foundation of 419 Zhejiang Province (Grant number. LR23A020001), and the Pioneer R&D Program of Zhejiang (Grant 420 number. 2023C03007).
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SupplementaryMovie1.mp4  SupplementaryMovie2.mp4  SupplementaryMovie3.mp4  SupportingInformation.docx
+
+<--- Page Split --->

preprint/preprint__2af9ac54537a84d7e65fbe5670f2af99cc7c43c1b79b4d127a1a2be44b23db02/preprint__2af9ac54537a84d7e65fbe5670f2af99cc7c43c1b79b4d127a1a2be44b23db02_det.mmd

@@ -0,0 +1,355 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 853, 208]]<|/det|>
+# Gradient All-Nanostructured Aerogel Fibers for Enhanced Thermal Insulation and Mechanical Properties  
+
+<|ref|>text<|/ref|><|det|>[[44, 230, 212, 275]]<|/det|>
+Dongdong Ye ydd@whu.edu.cn  
+
+<|ref|>text<|/ref|><|det|>[[44, 302, 662, 322]]<|/det|>
+Anhui Agricultural University https://orcid.org/0000- 0002- 3377- 0656  
+
+<|ref|>text<|/ref|><|det|>[[44, 327, 303, 367]]<|/det|>
+Xiaotong Fu Anhui Agricultural University  
+
+<|ref|>text<|/ref|><|det|>[[44, 373, 270, 413]]<|/det|>
+Lianmeng Si Xi'an Jiaotong University  
+
+<|ref|>text<|/ref|><|det|>[[44, 419, 225, 459]]<|/det|>
+Zhaoxin Zhang Zhejiang University  
+
+<|ref|>text<|/ref|><|det|>[[44, 465, 303, 505]]<|/det|>
+Tingting Yang Anhui Agricultural University  
+
+<|ref|>text<|/ref|><|det|>[[44, 511, 321, 551]]<|/det|>
+Jianwei Song Xi'an Jiaotong University, Xi'an  
+
+<|ref|>text<|/ref|><|det|>[[44, 557, 584, 600]]<|/det|>
+Shuze Zhu Zhejiang University https://orcid.org/0000- 0001- 7849- 6067  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 643, 103, 660]]<|/det|>
+## Article  
+
+<|ref|>title<|/ref|><|det|>[[44, 680, 135, 698]]<|/det|>
+# Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[44, 717, 321, 737]]<|/det|>
+Posted Date: August 29th, 2024  
+
+<|ref|>text<|/ref|><|det|>[[44, 755, 475, 775]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 4912597/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 792, 914, 835]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[44, 853, 535, 872]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 908, 930, 951]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on March 10th, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 57646- 4.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[163, 89, 834, 142]]<|/det|>
+# Gradient All-Nanostructured Aerogel Fibers for Enhanced Thermal Insulation and Mechanical Properties  
+
+<|ref|>text<|/ref|><|det|>[[112, 165, 875, 213]]<|/det|>
+Xiaotong \(\mathrm{Fu}^{1,4}\) , Lianmeng \(\mathrm{Si}^{2,4}\) , Zhaoxin Zhang \(^{3}\) , Tingting Yang \(^{1}\) , Jianwei Song \(^{*2}\) , Shuze Zhu \(^{*3}\) , Dongdong \(\mathrm{Ye}^{*1}\)  
+
+<|ref|>text<|/ref|><|det|>[[112, 237, 883, 283]]<|/det|>
+\(^{1}\) School of Materials and Chemistry, Anhui Agricultural University, Hefei, Anhui Province 230036, China  
+
+<|ref|>text<|/ref|><|det|>[[112, 296, 883, 344]]<|/det|>
+\(^{2}\) State Key Laboratory for Strength and Vibration of Mechanical Structures, School of Aerospace Engineering, Xi'an Jiaotong University, Xi'an, 710049, China  
+
+<|ref|>text<|/ref|><|det|>[[112, 357, 883, 404]]<|/det|>
+\(^{3}\) Department of Engineering Mechanics, Institute of Applied Mechanics, Zhejiang University, Hangzhou 310000, China  
+
+<|ref|>text<|/ref|><|det|>[[112, 417, 492, 437]]<|/det|>
+\(^{4}\) These authors contributed equally to this work  
+
+<|ref|>text<|/ref|><|det|>[[112, 458, 870, 479]]<|/det|>
+Corresponding authors' emails: ydd@whu.edu.cn; shuezhu@zju.edu.cn; songjianwei@xjtue.edu.cn  
+
+<|ref|>text<|/ref|><|det|>[[111, 504, 884, 894]]<|/det|>
+Abstract: Lightweight, nanoporous aerogel fibers are crucial for personal thermal management and specialized heat protection. However, wet-spinning methods, exemplified by aramid aerogels, inevitably form a dense outer layer, significantly reducing the volume fraction of efficient thermal barrier nanovoids and limiting the development of ultimate thermal resistance in fibers. Herein, we develop a microfluidic spinning method to prepare gradient all-nanostructure aramid aerogel fibers (GAFs). Benefiting from the simultaneous shear alignment and diffusion dilution of a good solvent within the channels, the precursor gel fibers assemble into a structure with a sparse exterior and dense interior, which reverses during supercritical drying to form sheath and core layers with average pore diameters of 150 nm and 600 nm, respectively. Experiments and simulations reveal that the gradient nanostructure creates high interfacial thermal resistance at heat transfer interfaces, resulting in a GAF radial thermal conductivity as low as 0.0228 W m \(^{-1}\) K \(^{-1}\) , far below that of air and wet-spun aerogel fibers. Moreover, GAF's unique nano-entangled network efficiently dissipates stress, achieving exceptionally high tensile strength (29.5 MPa) and fracture strain (39.2%). This work establishes a correlation between multiscale nanostructures and superlative performance, thereby expanding the scope of aerogel applications in intricate environments.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[113, 90, 225, 108]]<|/det|>
+## Introduction  
+
+<|ref|>text<|/ref|><|det|>[[111, 115, 885, 664]]<|/det|>
+Thermal management systems within buildings and vehicles represent a substantial portion of global energy consumption, accounting for more than one- third of the total \(^{1}\) . This significant energy expenditure notably contributes to \(\mathrm{CO_2}\) emissions, thereby intensifying global carbon footprint \(^{2,3}\) . Developing advanced insulation materials presents a viable solution to mitigate heat loss in various applications, including building walls \(^{4}\) , roofs, vehicle exteriors, and apparel \(^{5}\) , thereby enhancing user comfort and curbing carbon emissions \(^{6}\) . Actually, materials exhibiting high porosity \((>0.9)\) are particularly effective in enhancing thermal insulation by limiting collisions between air molecules, elongating heat transfer pathways, and augmenting interface thermal resistance \(^{7- 9}\) . Distinguished by their abundant nanoporous structures, aerogels can be transformed into fibrous forms to preserve superior thermal insulation properties while attaining flexibility and fabricability, making them ideal for personal thermal management textiles. Currently, inorganic, polymeric, and composite aerogel fibers, including those made from silica, polyimide, cellulose, and hybrid materials, have shown promising insulation characteristics \(^{10- 13}\) . Nonetheless, challenges remain, such as inadequate mechanical strength, limited thermal insulation performance, and elevated manufacturing costs, primarily due to the inconsistent control over micro- and nanoscale structures.  
+
+<|ref|>text<|/ref|><|det|>[[112, 675, 885, 905]]<|/det|>
+Aramid nanofibers (ANFs) are regarded as ideal nanostructured units due to their inherent high strength, environmental resilience, active sites, and excellent dispersibility, derived from polyphenylene terephthalate (PPTA) \(^{14,15}\) . Aerogels assembled by ANFs exhibit outstanding thermal insulation properties to withstand extreme temperatures through three mechanisms: 1) reduced thermal conduction via the solid skeleton \(^{16}\) , 2) restricted thermal convection via the porous structure \(^{17}\) , and 3) multiple thermal radiation \(^{18}\) . Currently, representative aerogel fiber manufacturing methods such as wet spinning \(^{19}\) , reactive spinning \(^{20}\) , and confined spinning \(^{21}\) , while
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 884, 318]]<|/det|>
+enhancing scalability, inevitably reduce the precision of nanostructure control. This results in fibers exhibiting a dense skin layer on the exterior and a porous interior<sup>22,23</sup>, along with a wide range of thermal conductivity (0.027- 0.5 W m<sup>- 1</sup> K<sup>- 1</sup>)<sup>24</sup>. Previous researches indicate that the nanopore structure of aerogels significantly affects thermal conduction; transforming the dense packing structure of the shell layer into a loose nanostructure and optimizing the radial nanostructure distribution is expected to enhance the thermal insulation capabilities of aerogel fibers.  
+
+<|ref|>text<|/ref|><|det|>[[111, 333, 885, 878]]<|/det|>
+In this study, we unveil the development of gradient nanostructured aerogel fibers (GAFs), which exhibit exceptional thermal insulation properties, as corroborated by molecular simulations. Through the design of skin- core and gradient nanostructure models, we demonstrate that the outer- layer nanostructure of GAFs manifests a significantly superior thermal storage capacity and interfacial thermal resistance compared to skin- core aerogel fibers (SAFs), resulting in a \(67\%\) reduction in simulated thermal conductivity. Moreover, we have pioneered a manufacturing technique for producing GAFs using continuous microfluidic spinning technology. Coupled with supercritical drying, we have successfully fabricated aramid nanofiber aerogel fibers by precisely controlling solvent shear, diffusion, and sol- gel transition rates within the spinning channel. Compared to wet- spun aerogel fibers, GAFs exhibit increased porosity (from \(98\%\) to \(98.6\%\) ), reduced density (from \(20.5 \mathrm{kg / m}^3\) to \(15.7 \mathrm{kg / m}^3\) , enhanced tensile strength (from \(10.9 \mathrm{MPa}\) to \(29.5 \mathrm{MPa}\) ), and significantly reduced thermal conductivity (from \(0.0327 \mathrm{W m}^{- 1} \mathrm{K}^{- 1}\) to \(0.0228 \mathrm{W m}^{- 1} \mathrm{K}^{- 1}\) ). Notably, the superior durability of aramid ensures stability under extreme cold, high temperatures, and vacuum conditions. The excellent comprehensive performance of GAF significantly broadens its potential applications in textiles, construction, firefighting equipment, and even extreme aerospace scenarios, offering significant environmental benefits.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[113, 90, 341, 111]]<|/det|>
+## Results and Discussion  
+
+<|ref|>text<|/ref|><|det|>[[110, 116, 886, 775]]<|/det|>
+The thermal insulation mechanisms of aerogel materials, encompassing multiple reflected radiation, reduced conduction, and restricted convection, are intricately linked to pore size and spatial distribution. Intriguingly, natural animal hair, exemplified by polar bears<sup>25</sup>, and synthetic aerogel fibers, typically fabricated through wet spinning, exhibit dense outer layers, porous inner cores, and superior thermal insulation performance. To investigate the correlation between the nanostructure and thermal insulation, we employed coarse- grained molecular dynamics simulations under the same conditions to assess the radial heat transfer properties of two fibrous structural models with skin- core and gradient nanostructures (Fig. 1a; Supplementary Fig. S1). The results reveal that the two nanostructured fibers exhibit markedly different thermal distribution characteristics under the imposed flux from heat source with temperature 2 (Lennard- Jones unit) to heat sink with temperature 1 (Lennard- Jones unit). Unlike the skin- core fibers, whose dense outer layer nanostructure results in low interfacial thermal resistance, the gradient nanostructured fibers possess a loose nanoporous outer layer, which demonstrates high heat storage capacity and forms substantial interfacial thermal resistance at the gradient interface, thereby restricting radial heat transfer. Furthermore, quantitative simulations of the dimensionless thermal conductivity for both fiber types indicate a \(40\%\) reduction in thermal conductivity for the gradient nanostructured fibers compared to the skin- core fibers. These findings provide theoretical evidence that regulating aerogels' radial nanostructure size and spatial distribution can significantly enhance thermal insulation performance.  
+
+<|ref|>text<|/ref|><|det|>[[113, 783, 884, 874]]<|/det|>
+Guided with the simulation, we developed the gradient all- nanostructured aramid aerogel fibers (GAFs) from aramid nanofibers (ANFs) dispersion based on a spinning system integrated with nanoexfoliation, microfluidic spinning, sol- gel transition, and supercritical \(\mathrm{CO_2}\) drying
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 87, 880, 536]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 556, 883, 682]]<|/det|>
+<center>Fig. 1 | Thermal simulation and preparation of gradient nanostructured aerogel fibers (GAFs). a Coarse-Grained molecular dynamics simulations of thermal insulation for skin-core \((\mathbf{a}_1)\) and gradient nanostructured \((\mathbf{a}_2)\) models, with the radial temperature distribution map within the region marked by dashed box placed at the bottom, and the simulated thermal conductivities (in Lennard-Jones unit) through the gradient interface of the gradient nanostructured model and the bulk of the skin-core model on the right \((\mathbf{a}_3)\) ; b Sketch map for the preparation of GAFs, including nanoexfoliation, microfluidic spinning, sol-gel translation, and supercritical drying; c, d Representative SEM images of skin-core and gradient nanostructured aerogel fibers; e, f). Scalability of GAFs and textiles. </center>  
+
+<|ref|>text<|/ref|><|det|>[[112, 696, 884, 856]]<|/det|>
+process (Fig. 1b). Contrary to the skin- core aerogel fibers (SAFs) with dense shell and porous core configuration engendered by wet spinning (Fig. 1c), GAF is devoid of a dense exterior skin layer (Fig. 1d). ANF interweaves to forge a wholly nanoporous structure, congruent with the theoretical blueprint and rendering it an optimal candidate for thermal insulation. Additionally, the spinning system demonstrates the feasibility for large- scale fiber production (Fig. 1e,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 883, 144]]<|/det|>
+Supplementary Movie 1), which can be further woven into fabrics (Figure 1f), owing to improved mechanical properties (discussed in detail later).  
+
+<|ref|>text<|/ref|><|det|>[[111, 158, 884, 599]]<|/det|>
+The formation of GAFs undergoes three key stages (Fig. 2a): 1) Heterogeneous spatial distribution of ANF induced by complex flow in a microfluidic chip: When the sheath flow of DMSO solvent is injected at an angle into the central channel containing ANF dispersion (DMSO/KOH), it not only enhances the shear orientation of ANF but also dilutes the ANF solution from the outside in, creating a concentration gradient with lower concentration at the exterior and higher concentration at the center due to shear orientation along the channel and diffusion dilution perpendicular to the channel; 2) Sol- gel transition induced by protonation: Upon entering the acidic coagulation bath, the ANF undergoes protonation, rapidly triggering gelation and freezing the gradient structure, resulting in gel fibers with loose pores on the exterior and smaller pores on the interior (Supplementary Fig. S2); 3) Structural inversion during supercritical drying: During the solvent exchange process to form alcohol gel fibers, the large pores with high surface porosity collapse into smaller pores (Supplementary Figs. S3), leading to the formation of gradient all- nanostructured aerogel fibers after supercritical \(\mathrm{CO_2}\) drying (Fig. 2b).  
+
+<|ref|>text<|/ref|><|det|>[[111, 610, 884, 876]]<|/det|>
+We subsequently compared the structural assembly differences between SAFs produced by wet- spinning and GAFs constructed via microfluidic spinning (Fig. 2c). SAF exhibited a nanoporous core (Fig. 2c1) and a distinct dense skin layer approximately \(100 \mathrm{nm}\) thick (Fig. 2c2, 2c3), as confirmed by cross- sectional SEM images. The GAF core displays a similar nanoporous structure to SAF (Fig. 2c4), with an average pore size of about \(500 \mathrm{nm}\) (Supplementary Fig. S4), while the outer layer maintains a nanoporous structure with an average pore size of \(160 \mathrm{nm}\) (Fig. 2c5; Supplementary Fig. S5) and shows a pronounced gradient interfacial layer (Fig. 2c6). Notably, altering the flow rate of DMSO from \(1200 \mathrm{to} 200 \mu \mathrm{L} / \mathrm{min}\) enables the regulation of the gradient
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 88, 880, 590]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 598, 884, 712]]<|/det|>
+<center>Fig. 2 | Forming mechanism and morphological characterization of GAFs. a Schematic illustration of nanostructure changes during GAFs preparation, including ANF dispersion in the microfluidic system, sol-gel transition in the coagulation bath, and ethanol solvent exchange; b Photograph of GAFs; c Cross-sectional SEM images of SAF and GAF at various locations during wet-spinning and microfluidic spinning, including center positions (c1, c4), edge positions (c2, c3, and c5), and gradient interface position (c6); d Schematic of Raman line imaging along the diameter (d1), with representative Raman spectra selected at equal intervals from the exterior to the interior of SAF (d2) and GAF (d3). </center>  
+
+<|ref|>text<|/ref|><|det|>[[113, 725, 883, 815]]<|/det|>
+nanoporous outer layer structure (from dense to loose) and thickness (from 1.5 to \(35 \mu \mathrm{m}\) ), showcasing the excellent customizable structural capabilities of the microfluidic spinning system (Supplementary Fig. S6).  
+
+<|ref|>text<|/ref|><|det|>[[113, 829, 883, 885]]<|/det|>
+Furthermore, we characterized the variation in the intensity of the characteristic peak of the aramid I band (peak position at \(1610 \mathrm{cm}^{- 1}\) ) along the diameter of the aerogel fibers (Fig. 2d1),
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 87, 884, 319]]<|/det|>
+which visually illustrates the differences in ANF packing density between the two types of aerogel fibers. The results indicate that the Raman scattering intensity on the surface of SAF is significantly higher than that within the fiber (Fig. 2d₂), suggesting the presence of a highly compact thin layer on the surface. In contrast, the GAF shows a uniformly decreasing scattering intensity from the exterior to the interior (Fig. 2d₃), indicating that the packing density outside the fiber is slightly higher than that inside, exhibiting a gradient transition. Raman imaging further confirms the distinct assembly structures of the two fiber types.  
+
+<|ref|>text<|/ref|><|det|>[[110, 333, 886, 844]]<|/det|>
+To elucidate the impact of nanostructures on the properties of aerogel fibers, we initially assessed the mechanical properties of SAFs and GAFs. The results indicate that SAF with a dense skin layer exhibits a higher Young's modulus, reaching 112.74 MPa, a 37.7% increase over GAF (81.9 MPa) (Fig. 3a; Supplementary Fig. S7). However, SAF's tensile strength (10.9 MPa), fracture strain (17.8%), and toughness (1.06 MJ m⁻³) are significantly lower than those of GAF (29.5 MPa, 39.2%, and 5.7 MJ m⁻³), demonstrating the potential of gradient nanostructures in enhancing aerogel mechanical properties. Notably, GAFs developed in this work show superior mechanical properties compared to classical ANF- based fibers²⁹, ³⁰ and films³¹ (Fig. 3b). As a performance demonstration, a single GAF can bear a weight of 50 g (Fig. 3c). To explore the deeper mechanisms of gradient nanostructure reinforcement and toughening, we employed molecular dynamics simulations to evaluate the deformation processes of SAF and GAF (Supplementary Fig. S8). The simulations revealed distinct fracture behaviors: as the simulated strain (ε) increased from 2 to 5, due to the small thickness of the dense nanofiber layer on the surface of SAF, at microscopic large deformation the dense layer can no longer maintain structure integrity, leading to micro- crack formation that exposes the internal nanostructure (Fig. 3d₁, Supplementary Movie 2); in contrast,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[57, 90, 883, 110]]<|/det|>
+171 GAF maintained structural integrity during stretching, with outer nanopores aligning and  
+
+<|ref|>text<|/ref|><|det|>[[57, 126, 390, 144]]<|/det|>
+172 deforming with the fiber, avoiding  
+
+<|ref|>image<|/ref|><|det|>[[115, 160, 880, 730]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 740, 883, 853]]<|/det|>
+<center>Fig. 3 | Study on the mechanical properties and deformation mechanisms of GAF. a Stress-strain curves of SAF and GAF; b Comparison of toughness and tensile strength based on ANF aerogel materials<sup>28</sup>-<sup>32</sup>; c Photograph of a single GAF suspending a \(50 \mathrm{g}\) weight; d Simulation trajectories (d1) and simulated stress-strain curves (d2) for SAF and GAF models during tensile processes; e Raman imaging technique detecting internal stress levels in poly(p-phenylene terephthalate) based on characteristic Raman shift of C=O group (e1) and the representative 2D Raman images and Raman shifts of Kevlar fiber (e2, e3), SAF (e3, e6), and GAF (e4, e7) under \(5\%\) tensile strain. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 884, 144]]<|/det|>
+severe damage and retaining good mechanical properties (Fig. 3d₁, Supplementary Movie 3). The simulated stress- strain curves also reflected this trend (Fig. 3d₂).  
+
+<|ref|>text<|/ref|><|det|>[[111, 157, 885, 739]]<|/det|>
+Previous researches indicate that the specific functional group deformation under external force induces a redshift (Δv) in characteristic Raman peaks \(^{26,27}\) . Thus, we employed Raman imaging to investigate the relationship between the Raman shifts of the amide I band C=O stretching vibration in 5% deformed Kevlar fiber, SAF, GAF and even isotropic ANF- assembled membrane (Supplementary Fig. S9), aiming to elucidate the correlation between nanostructure and internal stress during deformation. We first tested the Raman spectra of isotropic ANF- assembled membrane prepared by the vacuum filtration method, confirming that the characteristic Raman shift of C=O in the relaxed state of PPTA molecular chains is 1656.5 cm\(^{-1}\) (Supplementary Fig. S9a). Comparing the 2D Raman imaging of the three types of fibers, we found that the Kevlar fiber mainly exhibits a low Raman shift in the red distribution (Fig. 3e₂; Supplementary Fig. S9b), with an average Δv of 5.6 cm\(^{-1}\) (Fig. 3e₅). In contrast, SAF shows significant stress concentration only in the densely packed skin layer (Fig. 3e₃; Supplementary Fig. S9c), with a Δv of 1.2 cm\(^{-1}\) (Fig. 3e₆), leading to initial high modulus and outer layer failure due to high internal stress during stretching. Interestingly, GAF shows the smallest overall Raman red shift (Fig. 3e₄; Supplementary Fig. S9d), with a Δv of only 0.3 cm\(^{-1}\) (Fig. 3e₇), indicating that the nanostructure effectively dissipates stress during stretching, resulting in the lowest internal stress and superior mechanical properties.  
+
+<|ref|>text<|/ref|><|det|>[[111, 765, 884, 890]]<|/det|>
+Subsequently, we evaluated the thermal insulation properties of skin- core and gradient- structured aerogel fibers and fabrics. Initially, we monitored the temperature changes atop a hot stage using an infrared thermal imager for both SAF and GAF specimens (Supplementary Fig. S10). The results indicated that when the hot stage was set at 200 °C, the top temperature of the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 886, 670]]<|/det|>
+GAF sample was only \(190.3^{\circ}\mathrm{C}\) , significantly lower than that of the SAF specimen (193.5 \(^{\circ}\mathrm{C}\) ), demonstrating superior insulation capabilities. Further, we weaved the two aerogel fibers into fabrics and compared their thermal insulation properties under the same weight or thickness conditions (Fig. 4a). The findings revealed enhanced performance across various metrics (Figs. 4b, c). Additionally, we examined the thermal insulation properties of GAF fabrics of varying thicknesses across a broad temperature range (- 30 \(^{\circ}\mathrm{C}\) to 270 \(^{\circ}\mathrm{C}\) ) (Supplementary Figs. S11a, c). The results showed that a 0.5 mm GAF fabric could achieve temperature differentials of 33.4 \(^{\circ}\mathrm{C}\) and 63 \(^{\circ}\mathrm{C}\) at cold and heat templates, respectively (Supplementary Figs. S11b, d); when the fabric thickness increased to 1mm, the insulation effect improved, yielding temperature differentials of 45.4 \(^{\circ}\mathrm{C}\) and 113 \(^{\circ}\mathrm{C}\) at cold and heat templates, respectively (Supplementary Figs. S11b, d). Notably, the insulation effect of GAF fibers based on a microfluidic system surpassed that of most previously reported thermal insulation fibers, including synthetic fibers \(^{33,34}\) (e.g., PVC, PI, porous PI, PS, PET), natural fibers \(^{7,34 - 37}\) (e.g., silk, wood, wool, cotton, cotton aerogel, TCNF), inorganic fibers \(^{38 - 40}\) (e.g., glass, rock wool, ceramic, silicon), and ANF aerogels \(^{28 - 30}\) (KNF fibers, KNF films, ANF films) (Fig. 4d). Therefore, GAF, integrating mechanical properties, weavability, time and economic costs, and thermal insulation, exhibits superior comprehensive performance compared to skin-core, nanowood, and cellulose aerogel fibers (Fig. 4e).  
+
+<|ref|>text<|/ref|><|det|>[[111, 682, 886, 876]]<|/det|>
+To investigate the relationship between GAF- enhanced thermal insulating performance and nanostructures, we established simplified models: a uniform density model and a gradient density model to examine the temperature field spatial distribution during heat transfer (Supplementary Fig. S12) at simulated time \(10000\tau\) . Under a simulated temperature gradient of \(450\mathrm{K}\) (left) to \(300\mathrm{K}\) (right), the uniform density model (representing SAF) showed a smoothly varying temperature trend (Fig. 4f). Conversely, the gradient density model (representing GAF) exhibited a noticeable
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 90, 880, 700]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 710, 883, 888]]<|/det|>
+<center>Fig. 4 | Investigation of thermal insulation properties and mechanism of GAFs. a Heat plate experiment with equal-weight fabrics: Infrared thermal images of Kevlar, SAF, and GAF fabrics on a heat source (a1); Temperature drop between the heat plate and Kevlar, SAF, and GAF fabrics (a2); b Heat plate experiment with 0.5 mm thick fabrics: Kevlar, SAF, and GAF; c Temperature-heating time curves for Kevlar, SAF, and GAF fabrics on a heat source; d Comparison of thermal conductivity of GAF with other natural and synthetic fibers<sup>28-30,33-37</sup>; e Radar plot comparing thermal insulating materials on mechanical properties, thermal insulation, weavability, cost, and processing efficiency; f Simulated temperature profiles at t = 10000τ for a uniform density model under identical heat source and sink conditions; g Simulated temperature profiles at t = 10000τ for a gradient density model under identical heat source and sink conditions; h Energy-time curves at heat sources and sinks for uniform density and gradient density models during simulation; i Simulated thermal conductivities (in Lennard-Jones unit) of SAF and GAF. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 885, 423]]<|/det|>
+temperature drop at the density interface (Fig. 4g), indicating greater thermal resistance at this interface. We calculated the energy accumulation of heat sources and sinks for both models. Fig. 4h shows two models as a function of time. At a given time, less energy is transmitted in the gradient density model due to the interface thermal resistance. Furthermore, the simulated thermal conductivity shows the contrast in energy supplied from heat sources and energy drained to heat sinks for the (Fig. 4i) of the uniform density structure was 11.7 (Lennard- Jones unit, or 0.054 W·m<sup>-1</sup>·K<sup>-1</sup> in real unit), while that of the gradient density structure was 7.0 (Lennard- Jones unit, or 0.032 W·m<sup>-1</sup>·K<sup>-1</sup> in real unit), whose magnitudes and decreasing trend agree with experimental measurements. Both the computational insights and the experimental results confirm the feasibility and necessity of gradient nanostructures in enhancing thermal barrier performance.  
+
+<|ref|>sub_title<|/ref|><|det|>[[114, 438, 220, 456]]<|/det|>
+## Conclusions  
+
+<|ref|>text<|/ref|><|det|>[[111, 470, 885, 910]]<|/det|>
+In summary, we develop a gradient all- nanostructure aramid aerogel fiber (GAF) facilitated by microfluidics. The concurrent shear alignment and solvent diffusion within the microfluidic channels enable the precursor gel fibers to form a structure characterized by a sparse outer layer and a dense inner core, which undergoes a reversal during supercritical drying, resulting in sheath and core layers with average pore diameters of \(150 \mathrm{nm}\) and \(600 \mathrm{nm}\) , respectively. Both experimental data and simulations indicate that this gradient nanostructure induces significant interfacial thermal resistance at heat transfer interfaces, leading to a remarkably low radial thermal conductivity of \(0.0228 \mathrm{W m^{-1} K^{-1}}\) for GAF, significantly lower than that of air and wet- spun aerogel fibers. Additionally, the distinctive nano- entangled network within GAF effectively dissipates stress, yielding extraordinary tensile strength (29.5 MPa) and fracture strain (39.2%). This research demonstrates a link between gradient nanostructures and exceptional mechanical and thermal performance, thereby broadening the potential applications of high- performance aerogels in construction, transportation, and specialized textiles.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[63, 90, 225, 110]]<|/det|>
+## 263 References  
+
+<|ref|>text<|/ref|><|det|>[[58, 123, 888, 870]]<|/det|>
+1. Zhong, X. et al. Global greenhouse gas emissions from residential and commercial building materials and mitigation strategies to 2060. Nat. Commun.12, 6126-6136 (2021).  
+2. Tian, X., Stranks, S. D. & You, F. Life cycle energy use and environmental implications of high-performance perovskite tandem solar cells. Sci. Adv. 6, eabb0055 (2020).  
+3. Hill, C., Norton, A. & Dibdiakova, J. A comparison of the environmental impacts of different categories of insulation materials. Energy Build. 162, 12-20 (2018).  
+4. Hua, S., Okello, M. O. & Zhang, J. Development of a vibration-damping, sound-insulating, and heat-insulating porous sphere foam system and its application in green buildings. Sci. Rep. 14, 14277-14299 (2024).  
+5. Wu, R. et al. Spectrally engineered textile for radiative cooling against urban heat islands. Science. 384, 1203-1212 (2024).  
+6. Li, Z. et al. Solution-shearing of dielectric polymer with high thermal conductivity and electric insulation. Sci. Adv. 7, eabi7410 (2021).  
+7. Zhao, X. et al. A scalable high-porosity wood for sound absorption and thermal insulation. Nat. Sustain. 6, 306-315 (2023).  
+8. Liu, H. & Zhao, X. P. Thermal conductivity analysis of high porosity structures with open and closed pores. Int. J. Heat Mass Transfer. 183, 122089-122100 (2022).  
+9. Zhao, X. P., Brozena, A. H. & Hu, L. B. Critical roles of pores and moisture in sustainable nanocellulose-based super-thermal insulators. Matter. 4, 769-772 (2021).  
+10. Ahankari, S., Paliwal, P., Subhedar, A. & Kargarzadeh, H. Recent developments in nanocellulose-based aerogels in thermal applications: a review. ACS Nano. 15, 3849-3874 (2021).  
+11. Tafreshi, O. A. et al. A review on multifunctional aerogel fibers: processing, fabrication, functionalization, and applications. Mater. Today Chem. 23, 100736-110761 (2022).  
+12. Liu, Q. et al. Recent advances in novel aerogels through the hybrid aggregation of inorganic nanomaterials and polymeric fibers for thermal insulation. Aggregate. 2, e30 (2021).  
+13. Li, C. et al. Silica aerogels: from materials research to industrial applications. Int. Mater. Rev. 68, 862-900 (2023).  
+14. Yang, B. et al. Fabrication, Applications, and Prospects of Aramid Nanofiber. Adv. Funct. Mater. 30, 2000186-2000214 (2020).  
+15. He, H. et al. Ultrastrong and multifunctional aerogels with hyperconnective network of composite polymeric nanofibers. Nat. Commun. 13, 4242-4250 (2022).  
+16. Zhao, H. Y. et al. Bioinspired Intelligent Solar-Responsive Thermally Conductive Pyramidal Phase Change Composites with Radially Oriented Layered Structures toward Efficient Solar–Thermal–Electric Energy Conversion. Adv. Funct. Mater. 33, 2302527-2302540 (2023).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[57, 90, 884, 127]]<|/det|>
+17. Guo, Y., Ruan, K. & Gu, J. Controllable thermal conductivity in composites by constructing thermal conduction networks. Mater. Today Phys. 20, 100449-100458 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[57, 131, 884, 170]]<|/det|>
+18. Shoaib, M. et al. Numerical investigation for rotating flow of MHD hybrid nanofluid with thermal radiation over a stretching sheet. Sci. Rep. 10, 18533-18447 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[57, 175, 884, 214]]<|/det|>
+19. Tang, Y., Chen, W. & Zhu, P. Towards superior thermal insulation: Flexible Kevlar aerogel fibers with ultrafine size. Mater. Today Sustain. 27, 100902-100911 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[57, 220, 884, 259]]<|/det|>
+20. Wei, P. et al. Reactive spinning to achieve nanocomposite gel fibers: from monomer to fiber dynamically with enhanced anisotropy. Mater. Horiz. 7, 811-819 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[57, 263, 884, 303]]<|/det|>
+21. Hu, P. et al. Robust and Flame-Retardant Zylon Aerogel Fibers for Wearable Thermal Insulation and Sensing in Harsh Environment. Adv. Mater. 36, e2310023 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[57, 308, 884, 346]]<|/det|>
+22. Sheng, Z. et al. The Rising Aerogel Fibers: Status, Challenges, and Opportunities. Adv. Sci. 10, e2205762 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[57, 351, 884, 391]]<|/det|>
+23. Bao, Y., Lyu, J., Liu, Z., Ding, Y. & Zhang, X. Bending Stiffness-Directed Fabricating of Kevlar Aerogel-Confined Organic Phase-Change Fibers. ACS Nano. 15, 15180-15190 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[57, 395, 884, 435]]<|/det|>
+24. Li, M. et al. Facile Preparation of Continuous and Porous Polyimide Aerogel Fibers for Multifunctional Applications. ACS Appl. Mater. Interfaces. 13, 10416-10427 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[57, 439, 884, 459]]<|/det|>
+25. Sheng, Z. & Zhang, X. Mimicking polar bear hairs in aerogel fibers. Science. 382, 1358-1359 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[57, 465, 884, 523]]<|/det|>
+26. Fu, X. et al. Aesthetic Cellulose Filaments with Water-Triggered Switchable Internal Stress and Customizable Polarized Iridescence Toward Green Fashion Innovation. ACS Nano. 18, 7496-7503 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[57, 528, 884, 568]]<|/det|>
+27. Lin, Z. Y. & Dongdong. Research progress in structural design of biomass materials aided by Raman imaging technology based on spectral information. Basic Sci. J. Textile Univ. 37, 1-10 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[57, 572, 884, 612]]<|/det|>
+28. Hu, Y. et al. Proton Donor-Regulated Mechanically Robust Aramid Nanofiber Aerogel Membranes for High-Temperature Thermal Insulation. ACS Nano. 16, 5984-5993 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[57, 617, 884, 657]]<|/det|>
+29. Li, J. et al. Highly compressible, heat-insulating and self-extinguishing cellulose nanofiber/aramid nanofiber nanocomposite foams. Carbohydr. Polym. 261, 117837-117846 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[57, 661, 884, 701]]<|/det|>
+30. Li, L. et al. Folk arts-inspired twice-coagulated configuration-editable tough aerogels enabled by transformable gel precursors. Nat. Commun. 14, 8450-8461 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[57, 705, 884, 765]]<|/det|>
+31. Li, X., Dong, G., Liu, Z. & Zhang, X. Polyimide Aerogel Fibers with Superior Flame Resistance, Strength, Hydrophobicity, and Flexibility Made via a Universal Sol-Gel Confined Transition Strategy. ACS Nano. 15, 4759-4768 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[57, 770, 884, 810]]<|/det|>
+32. Lyu, J., Sheng, Z., Xu, Y., Liu, C. & Zhang, X. Nanoporous Kevlar Aerogel Confined Phase Change Fluids Enable Super-Flexible Thermal Diodes. Adv. Funct. Mater. 32, 2200137-2200147 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[57, 814, 884, 854]]<|/det|>
+33. Dos Santos, W. N., de Sousa, J. A. & Gregorio, R. Thermal conductivity behaviour of polymers around glass transition and crystalline melting temperatures. Polymer Testing. 32, 987-994 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[57, 858, 884, 898]]<|/det|>
+34. Xue, T. et al. Polyimide Aerogel Fibers with Controllable Porous Microstructure for Super-Thermal Insulation Under Extreme Environments. Adv. Fiber Mater. 4, 1118-1128 (2022).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 88, 884, 126]]<|/det|>
+35. Jiang, Y. et al. Contrasting soil thermal responses to fire in Alaskan tundra and boreal forest. J. Geophys. Res.: Earth Surf. 120, 363-378 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[110, 133, 884, 170]]<|/det|>
+36. Xue, Y., Lofland, S. & Hu, X. Thermal Conductivity of Protein-Based Materials: A Review. Polymers. 11, 456-464 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[110, 178, 883, 214]]<|/det|>
+37. Tong, T. et al. Thermal conductivity of single silk fibroin fibers measured from the 3ω method. Int. J. Therm. Sci. 185, 108057-108064 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[110, 222, 883, 258]]<|/det|>
+38. Kono, J., Goto, Y., Ostermeyer, Y., Frischknecht, R. & Wallbaum, H. Factors for eco-efficiency improvement of thermal insulation materials. Key Eng. Mater. 678, 1-13 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[110, 265, 883, 302]]<|/det|>
+39. Uma, S., McConnell, A. D., Asheghi, M., Kurabayashi, K. & Goodson, K. E. Conduction and radiation in thermal analysis. Int. J. Thermophys. 22, 605-616 (2001).  
+
+<|ref|>text<|/ref|><|det|>[[110, 309, 883, 346]]<|/det|>
+40. Shintani, Y., Nagaoka, T., Deguchi, Y., Ido, K. & Harada, K. Simple method to predict downward heat flux from flame to floor. Fire Sci. Technol. 33, 17-34 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[110, 354, 883, 390]]<|/det|>
+41. Si, L. et al. A superstable, flexible, and scalable nanofluidic ion regulation composite membrane. Sci. Bull. 68, 2344-2353 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[110, 398, 883, 435]]<|/det|>
+42. Kremer, K. & Grest, G. S. Erratum: Dynamics of entangled polymer melts: A molecular-dynamics simulation. J. Chem. Phys. 94, 4103 (1991).  
+
+<|ref|>text<|/ref|><|det|>[[110, 442, 883, 479]]<|/det|>
+43. Thompson, A. P. et al. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Comput. Phys. Commun. 271, 108171-108205 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[110, 486, 883, 523]]<|/det|>
+44. Stukowski, A. Visualization and analysis of atomistic simulation data with OVITO-the Open Visualization Tool. Modelling Simul. Mater. Sci. Eng. 18, 015012 (2010).  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 561, 313, 579]]<|/det|>
+## Materials and Methods  
+
+<|ref|>text<|/ref|><|det|>[[110, 589, 884, 704]]<|/det|>
+Materials. Kevlar was purchased from Dupont company. Dimethyl sulfoxide (DMSO), sodium hydroxide (NaOH), acetic acid, and Ethanol were all purchased from Aladdin Biochemical Technology Co., Ltd (Shanghai, China). Deionized water (3.8 μS/cm) were prepared by the Millipore Milli-Q system. All other chemicals were utilized without additional purification.  
+
+<|ref|>text<|/ref|><|det|>[[110, 717, 884, 800]]<|/det|>
+Preparation of ANF Dispersion. The synthesis of ANF involved a deprotonation method<sup>41</sup>. In this procedure, Kevlar yarn (10 g) and KOH (15 g) were added to a mixture containing DMSO and water (500:25). The mixture was vigorously stirred for a week, resulting in a dark red dispersion of ANF.  
+
+<|ref|>text<|/ref|><|det|>[[110, 813, 884, 898]]<|/det|>
+Preparation of GAF. A customized microfluidic chip with two phases and three channels, which consisted of a core flow containing ANF solution and a sheath flow of DMSO, was utilized for the production of gel fibril. The flow rates were meticulously regulated, with the core flow operating at 80 μL/min and the sheath
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 884, 235]]<|/det|>
+flow ranging from 200 to \(1200\mu \mathrm{L / min}\) , achieved by employing two sets of flow pumps. The as- prepared gel fibril was wound onto a plastic roller at a constant speed of \(25\mathrm{rpm}\) . It was then subjected to an extensive washing procedure using deionized water to effectively remove any residual DMSO, KOH, and Acetic acid. Subsequent solvent exchange involved transferring the material into a blend of Ethanol for a duration of 24 h and repeated three times. The GAF was finally prepared by supercritical drying for \(48\mathrm{h}\) .  
+
+<|ref|>text<|/ref|><|det|>[[111, 247, 884, 748]]<|/det|>
+Characterization. The morphology of GAF and SAF samples, coated with gold vapor, was characterized using a field emission scanning electron microscope (SIGMA 500, Zeiss, Germany) operated at an accelerating voltage of \(5\mathrm{kV}\) . The pore size distribution in GAF was quantified using Image J software. Raman spectroscopic analysis, including Raman spectra and spatial Raman imaging, was conducted using a Renishaw inVia Raman spectrometer equipped with a \(532\mathrm{nm}\) laser. The concentration gradient of ANF was visualized through multivariate curve resolution, facilitated by supplementary software. Mechanical properties of GAF and SAF were evaluated using a nanomechanical actuating transducer (NMAT) integrated into a Universal Testing Machine (Nano UTM T150, KLA- Tencor, USA) at a strain rate of \(0.0027\mathrm{s}^{- 1}\) . Thermal insulation properties of the fabrics were assessed using a thermal infrared camera (InfReC R550, Avio, Japan) in an environment maintained at constant relative humidity ( \(35\%\) ) and controlled specific temperatures, achieved via a heating stage (LINKAM, LTS420, UK). The camera captured thermal images at a frequency of one per second, monitoring temperature changes across the test area at a rate of \(6^{\circ}\mathrm{C}\mathrm{s}^{- 1}\) . Thermal conductivity was measured using a thermal conductivity analyzer (HS- DR- 5, HESON Technology, China), employing a transient plane sensor with a \(6.4\mathrm{mm}\) radius positioned between two identical ANF aerogel fabrics. A compression device ensured optimal thermal contact between the sensor and the fabrics. The heating power was set at \(20\mathrm{mW}\) , with each measurement lasting \(10\mathrm{s}\) .  
+
+<|ref|>text<|/ref|><|det|>[[112, 758, 883, 810]]<|/det|>
+Coarse- Grained Molecular Dynamics Simulation. The effective thermal conductivity \(\Lambda \mathrm{e}\) of the thermal conduction region was calculated using the following formula:  
+
+<|ref|>equation<|/ref|><|det|>[[368, 821, 629, 857]]<|/det|>
+\[\mathrm{q} = \Lambda_{\mathrm{e}}\frac{\Delta\mathrm{T}}{\Delta\mathrm{L}} = \frac{1}{2\mathrm{A}}\left(\left|\frac{\Delta\mathrm{E}_{\mathrm{Hot}}}{\Delta\mathrm{t}}\right| + \left|\frac{\Delta\mathrm{E}_{\mathrm{Cold}}}{\Delta\mathrm{t}}\right|\right)\]
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[112, 88, 884, 181]]<|/det|>
+where \(\mathbf{q}\) is the heat flux, A is the cross- sectional area perpendicular to the direction of heat conduction, \(\Delta \mathrm{t}\) is the time step, \(\mathrm{E}_{\mathrm{Source}}\) and \(\mathrm{E}_{\mathrm{Sink}}\) are respectively the heat added to the source and subtracted from the sink, \(\Delta \mathrm{L}\) is the length of the thermal conduction region and \(\Delta \mathrm{T}\) is the temperature difference in that region.  
+
+<|ref|>text<|/ref|><|det|>[[112, 193, 884, 309]]<|/det|>
+We utilize coarse- grained molecular dynamics simulations to validate the impact of pores on the mechanical properties of aramid fibers. The coarse- grained bead- spring \(\mathrm{KG}^{42}\) model is employed in our simulations. Lennard- Jones units are used. Energy, length, and mass units are respectively set as \(\epsilon\) , \(\sigma\) , and \(\mathrm{m}\) . The pair- wise interaction forces between beads are represented by the Lennard- Jones potential:  
+
+<|ref|>equation<|/ref|><|det|>[[286, 336, 707, 373]]<|/det|>
+\[\mathrm{E}_{\mathrm{ij}} = 4 \epsilon_{\mathrm{ij}} \left[ \left(\frac{\sigma_{\mathrm{i}}}{\mathrm{r}}\right)^{12} - \left(\frac{\sigma_{\mathrm{i}}}{\mathrm{r}}\right)^{6} \right] - 4 \epsilon_{\mathrm{ij}} \left[ \left(\frac{\sigma_{\mathrm{i}}}{\mathrm{r_{\mathrm{c}}}}\right)^{12} - \left(\frac{\sigma_{\mathrm{i}}}{\mathrm{r_{\mathrm{c}}}}\right)^6 \right] \qquad r < \mathrm{r_{\mathrm{c}}} = 2.5,\]  
+
+<|ref|>text<|/ref|><|det|>[[112, 399, 884, 584]]<|/det|>
+Where \(\mathrm{r}\) is the distance between two beads, \(\mathrm{r}_{\mathrm{c}} = 2.5\) is the cutoff of interaction. \(\epsilon_{ij}\) is the potential strength and \(\sigma_{i}\) is the size of the bead. For intrachain pair- wise interaction forces, \(\epsilon_{ij} = 1\epsilon\) and \(\sigma_{i} = 1\sigma\) . For interchain pair- wise interaction forces, different \(\epsilon_{inter}\) values are used to tune the density of fibers. Arithmetic mixing is used for pair- wise interaction between core and skin. All \(\sigma_{i}\) values are uniformly set to \(1\sigma\) . The Unbreakable Finitely Extensible Nonlinear Elastic (FENE) potential is employed to describe the chemical bond interactions between particles.  
+
+<|ref|>equation<|/ref|><|det|>[[377, 613, 617, 647]]<|/det|>
+\[\mathrm{U}_{\mathrm{FENE}}(\mathrm{r}) = -0.5\mathrm{KR}_{0}^{2}\ln \left[1 - \left(\frac{\mathrm{r}}{\mathrm{R}_{0}}\right)^{2}\right];\]  
+
+<|ref|>text<|/ref|><|det|>[[112, 660, 884, 897]]<|/det|>
+With \(R_{0} = 1.5\sigma\) is the maximum extent of the bond. \(\mathrm{K}\) is the strength of bond interaction, \(\mathrm{K} = 30 \epsilon /\sigma^{2}\) . The bond bending potential is harmonic style. \(U_{angle} = \frac{K_{\theta}}{2} (\theta - \theta_{0})^{2}\) with \(\mathrm{K}_{\theta} = 10\epsilon\) per radian \(^{2}\) is the strength of the interaction and \(\theta_{0} = 180^{\circ}\) is the equilibrium bond angle. To scale the simulation units to real units, Set the LJ temperature unit \(\mathrm{T}^{*} = 1\) to correspond to the real unit of \(300\mathrm{K}\) , so that \(1\epsilon = 4.14\mathrm{E}^{- 21}\mathrm{J}\) . Set two consecutive beads to represent a \(\mathrm{C}_{14}\mathrm{H}_{10}\mathrm{O}_{2}\mathrm{N}_{2}\) monomer (the chemical formula for each bead is \(\mathrm{C}_{7}\mathrm{H}_{5}\mathrm{ON}\) ), so that \(1\sigma = 6.57\mathrm{E}^{- 10}\mathrm{m}\) , and \(1\mathrm{m} = 0.11912\mathrm{kg / NA}\) where \(\mathrm{NA}\) is the avagardo number. The simulation time unit is \(\tau = \sqrt{\mathrm{m} \frac{\sigma^{2}}{\epsilon}}\) . The simulated thermal conductivity unit is \(\epsilon /(\tau \sigma \mathrm{T}^{*})\) . All simulations are conducted using
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[60, 88, 884, 108]]<|/det|>
+413 the LAMMPS<sup>43</sup> simulation package and visualization is performed using the OVITO<sup>44</sup> software package.  
+
+<|ref|>text<|/ref|><|det|>[[60, 120, 789, 140]]<|/det|>
+414 Model setup and simulation details for each model can be found in Supplementary Materials.  
+
+<|ref|>sub_title<|/ref|><|det|>[[63, 186, 308, 207]]<|/det|>
+## Acknowledgements  
+
+<|ref|>text<|/ref|><|det|>[[60, 216, 884, 333]]<|/det|>
+417 This study was supported by the National Natural Science Foundation of China (Grant number. 52103124, 418 12322205, 12002304), the Distinguished Young Scientists Fund from the Natural Science Foundation of 419 Zhejiang Province (Grant number. LR23A020001), and the Pioneer R&D Program of Zhejiang (Grant 420 number. 2023C03007).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 42, 312, 71]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[43, 93, 768, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[59, 130, 330, 230]]<|/det|>
+SupplementaryMovie1.mp4  SupplementaryMovie2.mp4  SupplementaryMovie3.mp4  SupportingInformation.docx
+
+<--- Page Split --->

preprint/preprint__2b06af1324d99d554e163785b31fdb051bec082f852a83042e26b77edd7d2edd/images_list.json

@@ -0,0 +1,190 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1",
+    "footnote": [],
+    "bbox": [
+      [
+        118,
+        145,
+        880,
+        500
+      ]
+    ],
+    "page_idx": 20
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2.",
+    "footnote": [],
+    "bbox": [
+      [
+        128,
+        140,
+        880,
+        832
+      ]
+    ],
+    "page_idx": 22
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 24
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        140,
+        880,
+        641
+      ]
+    ],
+    "page_idx": 26
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        140,
+        777,
+        760
+      ]
+    ],
+    "page_idx": 28
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_7.jpg",
+    "caption": "Figure 7.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        140,
+        956,
+        567
+      ]
+    ],
+    "page_idx": 30
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1",
+    "footnote": [],
+    "bbox": [
+      [
+        53,
+        95,
+        930,
+        511
+      ]
+    ],
+    "page_idx": 32
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2",
+    "footnote": [],
+    "bbox": [
+      [
+        60,
+        101,
+        848,
+        821
+      ]
+    ],
+    "page_idx": 48
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3",
+    "footnote": [],
+    "bbox": [
+      [
+        62,
+        48,
+        901,
+        664
+      ]
+    ],
+    "page_idx": 49
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4",
+    "footnote": [],
+    "bbox": [
+      [
+        46,
+        45,
+        920,
+        620
+      ]
+    ],
+    "page_idx": 51
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5",
+    "footnote": [],
+    "bbox": [
+      [
+        42,
+        40,
+        950,
+        660
+      ]
+    ],
+    "page_idx": 53
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6",
+    "footnote": [],
+    "bbox": [
+      [
+        40,
+        45,
+        820,
+        787
+      ]
+    ],
+    "page_idx": 54
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_7.jpg",
+    "caption": "Figure 7",
+    "footnote": [],
+    "bbox": [
+      [
+        42,
+        316,
+        955,
+        780
+      ]
+    ],
+    "page_idx": 56
+  }
+]

preprint/preprint__2b0f6ea9e001b8d5fc2be38173edd65f51a0d7285b0182711c321e987e96610a/images_list.json

@@ -0,0 +1,244 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "b",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 42
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2 Compositional analysis of clinical scores highlighted two distinct skin manifestations in AD. a. Separation pattern by multidimensional scaling (MDS) on individual components of EASI across AD patients. Components that are correlated with each other (Pearson \\(r > 0.40\\) ) were connected with gray lines. Two major clusters were identified in the aspect of key signs of eczema, among which erythema and induration/papulation are two primary skin manifestations that bear the distinction. b,c. Clinical pictures (b) and immunohistochemistry of skin tissue for CD4 (c, target protein was stained in red) in two representative patients who have a score composition that are skewed to either of erythema (upper) or induration/papulation (lower). Upper: a 51-year-old male patient who has erythema-skewed EASI composition (total = 19.6, erythema = 5.2, papulation = 3.4). Lower: a 50-year-old male patient who has papulation-skewed EASI composition (total = 21.0, erythema = 3.0, papulation = 8.4).",
+    "footnote": [],
+    "bbox": [
+      [
+        164,
+        440,
+        933,
+        755
+      ]
+    ],
+    "page_idx": 42
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3 General transcriptional characteristics of skin and PBMC in AD.",
+    "footnote": [],
+    "bbox": [
+      [
+        140,
+        150,
+        916,
+        820
+      ]
+    ],
+    "page_idx": 42
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4 Inference in ligand-receptor coupling suggests augmented skin-PBMC crosstalk in AD patients. a. Connection map of cytokine-receptor coupling across skin and PBMC in a representative healthy control (left) and AD patient (right). Genes that code cytokines and receptors are aligned along the perimeter of the circles. From the outer layer to the center is the tissue expressing the genes (either skin or PBMC), inferred cell specificity, classification of cytokine or receptor, and the connections between cytokines and its matching receptors. The connections were indicated in different colors according to the classification of direction, i.e. in which tissue the cytokines are produced and on which tissue they act. b. Number of active connections between cytokines and receptors. Connections were enumerated according to 4 classes defined by a sender organ and a receiver organ. Boxplots show median and first and third quartiles, whiskers extending to the highest and lowest values no further than 1.5\\*interquartile range. Brunner-Munzel rank test, \\(^{**}p < 0.01\\) , NS: not significant. c. Pearson correlation between number of active connections and clinical index.",
+    "footnote": [],
+    "bbox": [
+      [
+        55,
+        30,
+        933,
+        780
+      ]
+    ],
+    "page_idx": 42
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "<b>b</b>",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 43
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "GO description",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        216,
+        460,
+        475
+      ]
+    ],
+    "page_idx": 43
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_3.jpg",
+    "caption": "Standardized PC1 value",
+    "footnote": [],
+    "bbox": [
+      [
+        61,
+        500,
+        850,
+        927
+      ]
+    ],
+    "page_idx": 44
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Fig. 5 Identification and characterization of transcriptional modules from skin/PBMC RNA-seq data. a. Cluster dendrograms of transcripts produced by implementation of WGCNA. Color indicates separation of transcriptional module. b, d. Cell type expression and GO enrichment in skin tissue (b) and PBMC (d) analyzed by referring public database. c, e. Visualization of gene-gene networks in PC1 top 30 genes from each transcriptome module in skin (c) and PBMC (e). Genes that have eigengene-based connectivity \\(> 0.65\\) were connected with lines. sModu: skin transcriptome module, pModu: PBMC transcriptome module.",
+    "footnote": [],
+    "bbox": [
+      [
+        100,
+        60,
+        950,
+        884
+      ]
+    ],
+    "page_idx": 45
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Fig. 6 Regression analysis revealed differential patterns of modular involvement in erythema and papulation skin manifestation in AD. a. Regression models for the prediction of clinical phenotypes. b. Predicted dysregulated networks of blood tests and skin/PBMC transcriptome modules contributing to distinct phenotypes. Node size and node frame color represent size and the sign of coefficients for each variable predicted by elastic net regression.",
+    "footnote": [],
+    "bbox": [
+      [
+        60,
+        75,
+        925,
+        732
+      ]
+    ],
+    "page_idx": 46
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_7.jpg",
+    "caption": "Fig. 7 Prediction performance of regression models on longitudinal data set.",
+    "footnote": [],
+    "bbox": [
+      [
+        55,
+        70,
+        890,
+        432
+      ]
+    ],
+    "page_idx": 46
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_4.jpg",
+    "caption": "Fig. S1 Skin sample filtering by expression of hair follicle-related genes. UMAP plot of expression of pilosebaceous unit-related genes in skin samples from both AD patients and healthy controls. Sixty-five skin samples (AD: 45, healthy controls: 20) forming a cluster with extremely strong signatures of the hair follicle gene set were considered to be occupied by hair follicles along with incidental sebaceous glands, and therefore excluded from this study.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 46
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_5.jpg",
+    "caption": "Fig. S2 Identification of the optimal number of clusters for individual EASI partial scores. a,b. Silhouette width plot (a) and elbow plot (b) for identifying the optimal number of clusters for the correlation matrix of EASI partial score. The X-axes indicate the number of clusters ranging from 1 to 10 and the Y-axes indicate average silhouette width (a) or total within sum of square (b). Both of a highest average silhouette coefficient in the silhouette plot and a large drop in the elbow plot were observed at the number of clusters of 2.",
+    "footnote": [],
+    "bbox": [
+      [
+        90,
+        175,
+        884,
+        375
+      ]
+    ],
+    "page_idx": 46
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_6.jpg",
+    "caption": "Fig. S3 Immunohistochemi I analysis revealed shared and differential characteristics in the skin tissue of erythema and papulation-skewed AD patients. Immunohistochemistry of skin tissue in two representative AD patients who have a score composition that are highly skewed to either of erythema (upper) or induration/papulation (lower) as well as healthy control. Left: a 51-year-old male patient who has erythema-skewed EASI composition. Middle: a 50-year-old male patient who has papulation-skewed EASI composition (total = 21.0, erythema = 3.0, papulation = 8.4). Right: a 52-year-old female who have no eczema nor skin diseases. Target proteins were stained in red. Bars = 100 μm. MPO: myeloperoxidase, MBP: major basic protein, K16: keratin-16, FLG: filaggrin.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 47
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_7.jpg",
+    "caption": "Fig. S4 Statistics of transcriptome modules associated with AD. Variance explained by PC1-PC20 in PCA on expression level of transcriptome modules in skin (upper) and PPBMC (lower) across patients.",
+    "footnote": [],
+    "bbox": [
+      [
+        163,
+        160,
+        684,
+        789
+      ]
+    ],
+    "page_idx": 48
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_8.jpg",
+    "caption": "Fig. S5 Prediction accuracy of personalized disease course. The histogram indicates the distribution of Pearson correlation coefficient between observed EASI and predicted EASI in intra-patient dynamics. There was no significant difference in the size of the coefficient among patients regarding treatment classes (Kruskal-Wallis rank sum test, \\(p = 0.57\\) ).",
+    "footnote": [],
+    "bbox": [
+      [
+        105,
+        216,
+        925,
+        525
+      ]
+    ],
+    "page_idx": 50
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_9.jpg",
+    "caption": "Fig. S6 Association of omics features with clinical severity in a longitudinal setting. Trajectories of observed/predicted EASI (upper) and intensity of blood examination and PBMC transcriptome modules (lower) in one year in two representative patients.",
+    "footnote": [],
+    "bbox": [
+      [
+        201,
+        240,
+        927,
+        768
+      ]
+    ],
+    "page_idx": 51
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_10.jpg",
+    "caption": "Fig. S7 Time series features in 1-year monitoring of 30 AD patients. Hierarchical clustering of 7 types of time series features of clinical severity. RMS: root mean square, MAC: mean absolute change, CID: complexity-invariant distance.",
+    "footnote": [],
+    "bbox": [
+      [
+        147,
+        266,
+        797,
+        472
+      ]
+    ],
+    "page_idx": 59
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_11.jpg",
+    "caption": "Fig. S8 Comparison of time series features in top contributing factors among three patient clusters. Two classes of time series features: mean (upper) and mean absolute change (MAC, lower) of top contributing factors in principal component analysis of time series features across patients. Boxplots show median and first and third quartiles, whiskers extending to the highest and lowest values no further than 1.5\"interquartile range. Brunner-Munzel rank test. Multiple comparison tests were carried out using Kruskal-Wallis test. P-values less than 0.05 were considered as significant and subsequently tested for post-hoc comparison with Brunner-Munzel test with \\(p\\) -value correction by Holm's method. NS: not significant, \\(^{*}p < 0.05\\) , \\(^{**}p < 0.01\\) , \\(^{***}p < 0.001\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        45,
+        145,
+        940,
+        570
+      ]
+    ],
+    "page_idx": 60
+  }
+]

preprint/preprint__2b0f6ea9e001b8d5fc2be38173edd65f51a0d7285b0182711c321e987e96610a/preprint__2b0f6ea9e001b8d5fc2be38173edd65f51a0d7285b0182711c321e987e96610a.mmd

@@ -0,0 +1,694 @@
+
+# Multifaceted analysis of cross-tissue, cross-sectional and longitudinal transcriptomes reveals phenotype-endotype associations in atopic dermatitis.  
+
+Haruhiko Koseki ( haruhiko.koseki@riken.jp ) RIKEN Center for Integrative Medical Sciences https://orcid.org/0000- 0001- 8424- 5854  
+
+Aiko Sekita RIKEN Center for Integrative Medical Sciences  
+
+Hiroshi Kawasaki RIKEN  
+
+Ayano Fukushima- Nomura Keio University School of Medicine  
+
+Kiyoshi Yashiro Keio University School of Medicine  
+
+Keiji Tanese Keio University School of Medicine  
+
+Susumu Toshima RIKEN Center for Integrative Medical Sciences  
+
+Koichi Ashizaki RIKEN Information R&D and Strategy Headquarters  
+
+Tomohiro Miyai RIKEN Center for Integrative Medical Sciences https://orcid.org/0000- 0003- 1952- 0228  
+
+Junshi Yazaki RIKEN Center for Integrative Medical Sciences https://orcid.org/0000- 0002- 0697- 8320  
+
+Atsuo Kobayashi RIKEN Center for Integrative Medical Sciences  
+
+Shinichi Namba Osaka University https://orcid.org/0000- 0002- 7486- 3146  
+
+Tatsuhiko Naito Osaka University Graduate School of Medicine https://orcid.org/0000- 0002- 2779- 4600  
+
+Qingbo Wang Osaka University Graduate School of Medicine  
+
+Eiryo Kawakami RIKEN https://orcid.org/0000- 0001- 9955- 4342
+
+<--- Page Split --->
+
+Jun Seita RIKEN https://orcid.org/0000- 0002- 3008- 3615  
+
+Osamu Ohara Kazusa DNA Research Institute https://orcid.org/0000- 0002- 3328- 9571  
+
+Kazuhiro Sakurada RIKEN Information R&D and Strategy Headquarters  
+
+Yukinori Okada Osaka University Graduate School of Medicine https://orcid.org/0000- 0002- 0311- 8472  
+
+Masayuki Amagai Keio University School of Medicine  
+
+## Article  
+
+# Keywords:  
+
+Posted Date: September 26th, 2022  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 2006961/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License
+
+<--- Page Split --->
+
+1 Title 
+
+2 Multifaceted analysis of cross-tissue, cross-sectional and longitudinal transcriptomes 
+
+3 reveals phenotype–endotype associations in atopic dermatitis. 
+
+4 Authors 
+
+6 Aiko Sekita¹,², Hiroshi Kawasaki¹,², Ayano Fukushima-Nomura², Kiyoshi Yashiro², Keiji 
+
+7 Tanese², Susumu Toshima¹,², Koichi Ashizaki¹-³, Tomohiro Miyai¹,², Junshi Yazaki¹, Atsuo 
+
+8 Kobayashi¹, Shinichi Namba⁴,⁵, Tatsuhiko Naito¹,⁴,⁵, Qingbo Wang¹,⁴,⁵, Eiryo Kawakami³,⁶, 
+
+9 Jun Seita¹,³, Osamu Ohara⁷, Kazuhiro Sakurada³,⁸, Yukinori Okada¹,⁴,⁵*, Masayuki 
+
+10 Amagai¹,²*, Haruhiko Koseki¹,⁹* 
+
+11 
+
+12 1, RIKEN Center for Integrative Medical Sciences, Yokohama, Japan 
+
+13 2, Department of Dermatology, Keio University School of Medicine, Tokyo, Japan 
+
+14 3, Advanced Data Science Project, RIKEN Information R&D and Strategy Headquarters, 
+
+15 Tokyo, Japan 
+
+16 4, Department of Statistical Genetics, Osaka University Graduate School of Medicine, Osaka, 
+
+17 Japan. 
+
+18 5, Department of Genome Informatics, Graduate School of Medicine, the University of Tokyo, 
+
+19 Tokyo, Japan 
+
+20 6, Artificial Intelligence Medicine, Graduate School of Medicine, Chiba University, Chiba, 
+
+21 Japan. 
+
+22 7, Kazusa DNA Research Institute, Chiba, Japan 
+
+23 8, Department of Extended Intelligence for Medicine, Keio University School of Medicine,
+
+<--- Page Split --->
+
+24 Tokyo, Japan. 
+
+25 9, Cellular and Molecular Medicine, Advanced Research Departments, Graduate School of 
+
+26 Medicine, Chiba University, Chiba, Japan 
+
+27 *Correspondence: haruhiko.koseki@riken.jp (H. Koseki), amagai@keio.jp (M. Amagai), 
+
+28 yokada@sg.med.osaka-u.ac.jp (Y.Okada)
+
+<--- Page Split --->
+
+## Abstract  
+
+Atopic dermatitis (AD) is a skin disease heterogeneous both in terms of clinical manifestations and molecular profiles. It is increasingly recognized that AD is a systemic rather than a local disease and should be assessed in the context of whole- body level biology. In this study, we integrated RNA- seq data of both skin and PBMC along with clinical data from 115 AD patients and matched 14 healthy controls aiming to comprehensively capture the molecular signature associated with specific clinical presentation. Analysis of cross- tissue ligand- receptor coupling suggested increase of skin- PBMC interactions in AD patients compared to healthy controls. We built a regression model that predicts clinical phenotypes of AD using transcriptome modules identified from weighted gene co- expression network analysis of RNA- seq data in skin tissue and PBMC. We identified differential immunological signatures associated with two qualitatively differential skin manifestations of AD, erythema and papulation. Furthermore, we applied the regression model established in the cross- sectional dataset to a longitudinal dataset collected monthly from 30 AD patients for up to a year for personalized monitoring of disease trajectory and examined longitudinal heterogeneity in association with clinical presentation. Three patient clusters identified on the basis of longitudinal features of blood tests and PBMC transcriptome modules were found to be associated with longitudinal features in clinical severity as well as in the medication history. Our approach serves as a framework for effective clinical investigation to gain a holistic view of pathophysiology of complex human diseases by highlighting inter- and intra- patient heterogeneity.
+
+<--- Page Split --->
+
+## Introduction  
+
+Atopic dermatitis (AD) is one of the most common chronic inflammatory skin diseases worldwide and is characterized by a highly heterogeneous clinical phenotype [1, 2]. Causal factors, disease course and underlying immunological pathways of AD vary greatly among patients, making clinical management tremendously complicated [3, 4]. In spite of growing therapeutic options with a wave of development of novel targeted drugs such as an anti- IL- 4Rα antibody [5] and an anti- IL- 31Rα antibody [6], there is no consensus concerning therapeutic decisions for individual patients [7]. In order to provide optimal treatment for each patient with maximum cost- effectiveness, there is an urgent need to characterize patients in terms of endotypes that are potentially linked with disease course.  
+
+Although recent advances in biomedical technologies have enabled us to acquire an enormous amount of patient omics data including genome data, capturing fundamental endophenotypes of individual patients is still challenging. In the past decades, multiple attempts were made to uncover the biological features of skin tissues or peripheral blood mononuclear cells (PBMC) from AD patients using transcriptomic and proteomic approaches. Those studies have revealed important roles of Th2 or Th17 pathways both in skin and in PBMC along with altered skin barrier function in AD pathology [8- 10], and some of them further demonstrated how these pathways can be targeted by systemic treatment with immunosuppressants [11], anti- IL- 4Rα antibody [12, 13] and oral JAK inhibitors [14]. However, these observations in either skin tissue or blood only focus on alterations in a specific part of the body that could reflect just one aspect of a highly complex pathology. It is widely recognized that complex diseases should be assessed in the context of whole- body level biology since organs are communicating with each other [15- 17]. Projects such as GTEx
+
+<--- Page Split --->
+
+[18] and Human Cell Atlas [19] can be utilized for per- tissue/cell type characterization of human biology, as well as characterization of inter- tissue communications ("crosstalk"). Skin disorders including AD, which is now recognized as a systemic disease [20, 21], need special attention to such crosstalk between the originally damaged organ and the circulatory system [22, 23]. The importance of considering cross- tissue interactions in skin immunological regulation is also supported by the evidence of concurrent biological alterations in both skin tissue and blood after systemic treatment in AD [12, 24, 25] or in HIV infection, which frequently causes cutaneous malignancies or inflammation [26].  
+
+Other essential factors in AD pathology include the heterogeneous disease trajectories as characterized by repeated exacerbation and remission, with different cycles by patients. Correspondingly, most patients have their own medication history over time, based on their incidence of exacerbations. Accounting for such heterogeneity in disease trajectory has been extremely challenging in previous omics- based studies of AD.  
+
+In this study, we carried out two types of analysis, cross- sectional analysis and longitudinal analysis with observational datasets, aiming to capture biological signatures in the context of clinical profiles in the Japanese AD population (Fig. 1). For the cross- sectional analysis, we analyzed RNA- seq data of both skin and PBMC from AD patients as well as healthy controls and linked them to clinical data in order to capture a holistic view of the biology in individual AD patients that are associated with particular clinical phenotypes. By assessing crosstalk between skin and PBMC based on cytokine- receptor coupling, we found that there was enhanced interaction between skin and the circulatory system in AD patients compared to healthy controls, confirming the need for system- level investigation to understand AD pathophysiology. Therefore, we built regression models incorporating both skin and PBMC
+
+<--- Page Split --->
+
+transcriptome data that were preprocessed into interpretable transcriptome modules in search of factors contributing to clinical presentations across patients. For the longitudinal analysis, we applied the transcriptome modules along with the regression models established in the analysis of cross- sectional dataset to a time series dataset to monitor personalized disease courses and to examine inter- patient heterogeneity in longitudinal features. These multifaceted analyses of cross- tissue, cross- sectional and longitudinal transcriptomes highlighted the close association between phenotypes and endotypes in AD that was never clarified before. Our approach should serve as a framework for effective clinical investigation of heterogeneous and complex human diseases.
+
+<--- Page Split --->
+
+## Results  
+
+## Characterization of participants  
+
+For cross- sectional analysis, 165 AD patients (271 lesional skin samples and 194 PBMC samples) and 45 healthy controls (56 non- lesional skin samples and 45 PBMC samples) were extracted from the overall sample collection according to the criteria defined in the section Study design in Methods. RNA- seq data from samples either with low read count (total read count < 5 million) or with a strong batch effect attributable to inadequate sample processing were excluded. Consequently, 151 AD patients and 19 healthy controls that met the criteria for PBMC RNA- seq data were extracted. Patients were further filtered by gene expression intensity of pilosebaceous unit- related genes in skin samples (Fig. S1), resulting in 115 AD patients and 14 healthy controls as eligible samples (one sample per patient) for regression analysis using all of skin, PBMC and blood tests. All the samples (315 skin samples and 235 PBMC samples from both AD patients and healthy controls) that were assured for RNA- seq data quality by itself were included for transcriptome modules identification to increase power. Of these participants, 27.1% (30 AD patients and 5 healthy controls) were female. Their mean age was 41.3 years (AD: 40.5, healthy: 47.3, range, 21- 70 years).  
+
+For longitudinal analysis, time series dataset consisting of PBMC transcriptome, laboratory blood tests and clinical severity score from 30 AD patients on monthly basis up to a year (total 360 time points) were extracted, and after quality control, 284 data were considered as eligible and used for longitudinal analysis. Of these AD patients, 7 patients (23.3%) were female, and 30 patients (100%) and 17 patients (56.7%) overlapped with the cross- sectional population for PBMC only and PBMC+skin analysis, respectively.
+
+<--- Page Split --->
+
+For meta- analysis of clinical severity scores, we used a total of 1424 data points obtained during the period of November 2016 to July 2021 from the 151 AD patients who were included in cross- sectional and/or longitudinal analysis. All AD patients included in this observational study were under treatment with topical steroids and emollients as needed, as directed by dermatologists. Their history of systemic treatment was categorized as follows: intermittent use of immunosuppressant, antiallergic agents with continuous use (a total of more than 120 days/year), antiallergic agents with occasional use (a total of fewer than 120 days/year), and no use of these agents. Drugs used for systemic treatment in the overall AD population in this study are listed in the Table S2. Characteristic information of the participants is summarized in Table S3.  
+
+# Compositional analysis of clinical scores highlighted two distinct skin manifestations  
+
+The extent and severity of atopic dermatitis were measured using the Eczema Area and Severity Index (EASI) [27]. In this scoring system, severity is determined by grading the key signs of eczema (i.e. erythema, induration/papulation, excoriation, and lichenification) over the four anatomic divisions of the body (i.e. the head and neck, the trunk, the upper extremities, and the lower extremities) separately. The average severity of each sign in each of the four body regions was assigned a score of 0 to 3 (none, mild, moderate, and severe, respectively).  
+
+To capture the relationship between individual components of eczema severity, we performed multidimensional scaling (MDS) on the collection of partial scores across patients (Fig. 2a). Two major clusters were found in the aspect of key signs of eczema (Fig.S2); one consisted of erythema and lichenification and the other consisted of induration/papulation
+
+<--- Page Split --->
+
+and excoriation. These findings suggested that erythema and induration/papulation constitute two distinct skin manifestations, apt to be accompanied by lichenification and excoriation, respectively, as signs of progression or chronicity. From a regional perspective, the configuration of the scores for the four body regions was all in the same order in the MDS plot, i.e. from left to right are the lower extremities, the upper extremities, the trunk to the head and neck, leaving the head and neck distant from the other three regions. This finding is consistent with the recent view that head and neck erythema is a prominent form of AD [28, 29].  
+
+Based on these findings, we defined two distinct phenotypes in AD, an erythema form and a papulation form, using the summation of either erythema or papulation scores in all the body regions except for the head and neck, respectively. Meanwhile, we defined the general severity of AD as the summation of all the scores, i.e. EASI (total) as is conventionally used.  
+
+Fig. 2b shows clinical and histological images of representative patients who have a score composition that is highly skewed to either erythema or induration/papulation (partial score for the left patient; erythema = 9.6, papulation = 4.8, the right patient; erythema = 4.3, papulation = 8.6). Histopathological examination of lesional skin revealed intense epidermal hyperplasia in both patients. In contrast, the manner of lymphocytic infiltration in dermis appeared to be different; increased numbers of ectatic capillaries accompanied by perivascular cell infiltration were observed in the erythema- skewed patient while spotted but intense cell infiltration was observed in the papulation- skewed patient (Fig. 2b right panel).  
+
+Fig. 2b shows clinical and histological images of two representative patients who have a clinical score composition that is skewed to either erythema (AD#089, partial score; erythema = 9.6, papulation = 4.8) or papulation (AD#176, partial score; erythema = 4.3, papulation =
+
+<--- Page Split --->
+
+8.6). Histopathological examination of lesional skin revealed shared and differential characteristics in the skin tissue from the erythema and papulation-skewed AD patients (Fig. 2c, Fig. S3). In both skin samples, intense infiltration of immune cells including \(\mathrm{CD4^{+}}\) T cell (Fig. 2c), macrophage \((\mathrm{CD}206^{+})\) , myeloid dendritic cell \((\mathrm{CD}11\mathrm{c}^{+}, \mathrm{DC} - \mathrm{LAMP}^{+})\) and Langerhans cell \((\mathrm{CD}1\mathrm{a}^{+})\) , along with epidermal hyperplasia and diminished epidermal barrier (as observed by filaggrin expression) were commonly observed. However, the patterns for immune cell infiltration appeared to be different between two skin samples; the skin sample from the erythema-skewed patient were characterized by diffuse infiltration of immune cells in dermis, accompanied by epidermal lymphocytic infiltration (Fig. 2c right panel). On the other hand, the skin sample from the papulation-skewed patient was characterized by nodular infiltration of immune cells in dermis suggestive of geometrical heterogeneity over the lesion, as well as prominent hyperkeratosis. Neutrophil (myeloperoxidase: \(\mathrm{MPO^{+}}\) ) infiltration were substantially observed in the skin sample from the erythema-skewed patient but not in the skin sample from the papulation-skewed patient (Fig. S3).  
+
+## Transcriptional characteristics of skin tissue and PBMC typically found in AD  
+
+To identify transcriptome signatures enriched in AD patients, we firstly conducted differential gene expression analysis on RNA- seq data of skin and PBMC specimens. Accordingly, 272 and 33 differentially expressed genes for skin and PBMC, respectively, were identified (llog2 fold change \((\log 2\mathrm{FC}) \geq 2\) and false discovery rate \((\mathrm{FDR}) < 0.01\) for skin and \(\log 2\mathrm{FC} \geq 1\) and \(\mathrm{FDR} < 0.05\) for PBMC, Fig. 3a). Pathways enriched in skin of AD patients included antimicrobial peptides, chemokine and interleukin signaling genes and epidermal differentiation/keratinization (Fig. 3b left), which is largely consistent with previous reports
+
+<--- Page Split --->
+
+[10, 30]. Pathways enriched in PBMC of AD patients included neutrophil degranulation and immune system.  
+
+# Inference in ligand-receptor coupling suggests augmented skin-PBMC crosstalk in AD patients  
+
+The increased expression of inflammation- related genes in both skin and PBMC suggested that inflammation induced in skin tissue in turn triggered inflammatory responses in PBMC, or vice versa in some cases, presumably through secretion of soluble factors that can act on cells in the circulatory system [22]. In order to illuminate such potential crosstalk between skin tissue and PBMC, we integrated RNA- seq data from both sources and quantified ligand- receptor couplings that are particularly engaged in inflammatory signaling [31].  
+
+We defined active cytokine- receptor pairs as having concurrent expression of a cytokine gene and its matching receptor gene at a level of cytokine gene \(>1\) and receptor gene \(>0.5\) in value of variance stabilizing transformation (vst) applied to the expression values that were followed by normalization across the population. A total of 210 pairs of inflammatory cytokine and receptor genes were assessed in the skin and PBMCs of each AD patient and healthy control. The active cytokine- receptor pairs were enumerated according to classes defined by the combination of a sender organ that expressed a cytokine gene and a receiver organ that expressed a receptor gene (Fig. 4a; Methods). The total number of active cytokine- receptor pairs was significantly higher in AD patients than in healthy controls (mean \(= 50.9\) vs \(29.6\) ; \(p = 1.0\mathrm{E} - 3\) ).  
+
+Among these, the number of connections from skin to skin and the number of connections from skin to PBMC were significantly increased in AD patients compared to healthy controls
+
+<--- Page Split --->
+
+(mean = 24.6 vs 10.9 and 17.3 vs 10.6; \(p = 8.3E- 5\) and \(p = 2.8E- 3\) , respectively), while the number of connections from PBMC to either of skin or PBMC was not significantly different between AD patients and healthy controls (Fig. 4b). There were moderate correlations between the total number of cytokine-receptor connections and either EASI ( \(r = 0.32\) ; \(p = 2.5E- 4\) ) or serum TARC ( \(r = 0.35\) ; \(p = 6.6E- 5\) ) (Fig. 4c).  
+
+The most frequently observed pairs in AD were CCL22- CCR4/ and CCL17- CCR4 in skin, while in healthy controls they were IL37 (skin) - IL18R1/IL18RAP (PBMC) and IL34 (skin) - CSF1R (skin). The top two frequently observed pairs involving PBMC in AD were CCL18 (skin) - CCR8 (PBMC) and IL20 (skin) - IL20RB (PBMC) (Table S4). Cell types responsible for expression of these cytokine/receptor genes were estimated by referring to publicly available datasets that are suitable for analyzing cell type expression [32, 33]. The most frequently appearing cell types in AD were T cells and vascular endothelial cells (VEC) as cytokine- expressing cells, and myeloid cells and T cells as receptor- expressing cells, all of which were found in the skin. The most highly involved cell type in PBMC was the monocyte, for both cytokine and receptor expression (Table S5). Collectively, the indication of enhanced ligand- receptor coupling involving both skin and the circulatory system in AD patients suggested the need for a system- level investigation into AD pathology.  
+
+## Identification and characterization of transcriptional modules associated with AD  
+
+To illuminate the heterogeneity in the biological signature across AD patients, expression levels of not only DEGs between AD patients and healthy controls but also the extended range of gene sets that have potential association with AD pathology should be analyzed. Weighted gene co- expression network analysis (WGCNA) is a powerful technique to depict
+
+<--- Page Split --->
+
+functional subsystems by highlighting biologically relevant transcripts with reduced dimensionality across a population [34]. We applied WGCNA to our entire expression dataset, including AD patients and healthy controls for skin and PBMC, respectively to identify AD- related transcriptional modules. This procedure identified 21 skin transcriptional modules (sModus) and 15 PBMC transcriptional modules (pModus), each comprising 51- 774 genes (mean; 258.7 for skin, 191.8 for PBMC) that behave synchronously in a tissue, suggesting their biological relevance to each other (Fig. 5a).  
+
+As expected, genes in each module exhibited substantial cell type specificity in their expression as confirmed by referring to publicly available dataset of either sc- RNA- seq (skin, Fig. 5b) or sorted cell RNA- seq (PBMC, Fig. 5d). Relationships among the top 30 genes of the first principal component (PC1) from each module were visualized on the basis of gene- gene networks using thresholding of eigengene- based connectivity \(>0.65\) (Fig. 5c, 5e). This analysis revealed several notable signaling compartments in each tissue; compartments of acquired immune regulation (cytokine signaling), innate immune regulation (interferon signaling) and compartments of keratinization/formation of cornified envelope, in skin tissue. Additionally, three modules were found to be representing skin appendages; sebaceous gland (sModu01, GO: fatty acid metabolism) and sweat gland (sModu03 and sModu19, GO: ion channel transport and developmental biology, respectively). The intensity of these modules was not relevant to dermatitis, and was strongly biased by sampling regions. Therefore, we considered these two modules as noises, and excluded from the following analysis.  
+
+To obtain personalized profiles based on transcriptional modules, scores for each module and each patient were defined. Since identified modules consist of co- expressing genes,
+
+<--- Page Split --->
+
+expression patterns in each module became simple enough to be handled linearly, as verified by principal component analysis (PCA) with explanatory capability of PC1 reaching \(40 - 60\%\) (Fig. S4). Therefore, we used PC1 values followed by standardization as the index of intensity of gene expression of transcriptome modules in each patient.  
+
+## Regression analysis reveals differential patterns of modular involvement in erythematous and papular skin manifestations in AD  
+
+We next investigated how the AD phenotypes can be represented by transcriptome modules from both skin and PBMC, as well as by laboratory tests obtained at the same visits. Given the relatively high ratio of the number of variables to the sample size, we built regression models using elastic net, an algorithm for regularized regression and variable selection that is applicable to high dimensional data with multicollinearity [35].  
+
+To confirm the association of PBMC and skin transcriptomes, variables were added per category in a step wise manner to predict EASI (total, as a representation of general severity in AD) (Fig. 6a). Adding skin and PBMC transcriptome modules to basic information and blood test resulted in better prediction models as demonstrated by increased \(R^2\) , again suggesting an association of both skin and PBMC in AD pathology and, accordingly, the validity of using these variables to predict phenotypic severity in AD.  
+
+Using elastic net, we also built models for prediction of EASI (erythema) and EASI (papulation) which, as described above, were found to make a major distinction in skin manifestations. The model performance ( \(R^2\) ) and the set of significant features ( \(p < 0.05\) ) in each model as well as its biological characteristics are summarized in Table 1.
+
+<--- Page Split --->
+
+Both in erythema and in papulation skin manifestation, a decreased lymphocyte ratio in blood, an indication of increased proportion of myeloid cells (a populational summation of monocytes and neutrophils) and an increased eosinophil ratio in blood were found to be associated in symptom.  
+
+Erythema was characterized by a bolstered signature of immediate early genes (NR4A1, FOSL1, FOSB, ATF3, NR4A2) and immune system (CD163, C1QB, C1QC, THY1, MS4A7) that are inferentially expressed mainly in keratinocytes and myeloid cells, respectively, in skin tissue, along with Treg specific genes (CCR4, CNTNAP1, DUSP4, LMNA, PI16) in PBMC. In contrast, papulation was characterized by decreased B cell signature (FCRL1, MS4A1, PAX5, CD22, LINC00926) and increased naive CD4 signature (NELL2, LRRN3, OBSCN, CCR7, GRASP1) in PBMC along with enhanced signature of interferon signaling (MMP12, CCL18, IFI27, TYMP, COL6A6) and extracellular matrix (PI15, GREM1, COL4A1, TNFAIP6, NNMT), suggestive of altered activity in VEC and fibroblast in skin tissue. Dysregulated module networks contributing to distinct phenotypes were predicted based on the coefficient of each variable (Fig. 6b). These results suggest that pathologies underlying erythema and papulation are substantially different on a molecular basis.  
+
+## Personalized monitoring of trajectory of disease severity and molecular signatures  
+
+One of the most essential features of AD is that patients follow a disease course complicated by exacerbations and remissions throughout the years, thereby patients take individual treatment steps based on their condition at a given time [36, 37]. To provide an overview including symptom changes and use of systemic treatment in individual patients, we conducted monthly monitoring of PBMC transcriptomes, laboratory tests and severity scores
+
+<--- Page Split --->
+
+for 30 AD patients for up to a year. We leveraged transcriptomic modules generated in the cross- sectional patient dataset and profiled the dynamics of transcriptomic features as well as blood tests that were lastly analyzed in association with disease severity.  
+
+We first tested the performance of elastic net regression model trained with cross- sectional dataset when applied to the longitudinal dataset (Fig. 7a). Prediction performance for EASI (total) was higher in a model using all of basic information (age, age^2, sex), laboratory tests and PBMC transcriptome compared to a model using only basic information and laboratory tests ( \(R^2\) : 0.10 vs 0.26).  
+
+Taking a closer look at individual trajectories of disease course, we found substantial variability in prediction accuracy among patients. Personalized disease trajectories in two representative patients are shown as examples in Fig. 7b. In the first example, the prediction seemed successful (Pearson \(r = 0.81\) ; \(p = 2.4E- 3\) ), accurately capturing the disease flare (month 5). In contrast, prediction was unsuccessful in the second example as evident by Pearson \(r = - 0.44\) ( \(p = 0.20\) ). There was no significant difference in prediction accuracy of personalized trajectories of disease severity among patients regarding treatment classes (Fig. S5). We found that the time- course trajectory of the weights of TARC which was selected as the top predictive features in the elastic net model, strongly correlated with disease severity trajectory (Pearson \(r = 0.88\) ; \(p = 3.1E- 4\) ) in the first example, but not in the second example (Pearson \(r = - 0.047\) ; \(p = 0.91\) ) (Fig. S6). These observations suggest that the predominant features vary by patients, which could limit the performance of linear models assuming same feature weights across samples.
+
+<--- Page Split --->
+
+Close association between endotypic longitudinal features and phenotypic longitudinal features  
+
+Given that another factor that accounts for the endotypes in individual AD patients is longitudinal variability itself [36], just as in other chronic inflammatory diseases [38], it is important to evaluate time series features in clinical severity and transcriptome modules. Seven types of time series features, i.e. mean, minimum, maximum, root mean square (RMS), mean absolute change (MAC), approximate entropy, and complexity- invariant distance (CID) were extracted from three categories of datasets, i.e. blood tests, PBMC transcriptome modules and clinical severity (i.e. EASI) in individual patients at a monthly interval over 1 year using the Python module Tsfresh [39] (Fig. 8a).  
+
+Hierarchical clustering of those 7 features of clinical severity in 30 AD patients showed two major clusters; one includes mean, maximum, minimum and RMS, the other includes MAC, CID and approximate entropy (Fig. S7). Unsupervised k- means clustering on 30 AD patients based on time series features of PBMC transcriptome modules and blood tests, with number of clusters (k) determined using silhouette criterion (Fig. 8b), identified three patient clusters (Fig. 8c). We applied PCA to this dataset and evaluated the intensity of the top PC1/PC2 contributing factors. Cluster 1 (n = 2) was characterized by stably high levels of pModu07 (GO: neutrophil degranulation/Toll- like receptor signal), pModu09 (GO: neutrophil degranulation/interleukin signaling) and neutrophil (complete blood counts- derived ratio: CBC) and a stably low level of lymphocyte (CBC), whereas Cluster 2 (n = 7) showed volatile trajectories of all of those terms throughout the observation period, as observed by high values of MAC with a medium level of mean. An unstably high white blood cell (WBC) count was also observed. Meanwhile, Cluster 3 (n = 21) was characterized by relatively low levels
+
+<--- Page Split --->
+
+of all of those terms except for lymphocyte (CBC), which was relatively high in this patient cluster (Fig 8d, Fig. S8).  
+
+Remarkably, those patient clusters were found to show clinical phenotypes associated with endotypic longitudinal features. Cluster 1 showed severe and stable symptoms, Cluster 2 showed severe and unstable symptoms, and Cluster 3 showed mild symptoms (Fig. 8e). Additionally, this patient grouping was found to be closely linked with prescription status of systemic treatment (Fig. 8f). Cluster 2, which manifested severe and unstable symptoms, highly overlapped with the patients who were under systemic therapy with an oral immunosuppressant (5/7 patient overlap), while Cluster 1 and Cluster 3 were mostly managing the disease either with antihistamines only or without systemic treatment.  
+
+The dynamics of EASI (total), top PC1/PC2 contributing factors, as well as the treatment periods in representative patients are shown in Fig. 8g. One possible logic for the observation of patient overlap between disease severity/stability and systemic treatment is that only severe patients are supposed to be candidates for systemic immunosuppressant therapy that leads to rapid symptom mitigation and global transcriptome alterations [13], but could cause a flare at the time of drug cessation. Patients treated with immunosuppressants in this study were all administered with the drug intermittently as instructed by their dermatologists, considering their symptom improvement or the risk of side effects. Accordingly, some patients experienced disease flare during washout periods. We thus note that systemic immunosuppressant therapy could partially contribute to instability of disease severity trajectory as well as other personal time series features.
+
+<--- Page Split --->
+
+## Discussion  
+
+With an increase of therapeutic options expected in the coming years, understanding the heterogeneity in disease phenotypes and endotypes in AD is an urgent task in order to provide optimal treatment for individual patients. Phenotypic heterogeneity among AD patients, which has been empirically recognized though not yet clearly defined, includes variability in skin manifestation and longitudinal disease course. In this study, we sought to elucidate endotypic heterogeneity in association with these two aspects of phenotype, aiming at providing clinically significant and applicable insights in dermatology.  
+
+We profiled patients with transcriptome analysis on skin and blood biospecimens, each reflecting different aspects of disease state; skin for primary pathology at the site of ongoing or probable inflammation [10], and blood, a relatively homogeneous compartment, for systemic regulation of inflammation [40]. Although previous studies have reported patient stratification in AD based on single tissue data such as serum cytokine profiles [41], whole blood transcriptomes [42], or skin barrier profiles of comorbidity- stratified patient groups (with/without food allergy) [43], there are few reports on clinically significant endotypes regarding both skin and the circulatory system so far. He et al. demonstrated that patient groups defined on the basis of disease severity have differential molecular profiles in both non- lesional skin and serum [44]. Indeed, clinical manifestations in AD should be evaluated beyond the criterion of simple severity, given that several specific detailed signs of eczema have long been recognized in AD [27] including erythema and papulation, two distinct skin manifestations highlighted in our cross- sectional analysis. Exploring molecular involvement in such specific phenotypes using both skin and PBMC data should provide deeper insights
+
+<--- Page Split --->
+
+into the unique characteristics of individual patients than in the case where the focus is on conventional general severity or just the presence of disease (AD versus healthy controls).  
+
+Our combinatorial approach of WGCNA and elastic net regression enabled us to efficiently and jointly analyze high dimensional datasets of skin and PBMC transcriptomes. Our finding on skin manifestation- dependent molecular profiles suggests that endotypes in AD, as defined by transcriptome modules in skin and PBMC, are closely associated with the phenotypes of AD as defined by visual evaluation of skin. More fundamentally, this observation supports the assumption that the AD population comprises a variety of pathophysiological subtypes. To the best of our knowledge, this is the first report that demonstrates association between endotypes and phenotypes with granularity beyond general clinical severity in AD.  
+
+Furthermore, we assessed heterogeneity in personal longitudinal features in PBMC transcriptome modules and blood tests in association with clinical severity. We identified three patient clusters based on longitudinal blood- derived signatures that were found to be closely linked with disease course and medication history. In addition to serum TARC, LDH and eosinophil counts, all of which are well- recognized biomarkers in AD [45], newly defined PBMC transcriptome modules including pModule01 (inferred cell specificity: naïve CD4, PC1 top genes: NELL2, LRN3, OBSCN, CCR7 and GRASP1) and pModu04 (inferred cell specificity: Treg, PC1 top genes: MKI67, RRM2, TOP2A, ASPM and MYBL2) were identified as contributing factors in a personal disease course.  
+
+Although our study demonstrated integrative analysis of transcriptome data both from primary tissue where disease originated and from circulatory system is advantageous for understanding patient endotypes, such assessment could not be applied in routine medical
+
+<--- Page Split --->
+
+examination especially in the longitudinal contexts, since acquiring biospecimen other than blood requires invasive sampling. Our next task is to identify representative biomarkers that can predict system- level pathology in individual patients only by routine blood tests.  
+
+There are some limitations in this study. First, the clinical definition of skin phenotype manifestations is not totally objective. Scoring for severity of eczema was based on visual evaluation, which is strongly dependent on the expertise and experience of the dermatologists. The fact that most AD patients manifested multiple signs of eczema including erythema and papules simultaneously, with blurred boundaries, makes this issue even more of a problem. In the future, skin manifestations should be computationally and quantitatively evaluated, for example, through the abundance of hemoglobin or pigmentation in skin, as has been investigated in some other skin disorders [46, 47]. Second, our transcriptome data is from bulk RNA- seq which yields mixed signatures of different cell types in the tissue. Although we could infer cell type specificity for each molecular signature by deconvolution taking advantage of external scRNA- seq data, the resolution and accuracy is limited [48- 50]. Other limitations in our study includes limited sample size and population diversity, as is always the challenge in studies on complex human diseases. Studies with extended sample size and diversity may illuminate more profound heterogeneity in AD. Our study overall highlighted inter- and intra- patient heterogeneity in AD, and demonstrated the promises of personalized AD treatment.
+
+<--- Page Split --->
+
+## Methods  
+
+## Study design  
+
+This study was approved by the Keio University School of Medicine Ethics Committee (Approval Number 20150325, 20160225, 20160131 and 20160377) and the RIKEN Ethics Committee (Approval Number H28- 24) and conducted according to all relevant requirements from the Declaration of Helsinki. Written informed consent was obtained from all the participants. Diagnosis of AD was made according to diagnostic criteria of Hanifin and Rajka [51].  
+
+We enrolled 313 Japanese AD patients who visited Keio University hospital and 46 healthy controls for skin and blood sampling study between December 2016 and February 2020. Pregnant or breast- feeding women, patients with episodes of lidocaine allergy, prilocaine allergy, or complications of bleeding disorders were excluded from recruitment. For cross- sectional analysis, we extracted eligible sample population based on the following criteria:(1) 20 years of age or older, (2) not being under systemic therapy with anti- IL- 4Ra mAb nor JAK inhibitors, (3) having undergone biopsy from the back for skin samples. Accordingly, 188 AD patients (271 lesional skin samples and 190 PBMC samples) and 45 healthy controls (56 non- lesional skin samples and 45 PBMC samples) were extracted, and after data quality control as described in RNA- seq and data processing section as well as filtering with missing values in blood tests, 151 AD patients and 19 healthy controls were considered to be eligible for regression analysis on PBMC and blood tests, and 115 AD patients were considered to be eligible for regression analysis on all of skin, PBMC and blood tests.  
+
+For longitudinal analysis, samples from 30 AD patients who were enrolled in prospective observational study between December 2016 and September 2018 were analyzed. Time
+
+<--- Page Split --->
+
+series dataset consisting of PBMC transcriptome, laboratory blood tests and clinical severity score from 30 AD patients on monthly basis up to a year (total 360 time points), were extracted. After data quality control, 280 data were considered to be eligible and used for analysis.  
+
+All the patients included in two analyses were treated according to the Japanese Dermatological Association guidelines [28, 52], such as emollients, topical corticosteroids and/or tacrolimus, oral antihistamines and immunosuppressants. The Eczema Area and Severity Index (EASI) [27], assessed by two dermatology experts, was used for analysis as disease severity. Patient information including disease history, medication history (within 4 weeks for the cross- sectional dataset and 13 months for the longitudinal dataset), laboratory blood test data, and EASI were extracted and filed from electronic medical and patient questionnaires.  
+
+## Sample collection  
+
+For skin RNA- seq, lesional skin biopsy samples (1 mm punch) were obtained from the backs of the participants under local anesthesia with Emla cream (lidocaine 2.5% and prilocaine 2.5%, Sato Pharmaceutical) which was administered 1 hour before the performance of biopsy. Samples were placed in RNAlater (Life Technologies) overnight at \(4^{\circ}\mathrm{C}\) and stored at \(- 80^{\circ}\mathrm{C}\) until further processing. For immunohistochemistry, skin biopsy samples (1 mm punch) were taken from sites exhibiting similar skin conditions in close proximity (within 5 mm region) to the skin samples for RNA- seq, immediately snap- frozen and stored at \(- 80^{\circ}\mathrm{C}\) until further processing. For PBMC RNA- seq, PBMC were isolated from venous peripheral blood by density gradient purification using Vacutainer CPT tubes (Becton Dickinson) following the
+
+<--- Page Split --->
+
+manufacturer's instructions, suspended in RNAlater and stored at \(- 80^{\circ}C\) until further processing.  
+
+## Immunohistochemistry  
+
+Frozen skin samples were thawed and immediately embedded in O.C.T. compound (Sakura Finetech), snap- frozen and stored at \(- 80^{\circ}C\) until cryosectioning. Immunostaining was performed using the streptavidin- biotin complex/alkaline phosphatase method as previously described [53] with few modifications. Briefly, 10 μm- thick cryostat- cut tissue sections were fixed for 5 min in ice- cold acetone and rehydrated in Tris- buffered saline with \(0.1\%\) Triton- X followed by incubation with normal goat serum for 1 h. The sections were incubated with the primary antibodies (Table S6) diluted in blocking solution overnight at \(4^{\circ}C\) , followed by a biotinylated secondary antibody (either anti- mouse or anti- rabbit according to the primary antibodies) and thereafter with a streptavidin- biotin complex/alkaline phosphatase (Vectastain ABC- AP; Vector). Finally, the sections were developed with alkaline phosphatase substrate (ImmPACT Vector Red; Vector) and counterstained with hematoxylin. The images were captured using a digital image acquisition and analysis system (BX43 microscope, DP27 digital camera, cellSens v3.3 Software; Olympus).  
+
+## RNA-seq and data processing  
+
+For skin tissue RNA- seq, skin specimens were homogenized with BioMasher (Nippi) in TRIzol Reagent (Thermo), and RNA was isolated with Direct- Zol RNA Kit (ZYMO RESERCH). Library preparation was carried out using NEBNext Ultra RNA Library Prep Kit (New England Biolabs) following the manufacturer's instructions. For PBMC RNA- seq, RNA was isolated
+
+<--- Page Split --->
+
+using Maxwell 16 LEV simplyRNA Blood Kit and Maxwell 16 Instrument (Promega) and library preparation were carried out with SureSelect Strand- Specific RNA Library Prep Kit (Agilent). The libraries were pooled for skin tissue RNA- seq and PBMC RNA- seq, respectively, and sequenced on HiSeq1500 or HiSeq2500 (Illumina) to obtain 15- 20 million reads using the 50- bp single- end read configuration. Reads were aligned to the Ensembl GRCh38 human genome assembly using STAR [54] and feature counts were performed with the R package (version 3.6.2) Rsubread [55]. Genes were filtered by both of the following conditions: 1) expressed in more than 5% of the sample population, 2) maximum reads across the population \(>8\) . Samples were filtered with the criteria of total read count \(>5\) million. Genes coding hemoglobin proteins ("HBA2", "HBB", "HBA1") and ribosomal proteins were removed. The batch effects from each dataset attributable to difference in experimental periods or locations for sequencing were adjusted by ComBat- seq [56]. Differential gene expression analysis and vst normalization were conducted using the R package DESeq2 [57]. Since there is a chance where skin samples are occupied by considerable volume ratio of pilosebaceous unit in 1 mm punch biopsy, only biased by sampling regions, skin samples were also filtered by gene expression intensity of pilosebaceous unit- related gene set. A cluster that showed extremely strong signature of pilosebaceous unit- related genes in Uniform Manifold Approximation and Projection (UMAP) [58] were excluded. GO analysis and GSEA were performed with the R package clusterProfiler [59].  
+
+## Inference in ligand-receptor coupling  
+
+Since our datasets consist of bulk- derived samples, which represent mixed signatures of any cell type present in the tissue, we evaluated the degree of ligand- receptor coupling with a
+
+<--- Page Split --->
+
+binary scoring approach [31] and thereafter cell type specificity for individual active cytokines and receptors were inferred by using publicly available datasets of cell type- specific expression.  
+
+Ligand- receptor pairs that are classified into inflammatory response were extracted from the list of cytokine - cytokine receptor interactions in the KEGG pathway database (https://www.genome.jp/kegg/) [60]. Possible active cytokines were defined by their expression in the tissue over 0.5 in the value of vst normalization which accounts for the top \(14.2\%\) of the overall population, while possible active receptors were defined by their expression in the tissue over 0 in the value of vst normalization which accounts for the top \(48.6\%\) of the overall population. Possible active cytokine- receptor pairs were defined as concurrent presence of pairs of possible active cytokines and possible active receptors.  
+
+A total of 210 pairs of inflammatory cytokine and receptor genes were assessed in the skin and PBMC of each of AD patient and healthy control. The active cytokine- receptor pairs were enumerated according to classes defined by the combination of a sender organ that expressed a cytokine gene and a receiver organ that expressed a cognate receptor. Comparison of the number of active connections between cytokines and receptors between AD patients and healthy controls were carried out by a non- parametric Brunner- Munzel rank test [61] with R package lawstat [62], taking into account the nature of the data that showed non- normal and heteroscedastic distribution in two patient groups. \(P\) values less than 0.05 were considered significant.  
+
+For each of the cytokine and receptor genes, cell types responsible for the cytokine/receptor gene expression was estimated by referring to publicly available datasets (GSE147424; single- cell RNA- seq of skin tissue from AD patients and healthy controls [32],
+
+<--- Page Split --->
+
+Human Protein Atlas blood cell gene data; RNA- seq of 18 cell types sorted from human peripheral blood [33], for skin and PBMC RNA- seq data, respectively). Reference datasets were standardized among cell types by genes that were expressed at a level of \(z\) - score \(>2\) were deemed as cell- type specific genes and annotated to given cell types. Ligand- receptor connection were visualized using the R package circlize [63].  
+
+## Module detection and validation  
+
+Gene co- expression networks of skin and PBMC transcriptomes were constructed from the vst normalized matrix of variance top 10,000 genes in respective datasets using the R package WGCNA [64]. Modules were generated following the procedures recommended by the publication author, including determination of the algorithm's hyperparameters. Soft- thresholding power \((\beta)\) was chosen as the lowest power for which the scale- free topology fit index reached 0.80 with the minimum threshold of 6. As each module is composed of genes highly correlated with each other, the intensity of overall expression of a given module in a patient was represented as the first principal component of expression of all the genes in the module. Hub genes were defined using the signed KME function and transcriptome networks were visualized using the R package igraph [65]. Module characterization was performed based on two terms, cell type specificity and GO. Cell type specificity in its expression was determined by referring to the same external dataset used in the previous section, i.e. either sc- RNA- seq (skin) or sorted cell RNA- seq (PBMC). GO analysis were performed with the R package clusterProfiler.  
+
+## Regression analysis
+
+<--- Page Split --->
+
+Elastic net, a regularization and variable selection method that combines the L1 and L2 penalties of the lasso and ridge methods [35], was applied on cross- sectional datasets consisting of both skin and PBMC RNA- seq data along with blood tests (AD patients: \(\mathrm{n} = 115\) , healthy controls: \(\mathrm{n} = 14\) ) to determine the strength of the relationship between disease phenotypes and omics features in AD using the R package glmnet [66]. For each phenotype defined in compositional analysis of clinical scores, total samples labeled with the degree of specific skin conditions in continuous values were split into a training set (70%) and a testing set (30%). Models were built on the training set with optimization of the regularization parameter \(\lambda\) , which determines how much shrinkage is used to train the model, through tenfold cross validation with \(\alpha = 0.5\) . Then the model with the optimal parameters was applied to the test set to get the \(R^2\) value to evaluate how well the model fit to the observed data.  
+
+For longitudinal data analysis, the model was trained on a total cross- sectional dataset excluding 30 AD patients who are enrolled in the longitudinal cohort, and tested on the longitudinal dataset from 30 AD patients. Prediction performance on the test set was evaluated with \(R^2\) . Closeness of fit in personalized trajectory was evaluated with the Pearson correlation coefficient.  
+
+## Longitudinal data analysis  
+
+Time series data from blood tests, PBMC transcriptome modules and clinical severity were profiled by patients in date order. By using the Python (version 3.7.4) module Tsfresh [39], seven types of time series features, i.e., mean, minimum, maximum, root mean square (RMS), mean absolute change (MAC), approximate entropy, and complexity- invariant distance (CID) were extracted in individual patients. The values of time series features were standardized
+
+<--- Page Split --->
+
+among patients. PCA followed by unsupervised k- means clustering was conducted on longitudinal features of PBMC transcriptome modules and blood tests to identify patient clusters based on longitudinal endotypes.
+
+<--- Page Split --->
+
+## Data availability  
+
+All of the sequence data described in this study are available in the National Center for Biotechnology Information Gene Expression Omnibus (GEO) under accession numbers that are coming soon.  
+
+## Code availability  
+
+The source code to reproduce the presented results are available at the online code repository (the webpage address is coming soon).
+
+<--- Page Split --->
+
+618 References619 1. Weidinger S, Novak N: Atopic dermatitis. Lancet 2016, 387(10023):1109- 1122. 620 2. Yew YW, Thyssen JP, Silverberg JI: A systematic review and meta- analysis of the regional and age- related differences in atopic dermatitis clinical characteristics. Journal of the American Academy of Dermatology 2019, 80(2):390- 401. 621 3. Czarnowicki T, He H, Krueger JG, Guttman- Yassky E: Atopic dermatitis endotypes and implications for targeted therapeutics. J Allergy Clin Immunol 2019, 143(1):1- 11. 622 4. Bieber T: Atopic dermatitis: an expanding therapeutic pipeline for a complex disease. Nat Rev Drug Discov 2022, 21(1):21- 40. 623 5. Simpson EL, Bieber T, Guttman- Yassky E, Beck LA, Blauvelt A, Cork MJ, Silverberg JI, Deleuran M, Kataoka Y, Lacour JP et al: Two Phase 3 Trials of Dupilumab versus Placebo in Atopic Dermatitis. N Engl J Med 2016, 375(24):2335- 2348. 624 6. Ruzicka T, Hanifin JM, Furue M, Pulka G, Mlynarczyk I, Wollenberg A, Galus R, Etoh T, Mihara R, Yoshida H et al: Anti- Interleukin- 31 Receptor A Antibody for Atopic Dermatitis. N Engl J Med 2017, 376(9):826- 835. 625 7. Chun PIF, Lehman H: Current and Future Monoclonal Antibodies in the Treatment of Atopic Dermatitis. Clin Rev Allergy Immunol 2020, 59(2):208- 219. 626 8. Gittler JK, Shemer A, Suarez- Farinas M, Fuentes- Duculan J, Gulewicz KJ, Wang CQ, Mitsui H, Cardinale I, de Guzman Strong C, Krueger JG et al: Progressive activation of T(H)2/T(H)22 cytokines and selective epidermal proteins characterizes acute and chronic atopic dermatitis. J Allergy Clin Immunol 2012, 130(6):1344- 1354.
+
+<--- Page Split --->
+
+9. Suarez-Farinas M, Dhingra N, Gittler J, Shemer A, Cardinale I, Strong CD, Krueger JG, Guttman-Yassky E: Intrinsic atopic dermatitis shows similar T(H)2 and higher T(H)17 immune activation compared with extrinsic atopic dermatitis. J Allergy Clin Immunol 2013, 132(2):361-370.
+10. Tsoi LC, Rodriguez E, Degenhardt F, Baurecht H, Wehkamp U, Volks N, Szymczak S, Swindell WR, Sarkar MK, Raja K et al: Atopic Dermatitis Is an IL-13-Dominant Disease with Greater Molecular Heterogeneity Compared to Psoriasis. J Invest Dermatol 2019, 139(7):1480-1489.
+11. Khattri S, Brunner PM, Garcet S, Finney R, Cohen SR, Oliva M, Dutt R, Fuentes-Duculan J, Zheng X, Li X et al: Efficacy and safety of ustekinumab treatment in adults with moderate-to-severe atopic dermatitis. Exp Dermatol 2017, 26(1):28-35.
+12. Guttman-Yassky E, Bissonnette R, Ungar B, Suarez-Farinas M, Ardeleanu M, Esaki H, Suprun M, Estrada Y, Xu H, Peng X et al: Dupilumab progressively improves systemic and cutaneous abnormalities in patients with atopic dermatitis. J Allergy Clin Immunol 2019, 143(1):155-172.
+13. Mobus L, Rodriguez E, Harder I, Stolzl D, Boraczynski N, Gerdes S, Kleinheinz A, Abraham S, Heratizadeh A, Handrick C et al: Atopic dermatitis displays stable and dynamic skin transcriptome signatures. J Allergy Clin Immunol 2021, 147(1):213-223.
+14. Pavel AB, Song T, Kim HJ, Del Duca E, Krueger JG, Dubin C, Peng X, Xu H, Zhang N, Estrada YD et al: Oral Janus kinase/SYK inhibition (ASN002) suppresses
+
+<--- Page Split --->
+
+inflammation and improves epidermal barrier markers in patients with atopic dermatitis. J Allergy Clin Immunol 2019, 144(4):1011- 1024.  
+
+Kozawa S, Ueda R, Urayama K, Sagawa F, Endo S, Shizaki K, Kurosu H, Maria de Almeida G, Hasan SM, Nakazato K et al: The Body- wide Transcriptome Landscape of Disease Models. iScience 2018, 2:238- 268.  
+
+Priest C, Tontonoz P: Inter- organ cross- talk in metabolic syndrome. Nat Metab 2019, 1(12):1177- 1188.  
+
+Picollet- D'hahan N, Zuchowska A, Lemeunier I, Le Gac S: Multiorgan- on- a- Chip: A Systemic Approach To Model and Decipher Inter- Organ Communication. Trends Biotechnol 2021, 39(8):788- 810.  
+
+GTEx Consortium: Genetic effects on gene expression across human tissues. Nature 2017, 550(7675):204- 213.  
+
+Regev A, Teichmann SA, Lander ES, Amit I, Benoist C, Birney E, Bodenmiller B, Campbell P, Carninci P, Clatworthy M et al: The Human Cell Atlas. Elife 2017, 6.  
+
+Brunner PM, Silverberg JI, Guttman- Yassky E, Paller AS, Kabashima K, Amagai M, Luger TA, Deleuran M, Werfel T, Eyerich K et al: Increasing Comorbidities Suggest that Atopic Dermatitis Is a Systemic Disorder. J Invest Dermatol 2017, 137(1):18- 25.  
+
+Oliveira C, Torres T: More than skin deep: the systemic nature of atopic dermatitis. Eur J Dermatol 2019, 29(3):250- 258.  
+
+Hu Y, Zhu Y, Lian N, Chen M, Bartke A, Yuan R: Metabolic Syndrome and Skin Diseases. Front Endocrinol (Lausanne) 2019, 10:788.
+
+<--- Page Split --->
+
+23. Glickman JW, Dubin C, Renert-Yuval Y, Dahabreh D, Kimmel GW, Auyeung K, Estrada YD, Singer G, Krueger JG, Pavel AB et al: Cross-sectional study of blood biomarkers of patients with moderate to severe alopecia areata reveals systemic immune and cardiovascular biomarker dysregulation. J Am Acad Dermatol 2021, 84(2):370-380.
+24. Imai Y, Kusakabe M, Nagai M, Yasuda K, Yamanishi K: Dupilumab Effects on Innate Lymphoid Cell and Helper T Cell Populations in Patients with Atopic Dermatitis. JID Innov 2021, 1(1):100003.
+25. Mack MR, Brestoff JR, Berrien-Elliott MM, Trier AM, Yang TLB, McCullen M, Collins PL, Niu HX, Bodet ND, Wagner JA et al: Blood natural killer cell deficiency reveals an immunotherapy strategy for atopic dermatitis. Science Translational Medicine 2020, 12(532).
+26. Saluzzo S, Pandey RV, Gail LM, Dingelmaier-Hovorka R, Kleissl L, Shaw L, Reininger B, Atzmuller D, Strobl J, Touzeau-Romer V et al: Delayed antiretroviral therapy in HIV-infected individuals leads to irreversible depletion of skin- and mucosa- resident memory T cells. Immunity 2021, 54(12):2842-2858 e2845.
+27. Hanifin JM, Thurston M, Omoto M, Cherill R, Tofte SJ, Graeber M: The eczema area and severity index (EASI): assessment of reliability in atopic dermatitis. EASI Evaluator Group. Exp Dermatol 2001, 10(1):11-18.
+28. Yasuda-Sekiguchi F, Shiohama A, Fukushima A, Obata S, Mochimaru N, Honda A, Kawasaki H, Kubo A, Ebihara T, Amagai M et al: Single nucleotide variations in genes associated with innate immunity are enriched in Japanese adult cases of face and neck type atopic dermatitis. J Dermatol Sci 2021, 101(2):93-100.
+
+<--- Page Split --->
+
+de Wijs LEM, Nguyen NT, Kunkeler ACM, Nijsten T, Damman J, Hijnen DJ: Clinical and histopathological characterization of paradoxical head and neck erythema in patients with atopic dermatitis treated with dupilumab: a case series. Br J Dermatol 2020, 183(4):745- 749.  
+
+Suarez- Farinas M, Ungar B, da Rosa JC, Ewald DA, Rozenblit M, Gonzalez J, Xu H, Zheng XZ, Peng XY, Estrada YD et al: RNA sequencing atopic dermatitis transcriptome profiling provides insights into novel disease mechanisms with potential therapeutic implications. J Allergy Clin Immunol 2015, 135(5):1218- 1227.  
+
+Armingol E, Officer A, Harismendy O, Lewis NE: Deciphering cell- cell interactions and communication from gene expression. Nature Reviews Genetics 2021, 22(2):71- 88.  
+
+He HL, Suryawanshi H, Morozov P, Gay- Mimbrera J, Del Duca E, Kim HJ, Kameyama N, Estrada Y, Der E, Krueger JG et al: Single- cell transcriptome analysis of human skin identifies novel fibroblast subpopulation and enrichment of immune subsets in atopic dermatitis. J Allergy Clin Immunol 2020, 145(6):1615- 1628.  
+
+Uhlen M, Karlsson MJ, Zhong W, Tebani A, Pou C, Mikes J, Lakshmikanth T, Forsstrom B, Edfors F, Odeberg J et al: A genome- wide transcriptomic analysis of protein- coding genes in human blood cells. Science 2019, 366(6472):1471- +.  
+
+Stuart JM, Segal E, Koller D, Kim SK: A gene- coexpression network for global discovery of conserved genetic modules. Science 2003, 302(5643):249- 255.  
+
+Zou H, Hastie T: Regularization and variable selection via the elastic net (vol B 67, pg 301, 2005). J R Stat Soc B 2005, 67:768- 768.
+
+<--- Page Split --->
+
+Chovatiya R, Lei D, Ahmed A, Chavda R, Gabriel S, Silverberg JI: Clinical phenotyping of atopic dermatitis using combined itch and lesional severity: A prospective observational study. Ann Allergy Asthma Immunol 2021, 127(1):83- 90 e82.  
+
+Boguniewicz M, Alexis AF, Beck LA, Block J, Eichenfield LF, Fonacier L, Guttman- Yassky E, Paller AS, Pariser D, Silverberg JI et al: Expert Perspectives on Management of Moderate- to- Severe Atopic Dermatitis: A Multidisciplinary Consensus Addressing Current and Emerging Therapies. J Allergy Clin Immunol Pract 2017, 5(6):1519- 1531.  
+
+Nim HT, Connelly K, Vincent FB, Petitjean F, Hoi A, Koelmeyer R, Boyd SE, Morand EF: Novel Methods of Incorporating Time in Longitudinal Multivariate Analysis Reveals Hidden Associations With Disease Activity in Systemic Lupus Erythematosus. Front Immunol 2019, 10:1649.  
+
+Christ M, Braun N, Neuffer J, Kempa- Liehr AW: Time Series FeatuRe Extraction on basis of Scalable Hypothesis tests (tsfresh - A Python package). Neurocomputing 2018, 307:72- 77.  
+
+Mohr S, Liew CC: The peripheral- blood transcriptome: new insights into disease and risk assessment. Trends Mol Med 2007, 13(10):422- 432.  
+
+Bakker DS, Nierkens S, Knol EF, Giovannone B, Delemarre EM, van der Schaft J, van Wijk F, De Bruin- Weller MS, Drylewicz J, Thijs JL: Confirmation of multiple endotypes in atopic dermatitis based on serum biomarkers. J Allergy Clin Immunol 2021, 147(1):189- 198.
+
+<--- Page Split --->
+
+751 42. Mobus L, Rodriguez E, Harder I, Boraczynski N, Szymczak S, Hubenthal M, Stolzl D, Gerdes S, Kleinheinz A, Abraham S et al: Blood transcriptome profiling identifies two candidate endotypes of atopic dermatitis. J Allergy Clin Immunol 2022.
+754 43. Leung DYM, Calatroni A, Zaramela LS, LeBeau PK, Dyjack N, Brar K, David G, Johnson K, Leung S, Ramirez-Gama M et al: The nonlesional skin surface distinguishes atopic dermatitis with food allergy as a unique endotype. Sci Transl Med 2019, 11(480).
+755 44. He H, Del Duca E, Diaz A, Kim HJ, Gay-Mimbrera J, Zhang N, Wu JN, Beaziz J, Estrada Y, Krueger JG et al: Mild atopic dermatitis lacks systemic inflammation and shows reduced nonlesional skin abnormalities. J Allergy Clin Immunol 2021, 147(4):1369-1380.
+756 45. Thijs J, Krastev T, Weidinger S, Buckens CF, de Bruin-Weller M, Bruijnzeel-Koomen C, Flohr C, Hijnen D: Biomarkers for atopic dermatitis: a systematic review and meta-analysis. Curr Opin Allergy Clin Immunol 2015, 15(5):453-460.
+757 46. Abdlaty R, Doerwald-Munoz L, Madooei A, Sahli S, Yeh SCA, Zerubia J, Wong RKW, Hayward JE, Farrell TJ, Fang QY: Hyperspectral Imaging and Classification for Grading Skin Erythema. Front Phys-Lausanne 2018, 6.
+758 47. Romano RA, Rosa RGT, Salvio AG, Jo JA, Kurachi C: Multispectral autofluorescence dermoscope for skin lesion assessment. Photodiagn Photodyn 2020, 30.
+759 48. Alkon N, Bauer WM, Krausgruber T, Goh I, Griss J, Nguyen V, Reininger B, Bangert C, Staud C, Brunner PM et al: Single-cell analysis reveals innate lymphoid cell lineage infidelity in atopic dermatitis. J Allergy Clin Immunol 2022, 149(2):624-639.
+
+<--- Page Split --->
+
+49. Reynolds G, Vegh P, Fletcher J, Poyner EFM, Stephenson E, Goh I, Botting RA, Huang N, Olabi B, Dubois A et al: Developmental cell programs are co-opted in inflammatory skin disease. Science 2021, 371(6527):364-+.50. Nakamizo S, Dutertre CA, Khalilnezhad A, Zhang XM, Lim S, Lum J, Koh G, Foong C, Yong PJA, Tan KJ et al: Single-cell analysis of human skin identifies CD14(+) type 3 dendritic cells co-producing IL1B and IL23A in psoriasis. J Exp Med 2021, 218(9).51. Hanifin JM, Rajka G: Diagnostic Features of Atopic-Dermatitis. Acta Derm-Venereol 1980:44-47.52. Katayama I, Aihara M, Ohya Y, Saeki H, Shimojo N, Shoji S, Taniguchi M, Yamada H, Allergology JS: Japanese guidelines for atopic dermatitis 2017. Allergology International 2017, 66(2):230-247.53. Schlapbach C, Hanni T, Yawalkar N, Hunger RE: Expression of the IL-23/Th17 pathway in lesions of hidradenitis suppurativa. J Am Acad Dermatol 2011, 65(4):790-798.54. Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, Batut P, Chaisson M, Gingeras TR: STAR: ultrafast universal RNA-seq aligner. Bioinformatics 2013, 29(1):15-21.55. Liao Y, Smyth GK, Shi W: The R package Rsubread is easier, faster, cheaper and better for alignment and quantification of RNA sequencing reads. Nucleic Acids Res 2019, 47(8):e47.56. Zhang Y, Parmigiani G, Johnson WE: ComBat-seq: batch effect adjustment for RNA-seq count data. NAR Genom Bioinform 2020, 2(3):lqaa078.
+
+<--- Page Split --->
+
+797 57. Love MI, Huber W, Anders S: Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 2014, 15(12):550. 798 59. Yang Y, Sun H, Zhang Y, Zhang T, Gong J, Wei Y, Duan YG, Shu M, Yang Y, Wu D et al: Dimensionality reduction by UMAP reinforces sample heterogeneity analysis in bulk transcriptomic data. Cell Rep 2021, 36(4):109442. 799 60. Yu G, Wang LG, Han Y, He QY: clusterProfiler: an R package for comparing biological themes among gene clusters. OMICS 2012, 16(5):284- 287. 800 61. Ogata H, Goto S, Sato K, Fujibuchi W, Bono H, Kanehisa M: KEGG: Kyoto Encyclopedia of Genes and Genomes. Nucleic Acids Res 1999, 27(1):29- 34. 801 62. Brunner E, Munzel U: The nonparametric Behrens- Fisher problem: Asymptotic theory and a small- sample approximation. Biometrical J 2000, 42(1):17- 25. 802 63. Hui W, Gel YR, Gastwirth JL: lawstat: An R Package for Law, Public Policy and Biostatistics. J Stat Softw 2008, 28(3):1- 26. 803 64. Gu Z, Gu L, Eils R, Schlesner M, Brors B: circlize Implements and enhances circular visualization in R. Bioinformatics 2014, 30(19):2811- 2812. 804 65. Langfelder P, Horvath S: WGCNA: an R package for weighted correlation network analysis. Bmc Bioinformatics 2008, 9. 805 66. Cs- rdi Gb, Nepusz Ts: The igraph software package for complex network research. In: 2006; 2006. 806 67. Friedman J, Hastie T, Tibshirani R: Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Softw 2010, 33(1):1- 22.
+
+<--- Page Split --->
+
+## Acknowledgements  
+
+We would like to sincerely thank all the participants involved in this study. We thank H. Maeo, S. Shibata, R. Sato, M. Tanaka and E. Numazaki for supporting biospecimen sampling. We thank R. Ohashi, A. Hananoe, M. Ohtsuka, E. Okutsu, Y. Murahashi, A. Sugimoto and T. Takemori for supporting maintenance of the storage of human samples and data. We thank S. Koyasu, K. Yamamoto, K. Fujio, T. Endo and T. Ishikawa for helpful advice on analysis. This study was supported by AMED (18ek0410028h0003, 19ek0410046h0003, 21ek0410079h0002, 21ek0410058h0003), JST (JPMJIH1504) and Japan Society for the Promotion of Science (JSPS) KAKENHI (18K16072, 20K17333).
+
+<--- Page Split --->
+
+a 
+
+i) Cross-sectional analysis (n; AD = 115, healthy = 14) 
+
+![](images/Figure_unknown_0.jpg)
+
+ 
+
+ii) Longitudinal analysis (n; AD = 30) 
+
+![](images/Figure_2.jpg)
+
+ 
+
+b 
+
+i) Detailed evaluation of eczema severity 
+
+EASI partial score 
+
+![](images/Figure_3.jpg)
+
+ 
+
+ii) Disease course and medication history 
+
+![](images/Figure_4.jpg)
+
+ 
+
+Fig. 1 Summary of study design. 
+
+a. This study consists of two parts, i) a cross-sectional part (n; Atopic dermatitis: AD = 115, healthy = 14) and ii) a longitudinal part (n; AD = 30) to elucidate endotypes that are associated with phenotypes in AD. b. We focused of two classes of disease phenotypes highlighted by clinical data; i) skin manifestation and ii) longitudinal disease course along with medication history, that were examined in association with endotypes in cross-sectional and longitudinal analysis, respectively. EMR: electronic medical records, EASI: Eczema Area and Severity Index.
+
+<--- Page Split --->
+![](images/Figure_unknown_1.jpg)
+
+<center>b </center>  
+
+AD#089 (Erythemaskewed)  
+
+AD#176 (Papulation-skewed)  
+
+![](images/Figure_unknown_2.jpg)
+
+<center>Fig. 2 Compositional analysis of clinical scores highlighted two distinct skin manifestations in AD. a. Separation pattern by multidimensional scaling (MDS) on individual components of EASI across AD patients. Components that are correlated with each other (Pearson \(r > 0.40\) ) were connected with gray lines. Two major clusters were identified in the aspect of key signs of eczema, among which erythema and induration/papulation are two primary skin manifestations that bear the distinction. b,c. Clinical pictures (b) and immunohistochemistry of skin tissue for CD4 (c, target protein was stained in red) in two representative patients who have a score composition that are skewed to either of erythema (upper) or induration/papulation (lower). Upper: a 51-year-old male patient who has erythema-skewed EASI composition (total = 19.6, erythema = 5.2, papulation = 3.4). Lower: a 50-year-old male patient who has papulation-skewed EASI composition (total = 21.0, erythema = 3.0, papulation = 8.4). </center>
+
+<--- Page Split --->
+![](images/Figure_unknown_3.jpg)
+
+<center>Fig. 3 General transcriptional characteristics of skin and PBMC in AD. </center>  
+
+a. Volcano plot with significantly differentially expressed genes (llog2 fold change (log2FC) \(\geq 2\) and false discovery rate (FDR) \(< 0.01\) for skin and llog2FC \(\geq 1\) and FDR \(< 0.05\) for PBMC) highlighted in red (up-regulated in AD) and blue (down-regulated in AD) compared to healthy controls (n; AD = 115, healthy = 14). b. Pathways enriched in differentially expressed genes in AD (FDR \(< 0.1\) and Enrichment score \(>0.5\)). Enrichment score was obtained based on the size of a given gene set in gene ontology terms (see Methods).
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Fig. 4 Inference in ligand-receptor coupling suggests augmented skin-PBMC crosstalk in AD patients. a. Connection map of cytokine-receptor coupling across skin and PBMC in a representative healthy control (left) and AD patient (right). Genes that code cytokines and receptors are aligned along the perimeter of the circles. From the outer layer to the center is the tissue expressing the genes (either skin or PBMC), inferred cell specificity, classification of cytokine or receptor, and the connections between cytokines and its matching receptors. The connections were indicated in different colors according to the classification of direction, i.e. in which tissue the cytokines are produced and on which tissue they act. b. Number of active connections between cytokines and receptors. Connections were enumerated according to 4 classes defined by a sender organ and a receiver organ. Boxplots show median and first and third quartiles, whiskers extending to the highest and lowest values no further than 1.5\*interquartile range. Brunner-Munzel rank test, \(^{**}p < 0.01\) , NS: not significant. c. Pearson correlation between number of active connections and clinical index. </center>
+
+<--- Page Split --->
+![](images/Figure_6.jpg)
+
+ 
+
+![](images/Figure_7.jpg)
+
+ 
+
+<b>b</b> 
+
+![](images/Figure_unknown_4.jpg)
+
+ 
+
+<center>GO description</center> 
+
+Fatty acid metabolism
+Neuronal System
+Ion channel transport
+Extracellular matrix organization
+Cytokine signaling
+tRNA processing
+Innate immune System
+Cellular responses to stress
+Extracellular matrix organization
+Keratinization
+Mitotic cell cycle
+Keratinization
+Phospholipid metabolism
+Interferon signaling
+WNT ligand biogenesis and trafficking
+Extracellular matrix organization
+Formation of the cornified envelope
+Immune System
+Developmental Biology
+Muscle contraction
+Metabolism 
+
+![](images/Figure_unknown_5.jpg)
+
+ 
+
+<center>Standardized PC1 value</center>
+
+<--- Page Split --->
+![](images/Figure_unknown_6.jpg)
+
+<center>Fig. 5 Identification and characterization of transcriptional modules from skin/PBMC RNA-seq data. a. Cluster dendrograms of transcripts produced by implementation of WGCNA. Color indicates separation of transcriptional module. b, d. Cell type expression and GO enrichment in skin tissue (b) and PBMC (d) analyzed by referring public database. c, e. Visualization of gene-gene networks in PC1 top 30 genes from each transcriptome module in skin (c) and PBMC (e). Genes that have eigengene-based connectivity \(> 0.65\) were connected with lines. sModu: skin transcriptome module, pModu: PBMC transcriptome module. </center>
+
+<--- Page Split --->
+![](images/Figure_unknown_7.jpg)
+
+<center>Fig. 6 Regression analysis revealed differential patterns of modular involvement in erythema and papulation skin manifestation in AD. a. Regression models for the prediction of clinical phenotypes. b. Predicted dysregulated networks of blood tests and skin/PBMC transcriptome modules contributing to distinct phenotypes. Node size and node frame color represent size and the sign of coefficients for each variable predicted by elastic net regression. </center>
+
+<--- Page Split --->
+
+Table 1 Prediction variables extracted from regression models.
+
+<table><tr><td>Objective variables</td><td>Adjusted \(R^{2}\)</td><td>Predictors</td><td>Coefficient</td><td>P-value</td><td>Tissue</td><td>PC1 top 5 genes</td><td>Cell type specificity</td></tr><tr><td></td><td></td><td>Lymphocyte</td><td>-0.40</td><td>8.8E-04</td><td>Blood</td><td>-</td><td>Lymphocyte</td></tr><tr><td>EASI<br>(total)</td><td>Train 0.61<br>Test 0.43</td><td>Total IgE<br>Eosinophil<br>sModu10</td><td>0.25<br>0.21<br>0.27</td><td>0.042<br>0.043<br>0.081</td><td>Blood<br>Blood<br>Skin</td><td>-<br>-<br>S100A8, S100A9, KRT6C, SERPINB4, S100A7</td><td>-<br>Eosinophil<br>Keratinocyte</td></tr><tr><td></td><td></td><td>sModu08</td><td>0.22</td><td>2.6E-03</td><td>Skin</td><td>NR4A1, FOSL1, FOSB, ATF3, NR4A2</td><td>Keratinocyte</td></tr><tr><td>EASI<br>(erythema)</td><td>Train 0.63<br>Test 0.51</td><td>Lymphocyte<br>sModu18<br>pModu11<br>Eosinophil</td><td>-0.14<br>0.13<br>0.13<br>0.11</td><td>0.041<br>0.076<br>0.077<br>0.099</td><td>Skin<br>Blood<br>Blood<br>Blood</td><td>CD163, C1QB, C1QC, THY1, MS4A7<br>CCR4, CNTNAP1, DUSP4, LMNA, P116<br>-</td><td>Lymphocyte<br>Myeloids<br>Treg<br>Eosinophil</td></tr><tr><td></td><td></td><td>Lymphocyte</td><td>-0.39</td><td>7.5E-04</td><td>Blood</td><td>-</td><td>Lymphocyte</td></tr><tr><td>EASI<br>(papulation)</td><td>Train 0.54<br>Test 0.33</td><td>pModu06<br>Eosinophil<br>pModu01<br>ALT<br>BUN<br>sModu14<br>sModu16</td><td>-0.27<br>0.22<br>0.19<br>0.17<br>0.24<br>0.18</td><td>0.0029<br>0.0091<br>0.015<br>0.030<br>0.031<br>0.073</td><td>Blood<br>Blood<br>Blood<br>Blood<br>Blood<br>Skin</td><td>FCRL1, MS4A1, PAX5, CD22, LINC00926<br>NELL2, LRRN3, OBSCN, CCR7, GRASP1<br>-<br>MMP12, CCL18, IFI27, TYMP, COL6A6<br>PI15, GREM1, COL4A1, TNFAIP6, NNMT</td><td>B cell<br>Eosinophil<br>Naive CD4<br>-<br>-<br>VEC<br>Fibroblast</td></tr></table>
+
+<--- Page Split --->
+![](images/Figure_unknown_8.jpg)
+
+<center>Fig. 7 Prediction performance of regression models on longitudinal data set. </center>  
+
+a. Performance of elastic net regression models to predict EASI (total) as an index of general AD severity. Models were trained with cross-sectional patient data set and tested on longitudinal data set. b. Trajectories of observed and predicted EASI in two representative patients both with successful prediction outcome (left, Pearson \(r = 0.81\) , \(p = 2.4E-3\)) and with unsuccessful prediction outcome (right, Pearson \(r = -0.44\) , \(p = 0.20\) ).
+
+<--- Page Split --->
+![](images/Figure_unknown_9.jpg)
+
+
+<--- Page Split --->
+
+Fig. 8 Time series features of disease severity, clinical lab and PBMC transcriptome in each patient in association with history of systemic therapy.  
+
+a. Schematic of extraction of time series features in 30 AD patients. 
+b. Silhouette width plot for identifying the optimal number of patient clusters based on time series features. 
+c-f. PCA on 30 AD patients using time series features of blood tests and PBMC transcriptome modules. Color indicates patient clusters defined by k-means (c), the intensity of time series feature (upper; mean, lower; MAC) of 5 variables normalized among patients (d), time series features of clinical severity (e), and history of internal medication (f). 
+g. Dynamics of EASI (total), pModu07, pModu09, lymphocyte, neutrophil and WBC as well as period of internal medication in representative patients. MAC: mean absolute change, WBC: white blood cell.
+
+<--- Page Split --->
+
+# Supplementary data
+
+Table S1 Pilosebaceous unit-related gene set defined for filtering skin samples.Pilosebaceous unit-related genes were extracted by comparison between two patient clusters identified by applying unsupervised k-means clustering (k=2) on healthy controls.
+
+<table><tr><td>Gene</td></tr><tr><td>ALOX15B</td></tr><tr><td>ADGRL3</td></tr><tr><td>FABP7</td></tr><tr><td>THRSP</td></tr><tr><td>ACSBG1</td></tr><tr><td>CYP4F8</td></tr><tr><td>SEC14L6</td></tr><tr><td>FADS1</td></tr><tr><td>FAR2</td></tr><tr><td>SOAT1</td></tr><tr><td>CRAT</td></tr><tr><td>AWAT1</td></tr><tr><td>MGST1</td></tr><tr><td>CIDEA</td></tr><tr><td>ELOVL5</td></tr><tr><td>INSIG1</td></tr><tr><td>PNPLA5</td></tr><tr><td>APOC1</td></tr><tr><td>TMEM56</td></tr><tr><td>PM20D1</td></tr><tr><td>ELOVL3</td></tr><tr><td>KRT79</td></tr><tr><td>FADS2</td></tr><tr><td>GAL</td></tr><tr><td>AWAT2</td></tr><tr><td>AADACL3</td></tr><tr><td>DGAT2L6</td></tr></table>
+
+<--- Page Split --->
+
+Table S2 Drugs used for systemic treatment of AD in this study population.
+
+<table><tr><td>Drug name</td><td>Category</td></tr><tr><td>Cyclosporin A</td><td>Immunosuppressant</td></tr><tr><td>Bepotastine</td><td>Antiallergic drug</td></tr><tr><td>Bilrastine</td><td>Antiallergic drug</td></tr><tr><td>Clemastine Fumarate</td><td>Antiallergic drug</td></tr><tr><td>d-Chlorpheniramine Maleate</td><td>Antiallergic drug</td></tr><tr><td>Desloratadine</td><td>Antiallergic drug</td></tr><tr><td>Epinastine</td><td>Antiallergic drug</td></tr><tr><td>Fexofenadine</td><td>Antiallergic drug</td></tr><tr><td>Hydroxyzine</td><td>Antiallergic drug</td></tr><tr><td>Levocetirizine</td><td>Antiallergic drug</td></tr><tr><td>Olopatadine</td><td>Antiallergic drug</td></tr><tr><td>Rupatadine fumarate</td><td>Antiallergic drug</td></tr><tr><td>Suplatast tosylate</td><td>Antiallergic drug</td></tr></table>
+
+<--- Page Split --->
+
+Table S3 Characteristics of participants.
+
+<table><tr><td></td><td></td><td colspan="2">Cross-sectional</td><td>Longitudinal</td></tr><tr><td></td><td>Variable</td><td>AD</td><td>Healthy control</td><td>AD</td></tr><tr><td rowspan="4">General<br>information</td><td>Number of participants</td><td>115</td><td>14</td><td>30</td></tr><tr><td>Sex Male<br>Female</td><td>85(73.9%)<br>30(26.1%)</td><td>9(64.3%)<br>5(26.7%)</td><td>23(76.7%)<br>7(23.3%)</td></tr><tr><td>Age (years, mean±SD)</td><td>40.5±11.2</td><td>47.3±10.7</td><td>36.4±10.0</td></tr><tr><td>EASI (mean±SD)</td><td>17.7±10.8</td><td>-</td><td>14.0±7.96</td></tr><tr><td rowspan="3">Systemic<br>treatment</td><td>Immunosuppressant</td><td>3</td><td>-</td><td>9</td></tr><tr><td>Steroids</td><td>2</td><td>-</td><td>0</td></tr><tr><td>Antiallergic drugs</td><td>26</td><td>-</td><td>Constant:12<br>Occasional:2</td></tr></table>
+
+<--- Page Split --->
+
+Table S4 Combination of cytokine genes and receptor genes frequently observed in each participant group.
+
+<table><tr><td>Disease</td><td>Cytokine</td><td>Receptor</td><td>cytokine_type</td><td>receptor_type</td><td>#Connection</td><td>#Connection #Patients</td></tr><tr><td>AD</td><td>CCL22</td><td>CCR4</td><td>Skin</td><td>Skin</td><td>40</td><td>0.348</td></tr><tr><td>AD</td><td>CCL17</td><td>CCR4</td><td>Skin</td><td>Skin</td><td>39</td><td>0.339</td></tr><tr><td>AD</td><td>CCL13</td><td>CCR1</td><td>Skin</td><td>Skin</td><td>38</td><td>0.330</td></tr><tr><td>AD</td><td>CCL18</td><td>CCR8</td><td>Skin</td><td>Skin</td><td>38</td><td>0.330</td></tr><tr><td>AD</td><td>CCL19</td><td>CCR7</td><td>Skin</td><td>Skin</td><td>38</td><td>0.330</td></tr><tr><td>AD</td><td>IL13</td><td>IL4R</td><td>Skin</td><td>Skin</td><td>38</td><td>0.330</td></tr><tr><td>AD</td><td>1L13</td><td>IL13RA2</td><td>Skin</td><td>Skin</td><td>36</td><td>0.313</td></tr><tr><td>AD</td><td>IL36A</td><td>IL1RL2</td><td>Skin</td><td>Skin</td><td>36</td><td>0.313</td></tr><tr><td>A0</td><td>CXCL8</td><td>CXCR1</td><td>Skin</td><td>Skin</td><td>35</td><td>0.304</td></tr><tr><td>AD</td><td>CCL13</td><td>ACKR4</td><td>Skin</td><td>Skin</td><td>34</td><td>0.296</td></tr><tr><td>AD</td><td>CCL13</td><td>CCR2</td><td>Skin</td><td>Skin</td><td>34</td><td>0.296</td></tr><tr><td>AD</td><td>CCL2</td><td>CCR2</td><td>Skin</td><td>Skin</td><td>34</td><td>0.296</td><tr><td>AD</td><td>CXCL1</td><td>CXCR1</td><td>Skin</td><td>Skin</td><td>34</td><td>0.296</td></tr><tr><td>AD</td><td>IL36G</td><td>IL1RL2</td><td>Skin</td><td>Skin</td><td>34</td><td>0.296</td></tr><tr><td>AD</td><td>1L36RN</td><td>IL1RL2</td><td>Skin</td><td>Skin</td><td>34</td><td>0.29</td></tr><tr><td>AD</td><td>CCL18</td><td>CCR8</td><td>Skin</td><td>PBMC</td><td>33</td><td>0.287</td></tr><tr><td>AD</td><td>CCL2</td><td>CCR4</td><td>Skin</td><td>Skin</td><td>33</td><td>0.287</td></tr><tr><td>AD</td><td>CCL2</td><td>ACKR4</td><td>Skin</td><td>Skin</td><td>32</td><td>0.278</td></tr><tr><td>AD</td><td>IL20</td><td>IL20RB</td><td>Skin</td><td>PBMC</td><td>30</td><td>0.261</td></tr><tr><td>AD</td><td>CCL19</td><td>ACKR4</td><td>Skin</td><td>Skin</td><td>28</td><td>0.243</td></tr><tr><td>AD</td><td>IL13</td><td>IL4R</td><td>Skin</td><td>PBMC</td><td>28</td><td>0.243</td></tr><tr><td>AD</td><td>IL6</td><td>IL6R</td><td>Skin</td><td>Skin</td><td>28</td><td>0.243</td></tr><tr><td>AD</td><td>CXCL10</td><td>CXCR3</td><td>Skin</td><td>Skin</td><td>27</td><td>0.235</td></tr><tr><td>AD</td><td>IL20</td><td>IL20RB</td><td>Skin</td><td>Skin</td><td>27</td><td>0.235</td></tr><tr><td>AD</td><td>1L36G</td><td>IL1RAP</td><td>Skin</td><td>PBMC</td><td>27</td><td>0.235</td></tr><tr><td>AD</td><td>CCL1</td><td>CCR8</td><td>Skin</td><td>Skin</td><td>26</td><td>0.226</td></tr><tr><td>AD</td><td>CCL13</td><td>ACKR1</td><td>Skin</td><td>Skin</td><td>26</td><td>0.226</td></tr><tr><td>AD</td><td>CCL13</td><td>CCR3</td><td>Skin</td><td>Skin</td><td>26</td><td>0.226</td></tr><tr><td>AD</td><td>CCL17</td><td>CCR4</td><td>Skin</td><td>PBMC</td><td>26</td><td>0.226</td></tr><tr><td>AD</td><td>CCL22</td><td>CCR4</td><td>Skin</td><td>PBMC</td><td>26</td><td>0.22</td></tr><tr><td>AD</td><td>FPR3</td><td>ANXA1</td><td>Skin</td><td>Skin</td><td>26</td><td>0.226</td></tr><tr><td>AD</td><td>IL13</td><td>IL13RA1</td><td>Skin</td><td>Skin</td><td>26</td><td>0.226</td></tr><tr><td>A0</td><td>IL36A</td><td>IL1RAP</td><td>Skin</td><td>PBMC</td><td>26</td><td>0.226</td></tr><tr><td>AD</td><td>PBP</td><td>CXCR2</td><td>Skin</td><td>Skin</td><td>26</td><td>0.226</td></tr><tr><td>AD</td><td>CCL13</td><td>CCR1</td><td>Skin</td><td>PBMC</td><td>25</td><td>0.217</td></tr><tr><td>AD</td><td>CCL13</td><td>CCR3</td><td>Skin</td><td>PBMC</td><td>25</td><td>0.217</td></tr><tr><td>A0</td><td>CCL2</td><td>CCR2</td><td>Skin</td><td>PBMC</td><td>25</td><td>0.217</td></tr><tr><td>AD</td><td>CCL2</td><td>CCR4</td><td>Skin</td><td>PBMC</td><td>25</td><td>0.217</td></tr><tr><td>A0</td><td>CCL26</td><td>CCR3</td><td>Skin</td><td>Skin</td><td>25</td><td>0.217</td></tr><tr><td>AD</td><td>IL15</td><td>IL2RG</td><td>Skin</td><td>Skin</td><td>25</td><td>0.217</td></tr><tr><td>AD</td><td>1L36RN</td><td>IL1RAP</td><td>Skin</td><td>Skin</td><td>25</td><td>0.217</td></tr><tr><td>AD</td><td>IL6</td><td>IL6ST</td><td>Skin</td><td>Skin</td><td>25</td><td>0.217</td></tr><tr><td>AD</td><td>CXCL8</td><td>CXCR1</td><td>PBMC</td><td>PBMC</td><td>25</td><td>0.217</td></tr><tr><td>AD</td><td>CCL1</td><td>CCR8</td><td>Skin</td><td>PBMC</td><td>24</td><td>0.209</td></tr><tr><td>AD</td><td>CCL17</td><td>ACKR1</td><td>Skin</td><td>Skin</td><td>24</td><td>0.209</td></tr><tr><td>AD</td><td>CCL2</td><td>ACKR1</td><td>Skin</td><td>Skin</td><td>24</td><td>0.209</td><td></td></tr><tr><td>AD</td><td>CD40LG</td><td>CD40</td><td>Skin</td><td>Skin</td><td>24</td><td>0.209</td></tr><tr><td>AD</td><td>IL15</td><td>IL15RA</td><td>Skin</td><td>Skin</td><td>24</td><td>0.209</td></tr><tr><td>AD</td><td>1L36G</td><td>IL1RAP</td><td>Skin</td><td>Skin</td><td>24</td><td>0.209</td></tr><tr><td>AD</td><td>IL36RN</td><td>IL1RAP</td><td>Skin</td><td>PBMC</td><td>24</td><td>0.209</td></tr><tr><td>A0</td><td>IL6</td><td>IL6ST</td><td>Skin</td><td>PBMC</td><td>24</td><td>0.209</td></tr><tr><td>AD</td><td>CXCL14</td><td>CXCR4</td><td>PBMC</td><td>PBMC</td><td>24</td><td>0.209</td></tr><tr><td>AD</td><td>CCL8</td><td>CXCR1</td><td>PBMC</td><td>Skin</td><td>24</td><td>0.209</td></tr><tr><td>AD</td><td>CXCL8</td><td>CXCR2</td><td>PBMC</td><td>PBMC</td><td>24</td><td>0.209</td></tr><tr><td>AD</td><td>CCL13</td><td>CCR2</td><td>Skin</td><td>PBMC</td><td>23</td><td>0.200</td></tr><tr><td>AD</td><td>CCL21</td><td>CCR7</td><td>Skin</td><td>Skin</td><td>23</td><td>0.200</td></tr><tr><td>AD</td><td>IL20</td><td>IL22RA1</td><td>Skin</td><td>Skin</td><td>23</td><td>0.200</td></tr><tr><td>AD</td><td>OSM</td><td>OSMR</td><td>Skin</td><td>Skin</td><td>23</td><td>0.200</td></tr></table>
+
+<table><tr><td>Disease</td><td>Cytokine</td><td>Receptor</td><td>cytokine_type</td><td>receptor_type</td><td>#Connection</td><td>#Connection #Patients</td></tr><tr><td>normal</td><td>IL37</td><td>IL18R1</td><td>Skin</td><td>PBMC</td><td>13</td><td>0.922</td></tr><tr><td>normal</td><td>IL34</td><td>CSF1R</td><td>Skin</td><td>Skin</td><td>12</td><td>0.857</td></tr><tr><td>normal</td><td>IL37</td><td>IL18RAP</td><td>Skin</td><td>PBMC</td><td>11</td><td>0.786</td></tr><tr><td>normal</td><td>IL18</td><td>IL18R1</td><td>Skin</td><td>PBMC</td><td>8</td><td>0.571</td></tr><tr><td>normal</td><td>IL37</td><td>IL18R1</td><td>Skin</td><td>Skin</td><td>8</td><td>0.571</td></tr><tr><td>normal</td><td>CCL5</td><td>CCR5</td><td>PBMC</td><td>PBMC</td><td>7</td><td>0.500</td></tr><tr><td>normal</td><td>IL18</td><td>IL18RAP</td><td>Skin</td><td>PBMC</td><td>6</td><td>0.429</td></tr><tr><td>normal</td><td>IL1A</td><td>IL1R1</td><td>Skin</td><td>Skin</td><td>6</td><td>0.429</td></tr><tr><td>normal</td><td>IL37</td><td>IL18RAP</td><td>Skin</td><td>Skin</td><td>6</td><td>0.429</td></tr><tr><td>normal</td><td>CCL5</td><td>ACKR1</td><td>PBMC</td><td>Skin</td><td>6</td><td>0.429</td></tr><tr><td>normal</td><td>IL16</td><td>CD4</td><td>Skin</td><td>Skin</td><td>5</td><td>0.357</td></tr><tr><td>normal</td><td>IL18</td><td>IL18R1</td><td>Skin</td><td>Skin</td><td>5</td><td>0.357</td></tr><tr><td>normal</td><td>1L1A</td><td>IL1R2</td><td>Skin</td><td>PBMC</td><td>5</td><td>0.357</td></tr><tr><td>normal</td><td>IL1F10</td><td>IL1RL2</td><td>Skin</td><td>Skin</td><td>5</td><td>0.357</td></tr><tr><td>normal</td><td>CCL5</td><td>CCR3</td><td>PBMC</td><td>Skin</td><td>5</td><td>0.357</td></tr><tr><td>normal</td><td>CC12B</td><td>CCR3</td><td>Skin</td><td>Skin</td><td>4</td><td>0.286</td></tr><tr><td>normal</td><td>IL1A</td><td>IL1R2</td><td>Skin</td><td>Skin</td><td>4</td><td>0.286</td></tr><tr><td>normal</td><td>1L1A</td><td>IL1R1</td><td>Skin</td><td>PBMC</td><td>4</td><td>0.286</td></tr><tr><td>normal</td><td>LIF</td><td>IL6ST</td><td>Skin</td><td>Skin</td><td>4</td><td>0.286</td></tr><tr><td>normal</td><td>CCL4L1</td><td>CCR5</td><td>PBMC</td><td>PBMC</td><td>4</td><td>0.286</td></tr><tr><td>normal</td><td>CXCL5</td><td>CXCR1</td><td>PBMC</td><td>PBMC</td><td>4</td><td>0.286</td></tr><tr><td>norma1</td><td>CXCL2</td><td>XCR1</td><td>PBMC</td><td>Skin</td><td>4</td><td>0.286</td></tr><tr><td>normal</td><td>BMP5</td><td>BMPR1B</td><td>Skin</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>BMP5</td><td>BMPR1B</td><td> Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>BMP6</td><td>ACVR2A</td><td>Skin</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>BMP6</td><td>BMPR2</td><td>Skin</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>CCL21</td><td>ACKR4</td><td>Skin</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>CCL21</td><td>CCR7</td><td>Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>CXCL5</td><td>CXCR1</td><td>Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>norma1</td><td>CXCL5</td><td>CXCR2</td><td>Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>EPO</td><td>EPOR</td><td>Skin</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>IL17C</td><td>IL17RE</td><td>Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>IL18</td><td>IL18RAP</td><td>Skin</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>IL1A</td><td>IL1RAP</td><td>Skin</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>norma1</td><td>IL1F10</td><td>IL1RAP</td><td>Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>IL24</td><td>IL20RB</td><td>Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>norma1</td><td>IL33</td><td>IL1RL1</td><td>Skin</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>IL33</td><td>IL1RAP</td><td>Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>IL34</td><td>CSF1R</td><td>Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>norma1</td><td>BMP6</td><td>ACVR2A</td><td>PBMC</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>BMP6</td><td>BMPR1A</td><td>PBMC</td><td>Skin</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>CXCL5</td><td>CXCR2</td><td>PBMC</td><td>PBMC</td><td>3</td><td>0.214</td></tr><tr><td>normal</td><td>PBPB</td><td>CXCR2</td><td>PBMC</td><td>PBMC</td><td>3</td><td>0,214</td></tr><tr><td>normal</td><td>XCL1</td><td>XCR1</td><td>Skin</td><td>PBMC</td><td>3</td><td>0.214</td></tr></table>
+
+<--- Page Split --->
+
+Table S5 Summary of cell types highly involved in cytokine-receptor coupling in each participant group.
+
+<table><tr><td rowspan="2"></td><td colspan="3">Cytokine</td><td colspan="3">Receptor</td></tr><tr><td>Tissue</td><td>Cell type</td><td>#Connection</td><td>Tissue</td><td>Cell type</td><td>#Connection</td></tr><tr><td rowspan="10">AD</td><td>Skin</td><td>Tcell</td><td>119</td><td>PBMC</td><td>non-specific</td><td>116</td></tr><tr><td>Skin</td><td>VEC</td><td>90</td><td>Skin</td><td>Myeloids</td><td>87</td></tr><tr><td>PBMC</td><td>non-specific</td><td>58</td><td>Skin</td><td>Tcell</td><td>69</td></tr><tr><td>Skin</td><td>Myeloids</td><td>57</td><td>Skin</td><td>VEC</td><td>57</td></tr><tr><td>PBMC</td><td>Mono</td><td>56</td><td>PBMC</td><td>Mono</td><td>39</td></tr><tr><td>Skin</td><td>KC</td><td>56</td><td>Skin</td><td>FB</td><td>37</td></tr><tr><td>Skin</td><td>non-specific</td><td>46</td><td>Skin</td><td>non-specific</td><td>34</td></tr><tr><td>Skin</td><td>vSMC</td><td>23</td><td>PBMC</td><td>Treg</td><td>25</td></tr><tr><td>Skin</td><td>FB</td><td>18</td><td>Skin</td><td>KC</td><td>20</td></tr><tr><td>PBMC</td><td>DC_mye</td><td>16</td><td>PBMC</td><td>DC_mye</td><td>18</td></tr><tr><td>PBMC</td><td>NK</td><td>14</td><td>PBMC</td><td>CD4T</td><td>17</td></tr><tr><td>PBMC</td><td>CD8T</td><td>13</td><td>PBMC</td><td>T_MAIT</td><td>16</td></tr><tr><td>Skin</td><td>Sweat</td><td>9</td><td>PBMC</td><td>NK</td><td>14</td></tr><tr><td rowspan="10">Healthy<br>control</td><td>Skin</td><td>Tcell</td><td>33</td><td>PBMC</td><td>non-specific</td><td>44</td></tr><tr><td>Skin</td><td>VEC</td><td>29</td><td>Skin</td><td>Myeloids</td><td>35</td></tr><tr><td>PBMC</td><td>non-specific</td><td>27</td><td>Skin</td><td>VEC</td><td>19</td></tr><tr><td>Skin</td><td>non-specific</td><td>27</td><td>Skin</td><td>Tcell</td><td>17</td></tr><tr><td>Skin</td><td>KC</td><td>21</td><td>PBMC</td><td>Mono</td><td>12</td></tr><tr><td>Skin</td><td>Myeloids</td><td>21</td><td>Skin</td><td>FB</td><td>11</td></tr><tr><td>PBMC</td><td>Mono</td><td>18</td><td>Skin</td><td>non-specific</td><td>11</td></tr><tr><td>PBMC</td><td>CD8T</td><td>8</td><td>PBMC</td><td>DC_mye</td><td>10</td></tr><tr><td>Skin</td><td>FB</td><td>8</td><td>Skin</td><td>KC</td><td>10</td></tr><tr><td>PBMC</td><td>DC_mye</td><td>7</td><td>PBMC</td><td>Treg</td><td>8</td></tr><tr><td>Skin</td><td>Sweat</td><td>6</td><td>Skin</td><td>Melano</td><td>8</td></tr><tr><td>PBMC</td><td>NK</td><td>5</td><td>PBMC</td><td>DC_pls</td><td>7</td></tr><tr><td>Skin</td><td>vSMC</td><td>5</td><td>PBMC</td><td>T_MAIT</td><td>7</td></tr></table>
+
+<--- Page Split --->
+
+Table S6 Antibodies used for immunohistochemistry.
+
+<table><tr><td>Antibody</td><td>Manufacturer</td><td>Species</td><td>Clone</td><td>Isotype</td><td>Dilution</td></tr><tr><td>CD4</td><td>Novus biologicals</td><td>Rabbit</td><td>13B8.2</td><td>IgG</td><td>1/500</td></tr><tr><td>Myeloperoxidase</td><td>Dako</td><td>Rabbit</td><td>Polyclonal</td><td>Polyclonal</td><td>1/1000</td></tr><tr><td>Major basic protein</td><td>Bio-Rad</td><td>Mouse</td><td>BMK-13</td><td>IgG1</td><td>1/200</td></tr><tr><td>CD206</td><td>Novus biologicals</td><td>Mouse</td><td>15-2</td><td>IgG1</td><td>1/100</td></tr><tr><td>CD11c</td><td>BD</td><td>Mouse</td><td>B-ly6</td><td>IgG1</td><td>1/10</td></tr><tr><td>CD1a</td><td>Novus biologicals</td><td>Mouse</td><td>O10+C1A/711</td><td>IgG1</td><td>1/200</td></tr><tr><td>Keratin 16</td><td>LabVision</td><td>Mouse</td><td>LL025</td><td>IgG1</td><td>1/200</td></tr><tr><td>Filaggrin</td><td>GeneTex</td><td>Mouse</td><td>FLG01</td><td>IgG1</td><td>1/100</td></tr></table>
+
+<--- Page Split --->
+![](images/Figure_unknown_10.jpg)
+
+<center>Fig. S1 Skin sample filtering by expression of hair follicle-related genes. UMAP plot of expression of pilosebaceous unit-related genes in skin samples from both AD patients and healthy controls. Sixty-five skin samples (AD: 45, healthy controls: 20) forming a cluster with extremely strong signatures of the hair follicle gene set were considered to be occupied by hair follicles along with incidental sebaceous glands, and therefore excluded from this study. </center>
+
+<--- Page Split --->
+![](images/Figure_unknown_11.jpg)
+
+<center>Fig. S2 Identification of the optimal number of clusters for individual EASI partial scores. a,b. Silhouette width plot (a) and elbow plot (b) for identifying the optimal number of clusters for the correlation matrix of EASI partial score. The X-axes indicate the number of clusters ranging from 1 to 10 and the Y-axes indicate average silhouette width (a) or total within sum of square (b). Both of a highest average silhouette coefficient in the silhouette plot and a large drop in the elbow plot were observed at the number of clusters of 2. </center>
+
+<--- Page Split --->
+![PLACEHOLDER_61_0]
+
+
+<--- Page Split --->
+![PLACEHOLDER_62_0]
+
+<center>Fig. S3 Immunohistochemi I analysis revealed shared and differential characteristics in the skin tissue of erythema and papulation-skewed AD patients. Immunohistochemistry of skin tissue in two representative AD patients who have a score composition that are highly skewed to either of erythema (upper) or induration/papulation (lower) as well as healthy control. Left: a 51-year-old male patient who has erythema-skewed EASI composition. Middle: a 50-year-old male patient who has papulation-skewed EASI composition (total = 21.0, erythema = 3.0, papulation = 8.4). Right: a 52-year-old female who have no eczema nor skin diseases. Target proteins were stained in red. Bars = 100 μm. MPO: myeloperoxidase, MBP: major basic protein, K16: keratin-16, FLG: filaggrin. </center>
+
+<--- Page Split --->
+![PLACEHOLDER_63_0]
+
+<center>Fig. S4 Statistics of transcriptome modules associated with AD. Variance explained by PC1-PC20 in PCA on expression level of transcriptome modules in skin (upper) and PPBMC (lower) across patients. </center>
+
+<--- Page Split --->
+![PLACEHOLDER_64_0]
+
+<center>Fig. S5 Prediction accuracy of personalized disease course. The histogram indicates the distribution of Pearson correlation coefficient between observed EASI and predicted EASI in intra-patient dynamics. There was no significant difference in the size of the coefficient among patients regarding treatment classes (Kruskal-Wallis rank sum test, \(p = 0.57\) ). </center>
+
+<--- Page Split --->
+![PLACEHOLDER_65_0]
+
+<center>Fig. S6 Association of omics features with clinical severity in a longitudinal setting. Trajectories of observed/predicted EASI (upper) and intensity of blood examination and PBMC transcriptome modules (lower) in one year in two representative patients. </center>
+
+<--- Page Split --->
+![PLACEHOLDER_66_0]
+
+<center>Fig. S7 Time series features in 1-year monitoring of 30 AD patients. Hierarchical clustering of 7 types of time series features of clinical severity. RMS: root mean square, MAC: mean absolute change, CID: complexity-invariant distance. </center>
+
+<--- Page Split --->
+![PLACEHOLDER_67_0]
+
+<center>Fig. S8 Comparison of time series features in top contributing factors among three patient clusters. Two classes of time series features: mean (upper) and mean absolute change (MAC, lower) of top contributing factors in principal component analysis of time series features across patients. Boxplots show median and first and third quartiles, whiskers extending to the highest and lowest values no further than 1.5"interquartile range. Brunner-Munzel rank test. Multiple comparison tests were carried out using Kruskal-Wallis test. P-values less than 0.05 were considered as significant and subsequently tested for post-hoc comparison with Brunner-Munzel test with \(p\) -value correction by Holm's method. NS: not significant, \(^{*}p < 0.05\) , \(^{**}p < 0.01\) , \(^{***}p < 0.001\) . </center>
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+- nreditorialpolicychecklistNCOMMS2237312T.pdf- nrreportingsummaryNCOMMS2237312T.pdf
+
+<--- Page Split --->

preprint/preprint__2b38d2667b0ed62d397a8a20a8d846f75734d0824b7f39e0811c05ef40cb6eba/images_list.json

@@ -0,0 +1,145 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1. Developmental Trajectories of Aperiodic and Periodic Power Spectra (A) Example aperiodic power spectra. Offset is defined as power at 2.5Hz. (B) Example periodic power spectra. Peaks defined as maxima within a defined frequency range. Band power defined as the integral of the periodic power spectra between defined frequencies. (C) Longitudinal study enrollment. (D-F) Absolute, Aperiodic, and Periodic power spectra averaged across individuals within 8 age bins between 2 and 44 months. (G) Age-related changes in periodic power. (H-I) GAMMs modeled trajectories of aperiodic offset and slope for males (orange) and females (blue). Relative inflection points are shown with circular markers. Below, heatmaps show the standardized change in offset or slope per day, defined as [change per day]/ [standard deviation of measures across full age range]",
+    "footnote": [],
+    "bbox": [
+      [
+        97,
+        60,
+        889,
+        662
+      ]
+    ],
+    "page_idx": 7
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2. Transient and nonlinear changes in periodic power between 4 and 12 Hz. (A,B) Individual periodic power spectra for 2-4 months, and 6-8 months old. Red markers show peaks between 4 and 12 Hz. (C) Proportion of infants with two peaks identified between 4 and 12 Hz at each age bin. (D) Proportion of infants with an identified peak between 6.5 and 12Hz at each age bin. (E-I) GAMMs modeled trajectories for males (orange) and females (blue). Relative inflection points are shown with circular markers. Below, heatmaps show the standardized change in offset or slope per day, defined as [change per day]/[standard deviation of measures across full age range]. Both male and female heatmaps shown for models with significant age x sex interaction.",
+    "footnote": [],
+    "bbox": [
+      [
+        92,
+        65,
+        802,
+        715
+      ]
+    ],
+    "page_idx": 9
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3. Transient and nonlinear changes in periodic power between 12 and 35Hz. (A,B) Individual periodic power spectra 6-8 months, and 18-20 months old. Red markers show peaks between 12-20 Hz. Green markers show peaks between 20-35Hz. (C) Proportion of infants with peak identified between 12-20 Hz at each age bin. (D-H) GAMMS modeled trajectories for males (orange) and females (blue). Relative inflection points are shown with circular markers. Below, heatmaps show the standardized change in offset or slope per day, defined as [change per day]/[standard deviation of measures across full age range].",
+    "footnote": [],
+    "bbox": [
+      [
+        88,
+        60,
+        852,
+        732
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4. Increased anesthesia-induced alpha coherence in infants with identified low beta peak in baseline EEG. (A) Periodic power spectra of infants between 6 and 15 months old prior to receiving anesthesia (B) Mean alpha coherence during anesthesia in infants 7-12 months old, with (light blue) or without (white) an identified low beta peak. Ancova with sevoflurane levels as covariate: \\(\\mathrm{p}< 0.05\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        96,
+        75,
+        845,
+        355
+      ]
+    ],
+    "page_idx": 13
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5. FXS children have increased high beta peak that decreases with age. (A) Periodic power spectra of 35-48 month-old children with and without FXS. (B) Individual periodic power spectra of FXS children, with line hue corresponding to age in months.",
+    "footnote": [],
+    "bbox": [
+      [
+        93,
+        70,
+        682,
+        290
+      ]
+    ],
+    "page_idx": 17
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Supplemental Figure 2: Individual plots of the periodic spectrum averaged across the whole ROI for each age bin.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 32
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Supplemental Figure 3: GAMMs modeled trajectories of 8 power measures, with ROI, study, smoothed age, and sex as predictor terms.",
+    "footnote": [],
+    "bbox": [
+      [
+        91,
+        137,
+        890,
+        790
+      ]
+    ],
+    "page_idx": 33
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "A. Unedited FOOOF model estimates - Original Spectrum - FOOOF modeled Spectrum - Aperiodic Fit",
+    "footnote": [],
+    "bbox": [
+      [
+        84,
+        150,
+        895,
+        400
+      ]
+    ],
+    "page_idx": 34
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "B. Modified FOOOF model estimates - Original Spectrum - FOOOF modeled Spectrum - Aperiodic Fit",
+    "footnote": [],
+    "bbox": [
+      [
+        95,
+        440,
+        901,
+        692
+      ]
+    ],
+    "page_idx": 35
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "C. Original vs Edited FOOOF error",
+    "footnote": [],
+    "bbox": [
+      [
+        110,
+        734,
+        721,
+        880
+      ]
+    ],
+    "page_idx": 35
+  }
+]

preprint/preprint__2b38d2667b0ed62d397a8a20a8d846f75734d0824b7f39e0811c05ef40cb6eba/preprint__2b38d2667b0ed62d397a8a20a8d846f75734d0824b7f39e0811c05ef40cb6eba.mmd

@@ -0,0 +1,395 @@
+
+# Developmental trajectories of EEG aperiodic and periodic power: Implications for understanding the timing of thalamocortical development during infancy.  
+
+Carol Wilkinson  
+
+carol.wilkinson@childrens.harvard.edu  
+
+Boston Children's Hospital https://orcid.org/0000- 0002- 4694- 8564  
+
+Lisa Yankowitz Boston Children's Hospital  
+
+Jerry Chao Montefiore Medical Center  
+
+Rodrigo Gutierrez Massachusetts General Hospital  
+
+Jeff Rhoades Harvard University  
+
+Shlomo Shinnar Montefiore Medical Center  
+
+Patrick Purdon Massachusetts General Hospital, Harvard Medical School https://orcid.org/0000- 0003- 0080- 3340  
+
+Charles Nelson Harvard Medical School  
+
+## Article  
+
+Keywords:  
+
+Posted Date: September 18th, 2023  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 3215728/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.
+
+<--- Page Split --->
+
+Version of Record: A version of this preprint was published at Nature Communications on July 10th, 2024. See the published version at https://doi.org/10.1038/s41467-024-50204-4.
+
+<--- Page Split --->
+
+Title: Developmental trajectories of EEG aperiodic and periodic power: Implications for understanding the timing of thalamocortical development during infancy.  
+
+Authors: Carol L. Wilkinson\*1,2, Lisa Yankowitz\*1, Jerry Y. Chao3, Rodrigo Gutiérrez4,9, Jeff L. Rhoades5,6, Shlomo Shinnar7,8, Patrick L. Purdon9, Charles A. Nelson1,2,10  
+
+\*Designates co- first authors  
+
+## Affiliations  
+
+1 Division of Developmental Medicine, Boston Children's Hospital, Boston, MA, United States.  2 Harvard Medical School, Boston, MA, USA  3 Department of Anesthesiology, Montefiore Medical Center, Children's Hospital at Montefiore, Albert Einstein College of Medicine, Bronx, NY, USA.  4 Centro de Investigación Clínica Avanzada, Hospital Clínico de la Universidad de Chile, Santiago, Chile.  5 Department of Neurobiology, Harvard Medical School, Boston, MA, USA.  6 Program in Neuroscience, Division of Medical Sciences, Graduate School of Arts and Sciences, Harvard University, Cambridge, MA, USA.  7 The Saul R. Korey Department of Neurology, Montefiore Medical Center, Albert Einstein College of Medicine, Bronx, NY, USA.  8 Department of Epidemiology and Population Health, Albert Einstein College of Medicine, Bronx, NY, USA.  9 Department of Anesthesia, Critical Care and Pain Medicine, Massachusetts General Hospital, Boston, MA, United States.  10 Harvard Graduate School of Education, Cambridge, MA, United States.
+
+<--- Page Split --->
+
+## Abstract  
+
+The development of neural circuits over the first years of life has long- lasting effects on brain function, yet our understanding of early circuit development in humans remains limited. Here, aperiodic and periodic EEG power features were examined from longitudinal EEGs collected from 592 healthy 2–44 month- old infants, revealing age- dependent nonlinear changes suggestive of distinct milestones in early brain maturation. Consistent with the transient developmental progression of thalamocortical circuitry, we observe the presence and then absence of periodic alpha and high beta peaks across the three- year period, as well as the emergence of a low beta peak (12- 20Hz) after six months of age. We present preliminary evidence that the emergence of the low beta peak is associated with thalamocortical connectivity sufficient for anesthesia- induced alpha coherence. Together, these findings suggest that early age- dependent changes in alpha and beta periodic peaks may reflect the state of thalamocortical network development.
+
+<--- Page Split --->
+
+## Introduction  
+
+The infant brain undergoes dramatic structural and physiological change in the first year after birth. Rapid increases in brain volume coincide with expansive synaptogenesis \(^{1 - 3}\) , as well as interneuron migration, maturation and network integration \(^{4}\) . In particular, during this early period thalamocortical connections are established through an intricately choreographed sequence that plays a critical role in the development of sensory cortical networks \(^{5}\) . However, the detailed timing of interneuron and thalamocortical maturation in human development is largely unknown. In rodent models, the development of thalamocortical circuitry is notable for transient inhibitory connections that drive subsequent circuit formation and coincide with critical periods of plasticity present during the first 2- 3 postnatal weeks \(^{6}\) . In humans, longitudinal resting- state fMRI data suggest that while thalamus- sensorimotor connectivity networks are present at birth, other networks (e.g. thalamus- medial- visual, thalamus- default- mode) do not emerge until 1 year of age \(^{7}\) . However, MRI studies thus far have been limited to measuring annual changes in structural or functional connectivity, preventing a detailed understanding of rapid developmental change during this period. In contrast, electroencephalography (EEG) can provide frequent and non- invasive repeated measurement of brain oscillations that directly result from transient developmental changes in inhibitory networks and maturation of thalamocortical circuitry \(^{5,8,9}\) .  
+
+The EEG power spectrum is comprised of two physiologically distinct components reflecting underlying neuronal activity: aperiodic and periodic power. The aperiodic component defines the slope of the power spectrum, following a 1/f power law distribution (Fig 1A) and reflects non- oscillatory neuronal spiking activity. In addition, the aperiodic slope has been linked to the excitatory- inhibitory (E/I) balance of the underlying neuronal network, where a flattened, reduced slope is associated with increased excitation over inhibition, and a steeply more accelerated slope with increased inhibition over excitation \(^{10}\) . Longitudinal studies of child- to adulthood have observed decreases in aperiodic slope with age, suggestive of increases in E/I balance with age \(^{11 - 14}\) . Changes in the aperiodic component in early infancy are less well described, and we hypothesize they may be substantially different from those in childhood, as the first year after birth includes rapid increases in neuronal activity, synaptogenesis, and inhibitory neuron integration.  
+
+The periodic component of the power spectrum is defined as the portion of the absolute power spectrum rising above the aperiodic slope (Fig 1B). Periodic power reflects oscillatory activity occurring in narrow frequency bands that are highly correlated with various cognitive processes and behavioral states \(^{15,16}\) , and provide the foundation for both local and long- range communication within the brain. The majority of
+
+<--- Page Split --->
+
+neural oscillations observed in the power spectrum are the direct result of inhibitory and thalamocortical network responses to sensory input. Thus, as a measure, the EEG power spectrum is well positioned to shed light on the developmental timing of inhibitory and thalamocortical network maturation.  
+
+Thus far, developmental EEG studies have largely focused on theta/alpha oscillations which are modulated by thalamocortical interactions and are associated with cognitive functions of attention and memory \(^{17 - 19}\) . Multiple studies of the first two years of life have observed a shift in alpha peak frequency from \(5 \text{Hz}\) at 5 months to \(8 \text{Hz}\) at 2 years, coinciding with increases in alpha power across this period \(^{20 - 23}\) . This dominant peak frequency continues to increase into the mature \(10 \text{Hz}\) posterior alpha rhythm by adolescence \(^{9,24}\) . It is hypothesized that the gradual shift in dominant peak frequency is modulated by maturation of thalamocortical circuitry in concert with developmental gains in cognitive functions \(^{9}\) , however, the precise mechanisms remain unknown.  
+
+Unlike theta/alpha power, little is known about the early developmental changes in periodic beta power. In adults, beta oscillations are strongly associated with sensorimotor processing in addition to higher- order cognitive tasks such as working memory \(^{25}\) . Similar to alpha oscillations, the generation of beta oscillations relies on GABAergic interneuron networks and thalamocortical connectivity. In adults, low- dose GABA- modulating anesthetics induce a sedative state with 13- 25Hz beta oscillations, whereas higher doses used to maintain unconsciousness progressively slow these beta oscillations into coherent, frontal specific, alpha oscillations \(^{26 - 28}\) . However, GABA- dependent anesthesia- induced frontal alpha coherence does not emerge in infants until after 10 months of age and is not consistently present until 15- 20 months of age \(^{29 - 31}\) . Anesthesia- induced alpha coherence is hypothesized to involve GABA- dependent thalamocortical loops leading to hypersynchronization between thalamic and prefrontal cortices \(^{27,32,33}\) . Therefore, potential covariation of developing beta oscillations and anesthesia- induced counterparts may lend insight into the role and time course of developing inhibition in human thalamocortical circuit development.  
+
+Using longitudinal EEG data collected from 592 healthy infants (yielding a total of 1335 EEGs) from 2 to 44 months after birth, we characterize early developmental trajectories of EEG aperiodic and periodic power from 2- 50Hz and to identify potential ages relevant to sequential steps in inhibitory network and thalamocortical circuit development. Consistent with the transient and stepwise developmental progression of thalamocortical circuitry, we observe transient periodic peaks in alpha power at 2- 3 months and high beta power at 4- 18 months. A low beta peak (12- 20Hz) also begins to emerge in infants starting as early as 6 months of age. We hypothesized that emergence of this low beta peak reflects maturation of early connections between the thalamus and cortex. To test this hypothesis, we leveraged a smaller dataset consisting of a cohort of infants with EEG recordings before and during clinical
+
+<--- Page Split --->
+
+anesthesia. Consistent with our hypothesis, we find infants with an identifiable low beta peak have higher anesthesia- induced alpha coherence than those that do not, suggesting that the emergence of this peak is associated with thalamocortical loop maturation.  
+
+## RESULTS  
+
+Resting- state EEG were collected longitudinally from 592 healthy infants, aged 2- 44 months, across 4 studies occurring in the same laboratory (Fig 1C, Table 1). Whole brain power spectra for each individual were calculated by averaging across electrodes (Supplemental Figure 1. Individual spectra shown in Supplemental Figure 2). Spectra were then averaged across individuals within 8 age bins (Fig 1D). Notable nonlinear changes in aperiodic and periodic power spectra were observed between age bins, including transient peaks in the periodic spectrum across both alpha and beta frequency ranges (Fig 1E- G). To further characterize these developmental changes in the spectra, we used generalized additive mixed models (GAMMs) to model non- linear trajectories of power parameters. For each model an age- by- sex interaction was tested for significance. If not significant, the interaction term was removed and the model was refit using sex as an additive covariate. All models also included study as a covariate factor.  
+
+Aperiodic power increases most during first year of life  
+
+First, we assessed age- dependent changes in the aperiodic component and observed the largest developmental increases in aperiodic power between 2 and 8 months after birth (Fig 1E). Aperiodic offset, but not slope, significantly increased with age (FDR- adjusted \(q\) value \(< 0.01\) ), and age- by- sex interactions were present for both aperiodic offset ( \(F = 4.59\) , \(q = 0.01\) ) and slope ( \(F = 3.18\) , \(q = 0.02\) ). Modeled developmental trajectories of the aperiodic offset showed a sharp linear increase over the first year after birth for both males and females (Fig 1H). Modeled developmental trajectories of the aperiodic slope showed a gradual increase over the first year. These findings contrast with consistent reports of decreasing offset and slope across child and adulthood \(^{11 - 14}\) , and likely reflect the known increases in brain volume and synaptogenesis occurring across the first year of life. Differences in developmental trajectories between 4 regions of interest (ROI) (frontal, central, temporal, and posterior) were also assessed (Supplemental Figure 3). The posterior ROI had higher aperiodic offset than all other ROIs, with the greatest increase in offset occurring in the first year (Frontal \(F = - 26.32\) , \(q < 0.0001\) , Central \(F = - 40.63\) , \(q < 0.0001\) , Temporal - 31.85 , \(q < 0.0001\) ; Supplemental Fig 3B).
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1. Developmental Trajectories of Aperiodic and Periodic Power Spectra (A) Example aperiodic power spectra. Offset is defined as power at 2.5Hz. (B) Example periodic power spectra. Peaks defined as maxima within a defined frequency range. Band power defined as the integral of the periodic power spectra between defined frequencies. (C) Longitudinal study enrollment. (D-F) Absolute, Aperiodic, and Periodic power spectra averaged across individuals within 8 age bins between 2 and 44 months. (G) Age-related changes in periodic power. (H-I) GAMMs modeled trajectories of aperiodic offset and slope for males (orange) and females (blue). Relative inflection points are shown with circular markers. Below, heatmaps show the standardized change in offset or slope per day, defined as [change per day]/ [standard deviation of measures across full age range] </center>
+
+<--- Page Split --->
+
+Transient 9.5Hz alpha peak observed in 2- 4 month- old infantsAt the youngest age bin (2- 4 months) two peaks with similar amplitude are observed across the theta/alpha (4- 12Hz) range in the majority of infants (69%; Fig 2A,C). A lower frequency peak is observed in the theta (4- 6Hz) range at \(5.5 \pm 0.3\mathrm{Hz}\) , and higher frequency peak is observed in the alpha (6- 12Hz) range at \(9.5 \pm 0.45\mathrm{Hz}\) . However, by 6- months only 15% of infants have two peaks in this range, and for most infants it is the higher 9.5Hz peak that is no longer observed. At 6- months fewer than 40% of infants exhibit a dominant peak in the “alpha” (6.5- 12Hz) range (Fig 1D) and the average peak frequency in the theta/alpha range is \(6.3 \pm 1\mathrm{Hz}\) . This disappearance of the higher peak after 4- months of age may reflect a transient step in thalamocortical circuitry development. Previous research has observed a gradual shift in peak frequency from 5 to 8Hz from infancy to early childhood, however these studies started no earlier than 5 months of age \(^{20 - 23}\) . In order to assess whether an increase in peak frequency beginning at 5 months is present in our data set we modeled developmental trajectories of peak amplitude and frequency between 4- 12 Hz starting at 170 days, when the vast majority of EEGs exhibited a single dominant peak. No age- by- sex interactions were observed in GAMMs modeled trajectories, and consistent with previous studies peak frequency and peak amplitude significantly increased with age (frequency: \(\mathrm{F} = 12.9\) , q <0.0001 Fig 2E; amplitude: \(\mathrm{F} = 16.78\) , q < 0.0001, Fig 2F). Figure 2G- I show modeled trajectories for EEG power calculated over defined frequency bands commonly used in infant EEG research: theta (4- 6Hz), low alpha (6- 9Hz), and high alpha (9- 12Hz). An age- by- sex interaction was present for theta power, although qualitatively the shapes of trajectories were similar (Fig G; \(\mathrm{F} = 4.17\) , q < 0.01).
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure 2. Transient and nonlinear changes in periodic power between 4 and 12 Hz. (A,B) Individual periodic power spectra for 2-4 months, and 6-8 months old. Red markers show peaks between 4 and 12 Hz. (C) Proportion of infants with two peaks identified between 4 and 12 Hz at each age bin. (D) Proportion of infants with an identified peak between 6.5 and 12Hz at each age bin. (E-I) GAMMs modeled trajectories for males (orange) and females (blue). Relative inflection points are shown with circular markers. Below, heatmaps show the standardized change in offset or slope per day, defined as [change per day]/[standard deviation of measures across full age range]. Both male and female heatmaps shown for models with significant age x sex interaction. </center>
+
+<--- Page Split --->
+
+Several age- dependent transient changes are observed in the low beta (13- 20Hz) and high beta (2035Hz) range. First, the shape of the periodic power spectra in the low beta range is notable for a prominent trough prior to 1 year of age (Fig 3A), with only \(10\%\) of infants (24/222) exhibiting a low beta peak between 6- 8 months of age (Fig 3A- C). After 8 months, a low beta peak begins to emerge in some of infants, with \(48\%\) (52/107) showing a peak at 18- 20 months, and \(70\%\) (199/285) by 36 months (Fig 2C). As a low beta peak was not identified in many children across the age range, peak amplitude and frequency was not modeled. In contrast, virtually all \((99.5\%)\) of the infants had an identifiable high beta peak prior to 12 months of age (Fig 3A). However, notable nonlinear shifts in frequency and amplitude of the high beta peak were observed (Fig 1F, 3D and E). During the first year after birth, the high beta peak amplitude rapidly increases, peaking at 229 days (7.5 months), and then substantially decreases until 802 days(2.2 years). High beta peak frequency trajectories are also nonlinear, with peak frequency at its highest at 473 days (male 29.0Hz, female 29.4Hz), followed by a steady decline in frequency. Modeled trajectories of periodic power for commonly used frequency bands are shown in Fig 3F- H: low beta (12- 20Hz), high beta (20- 30Hz), and gamma (30- 45Hz).  
+
+The observed nonlinear changes across the beta range are striking. While many EEG infant studies group beta oscillations into a singular frequency range, the data presented here supports that low and high beta have distinct developmental origins. Specifically, between 6- 24 months we observe the gradual emergence of a low beta peak, and simultaneously the rise and fall of a prominent high beta peak, ultimately resolving into a broader beta peak by 36 months.  
+
+Traditionally, beta oscillations measured in children and adults are associated with sensory and motor processing, where reductions in beta power are observed during the preparation or execution of motor tasks<sup>25</sup>. However, beta activity has also been shown to be modulated during a wide range of nonmotor cognitive tasks<sup>25,34</sup>. The developmental emergence of low beta oscillations may represent sensorimotor skills (e.g., crawling, walking) gained during this period, but may also represent the developmental maturation of neurobiological circuitry. For example, GABAergic interneuron networks and thalamocortical connectivity are highly associated with the generation of cortical beta oscillations, as well as anesthesia- induced frontal alpha coherence, but neither are not fully established at birth<sup>29</sup>.
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Figure 3. Transient and nonlinear changes in periodic power between 12 and 35Hz. (A,B) Individual periodic power spectra 6-8 months, and 18-20 months old. Red markers show peaks between 12-20 Hz. Green markers show peaks between 20-35Hz. (C) Proportion of infants with peak identified between 12-20 Hz at each age bin. (D-H) GAMMS modeled trajectories for males (orange) and females (blue). Relative inflection points are shown with circular markers. Below, heatmaps show the standardized change in offset or slope per day, defined as [change per day]/[standard deviation of measures across full age range]. </center>
+
+<--- Page Split --->
+
+1 Low beta peak associated with developmental emergence of anesthesia- induced frontal alpha coherence in infants 2 We hypothesized that developmental changes in infant beta power measured in a resting state may represent concurrent maturation of GABAergic interneuron networks and thalamocortical connectivity. To explore this possibility, we assessed EEG recordings of healthy infants before and during exposure to GABA- modulating sevoflurane anesthesia<sup>35</sup>. All infants were undergoing elective procedures (eg. circumcision) and infants were excluded for prematurity, neurologic injury, epilepsy, or planned intracranial surgery. Here we hypothesized that the emergence of low beta oscillations (as measured by the presence of a low beta peak) before anesthesia would be associated with GABA- dependent anesthesia- induced frontal alpha coherence. EEG data from 36 infants (6- 15 months old), collected during the awake and anesthetized state were analyzed. Developmental changes in the aperiodic- adjusted power spectra in this smaller dataset were qualitatively similar to those described above (Figure 4A), with a low beta peak beginning to emerge after 7 months and present in roughly half the infants between 7- 12 months of age (11/21). As hypothesized, alpha coherence was significantly increased in those with a low beta peak compared to those without (ANCOVA, with sevoflurane level as covariate; p <0.05; Figure 4B).
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Figure 4. Increased anesthesia-induced alpha coherence in infants with identified low beta peak in baseline EEG. (A) Periodic power spectra of infants between 6 and 15 months old prior to receiving anesthesia (B) Mean alpha coherence during anesthesia in infants 7-12 months old, with (light blue) or without (white) an identified low beta peak. Ancova with sevoflurane levels as covariate: \(\mathrm{p}< 0.05\) . </center>  
+
+## Discussion  
+
+Here we present the largest- to- date longitudinal analysis of EEG data collected between 2 – 44 months of age. Findings provide insight into the developmental timing of inhibitory network and thalamocortical circuit maturation during human infancy. Several age- dependent findings in our study contrast to previous longitudinal studies of child and adulthood. First, we observe increases in both aperiodic offset and slope, especially during the first year, whereas decreases in both measures are observed starting as early as 4 years of age and continue to decrease with adulthood<sup>11–14</sup>. Second, while expected shifts in the dominant peak from the theta to alpha range were observed between 5 to 44 months, in the 2- 4 months age bin, a 9.5Hz peak was also transiently observed. Third, striking changes within the beta (12- 30Hz) range were observed, including the emergence of a low beta peak starting after 6- months of age, and age- dependent shifts in high beta peak frequency and amplitude - first increasing and then decreasing with age. Below we discuss how the above age- related changes may represent sequential developmental maturation in the inhibitory system and thalamocortical network connections.  
+
+The aperiodic offset is hypothesized to represent broad band neuronal firing<sup>36,37</sup>, and thus early increases in aperiodic offset are consistent with established increases in neuronal number, gray matter
+
+<--- Page Split --->
+
+volume, and synaptic number during the first year. Stabilization of aperiodic offset after 1- year of age is also consistent with MRI findings that gray matter volume doubles during the first postnatal year and then slows to \(20\%\) in its second year<sup>38- 40</sup>. Regionally, we also observe differences between posterior and frontal aperiodic offset trajectories, which either plateau after 1 year (posterior), or have a slow continued increase (frontal) beyond 1 year of age (Supplemental Fig 3). Consistent with this pattern, synaptogenesis differs across cortical regions, with the posterior visual cortex exhibiting a burst in synapse formation between 3- 4 months of age, whereas the prefrontal cortex shows peak synaptogenesis around 8 months of age and continued gains during the second year of life<sup>1</sup>.  
+
+Our observed age- dependent increases in aperiodic slope in infancy also contrast with multiple studies covering child to adulthood, where decreases in slope have consistently been reported. Schaworonskov et. al.<sup>41</sup> also reported decreased slope with age in infants from 1 to 7- months- old, however the parameterization of the spectra in that study was limited to 1- 10Hz due to excessive muscle noise in the data, and it is unclear how the shifts in 4- 12Hz periodic power described below may affect modeling of the underlying aperiodic component in this range. We hypothesize that observed increases in aperiodic slope reflect changes in inhibitory networks that are unique to early development. Indeed, aperiodic slope from EEG recorded from sleeping newborns is observed to increase with age during the first 7 weeks after birth<sup>42</sup>. Growing evidence suggests that aperiodic slope is modulated by the balance between excitation and inhibition, with increased slope associated with a reduction in E/I ratio<sup>10,12,43</sup>. An age- dependent reduction in E/I ratio during the first postnatal year is consistent with the prolonged developmental timing of inhibitory network maturation in humans. Unlike excitatory neurons which are well established by birth, during the first postnatal year GABAergic inhibitory neurons continue to migrate from ventral subregions of the brain to the cortices where they ultimately mature and integrate into neuronal networks<sup>44</sup>. In addition, during this first year GABAergic responses switch from being excitatory to inhibitory due to changes in the concentration of chloride channels on cell membranes<sup>45- 47</sup>. Overall, inhibitory neuron network integration and the excitatory- to- inhibitory GABA switch are unique to this developmental period and likely lead to increased inhibitory tone during the first year.  
+
+Observed changes in the periodic spectra may reflect sequential steps in inhibitory network and thalamocortical circuit development. Transient neural circuits are common in postnatal development and play critical steps in normal development of thalamocortical circuitry<sup>6</sup>. For example, transient circuits between sublate neurons (SPN) and thalamo- recipient layer 4 spiny stellate neurons help establish thalamocortical connections prior to the maturation of primary sensory cortices<sup>48</sup>. Studies of postmortem fetal monkey and human brains suggests that the SPN in primates and humans slowly begins to disappear in the 3<sup>rd</sup> trimester but may persist until 6 months, with an overlapping period in which the thalamus makes connections with both the SPN and cortical layer IV neurons<sup>6,49,50</sup>. We hypothesize that
+
+<--- Page Split --->
+
+the 9.5Hz peak observed at 2- 4 months, but not at 6 months, reflects this transient period when mature excitatory subplate neurons are still receiving and relaying thalamic input to cortices, resulting in higher frequency alpha oscillations. Additionally, newly established connections between the thalamus and layer IV produce lower frequency theta rhythms that will later become the dominant alpha rhythm.  
+
+The thalamus is thought to play a central role in the generation of the mature posterior alpha rhythm. A shift in dominant oscillatory frequency in the theta/alpha range (4- 12Hz) across early childhood has been observed now in many studies<sup>21,24,51</sup>. Here, we both confirm and extend those findings over the first 3 years after birth, with peak frequency increasing most between 4 and 18 months. What factors are potential contributors to this shift in peak frequency? The dynamic circuit motif model (DCM) proposes that cortical network rhythms result from a combination of the intrinsic resonant frequency of a neuronal population and the time course properties of the inhibitory inputs on the neuronal population<sup>52</sup>. Under the DMC model, prior to the maturation of both local inhibitory circuitry and thalamocortical feedback loops, peak frequency oscillations as measured by scalp EEG are more likely to represent the intrinsic properties of cortical networks, with thalamic inputs beginning to play a larger role with age. For example, lower frequency 4- 7Hz oscillations are intrinsically generated by isolated layer 5 cortical neurons, and the range of oscillations increases to 5- 12Hz when connections to other cortical layers remain intact<sup>53</sup>. Thalamic neurons in the lateral geniculate nucleus also fire across the theta and alpha range. In vitro slice experiments from cats suggest that cortical input to thalamus modulates whether theta versus alpha oscillations are dominant<sup>54,55</sup>. Thus, the developmental shift in peak frequency from the canonical theta to alpha range over the first three years after birth, may represent the integration and maturation of cortical inhibitory neurons, as well as the establishment and maturation of thalamocortical connections.  
+
+Finally, our study identified early age- dependent changes in periodic beta power that we hypothesize are associated with thalamocortical loop maturation. We observe the emergence of a low beta peak in infants older than 6- months of age and find that the presence of a low beta peak is associated with higher anesthesia- induced frontal alpha coherence. Biophysical models demonstrate that this frontal anesthesia- induced alpha coherence requires inputs from both the thalamus and cortex<sup>27</sup>. Together these findings suggest that low beta oscillations may directly reflect thalamocortical loop maturation. Beta rhythms are thought to be both generated locally in the cortex through pyramidal- interneuron loops, as well as through thalamus to cortical connections that also rely on inhibitory inputs<sup>25</sup>. The emergence of the low beta peak in awake infants may reflect the combination of newly established network connections between thalamic nuclei and cortical layers, as well as the maturation of interneurons within the thalamocortical pathways.
+
+<--- Page Split --->
+
+It is also possible that developmental changes in beta power are related to infant movement. During EEG acquisition, infants are held in their parent's lap and behavioral supports are in place to keep the infants calm. However, it is not possible to control the infants' movement, and movements both small (hand movements) and large (head turns, leg and arm movements) ubiquitously occur across recordings - likely increasing over the first year as infants become more mobile. Our preprocessing artifact removal pipelines (see Methods) includes several steps for removing high frequency noise from muscle artifact. However this would not remove EEG signal in response to sensorimotor processing. Infant jaw and upper limb movements have been shown to increase power between 9- 20Hz along frontal and occipital sites, while hand and lower limb movements do not have significant effects<sup>56</sup>. In our dataset, increases in low beta power were most prominent in central (not frontal or occipital) ROIs (Supplemental Fig 3K), suggesting that age- dependent changes in beta power more likely represent underlying brain maturation than sensorimotor processing or movement artifact during data acquisition. This is further supported by the consistency in age- dependent shifts of both low and high beta across individuals (see Supplemental Fig 2 for individual power spectra plots), and our observation that low beta is correlated with anesthesia- induced alpha coherence.  
+
+Transient high beta peaks were also observed across this early period of development. Specifically, we observed early increases in high beta peak amplitude, which reached a maximum at 7.5 months, followed by decreases in both high beta peak amplitude and frequency, such that by 36 months the low beta peak is the dominant peak across the 12- 30Hz range. The neurobiological mechanism of this high beta peak is unclear. As discussed above, administration of GABA agonists induces beta activity. In addition, several neurodevelopmental disorders associated with GABA receptor dysfunction show prominent beta peaks on EEG; individuals with Duplication 15q have a prominent beta peak at 23Hz<sup>57,58</sup>, and we have observed that children with Fragile X Syndrome (FXS), aged 3- 7 years, have a prominent 30Hz peak<sup>59</sup>. This 30hz peak observed in FXS children is qualitatively similar to the 30Hz peak observed in the present dataset at a much younger age. Further analysis of data previously published from FXS children shows that the observed high beta peak decreases with age (Figure 5), suggesting delayed brain development. Such observations highlight the value of the longitudinal EEG trajectories presented in this paper in placing findings from neurodevelopmental disorders in the broader context of developmental brain maturation.
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Figure 5. FXS children have increased high beta peak that decreases with age. (A) Periodic power spectra of 35-48 month-old children with and without FXS. (B) Individual periodic power spectra of FXS children, with line hue corresponding to age in months. </center>  
+
+In summary, our work highlights the dynamic developmental changes in neural activity occurring during the first three years after birth and provides insights in potential ways these age- dependent and sometimes transient changes may coincide with sequences in thalamocortical and inhibitory network maturation. Our findings help to ground cross- sectional work occurring at these early ages and provide a foundation to compare developmental trajectories of various neurodevelopmental disorders including autism, ADHD, and rare genetic disorders. Future studies examining early trajectories of functional connectivity and phase amplitude coupling across this age range will provide additional insight into the timing of critical periods in brain maturation.  
+
+## METHODS  
+
+Studies and Participants: Lab based EEGs for this paper were collected as part of four different studies occurring over 15 years conducted at our lab at Boston Children's Hospital (Figure 1A). Sample numbers for each age are shown in Table 1. Study 1, the Healthy Baby Study (IRB- P00019083), was a longitudinal study, enrolling infants starting at 2 months of age, from the Boston Children's Hospital Primary Care Center, which predominantly services families from low- income backgrounds. EEG was collected and developmental assessment using the Mullen Scales of Early Learning (MSEL) was performed at 2, 6, 9, 12, 24, and 36 months.  
+
+Study 2, The Infant Sibling Study (IRB- X06- 08- 0374), and Study 4, the Infant Screening Project (IRB- P00018377), were both prospective, longitudinal studies, enrolling infants with and without first degree family history of ASD starting as early as 3- months of age. For this analysis only infants without family
+
+<--- Page Split --->
+
+history of ASD were included. Study 4 also included a group of infants with elevated social communication concerns at 12 months of age, and they were also excluded from this analysis. EEGs and MSEL were performed at 3, 12, 18, 24, and 36 months for both studies, as well as 6 and 9 months for Study 2. Infants were specifically assessed for ASD using the Autism Diagnostic Observation Schedule (ADOS) in conjunction with clinical best estimate at 24- and 36- month visits.  
+
+Study 3, the Emotion Project (IRB- P00002876), was a cohort/longitudinal study. Infants were enrolled at either 5, 7, or 12 months, and then followed through 7 years of age. In addition to the first time point, EEG data was again collected at 3 years of age. There were no developmental assessments performed for this study, however parent questionnaires regarding child development, diagnoses (e.g., ASD), and therapies were collected.  
+
+All infants had a minimal gestational age of 36 weeks, no history of prenatal or postnatal medical or neurological problems, and no known genetic disorders. Infants who were later diagnosed with ASD (either by assessment during the study, or by community diagnosis disclosed by parents prior to age of 5) were not included in this study.  
+
+Sample characteristics across and within studies are shown in Table 1. The analysis included 1335 EEGs collected from 592 participants. While all studies took place in the same laboratory, participant demographics vary between studies as was expected given differences in recruitment and research aims of each study. In addition, studies differed in the age of enrollment and when subsequent visits were completed (Figure 1C). Combined, the sample remained predominantly white (74%).
+
+<--- Page Split --->
+
+1 Table 1: Sample Characteristics
+
+<table><tr><td></td><td>Combined<br>Studies<br>\(N=592\)</td><td>Study 1<br>\(N=49\)</td><td>Study 2<br>\(N=72\)</td><td>Study 3<br>\(N=363\)</td><td>Study 4<br>\(N=108\)</td></tr><tr><td>Sex, % Female (n)</td><td>46.3(274)</td><td>51.0(25)</td><td>45.8(33)</td><td>46.3(168)</td><td>44.4(48)</td></tr><tr><td>Ethnicity, % (n)</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Hispanic</td><td>8.6(51)</td><td>16.3(8)</td><td>1.4(1)</td><td>10.5(38)</td><td>3.7(4)</td></tr><tr><td>Non-Hispanic</td><td>90.2(534)</td><td>81.6(40)</td><td>97.2(70)</td><td>88.4(321)</td><td>95.4(103)</td></tr><tr><td>Not Answered</td><td>1.2(7)</td><td>2.0(1)</td><td>1.4(1)</td><td>1.1(4)</td><td>0.9(1)</td></tr><tr><td>Race, % (n)</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>White</td><td>74.0(438)</td><td>10.2(5)</td><td>86.1(62)</td><td>79.1(287)</td><td>77.8(84)</td></tr><tr><td>Black or African-American</td><td>6.4(38)</td><td>53.1(26)</td><td>1.4(1)</td><td>2.2(8)</td><td>2.8(3)</td></tr><tr><td>Asian</td><td>3.9(23)</td><td>4.1(2)</td><td>2.8(2)</td><td>4.1(15)</td><td>3.7(4)</td></tr><tr><td>Mixed Race</td><td>12.8(76)</td><td>14.3(7)</td><td>8.3(6)</td><td>12.9(47)</td><td>14.8(16)</td></tr><tr><td>Other</td><td>1.4(8)</td><td>12.2(6)</td><td>0</td><td>0.6(2)</td><td>0</td></tr><tr><td>Not Answered</td><td>1.5(9)</td><td>6.1(3)</td><td>1.4(1)</td><td>1.1(4)</td><td>0.9(1)</td></tr><tr><td>Income, % (n)*</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>&lt;\(35,000\)</td><td>5.4(32)</td><td>32.7(16)</td><td>4.2(3)</td><td>2.5(9)</td><td>3.7(4)</td></tr><tr><td>\(35,000-\)75,000</td><td>11.1(66)</td><td>18.4(9)</td><td>9.7(7)</td><td>13.2(48)</td><td>1.9(2)</td></tr><tr><td>\(>75,000\)</td><td>73.8(437)</td><td>18.4(9)</td><td>70.8(51)</td><td>76.3(277)</td><td>92.6(100)</td></tr><tr><td>Not Answered or Don't Know</td><td>9.6(57)</td><td>30.6(15)</td><td>15.3(11)</td><td>8.0(29)</td><td>1.9(2)</td></tr><tr><td colspan="6">Participant EEG Data Included in Analysis, n</td></tr><tr><td>2-4m</td><td>97</td><td>46</td><td>10</td><td>-</td><td>41</td></tr><tr><td>4-6m</td><td>119</td><td>-</td><td>2</td><td>113</td><td>4</td></tr><tr><td>6-8m</td><td>223</td><td>43</td><td>50</td><td>130</td><td>-</td></tr><tr><td>8-11m</td><td>95</td><td>37</td><td>58</td><td>-</td><td></td></tr><tr><td>11-15m</td><td>291</td><td>40</td><td>62</td><td>102</td><td>87</td></tr><tr><td>18-20m</td><td>107</td><td>-</td><td>40</td><td>-</td><td>67</td></tr><tr><td>23-30m</td><td>118</td><td>22</td><td>44</td><td>-</td><td>52</td></tr><tr><td>35-44m</td><td>285</td><td>18</td><td>48</td><td>174</td><td>45</td></tr></table>
+
+2 *2 infants reported to have income of &#36;30-39,000 were placed in &lt;35,000 category. 1 infant with reported income of &#36;70-79,000 placed in &gt;75,000 category.
+
+4
+
+5
+
+6 **Lab-based EEG data collection**: Baseline, non-task-related EEG data was collected using similar
+
+7 methods and rooms for all four studies. The infant was held by their seated caregiver in a dimly lit, sound
+
+8 attenuated room with a low-electrical-signal background. For Study 2, a research assistant ensured that
+
+<--- Page Split --->
+
+the infant remained calm by blowing bubbles and/or showing toys. For Studies 1 and 3, a video of infant toys was shown for 2- 5 minutes and 2 minutes, respectively. For Study 4, a video of abstract moving objects was shown for 2- 5 minutes. Continuous scalp EEG for Studies 1, 3, and 4 was recorded using a 128- channel Hydrocel Geodesic Sensor Nets (Electrical Geodesics, Inc., Eugene, OR) connected to a NetAmps 300 amplifier (Electrical Geodesic Inc.) and sampled at 500Hz. Study 2 included recordings using 64- channel Geodesic Sensor ( \(< 10\%\) of data) or a 128- channel Hydrocel Geodesic Sensor Nets (Electrical Geodesics, Inc., Eugene, OR), connected to either a NetAmps 200 or 300 amplifier (Electrical Geodesic Inc.) and sampled at either 250 or 500Hz. Additional statistical analysis related to differences in net and amps is described below in EEG Power analysis. For all studies, data was referenced online to a single vertex electrode (Cz) and impedances were kept below 100kΩ in accordance with the impedance capabilities of the high- impedance amplifiers inside the electrically shielded room. Electrooculographic electrodes were removed to improve the child's comfort.  
+
+EEG pre- processing: Raw Netstation (Electrical Geodesics, Inc) files were exported to MATLAB (version R2017a) for preprocessing and absolute power calculations using the Batch Automated Processing Platform (BEAPP \(^{60}\) ) with integrated Harvard Automated Preprocessing Pipeline for EEG (HAPPE \(^{61}\) ). For each EEG, a 1Hz high- pass and 100Hz low- pass filter were applied, data sampled at 500Hz were resampled to 250Hz, and then run through the HAPPE module consisting of 60Hz line noise removal, bad channel rejection, and artifact removal using combined wavelet- enhanced independent component analysis (ICA) and Multiple Artifact Rejection Algorithm (MARA \(^{5,6}\) ). The following channels, in addition to the 10- 20 electrodes, were used for MARA: 64- channel net – 16, 9, 8, 3, 58, 57, 21, 25, 18, 30, 43, 50, 53, 32, 33, 38, 41, 45; and 128- channel net - 28, 19, 4, 117, 13, 112, 41, 47, 37, 55, 87, 103, 98, 65, 67, 77, 90, 75. These electrodes were chosen as they evenly cover all brain regions of interest (Supplemental Figure 1). After artifact removal, channels removed during bad channel rejection were then interpolated, data were referenced to the average reference, detrended to the signal mean, and segmented into 2- second segments. Any segments with retained artifact were rejected using HAPPE's amplitude and joint probability criteria.  
+
+EEG rejection criteria: EEG recordings were rejected using the following HAPPE data quality measures: Fewer than 20 segments (40 seconds of total EEG), percent good channels \(< 80\%\) , percent independent components rejected \(>80\%\) , mean artifact probability of components kept \(>0.3\) , and percent variance retained \(< 25\%\) . Expected differences between studies in number of segments remaining post- segment rejection were observed, with Study 3 with the shortest resting state recording period having fewer segments. All other quality metrics were similar across studies (Table 2).
+
+<--- Page Split --->
+
+
+Table 2: EEG data quality metrics   
+
+<table><tr><td></td><td>Combined Studies<br>N = 1335</td><td>Study 1<br>N = 206</td><td>Study 2<br>N = 314</td><td>Study 3<br>N = 519</td><td>Study 4<br>N = 296</td></tr><tr><td colspan="6">EEG quality metrics, Mean±SD</td></tr><tr><td>Number of Segments</td><td>81±39.5</td><td>126±27.5</td><td>85.8±39.9</td><td>48.8±7.1</td><td>105.2±30.3</td></tr><tr><td>Percent Good Channels</td><td>92.2±4.5</td><td>92.2±4.8</td><td>92.6±4.4</td><td>92.0±4.6</td><td>93.5±4.3</td></tr><tr><td>Percent ICs Rejected</td><td>35.7±10.3</td><td>35.7±10.6</td><td>35.4±10.1</td><td>34.1±14.0</td><td>35.0±10.5</td></tr><tr><td>Percent Variance Kept of Post Wavelet Data</td><td>67.2±17.0</td><td>63.1±16.1</td><td>66.5±15.7</td><td>68.8±18.3</td><td>68.1±16.1</td></tr><tr><td>Mean Artifact Probability of Kept ICs.</td><td>0.12±0.05</td><td>0.12±0.04</td><td>0.11±0.04</td><td>0.12±0.05</td><td>0.12±0.05</td></tr></table>  
+
+EEG Power analysis: The power spectral density at each electrode, for each 2- second segment, was calculated in the BEAPP Power Spectral Density (PSD) module using a multitaper spectral analysis<sup>62</sup> and three orthogonal tapers. For each electrode, the PSD was averaged across segments, and then further averaged across all available electrodes, or frontal, temporal, central, and posterior regions of interest (Supplemental Figure 1C,D). The PSD was then further analyzed using a modified version of FOOOF v1.0.0<sup>8</sup>(https://github.com/fooof-tools/fooof; in Python v3.6.8) in order to model periodic and aperiodic components of the power spectra. FOOOF required modification for use in this age range, as power spectrum models for 2- 7 month ages showed poor model fit (increased mean squared error) for frequencies between 10- 20Hz. Specifically, the FOOOF modeled curves did not accurately capture the "trough" in the power spectra visually observed in this frequency range at younger ages (Supplemental Figure 4). To improve model fit, the robust_ap_fit function, which initially defines the aperiodic component, was modified so that the initial estimate of the flattened power spectra (flatspect) has a baseline elevated such that the lowest point is \(\geq 0\) , to avoid omitting data important across the 2- 7 month age range. This is in contrast to the original method of setting all negative points of the initial flattened power spectra equal to 0. In the original and modified scripts, this initial fit is combined with thresholding to render a more robust second round of aperiodic parameters. After these second aperiodic parameters have been defined, the fit function re- estimates the flattened spectra (spectrum_flat). At this point, prior to fitting spectra peaks, the modified code sets negative data in the flattened spectra equal to 0, similar to the approach of the original code during the initial aperiodic fit. In both versions, aperiodic parameters are fit a third and final time to the spectra with peaks removed (spectrum_peak_rm). The FOOOF model was used in the fixed mode (no spectral knee) with peak_width_limits set to [0.5, 18.0], max_n_peaks = 7, and peak_threshold = 2. Code is available (osf.io/u3gp4) which runs both the original and modified versions of FOOOF, and graphs the RMSE across frequencies for both versions, separate by age. Comparisons are presented in Supplemental Figure 4. Further analyses were subsequently restricted to
+
+<--- Page Split --->
+
+2.5- 50Hz given elevated error between 2- 2.5, and 50- 55Hz. Mean \(\mathsf{R}^2\) for the full sample using this modified version of FOOOF was 0.997 (STD 0.008; range 0.890- 0.999). Mean estimated error for the sample was 0.01 (STD 0.01, range 0.002- 0.09).  
+
+FOOF provides two parameters to describe the aperiodic 1/f signal: offset and slope. As the FOOOF- determined offset is extrapolated to the estimated aperiodic power at 0Hz, where there are high amounts of error, we instead calculated the aperiodic offset based on aperiodic power at 2.5hz (Figure 1B). The periodic power spectrum (Figure 1B, F) was determined by subtracting the FOOOF estimated aperiodic spectrum (Figure 1E) from the absolute power spectrum (Figure 1D). To further characterize peaks and troughs within the power spectra across development, the periodic spectrum was then smoothed using a savgol filter (scipy.signal.savgol_filter, window length = 101, polyorder = 8). Individual periodic power spectrum plots before and after savgol filter are shown in Supplemental Figure 2. We decided to use this method instead of using the FOOOF estimated peak_fit as the high beta peak appeared to have a non- gaussian shape at some ages, and thus peak_fit estimates did not accurately identify the high beta peak frequency. Using the smoothed periodic spectra, maxima were identified within the following frequency ranges: 4- 6.5 (theta), 4- 12Hz (theta/alpha), 12- 20Hz (low beta) and 20- 35Hz (high beta). A low beta trough was also identified based on the minima between 10- 20Hz. Aperiodic and periodic power across the following canonical frequency bands was calculated taking the integral of each parametrized spectra between the following frequency ranges: theta (4- 6Hz), low alpha (6- 9Hz), high alpha (9- 12Hz), low beta (12- 20Hz), high beta (20- 30Hz), and gamma (30- 45Hz).  
+
+As Study 2 collected data with 2 net types and 2 amplifiers, data from 6-, 9-, and 12- month age bins were assessed for spectra differences in total (2- 50hz) aperiodic and periodic power as well as aperiodic slope and intercept measure from central electrodes between either 64 and 128 channel nets, or NetAmps 200 or 300 amplifiers. Of the 24 analyses performed, 3 showed significant differences. Nettype differences were observed for 9- and 12- month central aperiodic slope ( \(p = 0.04\) for both) and an amplifier difference was observed for 12- month central offset ( \(p = 0.04\) ). None of these were significant after correcting for multiple comparisons.  
+
+Anesthesia cohort: EEG data was also collected from infants undergoing anesthesia as part of a prospective observational study approved by the institutional review board at Montefiore Medical Center, Albert Einstein College of Medicine<sup>35,63</sup>. Infants scheduled for elective surgical procedures (e.g., circumcision, hernia repair) were recruited. Infants were excluded for prematurity, known neurologic injury, epilepsy, or planned intracranial surgery. Infants less than 6 months of age, or those documented to be asleep or crying during baseline (pre- anesthesia) recordings were excluded from further analysis. All subjects received general anesthesia with sevoflurane.
+
+<--- Page Split --->
+
+EEG recordings were obtained using a Food and Drug Administration (FDA) approved 26 channel device recording from scalp locations designated by the International 10- 20 System, reference midline occipital channel (Oz) (microEEG System, Biosignal, Acton, MA)64. Data was collected from 21 electrodes (FpZ, Fp1, Fp2, Fz, F3, F4, F7, F8, Cz, C3, C4, T3, T4, Pz, P3, P4, T5, T6, Oz, O1, O2) both prior to (baseline) and during sevoflurane induction and maintenance. End tidal sevoflurane concentration was recorded, locked in time with EEG recording. Data were sampled at a frequency of 250 Hz.  
+
+Baseline EEGs (n = 45) were visually inspected and approximately 2 minutes of continuous EEG with minimal artifact was segmented and then processed using BEAPP/HAPPE60,61. 9 EEGs were excluded due to excessive artifact during visual inspection. No additional EEGs were excluded based on HAPPE data quality criteria. Final sample (n = 36) was an average age of 9.1 months (range 6- 15 months), and predominantly male (n=25). Central periodic power using Pz and Fz electrodes was calculated using using multitaper spectral analysis using three orthogonal tapers. PSD was then further analyzed using a modified version of FOOOF v1.0.0 as described above, in order to model periodic and aperiodic components of the power spectra.  
+
+## Alpha coherence analysis  
+
+Preprocessing of Anesthesia EEG data: We used a bipolar montage to analyze the data (F7- Fp1 and F8- Fp2). We developed an automatized method to exclude epochs with high- amplitude noise based on the standard deviation of the time series signal. Three members of the team (CW, JC, and RG) visually inspected the remaining epochs to select 30- second artifact- free segments that were used for the analysis. We selected epochs with a stable sevoflurane concentration defined as two consecutive minutes of end- tidal sevoflurane levels within 0.2% preceding the selection of an epoch of data. EEG data were band- pass filtered from 0.1 to 30 Hz.  
+
+For each subject, corresponding EEG data collected during anesthesia were inspected to identify 30 second segments under stable sevoflurane concentration and that were "artifact free" (eg. no motion or electrocautery artifacts). Epochs with high- amplitude noise in frontal electrodes (F7- Fp1; F8- Fp2) based on standard deviation of the time series were automatically excluded. Blinded visual inspection by members of the team (CW, JC, and RG) identified remaining epochs with stable sevoflurane concentration without other anesthesia (e.g. propofol bolus) interference. Stable sevoflurane concentration was defined as two consecutive minutes of end- tidal sevoflurane levels within 0.2% preceding the selection of an epoch of data.
+
+<--- Page Split --->
+
+calculated to visualize developmental changes within features. The modeled value of a feature at a given age (in days) was subtracted from the modeled value from the subsequent day, and this was divided by the standard deviation of the modeled values of that feature across the age range.  
+
+To assess the differences between regions of interest, GAMM models were fit with the following form: (3) Power Measure \(\sim\) s(age_days) + oSex + Study+ ROI+ s(age_days, New_ID, bs = 'fs')  
+
+where the terms have the same meanings as above, and ROI is a factor representing the four regions (frontal, central, temporal, posterior). Because prior literature and preliminary visual inspection of the data indicated that the posterior ROI is most unique in the time course of development, the posterior ROI was set as the reference factor. Thus, the effect and significance associated with each of the other ROIs is a measure of the difference between that ROI and the posterior ROI.  
+
+Anesthesia Statistical Analysis: ANCOVA, with sevoflurane levels as a covariate, was used to determine effects of presence of low beta peak on anesthesia induced alpha coherence.  
+
+Figures were created using Python v3.6.8 and python data visualization libraries (matplotlib(60) and Seaborn (https://seaborn.pydata.org/index.html) or in R (version 4.1.2).  
+
+Funding Support: This research was supported by the National Institutes of Health (R01- DC010290 and MH078829 to CAN, K23DC07983 and T32MH112510 to CLW, and UL1TR002556, KL2TR002558, and K23DA057499 to JYC. Research was also supported by the JPB Research Network on Toxic Stress.  
+
+## Author Contributions:  
+
+Carol Wilkinson: Conceptualization, Methodology, Software, Formal analysis, Data Curation, Writing – Original Draft, Visualization. Lisa Yankowitz: Methodology, Software, Formal analysis, Writing – Review & Editing, Visualization. Jerry Y. Chao: Methodology, Formal analysis, Investigation, Data Curation, Resources, Writing – Review & Editing. Rodrigo Gutiérrez: Methodology, Software, Formal analysis, Data Curation, Writing – Review & Editing. Jeff L. Rhoades: Software, Writing – Review & Editing. Shlomo Shinnar: Project administration, Supervision, Writing – Review & editing. Patrick L. Purdon: Methodology, Supervision, Writing – Review & Editing. Charles A. Nelson: Investigation, Resources, Writing – Review & Editing, Supervision, Project administration, Funding acquisition.
+
+<--- Page Split --->
+
+## REFERENCES  
+
+1. Huttenlocher, P. R. & Dabholkar, A. S. Regional differences in synaptogenesis in human cerebral cortex. J Comp Neurol 387, 167-178 (1997).  
+
+2. Glantz, L. A., Gilmore, J. H., Hamer, R. M., Lieberman, J. A. & Jarskog, L. F. Synaptophysin and postsynaptic density protein 95 in the human prefrontal cortex from mid-gestation into early adulthood. Neuroscience 149, 582-591 (2007).  
+
+3. Tau, G. Z. & Peterson, B. S. Normal Development of Brain Circuits. Neuropsychopharmacology 35, 147-168 (2010).  
+
+4. Kim, J. Y. & Paredes, M. F. Implications of extended inhibitory neuron development. International Journal of Molecular Sciences 22, 1-16 (2021).  
+
+5. Antón-Bolaños, N., Espinosa, A. & López-Bendito, G. Developmental interactions between thalamus and cortex: a true love reciprocal story. Curr Opin Neurobiol 52, 33-41 (2018).  
+
+6. Molnár, Z., Luhmann, H. J. & Kanold, P. O. Transient cortical circuits match spontaneous and sensory-driven activity during development. Science 370, eabb2153 (2020).  
+
+7. Alcauter, S. et al. Development of Thalamocortical Connectivity during Infancy and Its Cognitive Correlations. The Journal of Neuroscience 34, 9067 (2014).  
+
+8. Doesburg, S. M., Ward, L. M. & Ribary, U. The alpha-theta-gamma (ATG) switch: Toward unified principles of cortical processing. Current Trends in Neurology 9, 1-12 (2015).  
+
+9. Tröndle, M., Popov, T., Dziemian, S. & Langer, N. Decomposing the role of alpha oscillations during brain maturation. eLife 11, e77571 (2022).  
+
+10. Gao, R., Peterson, E. J. & Voytek, B. Inferring synaptic excitation/inhibition balance from field potentials. Neuroimage 158, 70-78 (2017).
+
+<--- Page Split --->
+
+11. Merkin, A. et al. Do age-related differences in aperiodic neural activity explain differences in resting EEG alpha? Neurobiol Aging 121, 78-87 (2023).  
+
+12. Voytek, B. et al. Age-related changes in 1/f neural electrophysiological noise. Journal of Neuroscience 35, 13257-13265 (2015).  
+
+13. Hill, A. T., Clark, G. M., Bigelow, F. J., Lum, J. A. G. & Enticott, P. G. Periodic and aperiodic neural activity displays age-dependent changes across early-to-middle childhood. Developmental Cognitive Neuroscience 54, 101076 (2022).  
+
+14. McSweeney, M. et al. Age-related trends in aperiodic EEG activity and alpha oscillations during early-to middle-childhood. Neuroimage 269, 119925 (2023).  
+
+15. Buzsáki, G., Logothetis, N. & Singer, W. Scaling brain size, keeping timing: evolutionary preservation of brain rhythms. Neuron 80, 751-764 (2013).  
+
+16. Donoghue, T. et al. Parameterizing neural power spectra into periodic and aperiodic components. Nature Neuroscience 23, 1655-1665 (2020).  
+
+17. Klimesch, W. Alpha-band oscillations, attention, and controlled access to stored information. Trends in Cognitive Sciences 16, 606-617 (2012).  
+
+18. Ball, G. et al. Thalamocortical Connectivity Predicts Cognition in Children Born Preterm. Cerebral Cortex 25, 4310-4318 (2015).  
+
+19. Foxe, J. & Snyder, A. The Role of Alpha-Band Brain Oscillations as a Sensory Suppression Mechanism during Selective Attention. Frontiers in Psychology 2, (2011).  
+
+20. Orekhova, E. V., Stroganova, T. A. & Posikera, I. N. Theta synchronization during sustained anticipatory attention in infants over the second half of the first year of life. International Journal of Psychophysiology 32, 151172 (1999).  
+
+21. Marshall, P. J., Bar-Haim, Y. & Fox, N. A. Development of the EEG from 5 months to 4 years of age. Clinical Neurophysiology 113, 1199-1208 (2002).
+
+<--- Page Split --->
+
+22. Lindsley, D. B. A Longitudinal Study of the Occipital Alpha Rhythm in Normal Children: Frequency and Amplitude Standards. The Pedagogical Seminary and Journal of Genetic Psychology 55, 197–213 (1939).  
+
+23. Smith, J. R. The Frequency Growth of the Human Alpha Rhythms During Normal Infancy and Childhood. The Journal of Psychology 11, 177–198 (1941).  
+24. Cellier, D., Riddle, J., Petersen, I. & Hwang, K. The development of theta and alpha neural oscillations from ages 3 to 24 years. Developmental Cognitive Neuroscience 50, 100969 (2021).  
+25. Spitzer, B. & Haegens, S. Beyond the Status Quo: A Role for Beta Oscillations in Endogenous Content (Re)Activation. eNeuro 4, (2017).  
+26. Purdon, P. L., Sampson, A., Pavone, K. J. & Brown, E. N. Clinical Electroencephalography for Anesthesiologists: Part I: Background and Basic Signatures. Anesthesiology 123, 937–960 (2015).  
+27. Ching, S., Cimenser, A., Purdon, P. L., Brown, E. N. & Kopell, N. J. Thalamocortical model for a propofol-induced α-rhythm associated with loss of consciousness. Proceedings of the National Academy of Sciences 107, 22665–22670 (2010).  
+28. McCarthy, M. M., Brown, E. N. & Kopell, N. Potential network mechanisms mediating electroencephalographic beta rhythm changes during propofol-induced paradoxical excitation. J Neurosci 28, 13488–13504 (2008).  
+29. Cornelissen, L., Kim, S.-E., Purdon, P. L., Brown, E. N. & Berde, C. B. Age-dependent electroencephalogram (EEG) patterns during sevoflurane general anesthesia in infants. Elife 4, e06513 (2015).  
+30. Cornelissen, L. et al. Electroencephalographic markers of brain development during sevoflurane anaesthesia in children up to 3 years old. British Journal of Anaesthesia 120, 1274–1286 (2018).  
+31. Akeju, O. et al. Age-dependency of sevoflurane-induced electroencephalogram dynamics in children. Br J Anaesth 115 Suppl 1, i66–i76 (2015).
+
+<--- Page Split --->
+
+32. Flores, F. J. et al. Thalamocortical synchronization during induction and emergence from propofol-induced unconsciousness. Proc Natl Acad Sci U S A 114, E6660–E6668 (2017).  
+
+33. Bastos, A. M. et al. Neural effects of propofol-induced unconsciousness and its reversal using thalamic stimulation. eLife 10, e60824 (2021).  
+
+34. Engel, A. K. & Fries, P. Beta-band oscillations--signalling the status quo? Curr Opin Neurobiol 20, 156–165 (2010).  
+
+35. Chao, J. Y. et al. Electroencephalographic Findings and Clinical Behavior During Induction of Anesthesia With Sevoflurane in Human Infants: A Prospective Observational Study. Anesthesia & Analgesia 130, e161 (2020).  
+
+36. Manning, J. R., Jacobs, J., Fried, I. & Kahana, M. J. Broadband shifts in local field potential power spectra are correlated with single-neuron spiking in humans. Journal of Neuroscience 29, 13613–13620 (2009).  
+
+37. Miller, K. J. Broadband spectral change: Evidence for a macroscale correlate of population firing rate? Journal of Neuroscience 30, 6477–6479 (2010).  
+
+38. Knickmeyer, R. C. et al. A structural MRI study of human brain development from birth to 2 years. Journal of Neuroscience 28, 12176–12182 (2008).  
+
+39. Gilmore, J. H. et al. Longitudinal development of cortical and subcortical gray matter from birth to 2 years. Cereb Cortex 22, 2478–2485 (2012).  
+
+40. Gilmore, J. H., Knickmeyer, R. C. & Gao, W. Imaging structural and functional brain development in early childhood. Nat Rev Neurosci 19, 123–137 (2018).  
+
+41. Schaworonkow, N. & Voytek, B. Longitudinal changes in aperiodic and periodic activity in electrophysiological recordings in the first seven months of life. Dev Cogn Neurosci 47, 100895 (2021).
+
+<--- Page Split --->
+
+42. Chini, M., Pfeffer, T. & Hanganu-Opatz, I. An increase of inhibition drives the developmental decorrelation of neural activity. eLife 11, e78811 (2022).
+43. Voytek, B. & Knight, R. T. Dynamic Network Communication as a Unifying Neural Basis for Cognition, Development, Aging, and Disease. Biological Psychiatry 77, 1089–1097 (2015).
+44. Paredes, M. F. et al. Extensive migration of young neurons into the infant human frontal lobe. Science 354, aaf7073 (2016).
+45. Mueller, A. L., Taube, J. S. & Schwartzkroin, P. A. Development of hyperpolarizing inhibitory postsynaptic potentials and hyperpolarizing response to gamma-aminobutyric acid in rabbit hippocampus studied in vitro. J Neurosci 4, 860–867 (1984).
+46. Cherubini, E., Gaiarsa, J. L. & Ben-Ari, Y. GABA: an excitatory transmitter in early postnatal life. Trends Neurosci 14, 515–519 (1991).
+47. Xu, G. et al. Late Development of the GABAergic System in the Human Cerebral Cortex and White Matter. J Neuropathol Exp Neurol 70, 841–858 (2011).
+48. Kanold, P. O. & Shatz, C. J. Subplate neurons regulate maturation of cortical inhibition and outcome of ocular dominance plasticity. Neuron 51, 627–638 (2006).
+49. Kostovic, I. & Rakic, P. Developmental history of the transient subplate zone in the visual and somatosensory cortex of the macaque monkey and human brain. Journal of Comparative Neurology 297, 441–470 (1990).
+50. McQuillen, P. S., Sheldon, R. A., Shatz, C. J. & Ferriero, D. M. Selective Vulnerability of Subplate Neurons after Early Neonatal Hypoxia-Ischemia. J Neurosci 23, 3308–3315 (2003).
+51. Orekhova, E. V., Stroganova, T. A., Posikera, I. N. & Elam, M. EEG theta rhythm in infants and preschool children. Clinical Neurophysiology 117, 1047–1062 (2006).
+52. Womelsdorf, T., Valiante, T. A., Sahin, N. T., Miller, K. J. & Tiesinga, P. Dynamic circuit motifs underlying rhythmic gain control, gating and integration. Nat Neurosci 17, 1031–1039 (2014).
+
+<--- Page Split --->
+
+53. Silva, L. R., Amitai, Y. & Connors, B. W. Intrinsic Oscillations of Neocortex Generated by Layer 5 Pyramidal Neurons. Science 251, 432 (1991).54. Hughes, S. W. & Crunelli, V. Thalamic mechanisms of EEG alpha rhythms and their pathological implications. The Neuroscientist: a review journal bringing neurobiology, neurology and psychiatry 11, 357-372 (2005).55. Hughes, S. W. & Crunelli, V. Just a phase they're going through: The complex interaction of intrinsic high-threshold bursting and gap junctions in the generation of thalamic \(\alpha\) and \(\theta\) rhythms. International Journal of Psychophysiology 64, 3-17 (2007).56. Georgieva, S. et al. Toward the Understanding of Topographical and Spectral Signatures of Infant Movement Artifacts in Naturalistic EEG. Frontiers in Neuroscience 14, (2020).57. Frohlich, J. et al. A Quantitative Electrophysiological Biomarker of Duplication 15q11.2-q13.1 Syndrome. (2016) doi:10.1371/journal.pone.0167179.58. Frohlich, J. et al. Mechanisms underlying the EEG biomarker in Dup15q syndrome. Molecular Autism 10, 29 (2019).59. Wilkinson, C. L. & Nelson, C. A. Increased aperiodic gamma power in young boys with Fragile X Syndrome is associated with better language ability. Molecular Autism 1-15 (2021) doi:10.1186/s13229-021-00425-x.60. Levin, A. R., Méndez Leal, A. S., Gabard-Durnam, L. J. & O'Leary, H. M. BEAPP: The Batch Electroencephalography Automated Processing Platform. Frontiers in Neuroscience 12, 513 (2018).61. Gabard-Durnam, L. J., Méndez Leal, A. S., Wilkinson, C. L. & Levin, A. R. The Harvard Automated Processing Pipeline for Electroencephalography (HAPPE): standardized processing software for developmental and high-artifact data. Frontiers in Neuroscience 12, 97 (2018).62. Babadi, B. & Brown, E. N. A review of multitaper spectral analysis. IEEE Transactions on Biomedical Engineering 61, 1555-1564 (2014).
+
+<--- Page Split --->
+
+63. Chao, J. Y. et al. Decreased Electroencephalographic Alpha Power During Anesthesia Induction Is Associated With EEG Discontinuity in Human Infants. Anesthesia & Analgesia 135, 1207–1216 (2022).
+64. Grant, A. C. et al. Diagnostic accuracy of microEEG: a miniature, wireless EEG device. Epilepsy Behav 34, 81–85 (2014).
+65. Bokil, H., Andrews, P., Kulkarni, J. E., Mehta, S. & Mitra, P. P. Chronux: a platform for analyzing neural signals. J Neurosci Methods 192, 146–151 (2010).
+66. Wood, S. N. Generalized Additive Models: An Introduction with R, Second Edition. (CRC Press, 2017).
+67. Benjamini, Y. & Hochberg, Y. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society. Series B (Methodological) 57, 289–300 (1995).
+
+<--- Page Split --->
+
+Supplemental Figure 1 - Electrode layout: (A) 128- channel Hydrogel Geodesic Sensor Net. (B) 64- channel Geodesic Sensor Net. Pink circles denote 10- 20 electrodes, and blue circles denote the additional electrodes included in ICA and MARA steps of pre- processing. (C), (D) Electrodes averaged for frontal (yellow), central (blue), temporal (orange), and posterior (green) regions of interest.  
+
+![](images/Figure_2.jpg)
+
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Supplemental Figure 2: Individual plots of the periodic spectrum averaged across the whole ROI for each age bin. </center>
+
+<--- Page Split --->
+![](images/Figure_unknown_0.jpg)
+
+<center>Supplemental Figure 3: GAMMs modeled trajectories of 8 power measures, with ROI, study, smoothed age, and sex as predictor terms. </center>
+
+<--- Page Split --->
+
+Supplemental Figure 4: For each age bin the average FOOOF modeled spectra (green) across participants in each age bin is graphed along with the averaged original power spectrum (red), and averaged FOOOF estimated aperiodic spectrum (blue). (A) Unedited FOOOF estimates. (B) Modified FOOOF estimates. (C) Comparison of squared error across frequencies of unedited (orange) and modified (blue) FOOOF.  
+
+![](images/Figure_unknown_1.jpg)
+
+<center>A. Unedited FOOOF model estimates - Original Spectrum - FOOOF modeled Spectrum - Aperiodic Fit </center>  
+
+![](images/Figure_unknown_2.jpg)
+
+<center>B. Modified FOOOF model estimates - Original Spectrum - FOOOF modeled Spectrum - Aperiodic Fit </center>  
+
+![PLACEHOLDER_35_2]
+
+<center>C. Original vs Edited FOOOF error </center>
+
+<--- Page Split --->

preprint/preprint__2b38d2667b0ed62d397a8a20a8d846f75734d0824b7f39e0811c05ef40cb6eba/preprint__2b38d2667b0ed62d397a8a20a8d846f75734d0824b7f39e0811c05ef40cb6eba_det.mmd

@@ -0,0 +1,497 @@
+<|ref|>title<|/ref|><|det|>[[44, 107, 923, 243]]<|/det|>
+# Developmental trajectories of EEG aperiodic and periodic power: Implications for understanding the timing of thalamocortical development during infancy.  
+
+<|ref|>text<|/ref|><|det|>[[44, 263, 181, 282]]<|/det|>
+Carol Wilkinson  
+
+<|ref|>text<|/ref|><|det|>[[52, 291, 466, 308]]<|/det|>
+carol.wilkinson@childrens.harvard.edu  
+
+<|ref|>text<|/ref|><|det|>[[50, 337, 645, 356]]<|/det|>
+Boston Children's Hospital https://orcid.org/0000- 0002- 4694- 8564  
+
+<|ref|>text<|/ref|><|det|>[[44, 362, 285, 400]]<|/det|>
+Lisa Yankowitz Boston Children's Hospital  
+
+<|ref|>text<|/ref|><|det|>[[44, 407, 290, 446]]<|/det|>
+Jerry Chao Montefiore Medical Center  
+
+<|ref|>text<|/ref|><|det|>[[44, 453, 343, 492]]<|/det|>
+Rodrigo Gutierrez Massachusetts General Hospital  
+
+<|ref|>text<|/ref|><|det|>[[44, 498, 220, 538]]<|/det|>
+Jeff Rhoades Harvard University  
+
+<|ref|>text<|/ref|><|det|>[[44, 544, 290, 583]]<|/det|>
+Shlomo Shinnar Montefiore Medical Center  
+
+<|ref|>text<|/ref|><|det|>[[44, 590, 919, 630]]<|/det|>
+Patrick Purdon Massachusetts General Hospital, Harvard Medical School https://orcid.org/0000- 0003- 0080- 3340  
+
+<|ref|>text<|/ref|><|det|>[[44, 636, 266, 675]]<|/det|>
+Charles Nelson Harvard Medical School  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 721, 103, 738]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 758, 135, 776]]<|/det|>
+Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[44, 796, 354, 815]]<|/det|>
+Posted Date: September 18th, 2023  
+
+<|ref|>text<|/ref|><|det|>[[44, 834, 475, 853]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 3215728/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 872, 912, 913]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 932, 535, 952]]<|/det|>
+Additional Declarations: There is NO Competing Interest.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 77, 910, 120]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on July 10th, 2024. See the published version at https://doi.org/10.1038/s41467-024-50204-4.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[78, 85, 894, 130]]<|/det|>
+Title: Developmental trajectories of EEG aperiodic and periodic power: Implications for understanding the timing of thalamocortical development during infancy.  
+
+<|ref|>text<|/ref|><|det|>[[78, 156, 828, 200]]<|/det|>
+Authors: Carol L. Wilkinson\*1,2, Lisa Yankowitz\*1, Jerry Y. Chao3, Rodrigo Gutiérrez4,9, Jeff L. Rhoades5,6, Shlomo Shinnar7,8, Patrick L. Purdon9, Charles A. Nelson1,2,10  
+
+<|ref|>text<|/ref|><|det|>[[80, 230, 305, 248]]<|/det|>
+\*Designates co- first authors  
+
+<|ref|>sub_title<|/ref|><|det|>[[80, 279, 177, 296]]<|/det|>
+## Affiliations  
+
+<|ref|>text<|/ref|><|det|>[[75, 300, 910, 664]]<|/det|>
+1 Division of Developmental Medicine, Boston Children's Hospital, Boston, MA, United States.  2 Harvard Medical School, Boston, MA, USA  3 Department of Anesthesiology, Montefiore Medical Center, Children's Hospital at Montefiore, Albert Einstein College of Medicine, Bronx, NY, USA.  4 Centro de Investigación Clínica Avanzada, Hospital Clínico de la Universidad de Chile, Santiago, Chile.  5 Department of Neurobiology, Harvard Medical School, Boston, MA, USA.  6 Program in Neuroscience, Division of Medical Sciences, Graduate School of Arts and Sciences, Harvard University, Cambridge, MA, USA.  7 The Saul R. Korey Department of Neurology, Montefiore Medical Center, Albert Einstein College of Medicine, Bronx, NY, USA.  8 Department of Epidemiology and Population Health, Albert Einstein College of Medicine, Bronx, NY, USA.  9 Department of Anesthesia, Critical Care and Pain Medicine, Massachusetts General Hospital, Boston, MA, United States.  10 Harvard Graduate School of Education, Cambridge, MA, United States.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[80, 62, 157, 78]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[75, 110, 916, 371]]<|/det|>
+The development of neural circuits over the first years of life has long- lasting effects on brain function, yet our understanding of early circuit development in humans remains limited. Here, aperiodic and periodic EEG power features were examined from longitudinal EEGs collected from 592 healthy 2–44 month- old infants, revealing age- dependent nonlinear changes suggestive of distinct milestones in early brain maturation. Consistent with the transient developmental progression of thalamocortical circuitry, we observe the presence and then absence of periodic alpha and high beta peaks across the three- year period, as well as the emergence of a low beta peak (12- 20Hz) after six months of age. We present preliminary evidence that the emergence of the low beta peak is associated with thalamocortical connectivity sufficient for anesthesia- induced alpha coherence. Together, these findings suggest that early age- dependent changes in alpha and beta periodic peaks may reflect the state of thalamocortical network development.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[80, 86, 188, 103]]<|/det|>
+## Introduction  
+
+<|ref|>text<|/ref|><|det|>[[75, 131, 920, 511]]<|/det|>
+The infant brain undergoes dramatic structural and physiological change in the first year after birth. Rapid increases in brain volume coincide with expansive synaptogenesis \(^{1 - 3}\) , as well as interneuron migration, maturation and network integration \(^{4}\) . In particular, during this early period thalamocortical connections are established through an intricately choreographed sequence that plays a critical role in the development of sensory cortical networks \(^{5}\) . However, the detailed timing of interneuron and thalamocortical maturation in human development is largely unknown. In rodent models, the development of thalamocortical circuitry is notable for transient inhibitory connections that drive subsequent circuit formation and coincide with critical periods of plasticity present during the first 2- 3 postnatal weeks \(^{6}\) . In humans, longitudinal resting- state fMRI data suggest that while thalamus- sensorimotor connectivity networks are present at birth, other networks (e.g. thalamus- medial- visual, thalamus- default- mode) do not emerge until 1 year of age \(^{7}\) . However, MRI studies thus far have been limited to measuring annual changes in structural or functional connectivity, preventing a detailed understanding of rapid developmental change during this period. In contrast, electroencephalography (EEG) can provide frequent and non- invasive repeated measurement of brain oscillations that directly result from transient developmental changes in inhibitory networks and maturation of thalamocortical circuitry \(^{5,8,9}\) .  
+
+<|ref|>text<|/ref|><|det|>[[75, 540, 916, 800]]<|/det|>
+The EEG power spectrum is comprised of two physiologically distinct components reflecting underlying neuronal activity: aperiodic and periodic power. The aperiodic component defines the slope of the power spectrum, following a 1/f power law distribution (Fig 1A) and reflects non- oscillatory neuronal spiking activity. In addition, the aperiodic slope has been linked to the excitatory- inhibitory (E/I) balance of the underlying neuronal network, where a flattened, reduced slope is associated with increased excitation over inhibition, and a steeply more accelerated slope with increased inhibition over excitation \(^{10}\) . Longitudinal studies of child- to adulthood have observed decreases in aperiodic slope with age, suggestive of increases in E/I balance with age \(^{11 - 14}\) . Changes in the aperiodic component in early infancy are less well described, and we hypothesize they may be substantially different from those in childhood, as the first year after birth includes rapid increases in neuronal activity, synaptogenesis, and inhibitory neuron integration.  
+
+<|ref|>text<|/ref|><|det|>[[77, 828, 914, 920]]<|/det|>
+The periodic component of the power spectrum is defined as the portion of the absolute power spectrum rising above the aperiodic slope (Fig 1B). Periodic power reflects oscillatory activity occurring in narrow frequency bands that are highly correlated with various cognitive processes and behavioral states \(^{15,16}\) , and provide the foundation for both local and long- range communication within the brain. The majority of
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[78, 61, 909, 130]]<|/det|>
+neural oscillations observed in the power spectrum are the direct result of inhibitory and thalamocortical network responses to sensory input. Thus, as a measure, the EEG power spectrum is well positioned to shed light on the developmental timing of inhibitory and thalamocortical network maturation.  
+
+<|ref|>text<|/ref|><|det|>[[77, 156, 905, 345]]<|/det|>
+Thus far, developmental EEG studies have largely focused on theta/alpha oscillations which are modulated by thalamocortical interactions and are associated with cognitive functions of attention and memory \(^{17 - 19}\) . Multiple studies of the first two years of life have observed a shift in alpha peak frequency from \(5 \text{Hz}\) at 5 months to \(8 \text{Hz}\) at 2 years, coinciding with increases in alpha power across this period \(^{20 - 23}\) . This dominant peak frequency continues to increase into the mature \(10 \text{Hz}\) posterior alpha rhythm by adolescence \(^{9,24}\) . It is hypothesized that the gradual shift in dominant peak frequency is modulated by maturation of thalamocortical circuitry in concert with developmental gains in cognitive functions \(^{9}\) , however, the precise mechanisms remain unknown.  
+
+<|ref|>text<|/ref|><|det|>[[75, 375, 916, 683]]<|/det|>
+Unlike theta/alpha power, little is known about the early developmental changes in periodic beta power. In adults, beta oscillations are strongly associated with sensorimotor processing in addition to higher- order cognitive tasks such as working memory \(^{25}\) . Similar to alpha oscillations, the generation of beta oscillations relies on GABAergic interneuron networks and thalamocortical connectivity. In adults, low- dose GABA- modulating anesthetics induce a sedative state with 13- 25Hz beta oscillations, whereas higher doses used to maintain unconsciousness progressively slow these beta oscillations into coherent, frontal specific, alpha oscillations \(^{26 - 28}\) . However, GABA- dependent anesthesia- induced frontal alpha coherence does not emerge in infants until after 10 months of age and is not consistently present until 15- 20 months of age \(^{29 - 31}\) . Anesthesia- induced alpha coherence is hypothesized to involve GABA- dependent thalamocortical loops leading to hypersynchronization between thalamic and prefrontal cortices \(^{27,32,33}\) . Therefore, potential covariation of developing beta oscillations and anesthesia- induced counterparts may lend insight into the role and time course of developing inhibition in human thalamocortical circuit development.  
+
+<|ref|>text<|/ref|><|det|>[[75, 716, 920, 927]]<|/det|>
+Using longitudinal EEG data collected from 592 healthy infants (yielding a total of 1335 EEGs) from 2 to 44 months after birth, we characterize early developmental trajectories of EEG aperiodic and periodic power from 2- 50Hz and to identify potential ages relevant to sequential steps in inhibitory network and thalamocortical circuit development. Consistent with the transient and stepwise developmental progression of thalamocortical circuitry, we observe transient periodic peaks in alpha power at 2- 3 months and high beta power at 4- 18 months. A low beta peak (12- 20Hz) also begins to emerge in infants starting as early as 6 months of age. We hypothesized that emergence of this low beta peak reflects maturation of early connections between the thalamus and cortex. To test this hypothesis, we leveraged a smaller dataset consisting of a cohort of infants with EEG recordings before and during clinical
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[77, 61, 919, 128]]<|/det|>
+anesthesia. Consistent with our hypothesis, we find infants with an identifiable low beta peak have higher anesthesia- induced alpha coherence than those that do not, suggesting that the emergence of this peak is associated with thalamocortical loop maturation.  
+
+<|ref|>sub_title<|/ref|><|det|>[[79, 158, 168, 175]]<|/det|>
+## RESULTS  
+
+<|ref|>text<|/ref|><|det|>[[75, 204, 920, 440]]<|/det|>
+Resting- state EEG were collected longitudinally from 592 healthy infants, aged 2- 44 months, across 4 studies occurring in the same laboratory (Fig 1C, Table 1). Whole brain power spectra for each individual were calculated by averaging across electrodes (Supplemental Figure 1. Individual spectra shown in Supplemental Figure 2). Spectra were then averaged across individuals within 8 age bins (Fig 1D). Notable nonlinear changes in aperiodic and periodic power spectra were observed between age bins, including transient peaks in the periodic spectrum across both alpha and beta frequency ranges (Fig 1E- G). To further characterize these developmental changes in the spectra, we used generalized additive mixed models (GAMMs) to model non- linear trajectories of power parameters. For each model an age- by- sex interaction was tested for significance. If not significant, the interaction term was removed and the model was refit using sex as an additive covariate. All models also included study as a covariate factor.  
+
+<|ref|>text<|/ref|><|det|>[[78, 470, 515, 488]]<|/det|>
+Aperiodic power increases most during first year of life  
+
+<|ref|>text<|/ref|><|det|>[[74, 515, 915, 825]]<|/det|>
+First, we assessed age- dependent changes in the aperiodic component and observed the largest developmental increases in aperiodic power between 2 and 8 months after birth (Fig 1E). Aperiodic offset, but not slope, significantly increased with age (FDR- adjusted \(q\) value \(< 0.01\) ), and age- by- sex interactions were present for both aperiodic offset ( \(F = 4.59\) , \(q = 0.01\) ) and slope ( \(F = 3.18\) , \(q = 0.02\) ). Modeled developmental trajectories of the aperiodic offset showed a sharp linear increase over the first year after birth for both males and females (Fig 1H). Modeled developmental trajectories of the aperiodic slope showed a gradual increase over the first year. These findings contrast with consistent reports of decreasing offset and slope across child and adulthood \(^{11 - 14}\) , and likely reflect the known increases in brain volume and synaptogenesis occurring across the first year of life. Differences in developmental trajectories between 4 regions of interest (ROI) (frontal, central, temporal, and posterior) were also assessed (Supplemental Figure 3). The posterior ROI had higher aperiodic offset than all other ROIs, with the greatest increase in offset occurring in the first year (Frontal \(F = - 26.32\) , \(q < 0.0001\) , Central \(F = - 40.63\) , \(q < 0.0001\) , Temporal - 31.85 , \(q < 0.0001\) ; Supplemental Fig 3B).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[97, 60, 889, 662]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[99, 674, 840, 805]]<|/det|>
+<center>Figure 1. Developmental Trajectories of Aperiodic and Periodic Power Spectra (A) Example aperiodic power spectra. Offset is defined as power at 2.5Hz. (B) Example periodic power spectra. Peaks defined as maxima within a defined frequency range. Band power defined as the integral of the periodic power spectra between defined frequencies. (C) Longitudinal study enrollment. (D-F) Absolute, Aperiodic, and Periodic power spectra averaged across individuals within 8 age bins between 2 and 44 months. (G) Age-related changes in periodic power. (H-I) GAMMs modeled trajectories of aperiodic offset and slope for males (orange) and females (blue). Relative inflection points are shown with circular markers. Below, heatmaps show the standardized change in offset or slope per day, defined as [change per day]/ [standard deviation of measures across full age range] </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[30, 60, 920, 568]]<|/det|>
+Transient 9.5Hz alpha peak observed in 2- 4 month- old infantsAt the youngest age bin (2- 4 months) two peaks with similar amplitude are observed across the theta/alpha (4- 12Hz) range in the majority of infants (69%; Fig 2A,C). A lower frequency peak is observed in the theta (4- 6Hz) range at \(5.5 \pm 0.3\mathrm{Hz}\) , and higher frequency peak is observed in the alpha (6- 12Hz) range at \(9.5 \pm 0.45\mathrm{Hz}\) . However, by 6- months only 15% of infants have two peaks in this range, and for most infants it is the higher 9.5Hz peak that is no longer observed. At 6- months fewer than 40% of infants exhibit a dominant peak in the “alpha” (6.5- 12Hz) range (Fig 1D) and the average peak frequency in the theta/alpha range is \(6.3 \pm 1\mathrm{Hz}\) . This disappearance of the higher peak after 4- months of age may reflect a transient step in thalamocortical circuitry development. Previous research has observed a gradual shift in peak frequency from 5 to 8Hz from infancy to early childhood, however these studies started no earlier than 5 months of age \(^{20 - 23}\) . In order to assess whether an increase in peak frequency beginning at 5 months is present in our data set we modeled developmental trajectories of peak amplitude and frequency between 4- 12 Hz starting at 170 days, when the vast majority of EEGs exhibited a single dominant peak. No age- by- sex interactions were observed in GAMMs modeled trajectories, and consistent with previous studies peak frequency and peak amplitude significantly increased with age (frequency: \(\mathrm{F} = 12.9\) , q <0.0001 Fig 2E; amplitude: \(\mathrm{F} = 16.78\) , q < 0.0001, Fig 2F). Figure 2G- I show modeled trajectories for EEG power calculated over defined frequency bands commonly used in infant EEG research: theta (4- 6Hz), low alpha (6- 9Hz), and high alpha (9- 12Hz). An age- by- sex interaction was present for theta power, although qualitatively the shapes of trajectories were similar (Fig G; \(\mathrm{F} = 4.17\) , q < 0.01).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[92, 65, 802, 715]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[100, 719, 836, 896]]<|/det|>
+<center>Figure 2. Transient and nonlinear changes in periodic power between 4 and 12 Hz. (A,B) Individual periodic power spectra for 2-4 months, and 6-8 months old. Red markers show peaks between 4 and 12 Hz. (C) Proportion of infants with two peaks identified between 4 and 12 Hz at each age bin. (D) Proportion of infants with an identified peak between 6.5 and 12Hz at each age bin. (E-I) GAMMs modeled trajectories for males (orange) and females (blue). Relative inflection points are shown with circular markers. Below, heatmaps show the standardized change in offset or slope per day, defined as [change per day]/[standard deviation of measures across full age range]. Both male and female heatmaps shown for models with significant age x sex interaction. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[75, 111, 912, 444]]<|/det|>
+Several age- dependent transient changes are observed in the low beta (13- 20Hz) and high beta (2035Hz) range. First, the shape of the periodic power spectra in the low beta range is notable for a prominent trough prior to 1 year of age (Fig 3A), with only \(10\%\) of infants (24/222) exhibiting a low beta peak between 6- 8 months of age (Fig 3A- C). After 8 months, a low beta peak begins to emerge in some of infants, with \(48\%\) (52/107) showing a peak at 18- 20 months, and \(70\%\) (199/285) by 36 months (Fig 2C). As a low beta peak was not identified in many children across the age range, peak amplitude and frequency was not modeled. In contrast, virtually all \((99.5\%)\) of the infants had an identifiable high beta peak prior to 12 months of age (Fig 3A). However, notable nonlinear shifts in frequency and amplitude of the high beta peak were observed (Fig 1F, 3D and E). During the first year after birth, the high beta peak amplitude rapidly increases, peaking at 229 days (7.5 months), and then substantially decreases until 802 days(2.2 years). High beta peak frequency trajectories are also nonlinear, with peak frequency at its highest at 473 days (male 29.0Hz, female 29.4Hz), followed by a steady decline in frequency. Modeled trajectories of periodic power for commonly used frequency bands are shown in Fig 3F- H: low beta (12- 20Hz), high beta (20- 30Hz), and gamma (30- 45Hz).  
+
+<|ref|>text<|/ref|><|det|>[[78, 470, 919, 587]]<|/det|>
+The observed nonlinear changes across the beta range are striking. While many EEG infant studies group beta oscillations into a singular frequency range, the data presented here supports that low and high beta have distinct developmental origins. Specifically, between 6- 24 months we observe the gradual emergence of a low beta peak, and simultaneously the rise and fall of a prominent high beta peak, ultimately resolving into a broader beta peak by 36 months.  
+
+<|ref|>text<|/ref|><|det|>[[77, 612, 916, 802]]<|/det|>
+Traditionally, beta oscillations measured in children and adults are associated with sensory and motor processing, where reductions in beta power are observed during the preparation or execution of motor tasks<sup>25</sup>. However, beta activity has also been shown to be modulated during a wide range of nonmotor cognitive tasks<sup>25,34</sup>. The developmental emergence of low beta oscillations may represent sensorimotor skills (e.g., crawling, walking) gained during this period, but may also represent the developmental maturation of neurobiological circuitry. For example, GABAergic interneuron networks and thalamocortical connectivity are highly associated with the generation of cortical beta oscillations, as well as anesthesia- induced frontal alpha coherence, but neither are not fully established at birth<sup>29</sup>.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[88, 60, 852, 732]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[101, 753, 896, 904]]<|/det|>
+<center>Figure 3. Transient and nonlinear changes in periodic power between 12 and 35Hz. (A,B) Individual periodic power spectra 6-8 months, and 18-20 months old. Red markers show peaks between 12-20 Hz. Green markers show peaks between 20-35Hz. (C) Proportion of infants with peak identified between 12-20 Hz at each age bin. (D-H) GAMMS modeled trajectories for males (orange) and females (blue). Relative inflection points are shown with circular markers. Below, heatmaps show the standardized change in offset or slope per day, defined as [change per day]/[standard deviation of measures across full age range]. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[30, 60, 920, 465]]<|/det|>
+1 Low beta peak associated with developmental emergence of anesthesia- induced frontal alpha coherence in infants 2 We hypothesized that developmental changes in infant beta power measured in a resting state may represent concurrent maturation of GABAergic interneuron networks and thalamocortical connectivity. To explore this possibility, we assessed EEG recordings of healthy infants before and during exposure to GABA- modulating sevoflurane anesthesia<sup>35</sup>. All infants were undergoing elective procedures (eg. circumcision) and infants were excluded for prematurity, neurologic injury, epilepsy, or planned intracranial surgery. Here we hypothesized that the emergence of low beta oscillations (as measured by the presence of a low beta peak) before anesthesia would be associated with GABA- dependent anesthesia- induced frontal alpha coherence. EEG data from 36 infants (6- 15 months old), collected during the awake and anesthetized state were analyzed. Developmental changes in the aperiodic- adjusted power spectra in this smaller dataset were qualitatively similar to those described above (Figure 4A), with a low beta peak beginning to emerge after 7 months and present in roughly half the infants between 7- 12 months of age (11/21). As hypothesized, alpha coherence was significantly increased in those with a low beta peak compared to those without (ANCOVA, with sevoflurane level as covariate; p <0.05; Figure 4B).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[96, 75, 845, 355]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[92, 388, 909, 485]]<|/det|>
+<center>Figure 4. Increased anesthesia-induced alpha coherence in infants with identified low beta peak in baseline EEG. (A) Periodic power spectra of infants between 6 and 15 months old prior to receiving anesthesia (B) Mean alpha coherence during anesthesia in infants 7-12 months old, with (light blue) or without (white) an identified low beta peak. Ancova with sevoflurane levels as covariate: \(\mathrm{p}< 0.05\) . </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[80, 525, 179, 542]]<|/det|>
+## Discussion  
+
+<|ref|>text<|/ref|><|det|>[[75, 570, 914, 857]]<|/det|>
+Here we present the largest- to- date longitudinal analysis of EEG data collected between 2 – 44 months of age. Findings provide insight into the developmental timing of inhibitory network and thalamocortical circuit maturation during human infancy. Several age- dependent findings in our study contrast to previous longitudinal studies of child and adulthood. First, we observe increases in both aperiodic offset and slope, especially during the first year, whereas decreases in both measures are observed starting as early as 4 years of age and continue to decrease with adulthood<sup>11–14</sup>. Second, while expected shifts in the dominant peak from the theta to alpha range were observed between 5 to 44 months, in the 2- 4 months age bin, a 9.5Hz peak was also transiently observed. Third, striking changes within the beta (12- 30Hz) range were observed, including the emergence of a low beta peak starting after 6- months of age, and age- dependent shifts in high beta peak frequency and amplitude - first increasing and then decreasing with age. Below we discuss how the above age- related changes may represent sequential developmental maturation in the inhibitory system and thalamocortical network connections.  
+
+<|ref|>text<|/ref|><|det|>[[75, 882, 900, 926]]<|/det|>
+The aperiodic offset is hypothesized to represent broad band neuronal firing<sup>36,37</sup>, and thus early increases in aperiodic offset are consistent with established increases in neuronal number, gray matter
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[75, 60, 920, 248]]<|/det|>
+volume, and synaptic number during the first year. Stabilization of aperiodic offset after 1- year of age is also consistent with MRI findings that gray matter volume doubles during the first postnatal year and then slows to \(20\%\) in its second year<sup>38- 40</sup>. Regionally, we also observe differences between posterior and frontal aperiodic offset trajectories, which either plateau after 1 year (posterior), or have a slow continued increase (frontal) beyond 1 year of age (Supplemental Fig 3). Consistent with this pattern, synaptogenesis differs across cortical regions, with the posterior visual cortex exhibiting a burst in synapse formation between 3- 4 months of age, whereas the prefrontal cortex shows peak synaptogenesis around 8 months of age and continued gains during the second year of life<sup>1</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[75, 277, 920, 704]]<|/det|>
+Our observed age- dependent increases in aperiodic slope in infancy also contrast with multiple studies covering child to adulthood, where decreases in slope have consistently been reported. Schaworonskov et. al.<sup>41</sup> also reported decreased slope with age in infants from 1 to 7- months- old, however the parameterization of the spectra in that study was limited to 1- 10Hz due to excessive muscle noise in the data, and it is unclear how the shifts in 4- 12Hz periodic power described below may affect modeling of the underlying aperiodic component in this range. We hypothesize that observed increases in aperiodic slope reflect changes in inhibitory networks that are unique to early development. Indeed, aperiodic slope from EEG recorded from sleeping newborns is observed to increase with age during the first 7 weeks after birth<sup>42</sup>. Growing evidence suggests that aperiodic slope is modulated by the balance between excitation and inhibition, with increased slope associated with a reduction in E/I ratio<sup>10,12,43</sup>. An age- dependent reduction in E/I ratio during the first postnatal year is consistent with the prolonged developmental timing of inhibitory network maturation in humans. Unlike excitatory neurons which are well established by birth, during the first postnatal year GABAergic inhibitory neurons continue to migrate from ventral subregions of the brain to the cortices where they ultimately mature and integrate into neuronal networks<sup>44</sup>. In addition, during this first year GABAergic responses switch from being excitatory to inhibitory due to changes in the concentration of chloride channels on cell membranes<sup>45- 47</sup>. Overall, inhibitory neuron network integration and the excitatory- to- inhibitory GABA switch are unique to this developmental period and likely lead to increased inhibitory tone during the first year.  
+
+<|ref|>text<|/ref|><|det|>[[75, 732, 914, 920]]<|/det|>
+Observed changes in the periodic spectra may reflect sequential steps in inhibitory network and thalamocortical circuit development. Transient neural circuits are common in postnatal development and play critical steps in normal development of thalamocortical circuitry<sup>6</sup>. For example, transient circuits between sublate neurons (SPN) and thalamo- recipient layer 4 spiny stellate neurons help establish thalamocortical connections prior to the maturation of primary sensory cortices<sup>48</sup>. Studies of postmortem fetal monkey and human brains suggests that the SPN in primates and humans slowly begins to disappear in the 3<sup>rd</sup> trimester but may persist until 6 months, with an overlapping period in which the thalamus makes connections with both the SPN and cortical layer IV neurons<sup>6,49,50</sup>. We hypothesize that
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[73, 61, 920, 154]]<|/det|>
+the 9.5Hz peak observed at 2- 4 months, but not at 6 months, reflects this transient period when mature excitatory subplate neurons are still receiving and relaying thalamic input to cortices, resulting in higher frequency alpha oscillations. Additionally, newly established connections between the thalamus and layer IV produce lower frequency theta rhythms that will later become the dominant alpha rhythm.  
+
+<|ref|>text<|/ref|><|det|>[[72, 180, 920, 584]]<|/det|>
+The thalamus is thought to play a central role in the generation of the mature posterior alpha rhythm. A shift in dominant oscillatory frequency in the theta/alpha range (4- 12Hz) across early childhood has been observed now in many studies<sup>21,24,51</sup>. Here, we both confirm and extend those findings over the first 3 years after birth, with peak frequency increasing most between 4 and 18 months. What factors are potential contributors to this shift in peak frequency? The dynamic circuit motif model (DCM) proposes that cortical network rhythms result from a combination of the intrinsic resonant frequency of a neuronal population and the time course properties of the inhibitory inputs on the neuronal population<sup>52</sup>. Under the DMC model, prior to the maturation of both local inhibitory circuitry and thalamocortical feedback loops, peak frequency oscillations as measured by scalp EEG are more likely to represent the intrinsic properties of cortical networks, with thalamic inputs beginning to play a larger role with age. For example, lower frequency 4- 7Hz oscillations are intrinsically generated by isolated layer 5 cortical neurons, and the range of oscillations increases to 5- 12Hz when connections to other cortical layers remain intact<sup>53</sup>. Thalamic neurons in the lateral geniculate nucleus also fire across the theta and alpha range. In vitro slice experiments from cats suggest that cortical input to thalamus modulates whether theta versus alpha oscillations are dominant<sup>54,55</sup>. Thus, the developmental shift in peak frequency from the canonical theta to alpha range over the first three years after birth, may represent the integration and maturation of cortical inhibitory neurons, as well as the establishment and maturation of thalamocortical connections.  
+
+<|ref|>text<|/ref|><|det|>[[72, 611, 923, 870]]<|/det|>
+Finally, our study identified early age- dependent changes in periodic beta power that we hypothesize are associated with thalamocortical loop maturation. We observe the emergence of a low beta peak in infants older than 6- months of age and find that the presence of a low beta peak is associated with higher anesthesia- induced frontal alpha coherence. Biophysical models demonstrate that this frontal anesthesia- induced alpha coherence requires inputs from both the thalamus and cortex<sup>27</sup>. Together these findings suggest that low beta oscillations may directly reflect thalamocortical loop maturation. Beta rhythms are thought to be both generated locally in the cortex through pyramidal- interneuron loops, as well as through thalamus to cortical connections that also rely on inhibitory inputs<sup>25</sup>. The emergence of the low beta peak in awake infants may reflect the combination of newly established network connections between thalamic nuclei and cortical layers, as well as the maturation of interneurons within the thalamocortical pathways.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[75, 60, 920, 416]]<|/det|>
+It is also possible that developmental changes in beta power are related to infant movement. During EEG acquisition, infants are held in their parent's lap and behavioral supports are in place to keep the infants calm. However, it is not possible to control the infants' movement, and movements both small (hand movements) and large (head turns, leg and arm movements) ubiquitously occur across recordings - likely increasing over the first year as infants become more mobile. Our preprocessing artifact removal pipelines (see Methods) includes several steps for removing high frequency noise from muscle artifact. However this would not remove EEG signal in response to sensorimotor processing. Infant jaw and upper limb movements have been shown to increase power between 9- 20Hz along frontal and occipital sites, while hand and lower limb movements do not have significant effects<sup>56</sup>. In our dataset, increases in low beta power were most prominent in central (not frontal or occipital) ROIs (Supplemental Fig 3K), suggesting that age- dependent changes in beta power more likely represent underlying brain maturation than sensorimotor processing or movement artifact during data acquisition. This is further supported by the consistency in age- dependent shifts of both low and high beta across individuals (see Supplemental Fig 2 for individual power spectra plots), and our observation that low beta is correlated with anesthesia- induced alpha coherence.  
+
+<|ref|>text<|/ref|><|det|>[[75, 444, 912, 775]]<|/det|>
+Transient high beta peaks were also observed across this early period of development. Specifically, we observed early increases in high beta peak amplitude, which reached a maximum at 7.5 months, followed by decreases in both high beta peak amplitude and frequency, such that by 36 months the low beta peak is the dominant peak across the 12- 30Hz range. The neurobiological mechanism of this high beta peak is unclear. As discussed above, administration of GABA agonists induces beta activity. In addition, several neurodevelopmental disorders associated with GABA receptor dysfunction show prominent beta peaks on EEG; individuals with Duplication 15q have a prominent beta peak at 23Hz<sup>57,58</sup>, and we have observed that children with Fragile X Syndrome (FXS), aged 3- 7 years, have a prominent 30Hz peak<sup>59</sup>. This 30hz peak observed in FXS children is qualitatively similar to the 30Hz peak observed in the present dataset at a much younger age. Further analysis of data previously published from FXS children shows that the observed high beta peak decreases with age (Figure 5), suggesting delayed brain development. Such observations highlight the value of the longitudinal EEG trajectories presented in this paper in placing findings from neurodevelopmental disorders in the broader context of developmental brain maturation.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[93, 70, 682, 290]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[100, 325, 756, 373]]<|/det|>
+<center>Figure 5. FXS children have increased high beta peak that decreases with age. (A) Periodic power spectra of 35-48 month-old children with and without FXS. (B) Individual periodic power spectra of FXS children, with line hue corresponding to age in months. </center>  
+
+<|ref|>text<|/ref|><|det|>[[75, 384, 912, 572]]<|/det|>
+In summary, our work highlights the dynamic developmental changes in neural activity occurring during the first three years after birth and provides insights in potential ways these age- dependent and sometimes transient changes may coincide with sequences in thalamocortical and inhibitory network maturation. Our findings help to ground cross- sectional work occurring at these early ages and provide a foundation to compare developmental trajectories of various neurodevelopmental disorders including autism, ADHD, and rare genetic disorders. Future studies examining early trajectories of functional connectivity and phase amplitude coupling across this age range will provide additional insight into the timing of critical periods in brain maturation.  
+
+<|ref|>sub_title<|/ref|><|det|>[[80, 625, 173, 642]]<|/det|>
+## METHODS  
+
+<|ref|>text<|/ref|><|det|>[[77, 670, 920, 835]]<|/det|>
+Studies and Participants: Lab based EEGs for this paper were collected as part of four different studies occurring over 15 years conducted at our lab at Boston Children's Hospital (Figure 1A). Sample numbers for each age are shown in Table 1. Study 1, the Healthy Baby Study (IRB- P00019083), was a longitudinal study, enrolling infants starting at 2 months of age, from the Boston Children's Hospital Primary Care Center, which predominantly services families from low- income backgrounds. EEG was collected and developmental assessment using the Mullen Scales of Early Learning (MSEL) was performed at 2, 6, 9, 12, 24, and 36 months.  
+
+<|ref|>text<|/ref|><|det|>[[77, 863, 896, 932]]<|/det|>
+Study 2, The Infant Sibling Study (IRB- X06- 08- 0374), and Study 4, the Infant Screening Project (IRB- P00018377), were both prospective, longitudinal studies, enrolling infants with and without first degree family history of ASD starting as early as 3- months of age. For this analysis only infants without family
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[31, 60, 911, 193]]<|/det|>
+history of ASD were included. Study 4 also included a group of infants with elevated social communication concerns at 12 months of age, and they were also excluded from this analysis. EEGs and MSEL were performed at 3, 12, 18, 24, and 36 months for both studies, as well as 6 and 9 months for Study 2. Infants were specifically assessed for ASD using the Autism Diagnostic Observation Schedule (ADOS) in conjunction with clinical best estimate at 24- and 36- month visits.  
+
+<|ref|>text<|/ref|><|det|>[[77, 205, 912, 320]]<|/det|>
+Study 3, the Emotion Project (IRB- P00002876), was a cohort/longitudinal study. Infants were enrolled at either 5, 7, or 12 months, and then followed through 7 years of age. In addition to the first time point, EEG data was again collected at 3 years of age. There were no developmental assessments performed for this study, however parent questionnaires regarding child development, diagnoses (e.g., ASD), and therapies were collected.  
+
+<|ref|>text<|/ref|><|det|>[[77, 350, 904, 441]]<|/det|>
+All infants had a minimal gestational age of 36 weeks, no history of prenatal or postnatal medical or neurological problems, and no known genetic disorders. Infants who were later diagnosed with ASD (either by assessment during the study, or by community diagnosis disclosed by parents prior to age of 5) were not included in this study.  
+
+<|ref|>text<|/ref|><|det|>[[77, 469, 918, 585]]<|/det|>
+Sample characteristics across and within studies are shown in Table 1. The analysis included 1335 EEGs collected from 592 participants. While all studies took place in the same laboratory, participant demographics vary between studies as was expected given differences in recruitment and research aims of each study. In addition, studies differed in the age of enrollment and when subsequent visits were completed (Figure 1C). Combined, the sample remained predominantly white (74%).
+
+<--- Page Split --->
+<|ref|>table_caption<|/ref|><|det|>[[40, 65, 380, 80]]<|/det|>
+1 Table 1: Sample Characteristics
+
+<|ref|>table<|/ref|><|det|>[[80, 91, 909, 770]]<|/det|>
+
+<table><tr><td></td><td>Combined<br>Studies<br>\(N=592\)</td><td>Study 1<br>\(N=49\)</td><td>Study 2<br>\(N=72\)</td><td>Study 3<br>\(N=363\)</td><td>Study 4<br>\(N=108\)</td></tr><tr><td>Sex, % Female (n)</td><td>46.3(274)</td><td>51.0(25)</td><td>45.8(33)</td><td>46.3(168)</td><td>44.4(48)</td></tr><tr><td>Ethnicity, % (n)</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Hispanic</td><td>8.6(51)</td><td>16.3(8)</td><td>1.4(1)</td><td>10.5(38)</td><td>3.7(4)</td></tr><tr><td>Non-Hispanic</td><td>90.2(534)</td><td>81.6(40)</td><td>97.2(70)</td><td>88.4(321)</td><td>95.4(103)</td></tr><tr><td>Not Answered</td><td>1.2(7)</td><td>2.0(1)</td><td>1.4(1)</td><td>1.1(4)</td><td>0.9(1)</td></tr><tr><td>Race, % (n)</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>White</td><td>74.0(438)</td><td>10.2(5)</td><td>86.1(62)</td><td>79.1(287)</td><td>77.8(84)</td></tr><tr><td>Black or African-American</td><td>6.4(38)</td><td>53.1(26)</td><td>1.4(1)</td><td>2.2(8)</td><td>2.8(3)</td></tr><tr><td>Asian</td><td>3.9(23)</td><td>4.1(2)</td><td>2.8(2)</td><td>4.1(15)</td><td>3.7(4)</td></tr><tr><td>Mixed Race</td><td>12.8(76)</td><td>14.3(7)</td><td>8.3(6)</td><td>12.9(47)</td><td>14.8(16)</td></tr><tr><td>Other</td><td>1.4(8)</td><td>12.2(6)</td><td>0</td><td>0.6(2)</td><td>0</td></tr><tr><td>Not Answered</td><td>1.5(9)</td><td>6.1(3)</td><td>1.4(1)</td><td>1.1(4)</td><td>0.9(1)</td></tr><tr><td>Income, % (n)*</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>&lt;\(35,000\)</td><td>5.4(32)</td><td>32.7(16)</td><td>4.2(3)</td><td>2.5(9)</td><td>3.7(4)</td></tr><tr><td>\(35,000-\)75,000</td><td>11.1(66)</td><td>18.4(9)</td><td>9.7(7)</td><td>13.2(48)</td><td>1.9(2)</td></tr><tr><td>\(>75,000\)</td><td>73.8(437)</td><td>18.4(9)</td><td>70.8(51)</td><td>76.3(277)</td><td>92.6(100)</td></tr><tr><td>Not Answered or Don't Know</td><td>9.6(57)</td><td>30.6(15)</td><td>15.3(11)</td><td>8.0(29)</td><td>1.9(2)</td></tr><tr><td colspan="6">Participant EEG Data Included in Analysis, n</td></tr><tr><td>2-4m</td><td>97</td><td>46</td><td>10</td><td>-</td><td>41</td></tr><tr><td>4-6m</td><td>119</td><td>-</td><td>2</td><td>113</td><td>4</td></tr><tr><td>6-8m</td><td>223</td><td>43</td><td>50</td><td>130</td><td>-</td></tr><tr><td>8-11m</td><td>95</td><td>37</td><td>58</td><td>-</td><td></td></tr><tr><td>11-15m</td><td>291</td><td>40</td><td>62</td><td>102</td><td>87</td></tr><tr><td>18-20m</td><td>107</td><td>-</td><td>40</td><td>-</td><td>67</td></tr><tr><td>23-30m</td><td>118</td><td>22</td><td>44</td><td>-</td><td>52</td></tr><tr><td>35-44m</td><td>285</td><td>18</td><td>48</td><td>174</td><td>45</td></tr></table>
+
+<|ref|>text<|/ref|><|det|>[[40, 770, 880, 808]]<|/det|>
+2 *2 infants reported to have income of &#36;30-39,000 were placed in &lt;35,000 category. 1 infant with reported income of &#36;70-79,000 placed in &gt;75,000 category.
+
+<|ref|>text<|/ref|><|det|>[[40, 819, 55, 830]]<|/det|>
+4
+
+<|ref|>text<|/ref|><|det|>[[40, 844, 55, 855]]<|/det|>
+5
+
+<|ref|>text<|/ref|><|det|>[[40, 866, 868, 880]]<|/det|>
+6 **Lab-based EEG data collection**: Baseline, non-task-related EEG data was collected using similar
+
+<|ref|>text<|/ref|><|det|>[[40, 890, 915, 904]]<|/det|>
+7 methods and rooms for all four studies. The infant was held by their seated caregiver in a dimly lit, sound
+
+<|ref|>text<|/ref|><|det|>[[40, 914, 907, 927]]<|/det|>
+8 attenuated room with a low-electrical-signal background. For Study 2, a research assistant ensured that
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[73, 60, 914, 345]]<|/det|>
+the infant remained calm by blowing bubbles and/or showing toys. For Studies 1 and 3, a video of infant toys was shown for 2- 5 minutes and 2 minutes, respectively. For Study 4, a video of abstract moving objects was shown for 2- 5 minutes. Continuous scalp EEG for Studies 1, 3, and 4 was recorded using a 128- channel Hydrocel Geodesic Sensor Nets (Electrical Geodesics, Inc., Eugene, OR) connected to a NetAmps 300 amplifier (Electrical Geodesic Inc.) and sampled at 500Hz. Study 2 included recordings using 64- channel Geodesic Sensor ( \(< 10\%\) of data) or a 128- channel Hydrocel Geodesic Sensor Nets (Electrical Geodesics, Inc., Eugene, OR), connected to either a NetAmps 200 or 300 amplifier (Electrical Geodesic Inc.) and sampled at either 250 or 500Hz. Additional statistical analysis related to differences in net and amps is described below in EEG Power analysis. For all studies, data was referenced online to a single vertex electrode (Cz) and impedances were kept below 100kΩ in accordance with the impedance capabilities of the high- impedance amplifiers inside the electrically shielded room. Electrooculographic electrodes were removed to improve the child's comfort.  
+
+<|ref|>text<|/ref|><|det|>[[75, 373, 919, 702]]<|/det|>
+EEG pre- processing: Raw Netstation (Electrical Geodesics, Inc) files were exported to MATLAB (version R2017a) for preprocessing and absolute power calculations using the Batch Automated Processing Platform (BEAPP \(^{60}\) ) with integrated Harvard Automated Preprocessing Pipeline for EEG (HAPPE \(^{61}\) ). For each EEG, a 1Hz high- pass and 100Hz low- pass filter were applied, data sampled at 500Hz were resampled to 250Hz, and then run through the HAPPE module consisting of 60Hz line noise removal, bad channel rejection, and artifact removal using combined wavelet- enhanced independent component analysis (ICA) and Multiple Artifact Rejection Algorithm (MARA \(^{5,6}\) ). The following channels, in addition to the 10- 20 electrodes, were used for MARA: 64- channel net – 16, 9, 8, 3, 58, 57, 21, 25, 18, 30, 43, 50, 53, 32, 33, 38, 41, 45; and 128- channel net - 28, 19, 4, 117, 13, 112, 41, 47, 37, 55, 87, 103, 98, 65, 67, 77, 90, 75. These electrodes were chosen as they evenly cover all brain regions of interest (Supplemental Figure 1). After artifact removal, channels removed during bad channel rejection were then interpolated, data were referenced to the average reference, detrended to the signal mean, and segmented into 2- second segments. Any segments with retained artifact were rejected using HAPPE's amplitude and joint probability criteria.  
+
+<|ref|>text<|/ref|><|det|>[[77, 730, 904, 871]]<|/det|>
+EEG rejection criteria: EEG recordings were rejected using the following HAPPE data quality measures: Fewer than 20 segments (40 seconds of total EEG), percent good channels \(< 80\%\) , percent independent components rejected \(>80\%\) , mean artifact probability of components kept \(>0.3\) , and percent variance retained \(< 25\%\) . Expected differences between studies in number of segments remaining post- segment rejection were observed, with Study 3 with the shortest resting state recording period having fewer segments. All other quality metrics were similar across studies (Table 2).
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[97, 94, 833, 308]]<|/det|>
+<|ref|>table_caption<|/ref|><|det|>[[94, 72, 355, 88]]<|/det|>
+Table 2: EEG data quality metrics   
+
+<table><tr><td></td><td>Combined Studies<br>N = 1335</td><td>Study 1<br>N = 206</td><td>Study 2<br>N = 314</td><td>Study 3<br>N = 519</td><td>Study 4<br>N = 296</td></tr><tr><td colspan="6">EEG quality metrics, Mean±SD</td></tr><tr><td>Number of Segments</td><td>81±39.5</td><td>126±27.5</td><td>85.8±39.9</td><td>48.8±7.1</td><td>105.2±30.3</td></tr><tr><td>Percent Good Channels</td><td>92.2±4.5</td><td>92.2±4.8</td><td>92.6±4.4</td><td>92.0±4.6</td><td>93.5±4.3</td></tr><tr><td>Percent ICs Rejected</td><td>35.7±10.3</td><td>35.7±10.6</td><td>35.4±10.1</td><td>34.1±14.0</td><td>35.0±10.5</td></tr><tr><td>Percent Variance Kept of Post Wavelet Data</td><td>67.2±17.0</td><td>63.1±16.1</td><td>66.5±15.7</td><td>68.8±18.3</td><td>68.1±16.1</td></tr><tr><td>Mean Artifact Probability of Kept ICs.</td><td>0.12±0.05</td><td>0.12±0.04</td><td>0.11±0.04</td><td>0.12±0.05</td><td>0.12±0.05</td></tr></table>  
+
+<|ref|>text<|/ref|><|det|>[[70, 365, 916, 940]]<|/det|>
+EEG Power analysis: The power spectral density at each electrode, for each 2- second segment, was calculated in the BEAPP Power Spectral Density (PSD) module using a multitaper spectral analysis<sup>62</sup> and three orthogonal tapers. For each electrode, the PSD was averaged across segments, and then further averaged across all available electrodes, or frontal, temporal, central, and posterior regions of interest (Supplemental Figure 1C,D). The PSD was then further analyzed using a modified version of FOOOF v1.0.0<sup>8</sup>(https://github.com/fooof-tools/fooof; in Python v3.6.8) in order to model periodic and aperiodic components of the power spectra. FOOOF required modification for use in this age range, as power spectrum models for 2- 7 month ages showed poor model fit (increased mean squared error) for frequencies between 10- 20Hz. Specifically, the FOOOF modeled curves did not accurately capture the "trough" in the power spectra visually observed in this frequency range at younger ages (Supplemental Figure 4). To improve model fit, the robust_ap_fit function, which initially defines the aperiodic component, was modified so that the initial estimate of the flattened power spectra (flatspect) has a baseline elevated such that the lowest point is \(\geq 0\) , to avoid omitting data important across the 2- 7 month age range. This is in contrast to the original method of setting all negative points of the initial flattened power spectra equal to 0. In the original and modified scripts, this initial fit is combined with thresholding to render a more robust second round of aperiodic parameters. After these second aperiodic parameters have been defined, the fit function re- estimates the flattened spectra (spectrum_flat). At this point, prior to fitting spectra peaks, the modified code sets negative data in the flattened spectra equal to 0, similar to the approach of the original code during the initial aperiodic fit. In both versions, aperiodic parameters are fit a third and final time to the spectra with peaks removed (spectrum_peak_rm). The FOOOF model was used in the fixed mode (no spectral knee) with peak_width_limits set to [0.5, 18.0], max_n_peaks = 7, and peak_threshold = 2. Code is available (osf.io/u3gp4) which runs both the original and modified versions of FOOOF, and graphs the RMSE across frequencies for both versions, separate by age. Comparisons are presented in Supplemental Figure 4. Further analyses were subsequently restricted to
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[77, 61, 890, 130]]<|/det|>
+2.5- 50Hz given elevated error between 2- 2.5, and 50- 55Hz. Mean \(\mathsf{R}^2\) for the full sample using this modified version of FOOOF was 0.997 (STD 0.008; range 0.890- 0.999). Mean estimated error for the sample was 0.01 (STD 0.01, range 0.002- 0.09).  
+
+<|ref|>text<|/ref|><|det|>[[75, 156, 919, 545]]<|/det|>
+FOOF provides two parameters to describe the aperiodic 1/f signal: offset and slope. As the FOOOF- determined offset is extrapolated to the estimated aperiodic power at 0Hz, where there are high amounts of error, we instead calculated the aperiodic offset based on aperiodic power at 2.5hz (Figure 1B). The periodic power spectrum (Figure 1B, F) was determined by subtracting the FOOOF estimated aperiodic spectrum (Figure 1E) from the absolute power spectrum (Figure 1D). To further characterize peaks and troughs within the power spectra across development, the periodic spectrum was then smoothed using a savgol filter (scipy.signal.savgol_filter, window length = 101, polyorder = 8). Individual periodic power spectrum plots before and after savgol filter are shown in Supplemental Figure 2. We decided to use this method instead of using the FOOOF estimated peak_fit as the high beta peak appeared to have a non- gaussian shape at some ages, and thus peak_fit estimates did not accurately identify the high beta peak frequency. Using the smoothed periodic spectra, maxima were identified within the following frequency ranges: 4- 6.5 (theta), 4- 12Hz (theta/alpha), 12- 20Hz (low beta) and 20- 35Hz (high beta). A low beta trough was also identified based on the minima between 10- 20Hz. Aperiodic and periodic power across the following canonical frequency bands was calculated taking the integral of each parametrized spectra between the following frequency ranges: theta (4- 6Hz), low alpha (6- 9Hz), high alpha (9- 12Hz), low beta (12- 20Hz), high beta (20- 30Hz), and gamma (30- 45Hz).  
+
+<|ref|>text<|/ref|><|det|>[[75, 563, 909, 727]]<|/det|>
+As Study 2 collected data with 2 net types and 2 amplifiers, data from 6-, 9-, and 12- month age bins were assessed for spectra differences in total (2- 50hz) aperiodic and periodic power as well as aperiodic slope and intercept measure from central electrodes between either 64 and 128 channel nets, or NetAmps 200 or 300 amplifiers. Of the 24 analyses performed, 3 showed significant differences. Nettype differences were observed for 9- and 12- month central aperiodic slope ( \(p = 0.04\) for both) and an amplifier difference was observed for 12- month central offset ( \(p = 0.04\) ). None of these were significant after correcting for multiple comparisons.  
+
+<|ref|>text<|/ref|><|det|>[[75, 755, 914, 920]]<|/det|>
+Anesthesia cohort: EEG data was also collected from infants undergoing anesthesia as part of a prospective observational study approved by the institutional review board at Montefiore Medical Center, Albert Einstein College of Medicine<sup>35,63</sup>. Infants scheduled for elective surgical procedures (e.g., circumcision, hernia repair) were recruited. Infants were excluded for prematurity, known neurologic injury, epilepsy, or planned intracranial surgery. Infants less than 6 months of age, or those documented to be asleep or crying during baseline (pre- anesthesia) recordings were excluded from further analysis. All subjects received general anesthesia with sevoflurane.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[75, 85, 920, 224]]<|/det|>
+EEG recordings were obtained using a Food and Drug Administration (FDA) approved 26 channel device recording from scalp locations designated by the International 10- 20 System, reference midline occipital channel (Oz) (microEEG System, Biosignal, Acton, MA)64. Data was collected from 21 electrodes (FpZ, Fp1, Fp2, Fz, F3, F4, F7, F8, Cz, C3, C4, T3, T4, Pz, P3, P4, T5, T6, Oz, O1, O2) both prior to (baseline) and during sevoflurane induction and maintenance. End tidal sevoflurane concentration was recorded, locked in time with EEG recording. Data were sampled at a frequency of 250 Hz.  
+
+<|ref|>text<|/ref|><|det|>[[75, 252, 911, 440]]<|/det|>
+Baseline EEGs (n = 45) were visually inspected and approximately 2 minutes of continuous EEG with minimal artifact was segmented and then processed using BEAPP/HAPPE60,61. 9 EEGs were excluded due to excessive artifact during visual inspection. No additional EEGs were excluded based on HAPPE data quality criteria. Final sample (n = 36) was an average age of 9.1 months (range 6- 15 months), and predominantly male (n=25). Central periodic power using Pz and Fz electrodes was calculated using using multitaper spectral analysis using three orthogonal tapers. PSD was then further analyzed using a modified version of FOOOF v1.0.0 as described above, in order to model periodic and aperiodic components of the power spectra.  
+
+<|ref|>sub_title<|/ref|><|det|>[[80, 494, 304, 511]]<|/det|>
+## Alpha coherence analysis  
+
+<|ref|>text<|/ref|><|det|>[[75, 516, 916, 680]]<|/det|>
+Preprocessing of Anesthesia EEG data: We used a bipolar montage to analyze the data (F7- Fp1 and F8- Fp2). We developed an automatized method to exclude epochs with high- amplitude noise based on the standard deviation of the time series signal. Three members of the team (CW, JC, and RG) visually inspected the remaining epochs to select 30- second artifact- free segments that were used for the analysis. We selected epochs with a stable sevoflurane concentration defined as two consecutive minutes of end- tidal sevoflurane levels within 0.2% preceding the selection of an epoch of data. EEG data were band- pass filtered from 0.1 to 30 Hz.  
+
+<|ref|>text<|/ref|><|det|>[[75, 707, 910, 896]]<|/det|>
+For each subject, corresponding EEG data collected during anesthesia were inspected to identify 30 second segments under stable sevoflurane concentration and that were "artifact free" (eg. no motion or electrocautery artifacts). Epochs with high- amplitude noise in frontal electrodes (F7- Fp1; F8- Fp2) based on standard deviation of the time series were automatically excluded. Blinded visual inspection by members of the team (CW, JC, and RG) identified remaining epochs with stable sevoflurane concentration without other anesthesia (e.g. propofol bolus) interference. Stable sevoflurane concentration was defined as two consecutive minutes of end- tidal sevoflurane levels within 0.2% preceding the selection of an epoch of data.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[80, 62, 916, 130]]<|/det|>
+calculated to visualize developmental changes within features. The modeled value of a feature at a given age (in days) was subtracted from the modeled value from the subsequent day, and this was divided by the standard deviation of the modeled values of that feature across the age range.  
+
+<|ref|>text<|/ref|><|det|>[[78, 156, 890, 202]]<|/det|>
+To assess the differences between regions of interest, GAMM models were fit with the following form: (3) Power Measure \(\sim\) s(age_days) + oSex + Study+ ROI+ s(age_days, New_ID, bs = 'fs')  
+
+<|ref|>text<|/ref|><|det|>[[78, 229, 916, 345]]<|/det|>
+where the terms have the same meanings as above, and ROI is a factor representing the four regions (frontal, central, temporal, posterior). Because prior literature and preliminary visual inspection of the data indicated that the posterior ROI is most unique in the time course of development, the posterior ROI was set as the reference factor. Thus, the effect and significance associated with each of the other ROIs is a measure of the difference between that ROI and the posterior ROI.  
+
+<|ref|>text<|/ref|><|det|>[[78, 372, 912, 417]]<|/det|>
+Anesthesia Statistical Analysis: ANCOVA, with sevoflurane levels as a covariate, was used to determine effects of presence of low beta peak on anesthesia induced alpha coherence.  
+
+<|ref|>text<|/ref|><|det|>[[78, 444, 878, 489]]<|/det|>
+Figures were created using Python v3.6.8 and python data visualization libraries (matplotlib(60) and Seaborn (https://seaborn.pydata.org/index.html) or in R (version 4.1.2).  
+
+<|ref|>text<|/ref|><|det|>[[77, 515, 904, 606]]<|/det|>
+Funding Support: This research was supported by the National Institutes of Health (R01- DC010290 and MH078829 to CAN, K23DC07983 and T32MH112510 to CLW, and UL1TR002556, KL2TR002558, and K23DA057499 to JYC. Research was also supported by the JPB Research Network on Toxic Stress.  
+
+<|ref|>sub_title<|/ref|><|det|>[[80, 638, 270, 655]]<|/det|>
+## Author Contributions:  
+
+<|ref|>text<|/ref|><|det|>[[77, 660, 904, 848]]<|/det|>
+Carol Wilkinson: Conceptualization, Methodology, Software, Formal analysis, Data Curation, Writing – Original Draft, Visualization. Lisa Yankowitz: Methodology, Software, Formal analysis, Writing – Review & Editing, Visualization. Jerry Y. Chao: Methodology, Formal analysis, Investigation, Data Curation, Resources, Writing – Review & Editing. Rodrigo Gutiérrez: Methodology, Software, Formal analysis, Data Curation, Writing – Review & Editing. Jeff L. Rhoades: Software, Writing – Review & Editing. Shlomo Shinnar: Project administration, Supervision, Writing – Review & editing. Patrick L. Purdon: Methodology, Supervision, Writing – Review & Editing. Charles A. Nelson: Investigation, Resources, Writing – Review & Editing, Supervision, Project administration, Funding acquisition.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[80, 61, 219, 79]]<|/det|>
+## REFERENCES  
+
+<|ref|>text<|/ref|><|det|>[[70, 130, 884, 191]]<|/det|>
+1. Huttenlocher, P. R. & Dabholkar, A. S. Regional differences in synaptogenesis in human cerebral cortex. J Comp Neurol 387, 167-178 (1997).  
+
+<|ref|>text<|/ref|><|det|>[[70, 205, 857, 300]]<|/det|>
+2. Glantz, L. A., Gilmore, J. H., Hamer, R. M., Lieberman, J. A. & Jarskog, L. F. Synaptophysin and postsynaptic density protein 95 in the human prefrontal cortex from mid-gestation into early adulthood. Neuroscience 149, 582-591 (2007).  
+
+<|ref|>text<|/ref|><|det|>[[70, 316, 890, 375]]<|/det|>
+3. Tau, G. Z. & Peterson, B. S. Normal Development of Brain Circuits. Neuropsychopharmacology 35, 147-168 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[70, 391, 895, 450]]<|/det|>
+4. Kim, J. Y. & Paredes, M. F. Implications of extended inhibitory neuron development. International Journal of Molecular Sciences 22, 1-16 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[70, 465, 912, 525]]<|/det|>
+5. Antón-Bolaños, N., Espinosa, A. & López-Bendito, G. Developmental interactions between thalamus and cortex: a true love reciprocal story. Curr Opin Neurobiol 52, 33-41 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[70, 539, 919, 599]]<|/det|>
+6. Molnár, Z., Luhmann, H. J. & Kanold, P. O. Transient cortical circuits match spontaneous and sensory-driven activity during development. Science 370, eabb2153 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[70, 613, 877, 672]]<|/det|>
+7. Alcauter, S. et al. Development of Thalamocortical Connectivity during Infancy and Its Cognitive Correlations. The Journal of Neuroscience 34, 9067 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[70, 687, 890, 747]]<|/det|>
+8. Doesburg, S. M., Ward, L. M. & Ribary, U. The alpha-theta-gamma (ATG) switch: Toward unified principles of cortical processing. Current Trends in Neurology 9, 1-12 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[70, 761, 900, 820]]<|/det|>
+9. Tröndle, M., Popov, T., Dziemian, S. & Langer, N. Decomposing the role of alpha oscillations during brain maturation. eLife 11, e77571 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[70, 835, 857, 894]]<|/det|>
+10. Gao, R., Peterson, E. J. & Voytek, B. Inferring synaptic excitation/inhibition balance from field potentials. Neuroimage 158, 70-78 (2017).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[30, 60, 916, 95]]<|/det|>
+11. Merkin, A. et al. Do age-related differences in aperiodic neural activity explain differences in resting EEG alpha? Neurobiol Aging 121, 78-87 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[30, 108, 920, 194]]<|/det|>
+12. Voytek, B. et al. Age-related changes in 1/f neural electrophysiological noise. Journal of Neuroscience 35, 13257-13265 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[30, 208, 916, 305]]<|/det|>
+13. Hill, A. T., Clark, G. M., Bigelow, F. J., Lum, J. A. G. & Enticott, P. G. Periodic and aperiodic neural activity displays age-dependent changes across early-to-middle childhood. Developmental Cognitive Neuroscience 54, 101076 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[30, 320, 916, 380]]<|/det|>
+14. McSweeney, M. et al. Age-related trends in aperiodic EEG activity and alpha oscillations during early-to middle-childhood. Neuroimage 269, 119925 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[30, 394, 911, 454]]<|/det|>
+15. Buzsáki, G., Logothetis, N. & Singer, W. Scaling brain size, keeping timing: evolutionary preservation of brain rhythms. Neuron 80, 751-764 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[30, 468, 896, 528]]<|/det|>
+16. Donoghue, T. et al. Parameterizing neural power spectra into periodic and aperiodic components. Nature Neuroscience 23, 1655-1665 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[30, 542, 912, 602]]<|/det|>
+17. Klimesch, W. Alpha-band oscillations, attention, and controlled access to stored information. Trends in Cognitive Sciences 16, 606-617 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[30, 616, 884, 675]]<|/det|>
+18. Ball, G. et al. Thalamocortical Connectivity Predicts Cognition in Children Born Preterm. Cerebral Cortex 25, 4310-4318 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[30, 689, 912, 749]]<|/det|>
+19. Foxe, J. & Snyder, A. The Role of Alpha-Band Brain Oscillations as a Sensory Suppression Mechanism during Selective Attention. Frontiers in Psychology 2, (2011).  
+
+<|ref|>text<|/ref|><|det|>[[30, 763, 912, 860]]<|/det|>
+20. Orekhova, E. V., Stroganova, T. A. & Posikera, I. N. Theta synchronization during sustained anticipatory attention in infants over the second half of the first year of life. International Journal of Psychophysiology 32, 151172 (1999).  
+
+<|ref|>text<|/ref|><|det|>[[30, 874, 896, 935]]<|/det|>
+21. Marshall, P. J., Bar-Haim, Y. & Fox, N. A. Development of the EEG from 5 months to 4 years of age. Clinical Neurophysiology 113, 1199-1208 (2002).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[30, 60, 920, 157]]<|/det|>
+22. Lindsley, D. B. A Longitudinal Study of the Occipital Alpha Rhythm in Normal Children: Frequency and Amplitude Standards. The Pedagogical Seminary and Journal of Genetic Psychology 55, 197–213 (1939).  
+
+<|ref|>text<|/ref|><|det|>[[30, 171, 920, 900]]<|/det|>
+23. Smith, J. R. The Frequency Growth of the Human Alpha Rhythms During Normal Infancy and Childhood. The Journal of Psychology 11, 177–198 (1941).  
+24. Cellier, D., Riddle, J., Petersen, I. & Hwang, K. The development of theta and alpha neural oscillations from ages 3 to 24 years. Developmental Cognitive Neuroscience 50, 100969 (2021).  
+25. Spitzer, B. & Haegens, S. Beyond the Status Quo: A Role for Beta Oscillations in Endogenous Content (Re)Activation. eNeuro 4, (2017).  
+26. Purdon, P. L., Sampson, A., Pavone, K. J. & Brown, E. N. Clinical Electroencephalography for Anesthesiologists: Part I: Background and Basic Signatures. Anesthesiology 123, 937–960 (2015).  
+27. Ching, S., Cimenser, A., Purdon, P. L., Brown, E. N. & Kopell, N. J. Thalamocortical model for a propofol-induced α-rhythm associated with loss of consciousness. Proceedings of the National Academy of Sciences 107, 22665–22670 (2010).  
+28. McCarthy, M. M., Brown, E. N. & Kopell, N. Potential network mechanisms mediating electroencephalographic beta rhythm changes during propofol-induced paradoxical excitation. J Neurosci 28, 13488–13504 (2008).  
+29. Cornelissen, L., Kim, S.-E., Purdon, P. L., Brown, E. N. & Berde, C. B. Age-dependent electroencephalogram (EEG) patterns during sevoflurane general anesthesia in infants. Elife 4, e06513 (2015).  
+30. Cornelissen, L. et al. Electroencephalographic markers of brain development during sevoflurane anaesthesia in children up to 3 years old. British Journal of Anaesthesia 120, 1274–1286 (2018).  
+31. Akeju, O. et al. Age-dependency of sevoflurane-induced electroencephalogram dynamics in children. Br J Anaesth 115 Suppl 1, i66–i76 (2015).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[30, 60, 900, 120]]<|/det|>
+32. Flores, F. J. et al. Thalamocortical synchronization during induction and emergence from propofol-induced unconsciousness. Proc Natl Acad Sci U S A 114, E6660–E6668 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[30, 135, 859, 194]]<|/det|>
+33. Bastos, A. M. et al. Neural effects of propofol-induced unconsciousness and its reversal using thalamic stimulation. eLife 10, e60824 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[30, 210, 920, 268]]<|/det|>
+34. Engel, A. K. & Fries, P. Beta-band oscillations--signalling the status quo? Curr Opin Neurobiol 20, 156–165 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[30, 283, 890, 380]]<|/det|>
+35. Chao, J. Y. et al. Electroencephalographic Findings and Clinical Behavior During Induction of Anesthesia With Sevoflurane in Human Infants: A Prospective Observational Study. Anesthesia & Analgesia 130, e161 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[30, 395, 886, 490]]<|/det|>
+36. Manning, J. R., Jacobs, J., Fried, I. & Kahana, M. J. Broadband shifts in local field potential power spectra are correlated with single-neuron spiking in humans. Journal of Neuroscience 29, 13613–13620 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[30, 506, 880, 565]]<|/det|>
+37. Miller, K. J. Broadband spectral change: Evidence for a macroscale correlate of population firing rate? Journal of Neuroscience 30, 6477–6479 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[30, 580, 891, 639]]<|/det|>
+38. Knickmeyer, R. C. et al. A structural MRI study of human brain development from birth to 2 years. Journal of Neuroscience 28, 12176–12182 (2008).  
+
+<|ref|>text<|/ref|><|det|>[[30, 654, 910, 712]]<|/det|>
+39. Gilmore, J. H. et al. Longitudinal development of cortical and subcortical gray matter from birth to 2 years. Cereb Cortex 22, 2478–2485 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[30, 727, 902, 786]]<|/det|>
+40. Gilmore, J. H., Knickmeyer, R. C. & Gao, W. Imaging structural and functional brain development in early childhood. Nat Rev Neurosci 19, 123–137 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[30, 801, 866, 899]]<|/det|>
+41. Schaworonkow, N. & Voytek, B. Longitudinal changes in aperiodic and periodic activity in electrophysiological recordings in the first seven months of life. Dev Cogn Neurosci 47, 100895 (2021).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[37, 60, 911, 940]]<|/det|>
+42. Chini, M., Pfeffer, T. & Hanganu-Opatz, I. An increase of inhibition drives the developmental decorrelation of neural activity. eLife 11, e78811 (2022).
+43. Voytek, B. & Knight, R. T. Dynamic Network Communication as a Unifying Neural Basis for Cognition, Development, Aging, and Disease. Biological Psychiatry 77, 1089–1097 (2015).
+44. Paredes, M. F. et al. Extensive migration of young neurons into the infant human frontal lobe. Science 354, aaf7073 (2016).
+45. Mueller, A. L., Taube, J. S. & Schwartzkroin, P. A. Development of hyperpolarizing inhibitory postsynaptic potentials and hyperpolarizing response to gamma-aminobutyric acid in rabbit hippocampus studied in vitro. J Neurosci 4, 860–867 (1984).
+46. Cherubini, E., Gaiarsa, J. L. & Ben-Ari, Y. GABA: an excitatory transmitter in early postnatal life. Trends Neurosci 14, 515–519 (1991).
+47. Xu, G. et al. Late Development of the GABAergic System in the Human Cerebral Cortex and White Matter. J Neuropathol Exp Neurol 70, 841–858 (2011).
+48. Kanold, P. O. & Shatz, C. J. Subplate neurons regulate maturation of cortical inhibition and outcome of ocular dominance plasticity. Neuron 51, 627–638 (2006).
+49. Kostovic, I. & Rakic, P. Developmental history of the transient subplate zone in the visual and somatosensory cortex of the macaque monkey and human brain. Journal of Comparative Neurology 297, 441–470 (1990).
+50. McQuillen, P. S., Sheldon, R. A., Shatz, C. J. & Ferriero, D. M. Selective Vulnerability of Subplate Neurons after Early Neonatal Hypoxia-Ischemia. J Neurosci 23, 3308–3315 (2003).
+51. Orekhova, E. V., Stroganova, T. A., Posikera, I. N. & Elam, M. EEG theta rhythm in infants and preschool children. Clinical Neurophysiology 117, 1047–1062 (2006).
+52. Womelsdorf, T., Valiante, T. A., Sahin, N. T., Miller, K. J. & Tiesinga, P. Dynamic circuit motifs underlying rhythmic gain control, gating and integration. Nat Neurosci 17, 1031–1039 (2014).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[37, 60, 900, 944]]<|/det|>
+53. Silva, L. R., Amitai, Y. & Connors, B. W. Intrinsic Oscillations of Neocortex Generated by Layer 5 Pyramidal Neurons. Science 251, 432 (1991).54. Hughes, S. W. & Crunelli, V. Thalamic mechanisms of EEG alpha rhythms and their pathological implications. The Neuroscientist: a review journal bringing neurobiology, neurology and psychiatry 11, 357-372 (2005).55. Hughes, S. W. & Crunelli, V. Just a phase they're going through: The complex interaction of intrinsic high-threshold bursting and gap junctions in the generation of thalamic \(\alpha\) and \(\theta\) rhythms. International Journal of Psychophysiology 64, 3-17 (2007).56. Georgieva, S. et al. Toward the Understanding of Topographical and Spectral Signatures of Infant Movement Artifacts in Naturalistic EEG. Frontiers in Neuroscience 14, (2020).57. Frohlich, J. et al. A Quantitative Electrophysiological Biomarker of Duplication 15q11.2-q13.1 Syndrome. (2016) doi:10.1371/journal.pone.0167179.58. Frohlich, J. et al. Mechanisms underlying the EEG biomarker in Dup15q syndrome. Molecular Autism 10, 29 (2019).59. Wilkinson, C. L. & Nelson, C. A. Increased aperiodic gamma power in young boys with Fragile X Syndrome is associated with better language ability. Molecular Autism 1-15 (2021) doi:10.1186/s13229-021-00425-x.60. Levin, A. R., Méndez Leal, A. S., Gabard-Durnam, L. J. & O'Leary, H. M. BEAPP: The Batch Electroencephalography Automated Processing Platform. Frontiers in Neuroscience 12, 513 (2018).61. Gabard-Durnam, L. J., Méndez Leal, A. S., Wilkinson, C. L. & Levin, A. R. The Harvard Automated Processing Pipeline for Electroencephalography (HAPPE): standardized processing software for developmental and high-artifact data. Frontiers in Neuroscience 12, 97 (2018).62. Babadi, B. & Brown, E. N. A review of multitaper spectral analysis. IEEE Transactions on Biomedical Engineering 61, 1555-1564 (2014).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[33, 61, 910, 460]]<|/det|>
+63. Chao, J. Y. et al. Decreased Electroencephalographic Alpha Power During Anesthesia Induction Is Associated With EEG Discontinuity in Human Infants. Anesthesia & Analgesia 135, 1207–1216 (2022).
+64. Grant, A. C. et al. Diagnostic accuracy of microEEG: a miniature, wireless EEG device. Epilepsy Behav 34, 81–85 (2014).
+65. Bokil, H., Andrews, P., Kulkarni, J. E., Mehta, S. & Mitra, P. P. Chronux: a platform for analyzing neural signals. J Neurosci Methods 192, 146–151 (2010).
+66. Wood, S. N. Generalized Additive Models: An Introduction with R, Second Edition. (CRC Press, 2017).
+67. Benjamini, Y. & Hochberg, Y. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society. Series B (Methodological) 57, 289–300 (1995).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[92, 66, 899, 135]]<|/det|>
+Supplemental Figure 1 - Electrode layout: (A) 128- channel Hydrogel Geodesic Sensor Net. (B) 64- channel Geodesic Sensor Net. Pink circles denote 10- 20 electrodes, and blue circles denote the additional electrodes included in ICA and MARA steps of pre- processing. (C), (D) Electrodes averaged for frontal (yellow), central (blue), temporal (orange), and posterior (green) regions of interest.  
+
+<|ref|>image<|/ref|><|det|>[[93, 152, 884, 789]]<|/det|>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[91, 147, 904, 789]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[90, 115, 830, 131]]<|/det|>
+<center>Supplemental Figure 2: Individual plots of the periodic spectrum averaged across the whole ROI for each age bin. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[91, 137, 890, 790]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[101, 96, 860, 128]]<|/det|>
+<center>Supplemental Figure 3: GAMMs modeled trajectories of 8 power measures, with ROI, study, smoothed age, and sex as predictor terms. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[100, 65, 828, 125]]<|/det|>
+Supplemental Figure 4: For each age bin the average FOOOF modeled spectra (green) across participants in each age bin is graphed along with the averaged original power spectrum (red), and averaged FOOOF estimated aperiodic spectrum (blue). (A) Unedited FOOOF estimates. (B) Modified FOOOF estimates. (C) Comparison of squared error across frequencies of unedited (orange) and modified (blue) FOOOF.  
+
+<|ref|>image<|/ref|><|det|>[[84, 150, 895, 400]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[90, 134, 768, 149]]<|/det|>
+<center>A. Unedited FOOOF model estimates - Original Spectrum - FOOOF modeled Spectrum - Aperiodic Fit </center>  
+
+<|ref|>image<|/ref|><|det|>[[95, 440, 901, 692]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[90, 416, 777, 432]]<|/det|>
+<center>B. Modified FOOOF model estimates - Original Spectrum - FOOOF modeled Spectrum - Aperiodic Fit </center>  
+
+<|ref|>image<|/ref|><|det|>[[110, 734, 721, 880]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[90, 714, 320, 728]]<|/det|>
+<center>C. Original vs Edited FOOOF error </center>
+
+<--- Page Split --->

preprint/preprint__2b43013de3bb82ae9a113872a06ac68ff47a1fe3e657a67c2b2adc2ace467e88/images_list.json

@@ -0,0 +1,77 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1 - Distribution of width and slope of tidal flats along the coast of China. The distribution of tidal flat width ( \\(N = 2538\\) ) (a) and slope ( \\(N = 1620\\) ) (b) with a resolution of 1 km). Dark gray areas indicate coastal areas, namely Liaoning (LN), Hebei (HB), Shandong (SD), Jiangsu (JS), Zhejiang",
+    "footnote": [],
+    "bbox": [
+      [
+        190,
+        364,
+        808,
+        814
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2 - Relations between environmental factors and tidal flats width (2014-2016) and slope in space. Tidal flat width for different classes of suspended sediment concentrations (SSC) (a) and tidal ranges (b). Tidal flat slope for different classes of SSC (c) and tidal ranges (d). The number displayed at the base of each bar represents the corresponding number of data points. The relationship between the spatial distribution of tidal width and SSC (e) and the relationship between the spatial distribution of tidal slope and tidal range (f) are derived from a stepwise linear regression model. \\(R^2\\) denotes the correlation coefficient of linear regression. The data points represent the distribution of the data, and the solid lines indicate the model fit, bounded by 95% confidence intervals",
+    "footnote": [],
+    "bbox": [
+      [
+        220,
+        80,
+        785,
+        622
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3 - Response of tidal flat width to changes in suspended sediment concentration along the coast of China. Distribution of inter-annual trend of tidal flat width (2002-2016) (a) and suspended sediment concentration (2003-2011) (b). Outliers that deviated from the median by more than three times the median absolute deviation are excluded. The values in the figure represent the average values of data in the box. \\(51.3\\%\\) of the tidal flat transects show a decreasing trend in width, while \\(54.9\\%\\) of the transects experience a decrease in SSC. The insets represent the probability distribution of the data points. (c)Temporal changes in the median value of tidal flat width (linear regression, \\(P = 0.1167\\) ) and (d) SSC along the coast of China (linear regression, \\(P = 0.0274\\) ).",
+    "footnote": [],
+    "bbox": [
+      [
+        171,
+        120,
+        816,
+        630
+      ]
+    ],
+    "page_idx": 10
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4 - Elastic response of tidal flats width changes to altered sediment supply changes. Relation between SSC change rate and flat width change rate based on observation (a) and model results (b). Model results are based on tidal flat width changes of exemplified sediment-rich and sediment-starving areas (see Methods) The numbers represent the number of tidal flats transects corresponding to each bar. (c) Diagram showing the state of sediment-rich areas is similar to a stretched spring that tends to restore to morphological equilibrium (dashed line) and resist to be stretched further. (d) Diagram showing the state of sediment-starving areas is similar to a compressed spring that tends to restore morphological equilibrium (dashed line) and resist being compressed further. The dashed lines indicate the equilibrium state in both cases. The convex and concave profiles reflect their high and low sediment input level, respectively \\(^{14,23}\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        182,
+        191,
+        870,
+        595
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Fig. 5 – Modelled and observed tidal flat width response to sediment supply changes. The contour map represents the tidal flat width variation corresponding to different interannual variation rates in suspended sediment concentration (SSC) modeled by DET-ESTMORF, whereas bubbles indicate representative data points of tidal flat width changes along China's coast (see methods). The number indicates the location of the data point on the Chinese coast (selection of data points see Methods). Red bubbles indicate tidal flats expansion, and green bubbles indicate tidal flats degradation. The area between two solid black lines indicates a neutral zone where the interannual variation of tidal flats width is within (-0.1% - 0.3%). The hot zones are in the upper left and lower right corners of the diagram, whereas there are no data points at the upper left corner, i.e., rapid SSC increase (> 1% per year) in sediment-rich systems.",
+    "footnote": [],
+    "bbox": [
+      [
+        320,
+        191,
+        740,
+        459
+      ]
+    ],
+    "page_idx": 14
+  }
+]

preprint/preprint__2b43013de3bb82ae9a113872a06ac68ff47a1fe3e657a67c2b2adc2ace467e88/preprint__2b43013de3bb82ae9a113872a06ac68ff47a1fe3e657a67c2b2adc2ace467e88.mmd

@@ -0,0 +1,388 @@
+
+# Dynamics and drivers of tidal flat morphology in China  
+
+Zhan Hu  
+
+huzh9@mail.sysu.edu.cn  
+
+Sun Yat- Sen University https://orcid.org/0000- 0002- 2809- 3718  
+
+Shuai Liu  
+
+Sun Yat- Sen University https://orcid.org/0009- 0005- 9202- 6607  
+
+Tim Grandjean  
+
+Royal Netherlands Institute for Sea Research https://orcid.org/0000- 0002- 2729- 3566  
+
+Zheng Bing Wang  
+
+Deltares & Delft University of Technology https://orcid.org/0000- 0002- 8787- 4530  
+
+Vincent T. M. Zelst  
+
+Deltares https://orcid.org/0000- 0003- 2923- 1745  
+
+Lin Qi  
+
+NOAA Center for Satellite Applications and Research  
+
+Tianping Xu  
+
+Sun Yat- Sen University  
+
+Jun Seo  
+
+Chonnam National University  
+
+Tjeerd Bouma  
+
+Royal Netherlands Institute for Sea Research  
+
+## Article  
+
+Keywords:  
+
+Posted Date: April 17th, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4209550/v1  
+
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.
+
+<--- Page Split --->
+
+Version of Record: A version of this preprint was published at Nature Communications on March 4th, 2025. See the published version at https://doi.org/10.1038/s41467-025-57525-y.
+
+<--- Page Split --->
+
+## Dynamics and drivers of tidal flat morphology in China  
+
+Shuai Liu \(^{1}\) , Zhan Hu \(^{1,2,3*}\) , Tim J. Grandjean \(^{4,5}\) , Zheng B. Wang \(^{6,7}\) , Vincent T. M. van Zels \(^{6,7}\) , Lin  
+
+Qi \(^{8}\) , Tianping Xu \(^{1}\) , Jun Y. Seo \(^{9}\) , Tjeerd J. Bouma \(^{4,5}\)  
+
+\(^{1}\) School of Marine Sciences, Sun Yat- Sen University, and Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai, China.  
+
+\(^{2}\) Guangdong Provincial Key Laboratory of Marine Resources and Coastal Engineering, Guangzhou, China.  
+
+\(^{3}\) Pearl River Estuary Marine Ecosystem Research Station, Ministry of Education, Zhuhai  
+
+\(^{4}\) Department of Estuarine and Delta Systems, Royal Netherlands Institute for Sea Research  
+
+(NIOZ), Yerseke, The Netherlands.  
+
+\(^{5}\) Faculty of Geosciences, Department of Physical Geography, Utrecht University, Utrecht, The Netherlands.  
+
+\(^{6}\) Faculty of Civil Engineering and Geosciences, Delft University of Technology, 2628 CN Delft, the Netherlands.  
+
+\(^{7}\) Deltares, 2600 MH Delft, the Netherlands.  
+
+\(^{8}\) NOAA Center for Satellite Applications and Research, College Park, MD 20740, USA.  
+
+\(^{9}\) Department of Oceanography, Chonnam National University, Gwangju, Republic of Korea.  
+
+\(^{*}\) Corresponding author.  
+
+E- mail address: huzh9@mail.sysu.edu.cn
+
+<--- Page Split --->
+
+## Abstract  
+
+Recent remote sensing analysis has revealed extensive loss of tidal flats, yet the mechanisms driving these large- scale changes remain enigmatic. This study traces the spatiotemporal variations of 2538 tidal flat transects across China to elucidate how their morphological features vary with external factors, including suspended sediment concentration (SSC), tidal range, and wave height. We observe a correlation between flat width and SSC distribution, and between flat slope and tidal range. A national- wide decline in flat width is observed together with SSC reduction between 2002 and 2016. Intriguingly, sediment- rich flats exhibit a more rapid response to SSC reduction compared to sediment- starving areas, but the converse is observed with SSC increase. These conditional responses stem from the morphodynamic tendency towards equilibrium, which is well explained by synthetical modeling. This finding suggests that tidal flats are resilient to sediment supply reduction, and nation- scale sediment allocation can assist in preserving valuable intertidal areas.
+
+<--- Page Split --->
+
+## MAIN  
+
+## Introduction  
+
+Tidal flats, alongside their adjacent coastal vegetation systems such as marshes and mangroves, constitute one of the most widespread coastal habitats globally. They provide essential ecosystem services, including storm protection, carbon sequestration, and serving as nurseries for fisheries \(^{1 - 4}\) . The resilience of coastal vegetation ecosystems to climate change has been a focus of attention \(^{5 - 7}\) , and previous research has shown that the fates of tidal flats and vegetated coastal systems are closely linked \(^{8 - 10}\) , underscoring the need for in- depth study of tidal flats dynamics for science- based management.  
+
+With the recent advancement in remote sensing, global scale tidal flat dynamics have been mapped in detail \(^{11,12}\) . Strikingly, approximately \(16\%\) of the tidal flat area has been lost during 1984- 2016, and more than \(56\%\) of tidal flat area dynamics were attributed to natural coastal processes and global climate changes, such as wave erosion and altered sediment supply, rather than direct human impact, e.g., reclamation \(^{12}\) . These natural processes and environmental changes often operate at large scales and originate far from sites with emerging dynamics. Previous studies have demonstrated that tidal range, wave height, and sediment supply can all influence tidal flat morphology at a local scale \(^{13 - 16}\) , but which factors dominate over large scales and how they lead to temporal changes remain unclear. For instance, sediment supply is widely regarded as a key factor influencing tidal flat morphology \(^{17,18}\) , and there have been dramatic alterations in global suspended sediment flux to the coast \(^{19,20}\) . In the global hydrologic north (north of \(\sim 20^{\circ}\mathrm{N}\) ), a \(49\%\) reduction in river sediment flux has been observed due to damming, while in the global hydrologic south (south of \(\sim 20^{\circ}\mathrm{S}\) ), river sediment flux has increased by approximately \(41\%\) since the 1980s, primarily due
+
+<--- Page Split --->
+
+to activities such as mining<sup>20</sup>. These systematic changes in global river sediment flux are expected to have a profound impact on large- scale tidal flat systems, which has yet to be revealed.  
+
+China possesses the \(2^{\mathrm{nd}}\) largest tidal flat area in the world, with a variety of tidal ranges, wave heights, and sediment supply<sup>11,21,22</sup>. Additionally, China's tidal flat system has experienced some of the most rapid loss globally (at \(2.6\%\) per year)<sup>11,12,17</sup>. These facts make China's tidal flats an ideal model system for studying the diversification of large- scale tidal flat morphology. In the current study, tidal flats are defined as mud or sand flats with regular tidal inundation<sup>11</sup>. We focus on tidal flat slope and width, as they are regarded as the representative characteristics of tidal flat morphology<sup>23</sup>, which are available from global datasets<sup>11,24</sup>. In total, we composed 2538 tidal flat transects along China's coastline for width analysis and 1620 transects for slope analysis, following independent and complete profiles for flat width and elevation continuous profile for flat slope (see Methods). Sites affected by direct human impact (e.g., reclamation) were excluded from the analysis. To explain the observed spatiotemporal changes, we included morphodynamic modeling by the DET- ESTMORF model, which is based on a dynamic equilibrium theory<sup>23,25</sup> (see Methods). By combining remote sensing observation and morphodynamic modeling, we were able to identify the main drivers influencing spatio- temporal variation of tidal flats morphology in China and understand the nonlinear elastic responses of tidal flats to ambient sediment supply changes in relation to their equilibrium. Our findings provide valuable insights into the large- scale tidal flat morphology and can serve as a key reference for coastal management with changing sediment supply.  
+
+## Results  
+
+## Spatial variation and drivers in tidal flat morphology
+
+<--- Page Split --->
+
+Remote sensing observation shows that tidal flat width ranges from \(30\mathrm{m}\) to \(2.6\times 10^{4}\mathrm{m}\) between 2002- 2016 (Fig. 1a). \(34.3\%\) of the tidal flats have a width of \(30 - 300\mathrm{m}\) , whereas \(32.4\%\) of the tidal flats are wider than \(1.0\times 10^{3}\mathrm{m}\) (Supplementary Fig. 1). The median value of tidal flat width is 536 m. The widest tidal flats are found in Jiangsu (JS) province with an average width of \(3088\mathrm{m}\) , while the smallest average width is observed in the northern Shandong Peninsula., being \(74\mathrm{m}\) (Fig. 1a). Notably, the average flats width in estuaries ( \(1212\mathrm{m}\) , \(\mathrm{N} = 275\) ) is greater than those on open coasts ( \(921\mathrm{m}\) , \(\mathrm{N} = 2263\) ) (Supplementary Fig. 2) \(^{26}\) , which is likely attributed to sediment availability.  
+
+![](images/Figure_1.jpg)
+
+<center>Fig. 1 - Distribution of width and slope of tidal flats along the coast of China. The distribution of tidal flat width ( \(N = 2538\) ) (a) and slope ( \(N = 1620\) ) (b) with a resolution of 1 km). Dark gray areas indicate coastal areas, namely Liaoning (LN), Hebei (HB), Shandong (SD), Jiangsu (JS), Zhejiang </center>
+
+<--- Page Split --->
+
+(ZJ), Fujian (FJ), Guangdong (GD), Guangxi (GX), Hainan (HN) and Taiwan (TW). Red boxes indicate regions of significant width and slope, and blue boxes indicate regions of minimal values. The values in the figure represent the average values of data in the box. The variation of tidal flat width and slope along latitude is displayed on the right panel. The shaded areas represent the range from the 25th percentile to the 75th percentile. (c) The relation between tidal flat width and median slope in each group. The number displayed at the base of each bar represents the number of data points in the group. The error bars represent the upper 75%-ile and the 25%-ile.  
+
+The slope of tidal flats in China ranges from \(3.6 \times 10^{- 5} - 1.2 \times 10^{- 1}\) (Fig. 1b), and the median slope is \(7.2 \times 10^{- 3}\) . The average slope of tidal flats in estuaries ( \(6.8 \times 10^{- 3}\) , \(\mathrm{N} = 153\) ) is smaller than those on open coasts ( \(1.2 \times 10^{- 2}\) , \(\mathrm{N} = 1467\) ). Specifically, the average slope value in the southern Fujian (FJ) is around \(1.8 \times 10^{- 2}\) , which is the highest in the country. In contrast, the tidal flats in southern Jiangsu (JS) exhibit the lowest slope, being \(2.3 \times 10^{- 3}\) (Fig. 1b). Furthermore, the tidal flat slope generally decreases as the tidal flat width increases (Fig. 1c).  
+
+To identify the main drivers shaping the tidal flat morphology, we analyze the distribution of width (Fig. 2a, b) and slope (Fig. 2c, d) in conjunction with local wave height, tidal range, and suspended sediment concentration (SSC) (details in Materials and Methods). Results show tidal flat width exhibits a positive correlation with SSC ( \(\mathrm{R} = 0.71\) , \(\mathrm{P} < 0.01\) ) and tidal range ( \(\mathrm{R} = 0.44\) , \(\mathrm{P} < 0.01\) ), but a negative correlation with wave height ( \(\mathrm{R} = - 0.20\) , \(\mathrm{P} < 0.05\) ). A stepwise linear regression analysis indicates that the combination of SSC (80% relative importance) and tidal range (20% relative importance) best explains the distribution of tidal flat width ( \(\mathrm{R}^{2} = 0.53\) ) (Fig. 2e), highlighting the primary role of local SSC level.
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Fig. 2 - Relations between environmental factors and tidal flats width (2014-2016) and slope in space. Tidal flat width for different classes of suspended sediment concentrations (SSC) (a) and tidal ranges (b). Tidal flat slope for different classes of SSC (c) and tidal ranges (d). The number displayed at the base of each bar represents the corresponding number of data points. The relationship between the spatial distribution of tidal width and SSC (e) and the relationship between the spatial distribution of tidal slope and tidal range (f) are derived from a stepwise linear regression model. \(R^2\) denotes the correlation coefficient of linear regression. The data points represent the distribution of the data, and the solid lines indicate the model fit, bounded by 95% confidence intervals </center>
+
+<--- Page Split --->
+
+Tidal flat slope exhibits a positive correlation with tidal range \(\mathrm{(R = 0.34,P< 0.01)}\) , but a negative correlation with SSC \(\mathrm{(R = - 0.26,P< 0.01)}\) . Other factors, such as wave height, are known to have an impact on tidal flat slope \(^{13,23}\) , but they are less important in the current dataset. The combination of tidal range (58% of relative importance) and SSC (42% of relative importance) best explains the distribution of tidal flat slope \(\mathrm{(R^{2} = 0.30)}\) (Fig. 2f).  
+
+These findings on the spatial distribution of flat width and slope are in agreement with DET- ESTMORF modelling \(^{14}\) . On tidal flats with a large tidal range, the bed shear stress generated by tidal currents has a negative gradient towards the landward direction, leading to erosion at the low tidal flats and deposition at the higher tidal flats (Supplementary Fig. 5). Such a gradient becomes more prominent as tidal range increases (Supplementary Fig. 6). Thus, a greater tidal range can lead to a steeper tidal flat profile as observed. On the relation between SSC and tidal flat width, larger SSC can result in sediment surplus that exceeds the level required for equilibrium. This leads to deposition on the tidal flat and seaward expansion. Consequently, tidal flat width exhibits a positive correlation with SSC.  
+
+## Temporal dynamics in tidal flat width and the underlying elastic responses  
+
+The time series of tidal flat width from 2002 to 2016 shows a clear trend of overall erosion (Fig. 3a, c). The nationwide median tidal flat width reduced from 596 m to 536 m from 2002 to 2016, i.e., reduced at - 0.8% per year (Fig. 3c). Such a trend is particularly prominent in Hainan Island and the Changjiang River Estuary, where the interannual change rate in tidal flat width was - 1.89% and - 1.90% per year, respectively (Fig. 3a). In contrast, a few sites (28.3%) exhibited an expansion trend of > 0.1% per year, such as the southern Shandong (SD) Peninsula and the coast of Fujian Province, with width increase rates of 0.76% and 0.58% per year, respectively (Fig. 3a).
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Fig. 3 - Response of tidal flat width to changes in suspended sediment concentration along the coast of China. Distribution of inter-annual trend of tidal flat width (2002-2016) (a) and suspended sediment concentration (2003-2011) (b). Outliers that deviated from the median by more than three times the median absolute deviation are excluded. The values in the figure represent the average values of data in the box. \(51.3\%\) of the tidal flat transects show a decreasing trend in width, while \(54.9\%\) of the transects experience a decrease in SSC. The insets represent the probability distribution of the data points. (c)Temporal changes in the median value of tidal flat width (linear regression, \(P = 0.1167\) ) and (d) SSC along the coast of China (linear regression, \(P = 0.0274\) ). </center>
+
+<--- Page Split --->
+
+Concurrently, there was a nationwide declining trend in sediment supply at \(- 0.4\%\) per year as observed in interannual SSC from 2002 to 2012 (Fig. 3d). K- means clustering analysis, a measure for partitioning data points into distinct, non- overlapping groups, reveals a strong correlation between the interannual SSC reduction and tidal flat width decline ( \(\mathrm{R} = 0.64\) , \(\mathrm{P} < 0.01\) ) (Supplementary Fig. 7). Additionally, \(56\%\) of all observation transects exhibit concurrent decline or increase of sediment supply with tidal flat width (Fig. 3a, b), whereas most of the rest sites ( \(86\%\) ) exhibit an insignificant change in SSC. These statistics suggest a close linkage between altered sediment supply and changes in tidal flat morphology.  
+
+Beyond the overall connection between SSC reduction and tidal flat width loss, our research further discovers diverse responses of sediment- starving (SSC \(< 50 \mathrm{mg / L}\) , average SSC of China's coastal tidal flats) and sediment- rich (SSC \(>75 \mathrm{mg / L}\) , average SSC of China's tidal flats in the estuary) systems to altered sediment supply. Notably, sediment- rich systems undergo more rapid width reduction than their sediment- starving counterparts when subjected to the same rate of SSC decrease ( \(0.27\%\) vs. \(0.24\%\) per year, Fig. 4a). Conversely, sediment- starving systems exhibit more pronounced width expansion in response to an equivalent increase in SSC ( \(0.55\%\) vs. \(0.30\%\) per year). To elucidate these divergent responses, we employed DET- ESTMORF modelling on two exemplified sites (Fig. 4b and Supplementary Fig. 8). An increase in SSC leads to an immediate rise in actual concentration, but only a modest rise in equilibrium concentration (Supplementary Fig. 9). The difference between the actual concentration and the equilibrium concentration leads to accumulation and width growth. In sediment- starving systems, the equilibrium concentration is inherently lower, and thus the disparity from the actual concentration is greater than in sediment- rich systems (Supplementary Fig. 9). Consequently, the momentum behind the width growth is stronger in sediment- starving systems during SSC increases, leading to faster width
+
+<--- Page Split --->
+
+growth. Conversely, during SSC reductions, sediment-rich systems exhibit a more accelerated width reduction (see Supplementary Text for a more detailed explanation).  
+
+![](images/Figure_4.jpg)
+
+<center>Fig. 4 - Elastic response of tidal flats width changes to altered sediment supply changes. Relation between SSC change rate and flat width change rate based on observation (a) and model results (b). Model results are based on tidal flat width changes of exemplified sediment-rich and sediment-starving areas (see Methods) The numbers represent the number of tidal flats transects corresponding to each bar. (c) Diagram showing the state of sediment-rich areas is similar to a stretched spring that tends to restore to morphological equilibrium (dashed line) and resist to be stretched further. (d) Diagram showing the state of sediment-starving areas is similar to a compressed spring that tends to restore morphological equilibrium (dashed line) and resist being compressed further. The dashed lines indicate the equilibrium state in both cases. The convex and concave profiles reflect their high and low sediment input level, respectively \(^{14,23}\) . </center>
+
+<--- Page Split --->
+
+These non- linear responses have not been reported previously. They can be explained by an analogy to the dynamics of elastic systems such as springs. These systems have their intrinsic equilibriums but can be stretched or compressed from equilibriums in response to external sediment input (Fig. 4c, d). Tidal flats in sediment- rich environments can be regarded as stretched springs due to the above- average sediment inputs. When facing sediment input rise, they are more resistant to being further stretched from their equilibrium, but when sediment input drops, they tend to shrink faster to approach their equilibriums (Fig. 4c). As a contrast, tidal flats in sediment- starving environments can be regarded as compressed springs. They are more resistant to being compressed further when facing sediment input reduction but are more responsive to sediment input increase to expand their width, thus also approaching their equilibriums (Fig. 4d).  
+
+Furthermore, we conducted a synthetical modeling experiment to explore the elastic mechanism of altered SSC on a national scale (Fig. 5). The results delineate there are two 'hot' zones and a 'neutral' zone in tidal flat width dynamics in response to sediment supply changes. The two 'hot' zones are located at both extremes, i.e., increased sediment inputs in sediment- starving systems and decreased sediment inputs in sediment- rich systems, where the tidal flats are expected to experience rapid widening or narrowing (>0.3% per year). The 'neutral' zone, situated between the two hot zones, covers the space of sediment reduction in sediment- starving systems, and a small part of sediment increase in the sediment- rich systems, where the change in tidal flat width is expected to be slow (see Supplementary Text and Supplementary Fig. 10 for a Mechanism explanation). This synthetical modeling aligns with the observed data points along China's coastline with various sediment supply conditions. It indicates that sites with higher rates of widening or narrowing mostly fall in the 'hot' zones, while sites with lower rates generally fall in the 'neutral' zone. It is noteworthy that there are no data points with a rapid SSC increase (> 1% per
+
+<--- Page Split --->
+
+year) in sediment-rich systems (Fig. 5), reflecting the fact that China's tidal flat system is currently experiencing a nationwide reduction in sediment flux.  
+
+![](images/Figure_5.jpg)
+
+<center>Fig. 5 – Modelled and observed tidal flat width response to sediment supply changes. The contour map represents the tidal flat width variation corresponding to different interannual variation rates in suspended sediment concentration (SSC) modeled by DET-ESTMORF, whereas bubbles indicate representative data points of tidal flat width changes along China's coast (see methods). The number indicates the location of the data point on the Chinese coast (selection of data points see Methods). Red bubbles indicate tidal flats expansion, and green bubbles indicate tidal flats degradation. The area between two solid black lines indicates a neutral zone where the interannual variation of tidal flats width is within (-0.1% - 0.3%). The hot zones are in the upper left and lower right corners of the diagram, whereas there are no data points at the upper left corner, i.e., rapid SSC increase (> 1% per year) in sediment-rich systems. </center>  
+
+## Discussion  
+
+Tidal flat systems are highly valued for their ecosystem services, but their morphological evolution on a grand scale remains enigmatic. In this study, we use China's tidal flat system as an example and elucidate that tidal range predominantly shapes the slope of these flats, while sediment supply
+
+<--- Page Split --->
+
+is the critical factor influencing their width. The overall reduction in sediment supply to the coast (reflected by nearshore SSC) emerges as the main cause of the national- scale tidal flat narrowing. These insights may provide global implications. For instance, the observed large intertidal area contraction in other countries in the temperate Northern Pacific region (e.g., the United States) \(^{6,27}\) is likely also caused by sediment supply reduction.  
+
+Interestingly, our observation, combined with synthetic modeling, unveils a new non- linear response in tidal flat morphology to altered sediment supply, with a fast response in two 'hot' zones and a slow response in the 'neutral' zone. This finding suggests that tidal flat morphology tends to move towards the equilibrium state rather than being driven further away from it. Sediment- starving tidal flats (often found on the open coast) are in a compressed state compared to the equilibrium, which is more sensitive to sediment flux increase, but less sensitive to further reduction in sediment supply. Thus, these systems are more resilient to further sediment reduction than previously thought, whereas management efforts that increase sediment supply, e.g., controlled river diversion \(^{28}\) and nourishment operations \(^{29}\) would effectively extend tidal flat width in these systems. In contrast, sediment- rich tidal flats are more susceptible to sediment supply declines. Therefore, SSC in these systems (typically in estuaries) should be cautiously monitored and maintained (if possible) to prevent unexpected fast intertidal area losses. On the other hand, actions to increase sediment inputs are not recommended in sediment- rich systems, as tidal flats in these systems resist being stretched further away from their equilibrium and would not accommodate more sediment. The redundant sediment is better allocated to sediment- starving systems and to avoid diminished light penetration to adjacent coastal ecosystems, such as macroalgae \(^{30}\) , seagrasses \(^{31}\) , and coral reefs \(^{32,33}\) , which is vital to their survival.
+
+<--- Page Split --->
+
+In the simulation and analysis of tidal flat evolution, we consciously omit anthropogenic disturbances, such as reclamation activities, as we mainly focus on understanding how tidal flats initiate their evolution from a dynamic equilibrium state when subjected to external disturbances. It is worth noting that due to the time scale (2002- 2016) we considered in the current study we did not include the impact of sea level rise, but over several decades to centuries, its impact can be pronounced<sup>34- 37</sup>. The overall erosion pattern may deteriorate if the impact of accelerated sea level rise is combined with SSC reduction, but sufficient sediment supply could enable tidal flats to adapt to rising sea levels, thus preserving ecosystem stability<sup>9</sup>.  
+
+In light of global sediment flux alterations<sup>19,20</sup> and widespread tidal flat distribution<sup>11,12</sup>, our findings underscore the imperative of developing sustainable sediment allocation strategies, in which sediment supply history and its changing trends should be considered. These strategies would pave the way towards practical guidelines to fine- tune the sediment fluxes to our coasts, enabling effective tidal flat conservations and restorations.  
+
+## Methods  
+
+## Tidal flat width and slope data  
+
+The width of the tidal flat is defined as the horizontal distance from mean high water to low water (Supplementary Fig. 11). The data of tidal flat area along China's coastline were obtained based on the global tidal flat maps<sup>11</sup>, which were analyzed using machine learning methods to identify over \(7.0 \times 10^{5}\) satellite images. The resultant global map of tidal flats exhibits a horizontal resolution of \(30 \mathrm{~m}\) , delineating the spatiotemporal alterations in the distribution of intertidal regions on a global scale. To obtain the width of tidal flats from the map of tidal flat area, we measured the distance of cross- shore transects along the shoreline in Open Street Map (OSM)<sup>38</sup>
+
+<--- Page Split --->
+
+with a resolution of 1 km (Supplementary Fig. 11). The overall accuracy of the tidal flat area map stands at \(82.2\%^{39}\) .  
+
+The slope of the tidal flats was ascertained from the Foreshore Assessment using Space Technology (FAST) database<sup>24</sup>, which provides a high- resolution intertidal bathymetry/elevation dataset. The bed level data has a 20 m horizontal resolution and typically a 30- 50 cm vertical accuracy. To obtain the tidal flat slope, we extracted the bed level data of the points on cross- shore transects along the shoreline. The transect endpoints are based on the range of tidal flats from the tidal flat maps, while transects with discontinuous elevations were excluded (917/2538) since these sites are not considered to have a continuous tidal flat transect. The slope of tidal flat was then calculated by the cross- shore horizontal length L of the transect and vertical height difference \(\Delta H\) between the endpoints (i.e., slope = \(\Delta H / L\) , Supplementary Fig. 12).  
+
+Employing distinct datasets for tidal flat width and slope, we circumvented potential systematic errors and autocorrelation in our statistical analyses. By integrating tidal flat width data (2014- 2016) and slope data (1997- 2017) with SSC (2016), tidal range (2010), and wave height (1979- 2021), we identified the key factors that sculpt the spatial morphology of tidal flats. Our analysis revealed a significant correlation between the spatial distribution of tidal flat width and SSC. Thus, to further reveal the driver of the temporal changes in tidal flat width, we traced the tidal flat width change data from 2002- 2016 and SSC change data from 2003- 2011, constrained by data availability. Changes in tidal flat width from 2002 to 2016 were obtained by intersecting shoreline cross- shore transects with tidal flats maps every three years (2002- 2004, 2005- 2007, 2008- 2010, 2011- 2013, and 2014- 2016)<sup>11</sup>.
+
+<--- Page Split --->
+
+## Tidal range data  
+
+We utilized the TPXO9- atlas dataset<sup>40</sup>, to obtain the mean tidal range for each cross- shore tidal flat transect along China's coast. The spatial resolution of the dataset is \(1 / 30^{\circ}\) . Employing the tide model driver (TMD) tool, we calculated the water level at each transect point for the year 2010 based on the amplitudes and periods of eight tidal constituents: k1, k2, m2, n2, o1, p1, q1, and s2, as provided by the dataset. The average tidal range data was then interpolated from the mean high tide level and the mean low tide level. The average tidal range values for the tidal flats along China's coastline varied between 0.59- 7.33 m, with an average value of 2.78 m.  
+
+## Wave data  
+
+The offshore wave conditions were derived from the ERA5 reanalysis database<sup>41</sup>, which encompasses comprehensive wave height and period data from 2002 to 2016. These data were firstly reprojected to align with the offshore positions of the tidal flat transects. Subsequently, the offshore wave height values were transferred to the wave height situated 3 km off the coast, equating to \(78\%\) of the offshore wave height<sup>42</sup>. In the absence of detailed nearshore bathymetry data, we presumed linear profiles from the points at 3 km off the coast to the mean low tide level<sup>3</sup>, while the intertidal bathymetry (low tide to high tide level) data is available from FAST<sup>24</sup>. This approach allowed for the estimation of wave heights in the immediate vicinity of the tidal flats at mean low tide level, accounting for wave shallowing and dissipation by friction<sup>43</sup>:  
+
+\[D_{w} = \frac{2}{3\pi}\rho f_{w}U_{\delta}^{3} \quad (1)\]  
+
+with
+
+<--- Page Split --->
+
+\[U_{\delta} = \frac{\pi H}{T\sinh kh} \quad (2)\]  
+
+where \(D_{w}\) is the wave height dissipation, \(\rho\) is the water density, \(f_{e}\) is the wave friction factor, which is assigned a value of \(0.055^{44}\) , \(U_{\delta}\) is the wave orbital velocity, \(H\) is the significant wave height, \(T\) is the wave period, \(k\) is the wave number, and \(h\) is the water depth. Wave refraction and diffraction processes are not accounted due to the lack of nearshore bathymetry data.  
+
+The average wave height attenuation from ERA5 to the depth near tidal flat profiles (mean low water level) is \(70\%\) . Wind waves are included in ERA5 wave heights (ca. \(30\mathrm{km}\) offshore), and in the wave reduction process till \(3\mathrm{km}\) off the coast based on the reference (by \(22\%\) ). Thus, the effect of wind waves has been considered till \(3\mathrm{km}\) off the shore, but not extended to the mean low water level.  
+
+To validate the applicability of the nearshore wave calculation method, we collected significant wave height data from tidal flats at various locations along China's coast, including the Yellow River mouth \(^{45}\) , Chongming Island \(^{46,47}\) , Leizhou Bay \(^{48}\) , Taiwan Longdong \(^{49}\) , and Jiangsu \(^{50}\) . The agreement between our estimated wave heights and periods with in-situ observations (Supplementary Table 1 and 2) confirms the viability of our approach.  
+
+## Suspended Sediment Concentration (SSC) data  
+
+The SSC data near monitoring tidal flat transects was obtained from the GlobColour project (https://www.globcolour.info). The data represent total suspended matter (TSM) in the water column and have been validated. The SSC was derived from MERIS and OLCI satellite measurements using a neural network algorithm \(^{51}\) , which was then validated using field data
+
+<--- Page Split --->
+
+collected from coastal waters. While different satellite sensors such as VIIRS provide SSC data products using different algorithms<sup>52</sup>, the selection of MERIS and OLCI data was because of their temporal coverage of 2004–2012 (MERIS) and 2016 (OLCI). The SSC data product has a horizontal resolution of 1/24 degree and a temporal resolution of one month. To investigate the drivers behind the spatial distribution of tidal flat morphology, we analyzed the monthly average SSC data obtained through the inversion of OLCI satellite data for the year 2016 to be synchronized with the tidal flat width (2002 - 2016) and slope data (1997 - 2017). Specifically, the Pearl River estuary exhibits the highest annual average SSC of 100 mg/L, while the Changjiang Delta, particularly the Hangzhou Bay area, shows the highest maximum SSC values, reaching 200 mg/L, which are in-line with previous field observations<sup>53,54</sup>. The interannual trend of SSC was determined using linear regression applied to monthly average SSC data based on MERIS satellite data from April 2004 to April 2012, with a horizontal resolution of 1/24 degree. To assign a local sediment concentration level to each tidal flat data point, the TSM data point closest to the target point was used.  
+
+Areas with an annual average SSC below 50 mg/L are categorized as sediment- starving systems. This threshold is determined based on the annual average SSC level of China's tidal flat coast. Conversely, areas with an annual average SSC higher than 75 mg/L are defined as sediment- rich systems, which is the annual average SSC in estuaries with typically high sediment supply. These thresholds can be considerably higher than those in other countries<sup>20</sup>, due to the overall higher SSC on China's coast. Thus, sediment- starving and sediment- rich systems should be defined differently in other regions of the world.  
+
+# Statistical analysis
+
+<--- Page Split --->
+
+General Linear Models (GLM) were used to identify the key drivers that shape the spatial distribution of tidal flat morphology. The statistical analyses were performed using the ‘R’ software. GLMs were created for tidal flat width and tidal flat slope, respectively. Given the low resolution of offshore wave height data (0.5°), we consolidated data points that corresponded to the same offshore wave height and computed the mean values for tidal flat width, significant wave height (Hs), suspended sediment concentration (SSC), and tidal range (TR) within each offshore wave height pixel. To address the asymmetric distribution of the data, we applied transformations to the tidal flat width, tidal range, and SSC data by taking their cube roots, and log- transformed the tidal flat slope data prior to their inclusion in the statistical models.  
+
+A stepwise linear regression was used to determine the set of predictor variables that most effectively explained the response variables. This process was guided by the selection of the model yielding the lowest Akaike information criterion (AIC) scores, a widely accepted criterion for model selection \(^{2,55,56}\) . For tidal flat width and slope, SSC and tidal range were included in the most streamlined final model.  
+
+\[width^{3}\sim SSC^{3} + TR^{3} \quad (3)\]  
+
+\[\log (slope)\sim SSC^{3} + TR^{3} \quad (4)\]  
+
+After determining the final predictor variables, ANOVA tables were used to identify significant variables in the model. P- values and R- squared values were calculated using the ‘summary’ package in R software, and the relative importance of each parameter in the linear model was calculated using the ‘relaimpo’ package \(^{56}\) .  
+
+## DET-ESTMORF model
+
+<--- Page Split --->
+
+To explore the relationship between sediment supply variations and the corresponding temporal shifts in tidal flat width, we utilized a one- dimensional tidal flats morphological model, DET- ESTMORF (Fig. 4 and Fig. 5) \(^{14}\) . This model is a hybrid approach that integrates morphological equilibrium \(^{23,25}\) with hydrodynamic and sediment transport processes, in contrast to fully process- based models \(^{15}\) . This model has been validated previously \(^{14}\) , and applied in intertidal biogeomorphology studies \(^{9}\) .  
+
+Tidal flats in morphological equilibrium are characterized by a consistent distribution of bottom bed shear stress and SSC throughout their profile. When the actual local bed shear stress surpasses the equilibrium shear stress, there is a tendency of erosion, requiring a higher SSC to counteract the tendency and maintain equilibrium, i.e., a higher local equilibrium concentration than average. Hence, the magnitude of bottom bed shear stress determines the tendency of tidal flats' geomorphic change, while the difference between actual SSC and local equilibrium SSC indicates the occurrence of geomorphic change, which is similar to the process- based models, e.g., Delft3D.  
+
+The balance between erosion and accretion in the vertical direction is maintained, which can be expressed as follows \(^{57,58}\) :  
+
+\[m_{e}(\frac{\tau_{E}}{\tau_{cr}} -1) = c_{E}w_{s} \quad (5)\]  
+
+Here, \(m_{e}\) represents the erosion coefficient \([\mathrm{kg} / (\mathrm{m}^{2}\mathrm{s})]\) , \(\tau_{E}\) is the uniform bed shear stress (Pa), \(\tau_{cr}\) denotes the critical shear stress (Pa), and \(w_{s}\) represents the sedimentation rate (m/s). In this model, \(c_{E}\) is the total SSC value (mg/L) at the outer sea boundary, and the local equilibrium concentration \(c_{e}\) is based on the ratio of actual bottom bed shear stress to \(\tau_{E}\) :
+
+<--- Page Split --->
+
+\[c_{e} = c_{E}\left(\frac{\tau_{90}}{\tau_{E}}\right)^{n} \quad (6)\]  
+
+\(\tau_{90}\) represents the 90th percentile of bottom bed shear stress in one tidal cycle (Pa). The actual spatial and temporal distributions of local SSC are obtained by solving the diffusion equation:  
+
+\[\frac{\partial(h c)}{\partial t} = w_{s}(c_{e} - c) + \frac{\partial}{\partial x} (D h\frac{\partial c}{\partial x}) \quad (7)\]  
+
+where \(c\) is the volume sediment concentration \((\mathrm{m}^{3} / \mathrm{m}^{3})\) , and \(D\) represents the diffusion coefficient \((\mathrm{m}^{2} / \mathrm{s})\) . The change in tidal flats morphology following each tidal cycle is calculated based on the difference between \(c\) and \(c_{e}\) :  
+
+\[\frac{\partial z}{\partial t} = \frac{1}{1 - p} w_{s}(c - c_{e}) \quad (8)\]  
+
+where \(z\) is the bed elevation (m), and \(p\) is the bed porosity.  
+
+Regarding the calculation of hydrodynamic processes, we considered both effects of wave and tidal shear stresses on the bottom bed. To obtain the tidal- induced bottom bed shear stress, we used the same method as the original DET- ESTMORF<sup>14</sup>. Concerning the bottom- bed shear stress induced by wave propagation, we did not utilize the SWAN model as in the original DET- ESTMORF model to compute wave height distribution but instead calculated the along- range distribution of wave height by computing the energy loss due to bottom- bed friction<sup>59</sup>. The attenuation of wave energy can be described using the continuum equation for wave energy flux as follows<sup>60</sup>:  
+
+\[\frac{d}{d x} (E c_{g}) = \frac{d}{d x} (\frac{1}{8}\rho g H^{2}c_{g}) = -D_{w} \quad (9)\]  
+
+where \(E\) is the wave energy \([\mathrm{kg} / (\mathrm{m}^{2}\mathrm{s}^{2})]\) , \(c_{g}\) is the group velocity \((\mathrm{s}^{2} / \mathrm{m})\) , and \(g\) is the acceleration of gravity \((\mathrm{m} / \mathrm{s}^{2})\) .
+
+<--- Page Split --->
+
+The attenuation of wave energy is given by Eq. 1 and Eq. 2. We also considered the effect of wave breaking in the wave propagation process, where the maximum wave height \(H\) is associated with the local water depth and the breaking height coefficient:  
+
+\[H = \min (H,\gamma_{b}h) \quad (10)\]  
+
+where \(\gamma_{b}\) is a constant value (0.5). The bed stress induced by waves \(\tau_{w}\) is quantified as \(^{61}\) :  
+
+\[\tau_{w} = \frac{1}{4}\rho f w U_{\delta}^{2} \quad (11)\]  
+
+where \(f_{w}\) is a friction factor estimated as:  
+
+\[f_{w} = 1.39(\frac{A_{\delta}}{k_{b}})^{-0.52} \quad (12)\]  
+
+Where the wave orbital semi- excursion at the bed \(A_{\delta} = U_{\delta}T\) , bed roughness \(k_{b} = 2\pi d_{50} / 12\) , and \(d_{50}\) is the median grain size of sediment particles.  
+
+To illustrate the mechanism governing the spatial distribution of tidal flat slope and width, we set up 50 sets of randomized model experiments. Here, we set up each set of experiments by establishing a random SSC (20–150 mg/L) and tide range (2- 7 m) to obtain an equilibrium profile from a linear profile calculated over 100 years.  
+
+To reveal the nonlinear response of tidal flat width to SSC change, we selected two sites for explanatory modelling, namely the Western Scheldt site in the Netherlands (sediment- starving) \(^{9}\) and Jiangsu in China (sediment- rich) \(^{62}\) . At the Western Scheldt site (SSC=58 mg/L), the DET- ESTMORF model has been applied in previous studies \(^{9,14}\) . At the Jiangsu site, we obtained the equilibrium profile by performing a 100- year modeling from the initial linear profile based on the tide level, while wave height, and offshore SSC data (134 mg/L) are based on the collected
+
+<--- Page Split --->
+
+datasets in the current study. We validated the model's accuracy by comparing the obtained equilibrium profile with the observed profile reported in the literature (Supplementary Fig. 13) \(^{62}\) . Building on this validation, we proceeded to calculate equilibrium profiles for both sites under varying interannual SSC change rates, including - 1.2%, - 0.8%, 0.8%, and 1.2%. This exercise was designed to ascertain the interannual rates of change in tidal flat width over a 13- year period, reflective of the observational timeframe spanning 2013 to 2015.  
+
+## Analysis of temporal changes in tidal flats width  
+
+To reveal the conditional response of flat width to SSC changes, we extracted tidal width change rate of sediment- rich ( \(>75 \mathrm{mg / L}\) ) and sediment- starving ( \(< 50 \mathrm{mg / L}\) ) tidal flats across different SSC change rate groups (- 3% \(\sim\) - 1.5%/year, - 1.5% \(\sim\) 0%/year, 0% \(\sim\) 1.5%/year, 1.5% \(\sim\) 3%/year). Within each group, we choose median tidal width change (mean of the 48% - 52% percentiles) as the representative value (Fig 4a). To minimize the influence of anthropogenic factors on the geomorphology of tidal flats, data points exhibiting interannual tidal width trends exceeding 20% per year were omitted.  
+
+We then employed the DET- ESTMORF model to simulate the variation of tidal flat width under different sediment supplies (20\~120 mg/L) and varying interannual rates of change of sediment concentration (- 3% \(\sim\) 3%/year). Subsequently, we performed explanatory simulations for the 34 extracted sites as well (Fig. 5). For a consistent basis of comparison, we initialized the model with a wider profile for tidal flats with higher sediment supply, maintaining uniform parameters across all scenarios except for the initial SSC level and the SSC variation rates. For a given SSC level, we first obtained the equilibrium profile from the initial linear profile after 100 years of simulation, after which the change in the tidal flat morphology is negligible \(^{9}\) . Based on the
+
+<--- Page Split --->
+
+500 obtained equilibrium profile, we simulated the tidal flat morphological variations with changing SSC over 13 years (2002 to 2016, as a reflection of the observation), and finally compared the simulated flat width changes resulting from various SSC changing rates.  
+
+## 504 References  
+
+505 1. Wang, X. et al. Rebound in China's coastal wetlands following conservation and restoration. Nat Sustain 4, 1076-1083 (2021). 506 2. Wang, F. et al. Global blue carbon accumulation in tidal wetlands increases with climate change. Natl Sci Rev 8, nwaa296 (2020). 507 3. van Zelst, V. T. M. et al. Cutting the costs of coastal protection by integrating vegetation in flood defences. Nat Commun 12, 6533 (2021). 508 4. Zheng, J. et al. Synergy between coastal ecology and disaster mitigation in China: Policies, practices, and prospects. Ocean & Coastal Management 245, 106866 (2023). 509 5. Schuerch, M. et al. Future response of global coastal wetlands to sea-level rise. Nature 561, 231-234 (2018). 510 6. Jankowski, K. L., Törnqvist, T. E. & Fernandes, A. M. Vulnerability of Louisiana's coastal wetlands to present-day rates of relative sea-level rise. Nat Commun 8, 14792 (2017). 511 7. Appeaning Addo, K. & Manson Incoom, A. Impacts of shoreline morphological change and sea level rise on mangroves: the case of the keta coastal zone. E3 Journal of Environmental Research and Management 4, 0359-0367 (2013). 512 8. Schuerch, M., Spencer, T. & Evans, B. Coupling between tidal mudflats and salt marshes affects marsh morphology. Marine Geology 412, 95-106 (2019). 513 9. Hu, Z. et al. Mechanistic Modeling of Marsh Seedling Establishment Provides a Positive Outlook for Coastal Wetland Restoration Under Global Climate Change. Geophysical Research Letters 48, (2021). 514 10. Bouma, T. J. et al. Short-term mudflat dynamics drive long-term cyclic salt marsh dynamics. Limnology and Oceanography 61, 2261-2275 (2016). 515 11. Murray, N. J. et al. The global distribution and trajectory of tidal flats. Nature 565, 222-225 (2019). 516 12. Murray, N. J. et al. High-resolution mapping of losses and gains of Earth's tidal wetlands. Science 376, 744-749 (2022). 517 13. Roberts, W., Le Hir, P. & Whitehouse, R. J. S. Investigation using simple mathematical models of the effect of tidal currents and waves on the profile shape of intertidal mudflats. Continental Shelf Research 20, 1079-1097 (2000). 518 14. Hu, Z., Wang, Z. B., Zitman, T. J., Stive, M. J. F. & Bouma, T. J. Predicting long-term and short-term tidal flat morphodynamics using a dynamic equilibrium theory. Journal of Geophysical Research: Earth Surface 120, 1803-1823 (2015). 519 15. Liu, X. J., Gao, S. & Wang, Y. P. Modeling profile shape evolution for accreting tidal flats composed of mud and sand: A case study of the central Jiangsu coast, China. Continental Shelf Research 31, 1750-1760 (2011). 520 16. Fan, D., Guo, Y., Wang, P. & Shi, J. Z. Cross-shore variations in morphodynamic processes of an open-coast mudflat in the Changjiang Delta, China: With an emphasis on storm impacts. Continental Shelf Research 26, 517-538 (2006).
+
+<--- Page Split --->
+
+17. Wang, X. et al. Tracking annual changes of coastal tidal flats in China during 1986–2016 through analyses of Landsat images with Google Earth Engine. Remote Sensing of Environment 238, 110987 (2020).18. Chen, Y. et al. Land claim and loss of tidal flats in the Yangtze Estuary. Sci Rep 6, 24018 (2016).19. Best, J. Anthropogenic stresses on the world’s big rivers. Nature Geosci 12, 7–21 (2019).20. Dethier, E. N., Renshaw, C. E. & Magilligan, F. J. Rapid changes to global river suspended sediment flux by humans. Science 376, 1447–1452 (2022).21. Zhou, Z. et al. Sediment concentration variations in the East China Seas over multiple timescales indicated by satellite observations. Journal of Marine Systems 212, 103430 (2020).22. Li, J., Pan, S., Chen, Y., Fan, Y.-M. & Pan, Y. Numerical estimation of extreme waves and surges over the northwest Pacific Ocean. Ocean Engineering 153, 225–241 (2018).23. Friedrichs, C. T. 3.06 - Tidal Flat Morphodynamics: A Synthesis. in Treatise on Estuarine and Coastal Science (eds. Wolanski, E. & McLusky, D.) 137–170 (Academic Press, Waltham, 2011). doi:10.1016/B978-0-12-374711-2.00307-7.24. Sala Calero, J. et al. FAST MI-SAFE PLATFORM: FORESHORE ASSESSMENT USING SPACE TECHNOLOGY. (2017). doi:10.5281/zenodo.1065940.25. Friedrichs, C. T. & Aubrey, D. G. Uniform Bottom Shear Stress and Equilibrium Hyposometry of Intertidal Flats. in Mixing in Estuaries and Coastal Seas 405–429 (American Geophysical Union (AGU), 1996). doi:10.1029/CE050p0405.26. Bi, S. et al. Distribution of heavy metals and environmental assessment of surface sediment of typical estuaries in eastern China. Marine Pollution Bulletin 121, 357–366 (2017).27. Ganju, N. K. et al. Spatially integrative metrics reveal hidden vulnerability of microtidal salt marshes. Nat Commun 8, 14156 (2017).28. Nittrouer, J. A. et al. Mitigating land loss in coastal Louisiana by controlled diversion of Mississippi River sand. Nature Geosci 5, 534–537 (2012).29. Baptist, M. J. et al. Beneficial use of dredged sediment to enhance salt marsh development by applying a ‘Mud Motor’. Ecological Engineering 127, 312–323 (2019).30. Glover, H. E. et al. Impacts of Suspended Sediment on Nearshore Benthic Light Availability Following Dam Removal in a Small Mountainous River: In Situ Observations and Statistical Modeling. Estuaries and Coasts 42, 1804–1820 (2019).31. Cabaço, S., Santos, R. & Duarte, C. M. The impact of sediment burial and erosion on seagrasses: A review. Estuarine, Coastal and Shelf Science 79, 354–366 (2008).32. McClenachan, L., O’Connor, G., Neal, B. P., Pandolfi, J. M. & Jackson, J. B. C. Ghost reefs: Nautical charts document large spatial scale of coral reef loss over 240 years. Science Advances 3, e1603155 (2017).33. Cramer, K. L. et al. Widespread loss of Caribbean acroporid corals was underway before coral bleaching and disease outbreaks. Science Advances 6, eaax9395 (2020).34. Marani, M., D’Alpaos, A., Lanzoni, S., Carniello, L. & Rinaldo, A. The importance of being coupled: Stable states and catastrophic shifts in tidal biomorphodynamics. Journal of Geophysical Research: Earth Surface 115, (2010).35. Kirwan, M. L. et al. Limits on the adaptability of coastal marshes to rising sea level. Geophysical Research Letters 37, (2010).36. Kirwan, M. L., Temmerman, S., Skeehan, E. E., Guntenspergen, G. R. & Fagherazzi, S. Overestimation of marsh vulnerability to sea level rise. Nature Clim Change 6, 253–260 (2016).
+
+<--- Page Split --->
+
+37. Mariotti, G. & Carr, J. Dual role of salt marsh retreat: Long-term loss and short-term resilience. Water Resources Research 50, 2963–2974 (2014).38. Athanasiou, P. et al. Global distribution of nearshore slopes with implications for coastal retreat. Earth System Science Data 11, 1515–1529 (2019).39. Murray, N. J. et al. High-resolution global maps of tidal flat ecosystems from 1984 to 2019. Sci Data 9, 542 (2022).40. Egbert, G. D. & Erofeeva, S. Y. Efficient Inverse Modeling of Barotropic Ocean Tides. Journal of Atmospheric and Oceanic Technology 19, 183–204 (2002).41. Hersbach, H. et al. The ERA5 global reanalysis. Q.J.R. Meteorol. Soc. 146, 1999–2049 (2020).42. Passaro, M. et al. Global coastal attenuation of wind-waves observed with radar altimetry. Nat Commun 12, 3812 (2021).43. Maan, D. C., van Prooijen, B. C., Wang, Z. B. & De Vriend, H. J. Do intertidal flats ever reach equilibrium? Journal of Geophysical Research: Earth Surface 120, 2406–2436 (2015).44. Le Hir, P. et al. Characterization of intertidal flat hydrodynamics. Continental Shelf Research 20, 1433–1459 (2000).45. Xie, W. et al. Impacts of a storm on the erosion process of a tidal wetland in the Yellow River Delta. CATENA 205, 105461 (2021).46. Yang, S. L., Shi, B. W., Bouma, T. J., Ysebaert, T. & Luo, X. X. Wave Attenuation at a Salt Marsh Margin: A Case Study of an Exposed Coast on the Yangtze Estuary. Estuaries and Coasts 35, 169–182 (2012).47. Yang, S. L. The Role ofScirpusMarsh in Attenuation of Hydrodynamics and Retention of Fine Sediment in the Yangtze Estuary. Estuarine, Coastal and Shelf Science 47, 227–233 (1998).48. Cao, H., Chen, Y., Tian, Y. & Feng, W. Field Investigation into Wave Attenuation in the Mangrove Environment of the South China Sea Coast. coas 32, 1417–1427 (2016).49. Sheng, Y. P., Paramygin, V. A., Terng, C.-T. & Chu, C.-H. Simulating Storm Surge and Inundation Along the Taiwan Coast During Typhoons Fanapi in 2010 and Soulik in 2013. Terr. Atmos. Ocean. Sci. 27, 965–979 (2016).50. Zhao, Y. et al. Rapid formation of marsh-edge cliffs, Jiangsu coast, China. Marine Geology 385, 260–273 (2017).51. Doerffer, R. & Schiller, H. The MERIS Case 2 water algorithm. International Journal of Remote Sensing 28, 517–535 (2007).52. Wei, J. et al. Global Estimation of Suspended Particulate Matter From Satellite Ocean Color Imagery. JGR Oceans 126, (2021).53. Zhan, W., Wu, J., Wei, X., Tang, S. & Zhan, H. Spatio-temporal variation of the suspended sediment concentration in the Pearl River Estuary observed by MODIS during 2003–2015. Continental Shelf Research 172, 22–32 (2019).54. Feng, L., Hu, C., Chen, X. & Song, Q. Influence of the Three Gorges Dam on total suspended matters in the Yangtze Estuary and its adjacent coastal waters: Observations from MODIS. Remote Sensing of Environment 140, 779–788 (2014).55. Nakagawa, S. & Schielzeth, H. A general and simple method for obtaining R2 from generalized linear mixed-effects models. Methods in Ecology and Evolution 4, 133–142 (2013).56. Ladd, C. J. T., Duggan-Edwards, M. F., Bouma, T. J., Pages, J. F. & Skov, M. W. Sediment Supply Explains Long-Term and Large-Scale Patterns in Salt Marsh Lateral Expansion and Erosion. Geophysical Research Letters 46, 11178–11187 (2019).57. Amoudry, L. O. & Souza, A. J. Deterministic Coastal Morphological and Sediment Transport Modeling: A Review and Discussion. Reviews of Geophysics 49, (2011).
+
+<--- Page Split --->
+
+58. Winterwerp, J. C. & Van Kesteren, W. G. M. Settling and Sedimentation. in Developments in Sedimentology vol. 56 121-159 (Elsevier, 2004).  
+
+59. Zhong, Z. & Hu, Z. The Impact of Reclamation on Tidal Flat Morphological Equilibrium. Front. Mar. Sci. 8, 769077 (2021).  
+
+60. Christoffersen, J. B. & Jonsson, I. G. Bed friction and dissipation in a combined current and wave motion. Ocean Engineering 12, 387-423 (1985).  
+
+61. Callaghan, D. P. et al. Hydrodynamic forcing on salt-marsh development: Distinguishing the relative importance of waves and tidal flows. Estuarine, Coastal and Shelf Science 89, 73-88 (2010).  
+
+62. Wang, Y. et al. Sand-Mud Tidal Flat Morphodynamics Influenced by Alongshore Tidal Currents. Journal of Geophysical Research: Oceans 124, 3818-3836 (2019).  
+
+## Acknowledgments  
+
+The authors would like to thank the members of M5 (Mudflat, Marsh, Mangrove, Measurement & Modeling) Lab at Sun Yat-Sen University for their assistance in the research progress. The scientific results and conclusions, as well as any views or opinions expressed herein, are those of the author(s) and do not necessarily reflect those of NOAA or the Department of Commerce.  
+
+## Funding:  
+
+This work was supported by:  
+
+The National Natural Science Foundation of China under contract No. 42176202 (ZH)  
+
+Innovation Group Project of Southern Marine Science and Engineering Guangdong Laboratory  
+
+(Zhuhai) under contract No. 311021004 (ZH)  
+
+Guangdong Provincial Department of Science and Technology under contract No. 2019ZT08G090 (ZH)  
+
+## Author contributions:  
+
+Conceptualization: SL, ZH  Methodology: SL, ZH, VVZ, LQ  Data analysis and plotting: SL, ZH  Supervision: ZH  Writing—original draft: SL, ZH  Writing—review & editing: ZH, TG, ZBW, VVZ, LQ, TX, JYS, TJB  
+
+## Competing interests:
+
+<--- Page Split --->
+
+The authors declare that they have no competing interests.  
+
+672  
+
+Data availability: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Information. Global tidal flat maps are freely available at http://intertidal.app. The TPXO9 dataset is available at https://www.tpxo.net/global/tpxo9- atlas. Wave height data is available from the ERA5 reanalysis dataset. SSC data is available at https://www.globcolour.info.
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+Chinatidalflat.csv supplementarymaterialsChinatidalflat.docx
+
+<--- Page Split --->

preprint/preprint__2b43013de3bb82ae9a113872a06ac68ff47a1fe3e657a67c2b2adc2ace467e88/preprint__2b43013de3bb82ae9a113872a06ac68ff47a1fe3e657a67c2b2adc2ace467e88_det.mmd

@@ -0,0 +1,511 @@
+<|ref|>title<|/ref|><|det|>[[44, 106, 877, 174]]<|/det|>
+# Dynamics and drivers of tidal flat morphology in China  
+
+<|ref|>text<|/ref|><|det|>[[44, 196, 117, 214]]<|/det|>
+Zhan Hu  
+
+<|ref|>text<|/ref|><|det|>[[52, 223, 300, 240]]<|/det|>
+huzh9@mail.sysu.edu.cn  
+
+<|ref|>text<|/ref|><|det|>[[44, 269, 612, 290]]<|/det|>
+Sun Yat- Sen University https://orcid.org/0000- 0002- 2809- 3718  
+
+<|ref|>text<|/ref|><|det|>[[44, 295, 125, 312]]<|/det|>
+Shuai Liu  
+
+<|ref|>text<|/ref|><|det|>[[52, 316, 612, 335]]<|/det|>
+Sun Yat- Sen University https://orcid.org/0009- 0005- 9202- 6607  
+
+<|ref|>text<|/ref|><|det|>[[44, 340, 175, 358]]<|/det|>
+Tim Grandjean  
+
+<|ref|>text<|/ref|><|det|>[[52, 362, 805, 381]]<|/det|>
+Royal Netherlands Institute for Sea Research https://orcid.org/0000- 0002- 2729- 3566  
+
+<|ref|>text<|/ref|><|det|>[[44, 386, 198, 404]]<|/det|>
+Zheng Bing Wang  
+
+<|ref|>text<|/ref|><|det|>[[52, 408, 770, 427]]<|/det|>
+Deltares & Delft University of Technology https://orcid.org/0000- 0002- 8787- 4530  
+
+<|ref|>text<|/ref|><|det|>[[44, 432, 208, 450]]<|/det|>
+Vincent T. M. Zelst  
+
+<|ref|>text<|/ref|><|det|>[[52, 455, 488, 473]]<|/det|>
+Deltares https://orcid.org/0000- 0003- 2923- 1745  
+
+<|ref|>text<|/ref|><|det|>[[44, 479, 95, 496]]<|/det|>
+Lin Qi  
+
+<|ref|>text<|/ref|><|det|>[[52, 500, 515, 519]]<|/det|>
+NOAA Center for Satellite Applications and Research  
+
+<|ref|>text<|/ref|><|det|>[[44, 525, 150, 543]]<|/det|>
+Tianping Xu  
+
+<|ref|>text<|/ref|><|det|>[[52, 548, 252, 565]]<|/det|>
+Sun Yat- Sen University  
+
+<|ref|>text<|/ref|><|det|>[[44, 572, 118, 588]]<|/det|>
+Jun Seo  
+
+<|ref|>text<|/ref|><|det|>[[52, 593, 311, 611]]<|/det|>
+Chonnam National University  
+
+<|ref|>text<|/ref|><|det|>[[44, 617, 168, 635]]<|/det|>
+Tjeerd Bouma  
+
+<|ref|>text<|/ref|><|det|>[[52, 640, 446, 658]]<|/det|>
+Royal Netherlands Institute for Sea Research  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 700, 103, 718]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 738, 137, 756]]<|/det|>
+Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[44, 775, 300, 794]]<|/det|>
+Posted Date: April 17th, 2024  
+
+<|ref|>text<|/ref|><|det|>[[44, 813, 475, 833]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 4209550/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 851, 914, 894]]<|/det|>
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[44, 911, 535, 931]]<|/det|>
+Additional Declarations: There is NO Competing Interest.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 45, 920, 88]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on March 4th, 2025. See the published version at https://doi.org/10.1038/s41467-025-57525-y.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[110, 88, 736, 115]]<|/det|>
+## Dynamics and drivers of tidal flat morphology in China  
+
+<|ref|>text<|/ref|><|det|>[[110, 130, 872, 151]]<|/det|>
+Shuai Liu \(^{1}\) , Zhan Hu \(^{1,2,3*}\) , Tim J. Grandjean \(^{4,5}\) , Zheng B. Wang \(^{6,7}\) , Vincent T. M. van Zels \(^{6,7}\) , Lin  
+
+<|ref|>text<|/ref|><|det|>[[110, 158, 512, 180]]<|/det|>
+Qi \(^{8}\) , Tianping Xu \(^{1}\) , Jun Y. Seo \(^{9}\) , Tjeerd J. Bouma \(^{4,5}\)  
+
+<|ref|>text<|/ref|><|det|>[[110, 211, 810, 258]]<|/det|>
+\(^{1}\) School of Marine Sciences, Sun Yat- Sen University, and Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai, China.  
+
+<|ref|>text<|/ref|><|det|>[[110, 263, 805, 310]]<|/det|>
+\(^{2}\) Guangdong Provincial Key Laboratory of Marine Resources and Coastal Engineering, Guangzhou, China.  
+
+<|ref|>text<|/ref|><|det|>[[110, 315, 820, 336]]<|/det|>
+\(^{3}\) Pearl River Estuary Marine Ecosystem Research Station, Ministry of Education, Zhuhai  
+
+<|ref|>text<|/ref|><|det|>[[110, 341, 835, 362]]<|/det|>
+\(^{4}\) Department of Estuarine and Delta Systems, Royal Netherlands Institute for Sea Research  
+
+<|ref|>text<|/ref|><|det|>[[110, 368, 395, 387]]<|/det|>
+(NIOZ), Yerseke, The Netherlands.  
+
+<|ref|>text<|/ref|><|det|>[[110, 393, 865, 438]]<|/det|>
+\(^{5}\) Faculty of Geosciences, Department of Physical Geography, Utrecht University, Utrecht, The Netherlands.  
+
+<|ref|>text<|/ref|><|det|>[[110, 444, 875, 490]]<|/det|>
+\(^{6}\) Faculty of Civil Engineering and Geosciences, Delft University of Technology, 2628 CN Delft, the Netherlands.  
+
+<|ref|>text<|/ref|><|det|>[[110, 496, 460, 516]]<|/det|>
+\(^{7}\) Deltares, 2600 MH Delft, the Netherlands.  
+
+<|ref|>text<|/ref|><|det|>[[110, 522, 812, 544]]<|/det|>
+\(^{8}\) NOAA Center for Satellite Applications and Research, College Park, MD 20740, USA.  
+
+<|ref|>text<|/ref|><|det|>[[110, 549, 844, 570]]<|/det|>
+\(^{9}\) Department of Oceanography, Chonnam National University, Gwangju, Republic of Korea.  
+
+<|ref|>text<|/ref|><|det|>[[110, 576, 306, 595]]<|/det|>
+\(^{*}\) Corresponding author.  
+
+<|ref|>text<|/ref|><|det|>[[110, 602, 441, 621]]<|/det|>
+E- mail address: huzh9@mail.sysu.edu.cn
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[63, 88, 190, 106]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[110, 120, 888, 528]]<|/det|>
+Recent remote sensing analysis has revealed extensive loss of tidal flats, yet the mechanisms driving these large- scale changes remain enigmatic. This study traces the spatiotemporal variations of 2538 tidal flat transects across China to elucidate how their morphological features vary with external factors, including suspended sediment concentration (SSC), tidal range, and wave height. We observe a correlation between flat width and SSC distribution, and between flat slope and tidal range. A national- wide decline in flat width is observed together with SSC reduction between 2002 and 2016. Intriguingly, sediment- rich flats exhibit a more rapid response to SSC reduction compared to sediment- starving areas, but the converse is observed with SSC increase. These conditional responses stem from the morphodynamic tendency towards equilibrium, which is well explained by synthetical modeling. This finding suggests that tidal flats are resilient to sediment supply reduction, and nation- scale sediment allocation can assist in preserving valuable intertidal areas.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[111, 90, 170, 107]]<|/det|>
+## MAIN  
+
+<|ref|>sub_title<|/ref|><|det|>[[111, 124, 223, 142]]<|/det|>
+## Introduction  
+
+<|ref|>text<|/ref|><|det|>[[110, 156, 886, 386]]<|/det|>
+Tidal flats, alongside their adjacent coastal vegetation systems such as marshes and mangroves, constitute one of the most widespread coastal habitats globally. They provide essential ecosystem services, including storm protection, carbon sequestration, and serving as nurseries for fisheries \(^{1 - 4}\) . The resilience of coastal vegetation ecosystems to climate change has been a focus of attention \(^{5 - 7}\) , and previous research has shown that the fates of tidal flats and vegetated coastal systems are closely linked \(^{8 - 10}\) , underscoring the need for in- depth study of tidal flats dynamics for science- based management.  
+
+<|ref|>text<|/ref|><|det|>[[110, 435, 886, 876]]<|/det|>
+With the recent advancement in remote sensing, global scale tidal flat dynamics have been mapped in detail \(^{11,12}\) . Strikingly, approximately \(16\%\) of the tidal flat area has been lost during 1984- 2016, and more than \(56\%\) of tidal flat area dynamics were attributed to natural coastal processes and global climate changes, such as wave erosion and altered sediment supply, rather than direct human impact, e.g., reclamation \(^{12}\) . These natural processes and environmental changes often operate at large scales and originate far from sites with emerging dynamics. Previous studies have demonstrated that tidal range, wave height, and sediment supply can all influence tidal flat morphology at a local scale \(^{13 - 16}\) , but which factors dominate over large scales and how they lead to temporal changes remain unclear. For instance, sediment supply is widely regarded as a key factor influencing tidal flat morphology \(^{17,18}\) , and there have been dramatic alterations in global suspended sediment flux to the coast \(^{19,20}\) . In the global hydrologic north (north of \(\sim 20^{\circ}\mathrm{N}\) ), a \(49\%\) reduction in river sediment flux has been observed due to damming, while in the global hydrologic south (south of \(\sim 20^{\circ}\mathrm{S}\) ), river sediment flux has increased by approximately \(41\%\) since the 1980s, primarily due
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 886, 144]]<|/det|>
+to activities such as mining<sup>20</sup>. These systematic changes in global river sediment flux are expected to have a profound impact on large- scale tidal flat systems, which has yet to be revealed.  
+
+<|ref|>text<|/ref|><|det|>[[110, 191, 884, 805]]<|/det|>
+China possesses the \(2^{\mathrm{nd}}\) largest tidal flat area in the world, with a variety of tidal ranges, wave heights, and sediment supply<sup>11,21,22</sup>. Additionally, China's tidal flat system has experienced some of the most rapid loss globally (at \(2.6\%\) per year)<sup>11,12,17</sup>. These facts make China's tidal flats an ideal model system for studying the diversification of large- scale tidal flat morphology. In the current study, tidal flats are defined as mud or sand flats with regular tidal inundation<sup>11</sup>. We focus on tidal flat slope and width, as they are regarded as the representative characteristics of tidal flat morphology<sup>23</sup>, which are available from global datasets<sup>11,24</sup>. In total, we composed 2538 tidal flat transects along China's coastline for width analysis and 1620 transects for slope analysis, following independent and complete profiles for flat width and elevation continuous profile for flat slope (see Methods). Sites affected by direct human impact (e.g., reclamation) were excluded from the analysis. To explain the observed spatiotemporal changes, we included morphodynamic modeling by the DET- ESTMORF model, which is based on a dynamic equilibrium theory<sup>23,25</sup> (see Methods). By combining remote sensing observation and morphodynamic modeling, we were able to identify the main drivers influencing spatio- temporal variation of tidal flats morphology in China and understand the nonlinear elastic responses of tidal flats to ambient sediment supply changes in relation to their equilibrium. Our findings provide valuable insights into the large- scale tidal flat morphology and can serve as a key reference for coastal management with changing sediment supply.  
+
+<|ref|>sub_title<|/ref|><|det|>[[112, 856, 177, 873]]<|/det|>
+## Results  
+
+<|ref|>sub_title<|/ref|><|det|>[[112, 889, 560, 910]]<|/det|>
+## Spatial variation and drivers in tidal flat morphology
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 886, 320]]<|/det|>
+Remote sensing observation shows that tidal flat width ranges from \(30\mathrm{m}\) to \(2.6\times 10^{4}\mathrm{m}\) between 2002- 2016 (Fig. 1a). \(34.3\%\) of the tidal flats have a width of \(30 - 300\mathrm{m}\) , whereas \(32.4\%\) of the tidal flats are wider than \(1.0\times 10^{3}\mathrm{m}\) (Supplementary Fig. 1). The median value of tidal flat width is 536 m. The widest tidal flats are found in Jiangsu (JS) province with an average width of \(3088\mathrm{m}\) , while the smallest average width is observed in the northern Shandong Peninsula., being \(74\mathrm{m}\) (Fig. 1a). Notably, the average flats width in estuaries ( \(1212\mathrm{m}\) , \(\mathrm{N} = 275\) ) is greater than those on open coasts ( \(921\mathrm{m}\) , \(\mathrm{N} = 2263\) ) (Supplementary Fig. 2) \(^{26}\) , which is likely attributed to sediment availability.  
+
+<|ref|>image<|/ref|><|det|>[[190, 364, 808, 814]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[110, 816, 888, 890]]<|/det|>
+<center>Fig. 1 - Distribution of width and slope of tidal flats along the coast of China. The distribution of tidal flat width ( \(N = 2538\) ) (a) and slope ( \(N = 1620\) ) (b) with a resolution of 1 km). Dark gray areas indicate coastal areas, namely Liaoning (LN), Hebei (HB), Shandong (SD), Jiangsu (JS), Zhejiang </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 886, 266]]<|/det|>
+(ZJ), Fujian (FJ), Guangdong (GD), Guangxi (GX), Hainan (HN) and Taiwan (TW). Red boxes indicate regions of significant width and slope, and blue boxes indicate regions of minimal values. The values in the figure represent the average values of data in the box. The variation of tidal flat width and slope along latitude is displayed on the right panel. The shaded areas represent the range from the 25th percentile to the 75th percentile. (c) The relation between tidal flat width and median slope in each group. The number displayed at the base of each bar represents the number of data points in the group. The error bars represent the upper 75%-ile and the 25%-ile.  
+
+<|ref|>text<|/ref|><|det|>[[110, 304, 884, 500]]<|/det|>
+The slope of tidal flats in China ranges from \(3.6 \times 10^{- 5} - 1.2 \times 10^{- 1}\) (Fig. 1b), and the median slope is \(7.2 \times 10^{- 3}\) . The average slope of tidal flats in estuaries ( \(6.8 \times 10^{- 3}\) , \(\mathrm{N} = 153\) ) is smaller than those on open coasts ( \(1.2 \times 10^{- 2}\) , \(\mathrm{N} = 1467\) ). Specifically, the average slope value in the southern Fujian (FJ) is around \(1.8 \times 10^{- 2}\) , which is the highest in the country. In contrast, the tidal flats in southern Jiangsu (JS) exhibit the lowest slope, being \(2.3 \times 10^{- 3}\) (Fig. 1b). Furthermore, the tidal flat slope generally decreases as the tidal flat width increases (Fig. 1c).  
+
+<|ref|>text<|/ref|><|det|>[[110, 550, 886, 812]]<|/det|>
+To identify the main drivers shaping the tidal flat morphology, we analyze the distribution of width (Fig. 2a, b) and slope (Fig. 2c, d) in conjunction with local wave height, tidal range, and suspended sediment concentration (SSC) (details in Materials and Methods). Results show tidal flat width exhibits a positive correlation with SSC ( \(\mathrm{R} = 0.71\) , \(\mathrm{P} < 0.01\) ) and tidal range ( \(\mathrm{R} = 0.44\) , \(\mathrm{P} < 0.01\) ), but a negative correlation with wave height ( \(\mathrm{R} = - 0.20\) , \(\mathrm{P} < 0.05\) ). A stepwise linear regression analysis indicates that the combination of SSC (80% relative importance) and tidal range (20% relative importance) best explains the distribution of tidal flat width ( \(\mathrm{R}^{2} = 0.53\) ) (Fig. 2e), highlighting the primary role of local SSC level.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[220, 80, 785, 622]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[111, 636, 884, 864]]<|/det|>
+<center>Fig. 2 - Relations between environmental factors and tidal flats width (2014-2016) and slope in space. Tidal flat width for different classes of suspended sediment concentrations (SSC) (a) and tidal ranges (b). Tidal flat slope for different classes of SSC (c) and tidal ranges (d). The number displayed at the base of each bar represents the corresponding number of data points. The relationship between the spatial distribution of tidal width and SSC (e) and the relationship between the spatial distribution of tidal slope and tidal range (f) are derived from a stepwise linear regression model. \(R^2\) denotes the correlation coefficient of linear regression. The data points represent the distribution of the data, and the solid lines indicate the model fit, bounded by 95% confidence intervals </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[108, 87, 886, 250]]<|/det|>
+Tidal flat slope exhibits a positive correlation with tidal range \(\mathrm{(R = 0.34,P< 0.01)}\) , but a negative correlation with SSC \(\mathrm{(R = - 0.26,P< 0.01)}\) . Other factors, such as wave height, are known to have an impact on tidal flat slope \(^{13,23}\) , but they are less important in the current dataset. The combination of tidal range (58% of relative importance) and SSC (42% of relative importance) best explains the distribution of tidal flat slope \(\mathrm{(R^{2} = 0.30)}\) (Fig. 2f).  
+
+<|ref|>text<|/ref|><|det|>[[108, 295, 886, 595]]<|/det|>
+These findings on the spatial distribution of flat width and slope are in agreement with DET- ESTMORF modelling \(^{14}\) . On tidal flats with a large tidal range, the bed shear stress generated by tidal currents has a negative gradient towards the landward direction, leading to erosion at the low tidal flats and deposition at the higher tidal flats (Supplementary Fig. 5). Such a gradient becomes more prominent as tidal range increases (Supplementary Fig. 6). Thus, a greater tidal range can lead to a steeper tidal flat profile as observed. On the relation between SSC and tidal flat width, larger SSC can result in sediment surplus that exceeds the level required for equilibrium. This leads to deposition on the tidal flat and seaward expansion. Consequently, tidal flat width exhibits a positive correlation with SSC.  
+
+<|ref|>sub_title<|/ref|><|det|>[[110, 645, 741, 666]]<|/det|>
+## Temporal dynamics in tidal flat width and the underlying elastic responses  
+
+<|ref|>text<|/ref|><|det|>[[108, 679, 886, 910]]<|/det|>
+The time series of tidal flat width from 2002 to 2016 shows a clear trend of overall erosion (Fig. 3a, c). The nationwide median tidal flat width reduced from 596 m to 536 m from 2002 to 2016, i.e., reduced at - 0.8% per year (Fig. 3c). Such a trend is particularly prominent in Hainan Island and the Changjiang River Estuary, where the interannual change rate in tidal flat width was - 1.89% and - 1.90% per year, respectively (Fig. 3a). In contrast, a few sites (28.3%) exhibited an expansion trend of > 0.1% per year, such as the southern Shandong (SD) Peninsula and the coast of Fujian Province, with width increase rates of 0.76% and 0.58% per year, respectively (Fig. 3a).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[171, 120, 816, 630]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[111, 640, 880, 867]]<|/det|>
+<center>Fig. 3 - Response of tidal flat width to changes in suspended sediment concentration along the coast of China. Distribution of inter-annual trend of tidal flat width (2002-2016) (a) and suspended sediment concentration (2003-2011) (b). Outliers that deviated from the median by more than three times the median absolute deviation are excluded. The values in the figure represent the average values of data in the box. \(51.3\%\) of the tidal flat transects show a decreasing trend in width, while \(54.9\%\) of the transects experience a decrease in SSC. The insets represent the probability distribution of the data points. (c)Temporal changes in the median value of tidal flat width (linear regression, \(P = 0.1167\) ) and (d) SSC along the coast of China (linear regression, \(P = 0.0274\) ). </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[108, 87, 888, 355]]<|/det|>
+Concurrently, there was a nationwide declining trend in sediment supply at \(- 0.4\%\) per year as observed in interannual SSC from 2002 to 2012 (Fig. 3d). K- means clustering analysis, a measure for partitioning data points into distinct, non- overlapping groups, reveals a strong correlation between the interannual SSC reduction and tidal flat width decline ( \(\mathrm{R} = 0.64\) , \(\mathrm{P} < 0.01\) ) (Supplementary Fig. 7). Additionally, \(56\%\) of all observation transects exhibit concurrent decline or increase of sediment supply with tidal flat width (Fig. 3a, b), whereas most of the rest sites ( \(86\%\) ) exhibit an insignificant change in SSC. These statistics suggest a close linkage between altered sediment supply and changes in tidal flat morphology.  
+
+<|ref|>text<|/ref|><|det|>[[108, 400, 888, 912]]<|/det|>
+Beyond the overall connection between SSC reduction and tidal flat width loss, our research further discovers diverse responses of sediment- starving (SSC \(< 50 \mathrm{mg / L}\) , average SSC of China's coastal tidal flats) and sediment- rich (SSC \(>75 \mathrm{mg / L}\) , average SSC of China's tidal flats in the estuary) systems to altered sediment supply. Notably, sediment- rich systems undergo more rapid width reduction than their sediment- starving counterparts when subjected to the same rate of SSC decrease ( \(0.27\%\) vs. \(0.24\%\) per year, Fig. 4a). Conversely, sediment- starving systems exhibit more pronounced width expansion in response to an equivalent increase in SSC ( \(0.55\%\) vs. \(0.30\%\) per year). To elucidate these divergent responses, we employed DET- ESTMORF modelling on two exemplified sites (Fig. 4b and Supplementary Fig. 8). An increase in SSC leads to an immediate rise in actual concentration, but only a modest rise in equilibrium concentration (Supplementary Fig. 9). The difference between the actual concentration and the equilibrium concentration leads to accumulation and width growth. In sediment- starving systems, the equilibrium concentration is inherently lower, and thus the disparity from the actual concentration is greater than in sediment- rich systems (Supplementary Fig. 9). Consequently, the momentum behind the width growth is stronger in sediment- starving systems during SSC increases, leading to faster width
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 857, 141]]<|/det|>
+growth. Conversely, during SSC reductions, sediment-rich systems exhibit a more accelerated width reduction (see Supplementary Text for a more detailed explanation).  
+
+<|ref|>image<|/ref|><|det|>[[182, 191, 870, 595]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[111, 611, 884, 866]]<|/det|>
+<center>Fig. 4 - Elastic response of tidal flats width changes to altered sediment supply changes. Relation between SSC change rate and flat width change rate based on observation (a) and model results (b). Model results are based on tidal flat width changes of exemplified sediment-rich and sediment-starving areas (see Methods) The numbers represent the number of tidal flats transects corresponding to each bar. (c) Diagram showing the state of sediment-rich areas is similar to a stretched spring that tends to restore to morphological equilibrium (dashed line) and resist to be stretched further. (d) Diagram showing the state of sediment-starving areas is similar to a compressed spring that tends to restore morphological equilibrium (dashed line) and resist being compressed further. The dashed lines indicate the equilibrium state in both cases. The convex and concave profiles reflect their high and low sediment input level, respectively \(^{14,23}\) . </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 88, 886, 430]]<|/det|>
+These non- linear responses have not been reported previously. They can be explained by an analogy to the dynamics of elastic systems such as springs. These systems have their intrinsic equilibriums but can be stretched or compressed from equilibriums in response to external sediment input (Fig. 4c, d). Tidal flats in sediment- rich environments can be regarded as stretched springs due to the above- average sediment inputs. When facing sediment input rise, they are more resistant to being further stretched from their equilibrium, but when sediment input drops, they tend to shrink faster to approach their equilibriums (Fig. 4c). As a contrast, tidal flats in sediment- starving environments can be regarded as compressed springs. They are more resistant to being compressed further when facing sediment input reduction but are more responsive to sediment input increase to expand their width, thus also approaching their equilibriums (Fig. 4d).  
+
+<|ref|>text<|/ref|><|det|>[[110, 470, 886, 911]]<|/det|>
+Furthermore, we conducted a synthetical modeling experiment to explore the elastic mechanism of altered SSC on a national scale (Fig. 5). The results delineate there are two 'hot' zones and a 'neutral' zone in tidal flat width dynamics in response to sediment supply changes. The two 'hot' zones are located at both extremes, i.e., increased sediment inputs in sediment- starving systems and decreased sediment inputs in sediment- rich systems, where the tidal flats are expected to experience rapid widening or narrowing (>0.3% per year). The 'neutral' zone, situated between the two hot zones, covers the space of sediment reduction in sediment- starving systems, and a small part of sediment increase in the sediment- rich systems, where the change in tidal flat width is expected to be slow (see Supplementary Text and Supplementary Fig. 10 for a Mechanism explanation). This synthetical modeling aligns with the observed data points along China's coastline with various sediment supply conditions. It indicates that sites with higher rates of widening or narrowing mostly fall in the 'hot' zones, while sites with lower rates generally fall in the 'neutral' zone. It is noteworthy that there are no data points with a rapid SSC increase (> 1% per
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 88, 884, 142]]<|/det|>
+year) in sediment-rich systems (Fig. 5), reflecting the fact that China's tidal flat system is currently experiencing a nationwide reduction in sediment flux.  
+
+<|ref|>image<|/ref|><|det|>[[320, 191, 740, 459]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[111, 473, 884, 728]]<|/det|>
+<center>Fig. 5 – Modelled and observed tidal flat width response to sediment supply changes. The contour map represents the tidal flat width variation corresponding to different interannual variation rates in suspended sediment concentration (SSC) modeled by DET-ESTMORF, whereas bubbles indicate representative data points of tidal flat width changes along China's coast (see methods). The number indicates the location of the data point on the Chinese coast (selection of data points see Methods). Red bubbles indicate tidal flats expansion, and green bubbles indicate tidal flats degradation. The area between two solid black lines indicates a neutral zone where the interannual variation of tidal flats width is within (-0.1% - 0.3%). The hot zones are in the upper left and lower right corners of the diagram, whereas there are no data points at the upper left corner, i.e., rapid SSC increase (> 1% per year) in sediment-rich systems. </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[112, 765, 205, 782]]<|/det|>
+## Discussion  
+
+<|ref|>text<|/ref|><|det|>[[111, 797, 884, 888]]<|/det|>
+Tidal flat systems are highly valued for their ecosystem services, but their morphological evolution on a grand scale remains enigmatic. In this study, we use China's tidal flat system as an example and elucidate that tidal range predominantly shapes the slope of these flats, while sediment supply
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 883, 248]]<|/det|>
+is the critical factor influencing their width. The overall reduction in sediment supply to the coast (reflected by nearshore SSC) emerges as the main cause of the national- scale tidal flat narrowing. These insights may provide global implications. For instance, the observed large intertidal area contraction in other countries in the temperate Northern Pacific region (e.g., the United States) \(^{6,27}\) is likely also caused by sediment supply reduction.  
+
+<|ref|>text<|/ref|><|det|>[[110, 295, 888, 875]]<|/det|>
+Interestingly, our observation, combined with synthetic modeling, unveils a new non- linear response in tidal flat morphology to altered sediment supply, with a fast response in two 'hot' zones and a slow response in the 'neutral' zone. This finding suggests that tidal flat morphology tends to move towards the equilibrium state rather than being driven further away from it. Sediment- starving tidal flats (often found on the open coast) are in a compressed state compared to the equilibrium, which is more sensitive to sediment flux increase, but less sensitive to further reduction in sediment supply. Thus, these systems are more resilient to further sediment reduction than previously thought, whereas management efforts that increase sediment supply, e.g., controlled river diversion \(^{28}\) and nourishment operations \(^{29}\) would effectively extend tidal flat width in these systems. In contrast, sediment- rich tidal flats are more susceptible to sediment supply declines. Therefore, SSC in these systems (typically in estuaries) should be cautiously monitored and maintained (if possible) to prevent unexpected fast intertidal area losses. On the other hand, actions to increase sediment inputs are not recommended in sediment- rich systems, as tidal flats in these systems resist being stretched further away from their equilibrium and would not accommodate more sediment. The redundant sediment is better allocated to sediment- starving systems and to avoid diminished light penetration to adjacent coastal ecosystems, such as macroalgae \(^{30}\) , seagrasses \(^{31}\) , and coral reefs \(^{32,33}\) , which is vital to their survival.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 884, 352]]<|/det|>
+In the simulation and analysis of tidal flat evolution, we consciously omit anthropogenic disturbances, such as reclamation activities, as we mainly focus on understanding how tidal flats initiate their evolution from a dynamic equilibrium state when subjected to external disturbances. It is worth noting that due to the time scale (2002- 2016) we considered in the current study we did not include the impact of sea level rise, but over several decades to centuries, its impact can be pronounced<sup>34- 37</sup>. The overall erosion pattern may deteriorate if the impact of accelerated sea level rise is combined with SSC reduction, but sufficient sediment supply could enable tidal flats to adapt to rising sea levels, thus preserving ecosystem stability<sup>9</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[110, 400, 860, 560]]<|/det|>
+In light of global sediment flux alterations<sup>19,20</sup> and widespread tidal flat distribution<sup>11,12</sup>, our findings underscore the imperative of developing sustainable sediment allocation strategies, in which sediment supply history and its changing trends should be considered. These strategies would pave the way towards practical guidelines to fine- tune the sediment fluxes to our coasts, enabling effective tidal flat conservations and restorations.  
+
+<|ref|>sub_title<|/ref|><|det|>[[111, 611, 190, 628]]<|/det|>
+## Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[112, 645, 382, 664]]<|/det|>
+## Tidal flat width and slope data  
+
+<|ref|>text<|/ref|><|det|>[[110, 679, 886, 910]]<|/det|>
+The width of the tidal flat is defined as the horizontal distance from mean high water to low water (Supplementary Fig. 11). The data of tidal flat area along China's coastline were obtained based on the global tidal flat maps<sup>11</sup>, which were analyzed using machine learning methods to identify over \(7.0 \times 10^{5}\) satellite images. The resultant global map of tidal flats exhibits a horizontal resolution of \(30 \mathrm{~m}\) , delineating the spatiotemporal alterations in the distribution of intertidal regions on a global scale. To obtain the width of tidal flats from the map of tidal flat area, we measured the distance of cross- shore transects along the shoreline in Open Street Map (OSM)<sup>38</sup>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 884, 143]]<|/det|>
+with a resolution of 1 km (Supplementary Fig. 11). The overall accuracy of the tidal flat area map stands at \(82.2\%^{39}\) .  
+
+<|ref|>text<|/ref|><|det|>[[110, 192, 886, 528]]<|/det|>
+The slope of the tidal flats was ascertained from the Foreshore Assessment using Space Technology (FAST) database<sup>24</sup>, which provides a high- resolution intertidal bathymetry/elevation dataset. The bed level data has a 20 m horizontal resolution and typically a 30- 50 cm vertical accuracy. To obtain the tidal flat slope, we extracted the bed level data of the points on cross- shore transects along the shoreline. The transect endpoints are based on the range of tidal flats from the tidal flat maps, while transects with discontinuous elevations were excluded (917/2538) since these sites are not considered to have a continuous tidal flat transect. The slope of tidal flat was then calculated by the cross- shore horizontal length L of the transect and vertical height difference \(\Delta H\) between the endpoints (i.e., slope = \(\Delta H / L\) , Supplementary Fig. 12).  
+
+<|ref|>text<|/ref|><|det|>[[110, 576, 886, 910]]<|/det|>
+Employing distinct datasets for tidal flat width and slope, we circumvented potential systematic errors and autocorrelation in our statistical analyses. By integrating tidal flat width data (2014- 2016) and slope data (1997- 2017) with SSC (2016), tidal range (2010), and wave height (1979- 2021), we identified the key factors that sculpt the spatial morphology of tidal flats. Our analysis revealed a significant correlation between the spatial distribution of tidal flat width and SSC. Thus, to further reveal the driver of the temporal changes in tidal flat width, we traced the tidal flat width change data from 2002- 2016 and SSC change data from 2003- 2011, constrained by data availability. Changes in tidal flat width from 2002 to 2016 were obtained by intersecting shoreline cross- shore transects with tidal flats maps every three years (2002- 2004, 2005- 2007, 2008- 2010, 2011- 2013, and 2014- 2016)<sup>11</sup>.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[112, 123, 261, 142]]<|/det|>
+## Tidal range data  
+
+<|ref|>text<|/ref|><|det|>[[110, 155, 886, 387]]<|/det|>
+We utilized the TPXO9- atlas dataset<sup>40</sup>, to obtain the mean tidal range for each cross- shore tidal flat transect along China's coast. The spatial resolution of the dataset is \(1 / 30^{\circ}\) . Employing the tide model driver (TMD) tool, we calculated the water level at each transect point for the year 2010 based on the amplitudes and periods of eight tidal constituents: k1, k2, m2, n2, o1, p1, q1, and s2, as provided by the dataset. The average tidal range data was then interpolated from the mean high tide level and the mean low tide level. The average tidal range values for the tidal flats along China's coastline varied between 0.59- 7.33 m, with an average value of 2.78 m.  
+
+<|ref|>sub_title<|/ref|><|det|>[[112, 437, 208, 455]]<|/det|>
+## Wave data  
+
+<|ref|>text<|/ref|><|det|>[[110, 468, 886, 771]]<|/det|>
+The offshore wave conditions were derived from the ERA5 reanalysis database<sup>41</sup>, which encompasses comprehensive wave height and period data from 2002 to 2016. These data were firstly reprojected to align with the offshore positions of the tidal flat transects. Subsequently, the offshore wave height values were transferred to the wave height situated 3 km off the coast, equating to \(78\%\) of the offshore wave height<sup>42</sup>. In the absence of detailed nearshore bathymetry data, we presumed linear profiles from the points at 3 km off the coast to the mean low tide level<sup>3</sup>, while the intertidal bathymetry (low tide to high tide level) data is available from FAST<sup>24</sup>. This approach allowed for the estimation of wave heights in the immediate vicinity of the tidal flats at mean low tide level, accounting for wave shallowing and dissipation by friction<sup>43</sup>:  
+
+<|ref|>equation<|/ref|><|det|>[[440, 784, 886, 828]]<|/det|>
+\[D_{w} = \frac{2}{3\pi}\rho f_{w}U_{\delta}^{3} \quad (1)\]  
+
+<|ref|>text<|/ref|><|det|>[[112, 850, 152, 866]]<|/det|>
+with
+
+<--- Page Split --->
+<|ref|>equation<|/ref|><|det|>[[433, 90, 886, 134]]<|/det|>
+\[U_{\delta} = \frac{\pi H}{T\sinh kh} \quad (2)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 150, 886, 279]]<|/det|>
+where \(D_{w}\) is the wave height dissipation, \(\rho\) is the water density, \(f_{e}\) is the wave friction factor, which is assigned a value of \(0.055^{44}\) , \(U_{\delta}\) is the wave orbital velocity, \(H\) is the significant wave height, \(T\) is the wave period, \(k\) is the wave number, and \(h\) is the water depth. Wave refraction and diffraction processes are not accounted due to the lack of nearshore bathymetry data.  
+
+<|ref|>text<|/ref|><|det|>[[110, 325, 886, 486]]<|/det|>
+The average wave height attenuation from ERA5 to the depth near tidal flat profiles (mean low water level) is \(70\%\) . Wind waves are included in ERA5 wave heights (ca. \(30\mathrm{km}\) offshore), and in the wave reduction process till \(3\mathrm{km}\) off the coast based on the reference (by \(22\%\) ). Thus, the effect of wind waves has been considered till \(3\mathrm{km}\) off the shore, but not extended to the mean low water level.  
+
+<|ref|>text<|/ref|><|det|>[[110, 535, 886, 695]]<|/det|>
+To validate the applicability of the nearshore wave calculation method, we collected significant wave height data from tidal flats at various locations along China's coast, including the Yellow River mouth \(^{45}\) , Chongming Island \(^{46,47}\) , Leizhou Bay \(^{48}\) , Taiwan Longdong \(^{49}\) , and Jiangsu \(^{50}\) . The agreement between our estimated wave heights and periods with in-situ observations (Supplementary Table 1 and 2) confirms the viability of our approach.  
+
+<|ref|>sub_title<|/ref|><|det|>[[110, 744, 525, 765]]<|/det|>
+## Suspended Sediment Concentration (SSC) data  
+
+<|ref|>text<|/ref|><|det|>[[110, 780, 886, 905]]<|/det|>
+The SSC data near monitoring tidal flat transects was obtained from the GlobColour project (https://www.globcolour.info). The data represent total suspended matter (TSM) in the water column and have been validated. The SSC was derived from MERIS and OLCI satellite measurements using a neural network algorithm \(^{51}\) , which was then validated using field data
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[108, 87, 886, 565]]<|/det|>
+collected from coastal waters. While different satellite sensors such as VIIRS provide SSC data products using different algorithms<sup>52</sup>, the selection of MERIS and OLCI data was because of their temporal coverage of 2004–2012 (MERIS) and 2016 (OLCI). The SSC data product has a horizontal resolution of 1/24 degree and a temporal resolution of one month. To investigate the drivers behind the spatial distribution of tidal flat morphology, we analyzed the monthly average SSC data obtained through the inversion of OLCI satellite data for the year 2016 to be synchronized with the tidal flat width (2002 - 2016) and slope data (1997 - 2017). Specifically, the Pearl River estuary exhibits the highest annual average SSC of 100 mg/L, while the Changjiang Delta, particularly the Hangzhou Bay area, shows the highest maximum SSC values, reaching 200 mg/L, which are in-line with previous field observations<sup>53,54</sup>. The interannual trend of SSC was determined using linear regression applied to monthly average SSC data based on MERIS satellite data from April 2004 to April 2012, with a horizontal resolution of 1/24 degree. To assign a local sediment concentration level to each tidal flat data point, the TSM data point closest to the target point was used.  
+
+<|ref|>text<|/ref|><|det|>[[110, 610, 886, 839]]<|/det|>
+Areas with an annual average SSC below 50 mg/L are categorized as sediment- starving systems. This threshold is determined based on the annual average SSC level of China's tidal flat coast. Conversely, areas with an annual average SSC higher than 75 mg/L are defined as sediment- rich systems, which is the annual average SSC in estuaries with typically high sediment supply. These thresholds can be considerably higher than those in other countries<sup>20</sup>, due to the overall higher SSC on China's coast. Thus, sediment- starving and sediment- rich systems should be defined differently in other regions of the world.  
+
+<|ref|>title<|/ref|><|det|>[[111, 890, 277, 908]]<|/det|>
+# Statistical analysis
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 85, 886, 388]]<|/det|>
+General Linear Models (GLM) were used to identify the key drivers that shape the spatial distribution of tidal flat morphology. The statistical analyses were performed using the ‘R’ software. GLMs were created for tidal flat width and tidal flat slope, respectively. Given the low resolution of offshore wave height data (0.5°), we consolidated data points that corresponded to the same offshore wave height and computed the mean values for tidal flat width, significant wave height (Hs), suspended sediment concentration (SSC), and tidal range (TR) within each offshore wave height pixel. To address the asymmetric distribution of the data, we applied transformations to the tidal flat width, tidal range, and SSC data by taking their cube roots, and log- transformed the tidal flat slope data prior to their inclusion in the statistical models.  
+
+<|ref|>text<|/ref|><|det|>[[110, 435, 886, 595]]<|/det|>
+A stepwise linear regression was used to determine the set of predictor variables that most effectively explained the response variables. This process was guided by the selection of the model yielding the lowest Akaike information criterion (AIC) scores, a widely accepted criterion for model selection \(^{2,55,56}\) . For tidal flat width and slope, SSC and tidal range were included in the most streamlined final model.  
+
+<|ref|>equation<|/ref|><|det|>[[411, 608, 886, 639]]<|/det|>
+\[width^{3}\sim SSC^{3} + TR^{3} \quad (3)\]  
+
+<|ref|>equation<|/ref|><|det|>[[400, 656, 886, 686]]<|/det|>
+\[\log (slope)\sim SSC^{3} + TR^{3} \quad (4)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 699, 886, 825]]<|/det|>
+After determining the final predictor variables, ANOVA tables were used to identify significant variables in the model. P- values and R- squared values were calculated using the ‘summary’ package in R software, and the relative importance of each parameter in the linear model was calculated using the ‘relaimpo’ package \(^{56}\) .  
+
+<|ref|>sub_title<|/ref|><|det|>[[111, 874, 320, 893]]<|/det|>
+## DET-ESTMORF model
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 85, 885, 283]]<|/det|>
+To explore the relationship between sediment supply variations and the corresponding temporal shifts in tidal flat width, we utilized a one- dimensional tidal flats morphological model, DET- ESTMORF (Fig. 4 and Fig. 5) \(^{14}\) . This model is a hybrid approach that integrates morphological equilibrium \(^{23,25}\) with hydrodynamic and sediment transport processes, in contrast to fully process- based models \(^{15}\) . This model has been validated previously \(^{14}\) , and applied in intertidal biogeomorphology studies \(^{9}\) .  
+
+<|ref|>text<|/ref|><|det|>[[110, 331, 886, 595]]<|/det|>
+Tidal flats in morphological equilibrium are characterized by a consistent distribution of bottom bed shear stress and SSC throughout their profile. When the actual local bed shear stress surpasses the equilibrium shear stress, there is a tendency of erosion, requiring a higher SSC to counteract the tendency and maintain equilibrium, i.e., a higher local equilibrium concentration than average. Hence, the magnitude of bottom bed shear stress determines the tendency of tidal flats' geomorphic change, while the difference between actual SSC and local equilibrium SSC indicates the occurrence of geomorphic change, which is similar to the process- based models, e.g., Delft3D.  
+
+<|ref|>text<|/ref|><|det|>[[110, 644, 884, 699]]<|/det|>
+The balance between erosion and accretion in the vertical direction is maintained, which can be expressed as follows \(^{57,58}\) :  
+
+<|ref|>equation<|/ref|><|det|>[[427, 720, 884, 758]]<|/det|>
+\[m_{e}(\frac{\tau_{E}}{\tau_{cr}} -1) = c_{E}w_{s} \quad (5)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 777, 885, 904]]<|/det|>
+Here, \(m_{e}\) represents the erosion coefficient \([\mathrm{kg} / (\mathrm{m}^{2}\mathrm{s})]\) , \(\tau_{E}\) is the uniform bed shear stress (Pa), \(\tau_{cr}\) denotes the critical shear stress (Pa), and \(w_{s}\) represents the sedimentation rate (m/s). In this model, \(c_{E}\) is the total SSC value (mg/L) at the outer sea boundary, and the local equilibrium concentration \(c_{e}\) is based on the ratio of actual bottom bed shear stress to \(\tau_{E}\) :
+
+<--- Page Split --->
+<|ref|>equation<|/ref|><|det|>[[445, 90, 886, 140]]<|/det|>
+\[c_{e} = c_{E}\left(\frac{\tau_{90}}{\tau_{E}}\right)^{n} \quad (6)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 159, 886, 215]]<|/det|>
+\(\tau_{90}\) represents the 90th percentile of bottom bed shear stress in one tidal cycle (Pa). The actual spatial and temporal distributions of local SSC are obtained by solving the diffusion equation:  
+
+<|ref|>equation<|/ref|><|det|>[[373, 228, 886, 275]]<|/det|>
+\[\frac{\partial(h c)}{\partial t} = w_{s}(c_{e} - c) + \frac{\partial}{\partial x} (D h\frac{\partial c}{\partial x}) \quad (7)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 291, 886, 384]]<|/det|>
+where \(c\) is the volume sediment concentration \((\mathrm{m}^{3} / \mathrm{m}^{3})\) , and \(D\) represents the diffusion coefficient \((\mathrm{m}^{2} / \mathrm{s})\) . The change in tidal flats morphology following each tidal cycle is calculated based on the difference between \(c\) and \(c_{e}\) :  
+
+<|ref|>equation<|/ref|><|det|>[[410, 397, 886, 444]]<|/det|>
+\[\frac{\partial z}{\partial t} = \frac{1}{1 - p} w_{s}(c - c_{e}) \quad (8)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 460, 590, 482]]<|/det|>
+where \(z\) is the bed elevation (m), and \(p\) is the bed porosity.  
+
+<|ref|>text<|/ref|><|det|>[[110, 494, 886, 760]]<|/det|>
+Regarding the calculation of hydrodynamic processes, we considered both effects of wave and tidal shear stresses on the bottom bed. To obtain the tidal- induced bottom bed shear stress, we used the same method as the original DET- ESTMORF<sup>14</sup>. Concerning the bottom- bed shear stress induced by wave propagation, we did not utilize the SWAN model as in the original DET- ESTMORF model to compute wave height distribution but instead calculated the along- range distribution of wave height by computing the energy loss due to bottom- bed friction<sup>59</sup>. The attenuation of wave energy can be described using the continuum equation for wave energy flux as follows<sup>60</sup>:  
+
+<|ref|>equation<|/ref|><|det|>[[361, 776, 886, 821]]<|/det|>
+\[\frac{d}{d x} (E c_{g}) = \frac{d}{d x} (\frac{1}{8}\rho g H^{2}c_{g}) = -D_{w} \quad (9)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 837, 886, 895]]<|/det|>
+where \(E\) is the wave energy \([\mathrm{kg} / (\mathrm{m}^{2}\mathrm{s}^{2})]\) , \(c_{g}\) is the group velocity \((\mathrm{s}^{2} / \mathrm{m})\) , and \(g\) is the acceleration of gravity \((\mathrm{m} / \mathrm{s}^{2})\) .
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 886, 179]]<|/det|>
+The attenuation of wave energy is given by Eq. 1 and Eq. 2. We also considered the effect of wave breaking in the wave propagation process, where the maximum wave height \(H\) is associated with the local water depth and the breaking height coefficient:  
+
+<|ref|>equation<|/ref|><|det|>[[434, 194, 886, 217]]<|/det|>
+\[H = \min (H,\gamma_{b}h) \quad (10)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 236, 852, 259]]<|/det|>
+where \(\gamma_{b}\) is a constant value (0.5). The bed stress induced by waves \(\tau_{w}\) is quantified as \(^{61}\) :  
+
+<|ref|>equation<|/ref|><|det|>[[435, 274, 886, 316]]<|/det|>
+\[\tau_{w} = \frac{1}{4}\rho f w U_{\delta}^{2} \quad (11)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 336, 458, 356]]<|/det|>
+where \(f_{w}\) is a friction factor estimated as:  
+
+<|ref|>equation<|/ref|><|det|>[[428, 371, 886, 414]]<|/det|>
+\[f_{w} = 1.39(\frac{A_{\delta}}{k_{b}})^{-0.52} \quad (12)\]  
+
+<|ref|>text<|/ref|><|det|>[[110, 433, 886, 490]]<|/det|>
+Where the wave orbital semi- excursion at the bed \(A_{\delta} = U_{\delta}T\) , bed roughness \(k_{b} = 2\pi d_{50} / 12\) , and \(d_{50}\) is the median grain size of sediment particles.  
+
+<|ref|>text<|/ref|><|det|>[[110, 540, 886, 666]]<|/det|>
+To illustrate the mechanism governing the spatial distribution of tidal flat slope and width, we set up 50 sets of randomized model experiments. Here, we set up each set of experiments by establishing a random SSC (20–150 mg/L) and tide range (2- 7 m) to obtain an equilibrium profile from a linear profile calculated over 100 years.  
+
+<|ref|>text<|/ref|><|det|>[[110, 714, 886, 909]]<|/det|>
+To reveal the nonlinear response of tidal flat width to SSC change, we selected two sites for explanatory modelling, namely the Western Scheldt site in the Netherlands (sediment- starving) \(^{9}\) and Jiangsu in China (sediment- rich) \(^{62}\) . At the Western Scheldt site (SSC=58 mg/L), the DET- ESTMORF model has been applied in previous studies \(^{9,14}\) . At the Jiangsu site, we obtained the equilibrium profile by performing a 100- year modeling from the initial linear profile based on the tide level, while wave height, and offshore SSC data (134 mg/L) are based on the collected
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 87, 888, 283]]<|/det|>
+datasets in the current study. We validated the model's accuracy by comparing the obtained equilibrium profile with the observed profile reported in the literature (Supplementary Fig. 13) \(^{62}\) . Building on this validation, we proceeded to calculate equilibrium profiles for both sites under varying interannual SSC change rates, including - 1.2%, - 0.8%, 0.8%, and 1.2%. This exercise was designed to ascertain the interannual rates of change in tidal flat width over a 13- year period, reflective of the observational timeframe spanning 2013 to 2015.  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 333, 533, 352]]<|/det|>
+## Analysis of temporal changes in tidal flats width  
+
+<|ref|>text<|/ref|><|det|>[[110, 365, 896, 595]]<|/det|>
+To reveal the conditional response of flat width to SSC changes, we extracted tidal width change rate of sediment- rich ( \(>75 \mathrm{mg / L}\) ) and sediment- starving ( \(< 50 \mathrm{mg / L}\) ) tidal flats across different SSC change rate groups (- 3% \(\sim\) - 1.5%/year, - 1.5% \(\sim\) 0%/year, 0% \(\sim\) 1.5%/year, 1.5% \(\sim\) 3%/year). Within each group, we choose median tidal width change (mean of the 48% - 52% percentiles) as the representative value (Fig 4a). To minimize the influence of anthropogenic factors on the geomorphology of tidal flats, data points exhibiting interannual tidal width trends exceeding 20% per year were omitted.  
+
+<|ref|>text<|/ref|><|det|>[[110, 644, 886, 910]]<|/det|>
+We then employed the DET- ESTMORF model to simulate the variation of tidal flat width under different sediment supplies (20\~120 mg/L) and varying interannual rates of change of sediment concentration (- 3% \(\sim\) 3%/year). Subsequently, we performed explanatory simulations for the 34 extracted sites as well (Fig. 5). For a consistent basis of comparison, we initialized the model with a wider profile for tidal flats with higher sediment supply, maintaining uniform parameters across all scenarios except for the initial SSC level and the SSC variation rates. For a given SSC level, we first obtained the equilibrium profile from the initial linear profile after 100 years of simulation, after which the change in the tidal flat morphology is negligible \(^{9}\) . Based on the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[57, 88, 884, 177]]<|/det|>
+500 obtained equilibrium profile, we simulated the tidal flat morphological variations with changing SSC over 13 years (2002 to 2016, as a reflection of the observation), and finally compared the simulated flat width changes resulting from various SSC changing rates.  
+
+<|ref|>sub_title<|/ref|><|det|>[[60, 211, 207, 228]]<|/det|>
+## 504 References  
+
+<|ref|>text<|/ref|><|det|>[[56, 240, 884, 911]]<|/det|>
+505 1. Wang, X. et al. Rebound in China's coastal wetlands following conservation and restoration. Nat Sustain 4, 1076-1083 (2021). 506 2. Wang, F. et al. Global blue carbon accumulation in tidal wetlands increases with climate change. Natl Sci Rev 8, nwaa296 (2020). 507 3. van Zelst, V. T. M. et al. Cutting the costs of coastal protection by integrating vegetation in flood defences. Nat Commun 12, 6533 (2021). 508 4. Zheng, J. et al. Synergy between coastal ecology and disaster mitigation in China: Policies, practices, and prospects. Ocean & Coastal Management 245, 106866 (2023). 509 5. Schuerch, M. et al. Future response of global coastal wetlands to sea-level rise. Nature 561, 231-234 (2018). 510 6. Jankowski, K. L., Törnqvist, T. E. & Fernandes, A. M. Vulnerability of Louisiana's coastal wetlands to present-day rates of relative sea-level rise. Nat Commun 8, 14792 (2017). 511 7. Appeaning Addo, K. & Manson Incoom, A. Impacts of shoreline morphological change and sea level rise on mangroves: the case of the keta coastal zone. E3 Journal of Environmental Research and Management 4, 0359-0367 (2013). 512 8. Schuerch, M., Spencer, T. & Evans, B. Coupling between tidal mudflats and salt marshes affects marsh morphology. Marine Geology 412, 95-106 (2019). 513 9. Hu, Z. et al. Mechanistic Modeling of Marsh Seedling Establishment Provides a Positive Outlook for Coastal Wetland Restoration Under Global Climate Change. Geophysical Research Letters 48, (2021). 514 10. Bouma, T. J. et al. Short-term mudflat dynamics drive long-term cyclic salt marsh dynamics. Limnology and Oceanography 61, 2261-2275 (2016). 515 11. Murray, N. J. et al. The global distribution and trajectory of tidal flats. Nature 565, 222-225 (2019). 516 12. Murray, N. J. et al. High-resolution mapping of losses and gains of Earth's tidal wetlands. Science 376, 744-749 (2022). 517 13. Roberts, W., Le Hir, P. & Whitehouse, R. J. S. Investigation using simple mathematical models of the effect of tidal currents and waves on the profile shape of intertidal mudflats. Continental Shelf Research 20, 1079-1097 (2000). 518 14. Hu, Z., Wang, Z. B., Zitman, T. J., Stive, M. J. F. & Bouma, T. J. Predicting long-term and short-term tidal flat morphodynamics using a dynamic equilibrium theory. Journal of Geophysical Research: Earth Surface 120, 1803-1823 (2015). 519 15. Liu, X. J., Gao, S. & Wang, Y. P. Modeling profile shape evolution for accreting tidal flats composed of mud and sand: A case study of the central Jiangsu coast, China. Continental Shelf Research 31, 1750-1760 (2011). 520 16. Fan, D., Guo, Y., Wang, P. & Shi, J. Z. Cross-shore variations in morphodynamic processes of an open-coast mudflat in the Changjiang Delta, China: With an emphasis on storm impacts. Continental Shelf Research 26, 517-538 (2006).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[58, 90, 888, 895]]<|/det|>
+17. Wang, X. et al. Tracking annual changes of coastal tidal flats in China during 1986–2016 through analyses of Landsat images with Google Earth Engine. Remote Sensing of Environment 238, 110987 (2020).18. Chen, Y. et al. Land claim and loss of tidal flats in the Yangtze Estuary. Sci Rep 6, 24018 (2016).19. Best, J. Anthropogenic stresses on the world’s big rivers. Nature Geosci 12, 7–21 (2019).20. Dethier, E. N., Renshaw, C. E. & Magilligan, F. J. Rapid changes to global river suspended sediment flux by humans. Science 376, 1447–1452 (2022).21. Zhou, Z. et al. Sediment concentration variations in the East China Seas over multiple timescales indicated by satellite observations. Journal of Marine Systems 212, 103430 (2020).22. Li, J., Pan, S., Chen, Y., Fan, Y.-M. & Pan, Y. Numerical estimation of extreme waves and surges over the northwest Pacific Ocean. Ocean Engineering 153, 225–241 (2018).23. Friedrichs, C. T. 3.06 - Tidal Flat Morphodynamics: A Synthesis. in Treatise on Estuarine and Coastal Science (eds. Wolanski, E. & McLusky, D.) 137–170 (Academic Press, Waltham, 2011). doi:10.1016/B978-0-12-374711-2.00307-7.24. Sala Calero, J. et al. FAST MI-SAFE PLATFORM: FORESHORE ASSESSMENT USING SPACE TECHNOLOGY. (2017). doi:10.5281/zenodo.1065940.25. Friedrichs, C. T. & Aubrey, D. G. Uniform Bottom Shear Stress and Equilibrium Hyposometry of Intertidal Flats. in Mixing in Estuaries and Coastal Seas 405–429 (American Geophysical Union (AGU), 1996). doi:10.1029/CE050p0405.26. Bi, S. et al. Distribution of heavy metals and environmental assessment of surface sediment of typical estuaries in eastern China. Marine Pollution Bulletin 121, 357–366 (2017).27. Ganju, N. K. et al. Spatially integrative metrics reveal hidden vulnerability of microtidal salt marshes. Nat Commun 8, 14156 (2017).28. Nittrouer, J. A. et al. Mitigating land loss in coastal Louisiana by controlled diversion of Mississippi River sand. Nature Geosci 5, 534–537 (2012).29. Baptist, M. J. et al. Beneficial use of dredged sediment to enhance salt marsh development by applying a ‘Mud Motor’. Ecological Engineering 127, 312–323 (2019).30. Glover, H. E. et al. Impacts of Suspended Sediment on Nearshore Benthic Light Availability Following Dam Removal in a Small Mountainous River: In Situ Observations and Statistical Modeling. Estuaries and Coasts 42, 1804–1820 (2019).31. Cabaço, S., Santos, R. & Duarte, C. M. The impact of sediment burial and erosion on seagrasses: A review. Estuarine, Coastal and Shelf Science 79, 354–366 (2008).32. McClenachan, L., O’Connor, G., Neal, B. P., Pandolfi, J. M. & Jackson, J. B. C. Ghost reefs: Nautical charts document large spatial scale of coral reef loss over 240 years. Science Advances 3, e1603155 (2017).33. Cramer, K. L. et al. Widespread loss of Caribbean acroporid corals was underway before coral bleaching and disease outbreaks. Science Advances 6, eaax9395 (2020).34. Marani, M., D’Alpaos, A., Lanzoni, S., Carniello, L. & Rinaldo, A. The importance of being coupled: Stable states and catastrophic shifts in tidal biomorphodynamics. Journal of Geophysical Research: Earth Surface 115, (2010).35. Kirwan, M. L. et al. Limits on the adaptability of coastal marshes to rising sea level. Geophysical Research Letters 37, (2010).36. Kirwan, M. L., Temmerman, S., Skeehan, E. E., Guntenspergen, G. R. & Fagherazzi, S. Overestimation of marsh vulnerability to sea level rise. Nature Clim Change 6, 253–260 (2016).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[58, 90, 888, 911]]<|/det|>
+37. Mariotti, G. & Carr, J. Dual role of salt marsh retreat: Long-term loss and short-term resilience. Water Resources Research 50, 2963–2974 (2014).38. Athanasiou, P. et al. Global distribution of nearshore slopes with implications for coastal retreat. Earth System Science Data 11, 1515–1529 (2019).39. Murray, N. J. et al. High-resolution global maps of tidal flat ecosystems from 1984 to 2019. Sci Data 9, 542 (2022).40. Egbert, G. D. & Erofeeva, S. Y. Efficient Inverse Modeling of Barotropic Ocean Tides. Journal of Atmospheric and Oceanic Technology 19, 183–204 (2002).41. Hersbach, H. et al. The ERA5 global reanalysis. Q.J.R. Meteorol. Soc. 146, 1999–2049 (2020).42. Passaro, M. et al. Global coastal attenuation of wind-waves observed with radar altimetry. Nat Commun 12, 3812 (2021).43. Maan, D. C., van Prooijen, B. C., Wang, Z. B. & De Vriend, H. J. Do intertidal flats ever reach equilibrium? Journal of Geophysical Research: Earth Surface 120, 2406–2436 (2015).44. Le Hir, P. et al. Characterization of intertidal flat hydrodynamics. Continental Shelf Research 20, 1433–1459 (2000).45. Xie, W. et al. Impacts of a storm on the erosion process of a tidal wetland in the Yellow River Delta. CATENA 205, 105461 (2021).46. Yang, S. L., Shi, B. W., Bouma, T. J., Ysebaert, T. & Luo, X. X. Wave Attenuation at a Salt Marsh Margin: A Case Study of an Exposed Coast on the Yangtze Estuary. Estuaries and Coasts 35, 169–182 (2012).47. Yang, S. L. The Role ofScirpusMarsh in Attenuation of Hydrodynamics and Retention of Fine Sediment in the Yangtze Estuary. Estuarine, Coastal and Shelf Science 47, 227–233 (1998).48. Cao, H., Chen, Y., Tian, Y. & Feng, W. Field Investigation into Wave Attenuation in the Mangrove Environment of the South China Sea Coast. coas 32, 1417–1427 (2016).49. Sheng, Y. P., Paramygin, V. A., Terng, C.-T. & Chu, C.-H. Simulating Storm Surge and Inundation Along the Taiwan Coast During Typhoons Fanapi in 2010 and Soulik in 2013. Terr. Atmos. Ocean. Sci. 27, 965–979 (2016).50. Zhao, Y. et al. Rapid formation of marsh-edge cliffs, Jiangsu coast, China. Marine Geology 385, 260–273 (2017).51. Doerffer, R. & Schiller, H. The MERIS Case 2 water algorithm. International Journal of Remote Sensing 28, 517–535 (2007).52. Wei, J. et al. Global Estimation of Suspended Particulate Matter From Satellite Ocean Color Imagery. JGR Oceans 126, (2021).53. Zhan, W., Wu, J., Wei, X., Tang, S. & Zhan, H. Spatio-temporal variation of the suspended sediment concentration in the Pearl River Estuary observed by MODIS during 2003–2015. Continental Shelf Research 172, 22–32 (2019).54. Feng, L., Hu, C., Chen, X. & Song, Q. Influence of the Three Gorges Dam on total suspended matters in the Yangtze Estuary and its adjacent coastal waters: Observations from MODIS. Remote Sensing of Environment 140, 779–788 (2014).55. Nakagawa, S. & Schielzeth, H. A general and simple method for obtaining R2 from generalized linear mixed-effects models. Methods in Ecology and Evolution 4, 133–142 (2013).56. Ladd, C. J. T., Duggan-Edwards, M. F., Bouma, T. J., Pages, J. F. & Skov, M. W. Sediment Supply Explains Long-Term and Large-Scale Patterns in Salt Marsh Lateral Expansion and Erosion. Geophysical Research Letters 46, 11178–11187 (2019).57. Amoudry, L. O. & Souza, A. J. Deterministic Coastal Morphological and Sediment Transport Modeling: A Review and Discussion. Reviews of Geophysics 49, (2011).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 88, 881, 125]]<|/det|>
+58. Winterwerp, J. C. & Van Kesteren, W. G. M. Settling and Sedimentation. in Developments in Sedimentology vol. 56 121-159 (Elsevier, 2004).  
+
+<|ref|>text<|/ref|><|det|>[[110, 124, 857, 160]]<|/det|>
+59. Zhong, Z. & Hu, Z. The Impact of Reclamation on Tidal Flat Morphological Equilibrium. Front. Mar. Sci. 8, 769077 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[110, 160, 872, 195]]<|/det|>
+60. Christoffersen, J. B. & Jonsson, I. G. Bed friction and dissipation in a combined current and wave motion. Ocean Engineering 12, 387-423 (1985).  
+
+<|ref|>text<|/ref|><|det|>[[110, 195, 880, 248]]<|/det|>
+61. Callaghan, D. P. et al. Hydrodynamic forcing on salt-marsh development: Distinguishing the relative importance of waves and tidal flows. Estuarine, Coastal and Shelf Science 89, 73-88 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[110, 247, 839, 283]]<|/det|>
+62. Wang, Y. et al. Sand-Mud Tidal Flat Morphodynamics Influenced by Alongshore Tidal Currents. Journal of Geophysical Research: Oceans 124, 3818-3836 (2019).  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 298, 270, 315]]<|/det|>
+## Acknowledgments  
+
+<|ref|>text<|/ref|><|det|>[[110, 323, 884, 425]]<|/det|>
+The authors would like to thank the members of M5 (Mudflat, Marsh, Mangrove, Measurement & Modeling) Lab at Sun Yat-Sen University for their assistance in the research progress. The scientific results and conclusions, as well as any views or opinions expressed herein, are those of the author(s) and do not necessarily reflect those of NOAA or the Department of Commerce.  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 459, 193, 477]]<|/det|>
+## Funding:  
+
+<|ref|>text<|/ref|><|det|>[[113, 485, 344, 503]]<|/det|>
+This work was supported by:  
+
+<|ref|>text<|/ref|><|det|>[[110, 510, 794, 530]]<|/det|>
+The National Natural Science Foundation of China under contract No. 42176202 (ZH)  
+
+<|ref|>text<|/ref|><|det|>[[110, 536, 864, 556]]<|/det|>
+Innovation Group Project of Southern Marine Science and Engineering Guangdong Laboratory  
+
+<|ref|>text<|/ref|><|det|>[[113, 562, 473, 580]]<|/det|>
+(Zhuhai) under contract No. 311021004 (ZH)  
+
+<|ref|>text<|/ref|><|det|>[[110, 587, 884, 632]]<|/det|>
+Guangdong Provincial Department of Science and Technology under contract No. 2019ZT08G090 (ZH)  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 666, 300, 684]]<|/det|>
+## Author contributions:  
+
+<|ref|>text<|/ref|><|det|>[[110, 692, 660, 840]]<|/det|>
+Conceptualization: SL, ZH  Methodology: SL, ZH, VVZ, LQ  Data analysis and plotting: SL, ZH  Supervision: ZH  Writing—original draft: SL, ZH  Writing—review & editing: ZH, TG, ZBW, VVZ, LQ, TX, JYS, TJB  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 876, 289, 894]]<|/det|>
+## Competing interests:
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[60, 88, 572, 108]]<|/det|>
+The authors declare that they have no competing interests.  
+
+<|ref|>text<|/ref|><|det|>[[60, 115, 88, 131]]<|/det|>
+672  
+
+<|ref|>text<|/ref|><|det|>[[60, 140, 884, 261]]<|/det|>
+Data availability: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Information. Global tidal flat maps are freely available at http://intertidal.app. The TPXO9 dataset is available at https://www.tpxo.net/global/tpxo9- atlas. Wave height data is available from the ERA5 reanalysis dataset. SSC data is available at https://www.globcolour.info.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 43, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[43, 93, 768, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[59, 131, 460, 177]]<|/det|>
+Chinatidalflat.csv supplementarymaterialsChinatidalflat.docx
+
+<--- Page Split --->

preprint/preprint__2b4640cbbd24fef518b88dcfeb69d338b82c53d043af3c97eb48d85e522c80ce/images_list.json

@@ -0,0 +1,130 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "Keywords: Dynamic knowledge graph, laboratory automation, digital twin, distributed laboratory, multi-agent system, goal-driven self-optimisation",
+    "footnote": [],
+    "bbox": [
+      [
+        140,
+        90,
+        808,
+        360
+      ]
+    ],
+    "page_idx": 3
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1: An overview of intertwined aspects of a chemical research laboratory that need to be represented by a connected lab digital twin, adapted from [34]. This paper focuses on the automation of chemical reaction optimisation, a task that can be viewed as part of the daily work of a research scientist.",
+    "footnote": [],
+    "bbox": [
+      [
+        330,
+        263,
+        664,
+        491
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2: An illustration of a distributed SDLs architecture.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3: A selection of concepts and relationships capturing different aspects in SDLs. Their namespaces correspond to the colour coding. For complete knowledge representation and namespace definitions see Supplementary Information.",
+    "footnote": [],
+    "bbox": [
+      [
+        167,
+        220,
+        832,
+        692
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4: A snapshot of reaction views from different perspectives. The knowledge graph representation puts chemical informatics into context, allowing for queries and answers from different layers of abstraction. The colour coding corresponds to the ontological expression.",
+    "footnote": [],
+    "bbox": [
+      [
+        175,
+        160,
+        822,
+        730
+      ]
+    ],
+    "page_idx": 9
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5: Autonomous workflow triggered in response to goal requests from scientists as information travelling within the knowledge graph.",
+    "footnote": [],
+    "bbox": [
+      [
+        170,
+        199,
+        866,
+        780
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "(a) Pareto front plot of the yield and cost objectives for the aldol condensation reaction collaboratively optimised by two distributed SDLs.",
+    "footnote": [],
+    "bbox": [
+      [
+        330,
+        167,
+        660,
+        401
+      ]
+    ],
+    "page_idx": 13
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "(b) Three-dimensional plot of the four sampled design variables colour coded for run material cost during the closed-loop optimisation. The size of the dots denotes the molar equivalents of 3 in each run.",
+    "footnote": [],
+    "bbox": [
+      [
+        186,
+        460,
+        480,
+        670
+      ]
+    ],
+    "page_idx": 15
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_3.jpg",
+    "caption": "(c) Three-dimensional plot of the four sampled design variables colour coded for yield during the closed-loop optimisation. The size of the dots denotes the molar equivalents of 3 in each run.",
+    "footnote": [],
+    "bbox": [
+      [
+        515,
+        460,
+        811,
+        670
+      ]
+    ],
+    "page_idx": 15
+  }
+]

preprint/preprint__2b4640cbbd24fef518b88dcfeb69d338b82c53d043af3c97eb48d85e522c80ce/preprint__2b4640cbbd24fef518b88dcfeb69d338b82c53d043af3c97eb48d85e522c80ce.mmd

@@ -0,0 +1,343 @@
+
+# From Platform to Knowledge Graph: Distributed Self-Driving Laboratories  
+
+Markus Kraft (mk306@cam.ac.uk) University of Cambridge https://orcid.org/0000- 0002- 4293- 8924  
+
+Jiaru Bai Sebastian Mosbach University of Cambridge https://orcid.org/0000- 0001- 7018- 9433  
+
+Connor Taylor Astex Pharmaceuticals  
+
+Dogancan Karan Cambridge Centre for Advanced, Research and Education in Singapore  
+
+Kok Foong Lee Cambridge Centre for Advanced, Research and Education in Singapore  
+
+Simon Rihm Cambridge Centre for Advanced, Research and Education in Singapore  
+
+Jethro Akroyd University of Cambridge  
+
+Alexei Lapkin University of Cambridge  
+
+## Article  
+
+Keywords: Dynamic knowledge graph, laboratory automation, digital twin, distributed laboratory, multiagent system, goal- driven self- optimisation  
+
+Posted Date: July 25th, 2023  
+
+DOI: https://doi.org/10.21203/rs.3. rs- 3141873/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on January 23rd, 2024. See the published version at https://doi.org/10.1038/s41467- 023- 44599- 9.
+
+<--- Page Split --->
+
+# 1. 1. 1. 1. 1. 1. 1. 2. 2. 2. 2. 2. 2. 2.
+
+<--- Page Split --->
+
+# From Platform to Knowledge Graph: Distributed Self-Driving Laboratories  
+
+Jiaru Bai \(^{1}\) , Sebastian Mosbach \(^{1,2}\) , Connor J. Taylor \(^{3,4}\) , Dogancan Karan \(^{2}\) , Kok Foong Lee \(^{5}\) , Simon D. Rihm \(^{1,2}\) , Jethro Akroyd \(^{1,2}\) , Alexei A. Lapkin \(^{1,2,4}\) ,  
+
+Markus Kraft \(^{1,2,6,7,*}\)  
+
+\*Corresponding author: mk306@cam.ac.uk  
+
+\(^{1}\) Department of Chemical Engineering and Biotechnology, University of Cambridge, Philippa Fawcett Drive, Cambridge CB3 0AS, United Kingdom \(^{2}\) CARES, Cambridge Centre for Advanced, Research and Education in Singapore, 1 Create Way, CREATE Tower, #05- 05, 138602 Singapore \(^{3}\) Astex Pharmaceuticals, 436 Cambridge Science Park Milton Road, Cambridge CB4 0QA, United Kingdom \(^{4}\) Innovation Centre in Digital Molecular Technologies, Yusuf Hamied Department of Chemistry, University of Cambridge, Lensfield Road, Cambridge CB2 1EW, United Kingdom \(^{5}\) CMCL Innovations, Sheraton House, Cambridge CB3 0AX, United Kingdom \(^{6}\) School of Chemical and Biomedical Engineering, Nanyang Technological University, 62 Nanyang Drive, 637459 Singapore \(^{7}\) The Alan Turing Institute, London NW1 2DB, United Kingdom  
+
+## Abstract  
+
+The ability to integrate resources and share knowledge across organisations empowers scientists to expedite the scientific discovery process. This is especially crucial in addressing emerging global challenges that require global solutions. In this work, we develop an architecture to enable distributed self- driving laboratories as part of The World Avatar project, which seeks to demonstrate how to create an all- encompassing digital twin based on a dynamic knowledge graph. Our approach utilises ontologies to capture the data and material flows involved in design- make- test- analyse cycles, and employs autonomous agents as executable knowledge components to carry out the experimentation workflow. All data provenance is recorded following the FAIR principles, ensuring its accessibility and interoperability. We demonstrate the practical application of our framework by linking two robots in Cambridge and Singapore to achieve a collaborative closed- loop optimisation for a pharmaceutically- relevant aldol condensation reaction in real time. The knowledge graph evolves autonomously while progressing towards the research goals set by the scientist. The two robots effectively produced a Pareto front for the cost- yield optimisation problem over the course of three days of operation.
+
+<--- Page Split --->
+![](images/Figure_unknown_0.jpg)
+
+<center>Keywords: Dynamic knowledge graph, laboratory automation, digital twin, distributed laboratory, multi-agent system, goal-driven self-optimisation </center>
+
+<--- Page Split --->
+
+## 1 Introduction  
+
+2 The concept of laboratory automation, recently reinterpreted as self- driving laboratories (SDLs) [1, 2], has been in existence since the 1960s, when Merrifield et al. [3] introduced the first automated chemistry hardware. Since then, SDLs have gained widespread adoption in chemistry [4- 7], materials science [8, 9], biotechnology [10, 11] and robotics [12], resulting in accelerated scientific discovery and societal development. However, the implementation of SDLs can be challenging and typically requires a highly specialised team of researchers with expertise in chemistry, engineering, and computer science. Consequently, studies are often conducted by large research groups within a single organisation. Even in cases where collaborations occur between research groups, the SDL is usually centralised within the same laboratory.  
+
+In response to the pressing global challenges of today, there is a growing consensus within the scientific community that a paradigm shift towards a globally collaborative research network is necessary [13- 15]. This shift requires decentralising SDLs to integrate different research groups to contribute their expertise towards solving emerging problems [16]. Such decentralisation also holds great potential in supporting human exploration in deep space [17]. Achieving this vision is not an easy task and entails three major challenges. The first challenge is efficiently orchestrating heterogeneous resources [18], which includes hardware from different vendors and diverse computing environments. The second challenge is sharing data across organisations [19], which requires standardising language in which the research is communicated [20]. During this process, the source and metadata of the research need to be tracked to facilitate reproducibility, which leads to the third challenge of data provenance recording following FAIR principles - Findable, Accessible, Interoperable and Reusable [21].  
+
+Many attempts have been made to tackle each of these challenges separately. For resource orchestration, middleware such as ChemOS [22], ESCALATE [23], and HELAO [24] exist to glue different components within an SDL and abstract the hardware resources. For data sharing, XDL [25, 26] and AnIML [27] are examples of standard protocols developed for synthesis and analysis respectively. For data provenance, Mitchell et al. [28] proposed a data pipeline to support the modelling of the COVID pandemic. Although these studies provide insights into building a collaborative research environment, they are developed in isolation with customised data interfaces. Enhancing interoperability between these systems is essential to establish a truly connected research network.  
+
+As discussed in our previous work [29, 30], semantic web technologies such as knowledge graphs [31] offer a viable path forward. Ontologies abstract both resources and data using the same notion, allowing for a common language between participants when allocating tasks and sharing results. The World Avatar [32, 33] is such a knowledge graph that aims to encompass all aspects of scientific research laboratories as shown in Fig. 1 in their entirety: The experiment itself, including its physical setup and underlying chemistry; moving handlers that can be of human or robotic nature; and the laboratory providing necessary infrastructure and resources [34]. The World Avatar goes beyond static knowledge representation by encoding software agents as executable knowledge components, enabling dynamicity and continuous incorporation of new concepts and data while preserving connections to existing information. As the knowledge graph expands, this characteristic allows for capturing data provenance from experimental processes as knowledge statements, effectively acting as a
+
+<--- Page Split --->
+
+1 living copy of the real world. By facilitating immediate dissemination of data between SDLs at its creation, the dynamic knowledge graph presents a promising holistic solution to the challenges aforementioned [30, 35] and supports the realisation of "AI Scientists" [34, 36]. 4 The purpose of this paper is to demonstrate a proof- of- concept for a distributed network of SDLs enabled by a dynamic knowledge graph. This signifies the first step towards digital research scientists (as shown in Fig. 1) collaborating autonomously. To illustrate the effectiveness of this approach, we present a demonstration using two robots in Cambridge and Singapore collaborating on a multi- objective closed- loop optimisation problem in response to a goal request from scientists.  
+
+![](images/Figure_1.jpg)
+
+<center>Figure 1: An overview of intertwined aspects of a chemical research laboratory that need to be represented by a connected lab digital twin, adapted from [34]. This paper focuses on the automation of chemical reaction optimisation, a task that can be viewed as part of the daily work of a research scientist. </center>  
+
+## 2 Results  
+
+### 2.1 Architecture of distributed SDLs  
+
+Closed- loop optimisation in SDLs is a dynamic process that revolves around design- make- test- analyse (DMTA) cycles [37, 38]. Compared to machine learning systems and scientific workflows that only capture data flows, SDLs offer an integrated approach by orchestrating both computational and physical resources. This involves the integration of data and material flows, as well as the interface that bridges the gap between the virtual and physical worlds. To this end, we propose a conceptual architecture of distributed SDLs that effectively incorporates all three flows, as illustrated in Fig. 2(a).  
+
+The proposed architecture presents a framework to enable scientists to set research goals and resource restrictions for a particular chemical reaction and have them trigger a closed- loop process in cyberspace. The process is initiated by the monitoring component, which parses the research goals and requests the iterations needed to achieve the objectives. The
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+  
+
+(a) Conceptual framework of components used to build a network of distributed SDLs for closed-loop optimisation. Information and materials from a chemical reaction flow autonomously across cyber and physical space until they accomplish the research goals set by scientists or the allocated resources have been consumed.  
+
+![](images/Figure_3.jpg)
+
+<center>Figure 2: An illustration of a distributed SDLs architecture. </center>  
+
+(b) Dynamic knowledge graph approach. The overall system comprises three layers: the real world where the hardware is located, the dynamic knowledge graph that hosts all information in cyberspace, and the layer of active agents that manages the knowledge graph.
+
+<--- Page Split --->
+
+iterating component collects prior information about the design space and passes it on to the component that designs the next experiment. The algorithm employed, as well as the availability of prior data, determines the combination of design variables to be proposed within the domain provided by the scientist. Similar to the scheduling of high- performance computing (HPC) jobs [39], the proposed physical experimentation is scheduled for execution in one of the available laboratories. The suggested conditions are translated to the machine- actionable recipe that enables the control of hardware for reaction and characterisation. In the physical world, this is reflected in the material flow between the two pieces of equipment. The data processing component is then responsible for computing the objectives by analysing the complete job information and raw data. If the resources are still available, a comparison of these objectives with the research goals determines whether the system should proceed to the next iteration.  
+
+This architecture liberates the scientists from routine work, however, it also poses challenges in the implementation in terms of ensuring robustness, scalability, maintainability, safety, and ethics. Ideally, the system should enable seamless integration of new devices, resources, and algorithms without disrupting the system's overall functioning. It is also critical to allow for dynamic adaption to changes in research goals and resource restrictions.  
+
+We believe dynamic knowledge graph technology can help with realising this design [30]. Specifically, as illustrated in Fig. 2(b), this technology represents the software components as agents that receive inputs and produce outputs, while the data flow between components is marked up as the agents' messages. Physical entities can be virtualised as digital twins in cyberspace, enabling real- time control and eliminating geospatial boundaries when multiple labs are involved. This reformulation of the closed- loop optimisation problem as information travelling through the knowledge graph and reflecting their changes in the real world offers a powerful framework for achieving true distributed SDLs. In this way, we can think of an occurrence of physical experimentation as a sequence of actions that dynamically generates information about a reaction experiment as it progresses in time, analogous to computational workflows [40].  
+
+This work is part of a series of papers introducing a holistic approach to lab automation by including all aspects of research laboratories (see Fig. 1) in an all- encompassing digital twin [34]. By employing dynamic knowledge graphs embedded with knowledge models across different domains, we can address the challenges related to interoperability and adaptability encountered in platform- based approaches [30]. The goal- driven top- down architecture enables reasoning across the knowledge base and ensures alignment. Consequently, only high- level, abstract goals need to be defined by humans, which are subsequently decomposed into concrete sub- goals and tasks. In this framework, humans play a dual role, acting both as handlers to execute and intervene in experiments. They can be represented in the knowledge graph in a similar manner to robots and are provided with instructions in a human- readable format. This enables the realisation of a hybrid and evolving digital laboratory, bridging potential "interim technology gaps" [41]. It is important to note that the work described in this paper relies fully on robotic handling only.
+
+<--- Page Split --->
+
+### 2.2 Chemical ontologies and digital twins  
+
+2.2 Chemical ontologies and digital twins2 The realisation of SDLs requires a connection between abstract chemistry knowledge and concrete hardware for execution [20]. This calls for a set of connected ontologies, as identified in our previous analysis on the gaps in current semantic representations for chemical digitalisation [30]. Figure 3 presents a selection of concepts and relationships as an effort to address these gaps. These concepts span various abstraction levels, ranging from the high- level research goals to the conceptual level of chemical reactions to the mathematical level of design of experiments to the physical realisation of reaction experiments and the laboratory digital twin. We describe below ontologies' cross- domain characteristics, for technical details on each ontology please see Supplementary Information section A.1.  
+
+For closed- loop optimisation in SDLs, we draw parallels between the pursuit of optimal objectives and the reasoning cycles involved in pursuing a goal [42, 43]. The multi- objective problem can be formulated as a GoalSet which comprises individual Goals. Each goal is associated with specific dimensional quantities that can be achieved by a Plan, which consists of multiple Steps to be carried out by corresponding agents. From the implementation perspective, this is akin to research in the scientific workflow community which seeks to manage workflow expressed as a directed cyclic graph (DCG) [44]. In this regard, we adopt the derived information framework [40], a knowledge- graph- native approach, to manage the iterative workflow.  
+
+In developing chemical ontologies for SDLs, we draw upon the lessons learnt in creating ontologies for chemical plants. One prominent example is the OntoCAPE material and chemical process system [45] ontology, which describes materials from three aspects: the ChemicalSpecies that reflects the intrinsic characteristics, Material as part of the phase system which describes macroscopic thermodynamic behaviour, and MaterialAmount that refers to a concrete occurrence of an amount of matter in the physical world. Building on this foundation, we introduce OntoReaction, an ontology that captures knowledge in wet- lab reaction experiments, and OntoDoE, an ontology for the design of experiments (DoE) in optimisation campaigns. As an effort to align with existing data, OntoReaction draws inspiration from established schemas used in chemical reaction databases like ORD [46] and UDM [47]. ReactionExperiment is a concrete realisation of a ChemicalReaction that is sampled at a set of ReactionConditions and measures certain PerformanceIndicators. When grouped together, they can form HistoricalData that are utilised by a DesignOfExperiment study to propose new experiments.  
+
+In the development of our hardware ontologies, we have utilised the Smart Applications REFerence (SAREF) ontology [48] wherever feasible, which is widely adopted in the field of the Internet of Things (IoT). We introduce OntoLab to represent the digital twin of a laboratory, comprising a group of LabEquipment and ChemicalContainers that contain ChemicalAmount. Furthermore, we create OntoVapourtec and OntoHPLC as ontologies for the equipment involved in this work, linking them to the concrete realisation aspect of OntoCAPE. We establish the link between abstract chemical knowledge and hardware by translating ReactionCondition to ParameterSetting, which can be combined to form EquipmentSettings for configuration.
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Figure 3: A selection of concepts and relationships capturing different aspects in SDLs. Their namespaces correspond to the colour coding. For complete knowledge representation and namespace definitions see Supplementary Information. </center>
+
+<--- Page Split --->
+
+### 2.3 Contextualised reaction informatics  
+
+2.3 Contextualised reaction informatics2 By utilising ontologies as blueprints, it becomes possible to instantiate reaction information while preserving connections with contextual recordings. The reaction we choose for demonstration is an aldol condensation reaction between benzaldehyde 1 (bold numbers for reference) and acetone 2, catalysed by sodium hydroxide 3 to yield the target product benzylideneacetone 4 [49]. There are also the reported side products dibenzylideneacetone 5 and further condensation products from acetone polymerisation. The target product is pharmaceutically relevant and can be used as an NK- 1 receptor inhibitor for the treatment of idiopathic vomiting [50].  
+
+Figure 4 provides an illustrative representation of the chosen reaction in the knowledge graph as viewed through various roles within a laboratory, each with its unique perspective on the same chemical. Taking the starting material benzaldehyde as an example, a chemist, more interested in conceptual description, focuses on its role as a reactant and seeks relevant species information. A data scientist, on the other hand, looks into its concentration to determine the appropriate usage of other chemicals when designing conditions for a particular reaction experiment. Meanwhile, a lab manager inspects via the 3D digital twin and ensures the availability of an internal standard that can be mixed with the physical existence of benzaldehyde to enable characterisation during the actual execution of the experiment. Lastly, the knowledge engineer seeks to represent and access this information in a machine- readable format following the proposed ontologies. In practice, the same individual might play several roles, and the emphasis here is on the cross- domain interoperability enabled by the knowledge graph.  
+
+The integration of chemical knowledge from PubChem, represented by OntoSpecies for unique species identification [51], serves as a critical link between these facets of chemicals. It enables the identification of potential input chemicals based on the reactant and solvent during DoE and allows for the selection of appropriate sources of starting materials from multiple chemical containers (see Supplementary Information section A.2). By combining various aspects into a unified representation, the knowledge graph encompasses information that is pertinent to diverse users while keeping humans in the loop [34]. For concrete examples of ontology instantiation see Supplementary Information section A.1.
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Figure 4: A snapshot of reaction views from different perspectives. The knowledge graph representation puts chemical informatics into context, allowing for queries and answers from different layers of abstraction. The colour coding corresponds to the ontological expression. </center>
+
+<--- Page Split --->
+
+### 2.4 Goal-driven knowledge dynamics  
+
+2 Figure 5 presents a high- level overview of the goal- driven self- evolution of the knowledge graph during closed- loop optimisation. The dynamicity of the knowledge graph is enabled by the presence of software agents that realise each component of the distributed architecture and facilitate the flow of information within the graph. The process begins with goal derivation when the scientist initiates a goal request. The Reaction Optimisation Goal (ROG) Agent is responsible for translating the goal request into a machine- readable statement that captures the scientist's intention. To accommodate all objectives, a goal set is formulated that considers each objective as a reward function for the agents' operations. For each participating laboratory, a Goal Iteration Derivation instance is created using the derived information framework [40] and requested for execution by the Reaction Optimisation Goal Iteration (ROGI) Agent.  
+
+13 The goal iteration stage plays a central role in the self- evolution of the dynamic knowledge graph. It involves the ROGI Agent initiating the flow of information among the participating agents towards achieving the goal set. This process begins with ROGI Agent creating tasks for the corresponding agents according to the DMTTA cycle, including the DoE Agent, Schedule Agent, and Post- Processing Agent. The DoE Agent perceives the knowledge graph to retrieve prior data and chemical stock available for experiments and then proposes a new experiment. The Schedule Agent evaluates the hardware available in the specified laboratory according to the proposed conditions and subsequently selects the most appropriate hardware to execute the experiment. This is accomplished by generating tasks for the agents responsible for managing the selected digital twin. These agents actuate the equipment to perform reaction and characterisation in the physical world. When the HPLC report is generated, the Post- Processing Agent analyses the chromatogram data to calculate the objectives.  
+
+25 During the third stage, the ROG Agent utilises the obtained results to determine whether the next iteration should be pursued. To do so, it checks if the Pareto front of the multi- objective fulfils the pre- defined goals and if the resources are still available. The reaction experiment performed in the current iteration is then considered historical data to be included as input for the succeeding round of the Goal Iteration Derivation across all participating SDLs. Afterwards, a new request will be made to the ROGI Agent to start a new iteration, forming a self- evolving feedback loop.  
+
+32 To ensure correct data dependencies and task execution, we employed the derived information framework [40] to manage the iterative workflow. We implemented each software agent using the derivation agent template provided by the framework. Once deployed, these agents autonomously update the knowledge graph to actively reflect and influence the state of the world.  
+
+37 This approach enables flexibility and extensibility in the system. As the digital twin of each lab is represented as a node in the knowledge graph, new hardware can be added or removed during the optimisation campaign by simply modifying the list of participating laboratories. The experimental allowance can also be updated when more chemicals become available. The system also supports data sharing across organisations at the very moment the data are generated. Details on the internal logic and technical aspects of the agents in the knowledge graph implementation are available in the Supplementary Information section A.2.
+
+<--- Page Split --->
+![](images/Figure_unknown_1.jpg)
+
+<center>Figure 5: Autonomous workflow triggered in response to goal requests from scientists as information travelling within the knowledge graph. </center>
+
+<--- Page Split --->
+
+## 2.5 Collaborative closed-loop optimisation  
+
+2 To demonstrate the scalability and modularity, the knowledge graph approach was applied to a real- time collaborative closed- loop optimisation distributed over two SDLs in Cambridge and Singapore. The objectives selected are run material cost and yield that were sampled for a search space of molar equivalents (relative to benzaldehyde 1) of acetone 2, NaOH 3, residence time and reaction temperature. The research goals and restrictions were populated in the knowledge graph via a web front end. As no prior experimental data was provided, the agents start experiments with random conditions and gradually update their beliefs using TSEMO algorithm [52]. Before running the optimisation, two control conditions were performed to validate the reproducibility across the two setups. For experimental details see Supplementary Information section A.3.  
+
+Figure 6(a) presents the cost- yield objectives consisting of 65 data points collected during the self- optimisation. Throughout the operation, two SDLs share the results with each other when proposing new experimental conditions. The real- time collaboration demonstrated faster advances in the Pareto front with the highest yield of \(93\%\) . The chemicals used in this study were obtained from different vendors compared to Jeraal et al. [49], the cost is therefore not directly comparable due to different prices. Although not considered in the optimisation, the environment factor and space- time yield were found to be highly correlated to the yield objective. The best values obtained are 26.17 and \(258.175 \mathrm{g L}^{- 1} \mathrm{h}^{- 1}\) when scaled to the same benzaldehyde injection volume (5 mL), both outperformed the previous study [49].  
+
+Figures 6(b) and 6(c) illustrate the influence of the continuous variables on the cost and yield objectives respectively. The cost is calculated to count for the molar amount of input chemicals sourced from the pumps for the reaction. Therefore, it increases linearly with the molar equivalents of the starting materials. Similarly as identified by Jeraal et al. [49], reaction temperature has a positive correlation with the yield of reaction, whereas the residence time shows a poor correlation. Upon examination of the molar equivalent of acetone 2, it can be observed that its further increase after 30 results in a reduction in yield. This decrease can be attributed to the formation of more side product 5 and other further condensation products of acetone and benzaldehyde.  
+
+Notably, the Singapore setup encountered an HPLC failure after running for approximately 10 hours. This caused peak shifting of the internal standard which resulted in a wrongly identified peak that gives more than \(3500\%\) yield. This point is considered abnormal by the agents and therefore not utilised in the following DoE. An email notification was sent to the developer for maintenance which took the hardware out of the campaign. The asynchronous and distributed design enabled the Cambridge side to further advance the Pareto front for the cost- yield trade- offs. It is also notable that the product peak was missed for one run at the Cambridge side due to a small shift of the peak which gives a yield of \(0\%\) . This point was taken into consideration in the DoE, but fortunately, it did not affect the final Pareto front as the corrected yield is still Pareto- dominated. The optimisation campaign was stopped since no more significant improvement was observed in terms of hypervolume, and also due to requests for repurposing the equipment for other projects. The complete provenance records (knowledge graph triples) are provided as Supplementary Data. An interactive animation of the optimisation progress is provided as a Supplementary Video.
+
+<--- Page Split --->
+![](images/Figure_unknown_2.jpg)
+
+<center>(a) Pareto front plot of the yield and cost objectives for the aldol condensation reaction collaboratively optimised by two distributed SDLs. </center>  
+
+![](images/Figure_unknown_3.jpg)
+
+<center>(b) Three-dimensional plot of the four sampled design variables colour coded for run material cost during the closed-loop optimisation. The size of the dots denotes the molar equivalents of 3 in each run. </center>  
+
+![PLACEHOLDER_15_2]
+
+<center>(c) Three-dimensional plot of the four sampled design variables colour coded for yield during the closed-loop optimisation. The size of the dots denotes the molar equivalents of 3 in each run. </center>  
+
+Figure 6: Plots for objectives and design variables of experiments conducted in the distributed closed-loop optimisation campaign. Each dot refers to a single run. The animation of the optimisation progress is available as Supplementary Video.
+
+<--- Page Split --->
+
+## 3 Discussion  
+
+2 In this contribution, we presented a dynamic knowledge graph approach to realise a conceptual architecture for distributed SDLs. We developed ontologies to represent various aspects of chemical knowledge and hardware digital twins involved in a closed- loop optimisation campaign. By employing autonomous agents as executable knowledge components to update and restructure the knowledge graph, we have enabled collaborative management of data and material flow across SDLs. Our approach allows scientists to initiate the autonomous workflow by setting up a goal request, which triggers the flow of information through the knowledge graph as the experimentation workflow progresses. Compared to contemporary designs, where a central orchestrator is responsible for all the data transfer, our approach emphasises information flows in the data layer, reducing the traffic stress on a single point in the network.  
+
+13 As a proof- of- concept demonstration, we applied the system to an aldol condensation reaction using two setups across different parts of the globe. Despite the differences in configurations, the reaction data produced by both machines were interoperable owing to the layered knowledge abstraction. Throughout the experiment, the system recorded all data provenance as the knowledge graph evolved autonomously, providing opportunities for informed machine learning [53]. Our collaborative approach resulted in faster data generation and advanced the Pareto front while exhibiting resilience to hardware failure.  
+
+20 The implementation of this work has provided valuable insights and identified areas for future improvement. In terms of orchestration, it is crucial for the system to be robust to network disruption, since it is distributed over the internet. We have implemented measures to ensure that agents deployed in the lab can handle internet cut- offs and resume operations once back online. To minimise downtime during reconnection, future developments could provide on- demand deployment of local agents to continue operation. For efficient optimisation and data quality, it is critical to have control conditions in place when adding new setups to the network, and only those generated results within the tolerance should be approved. To increase the system's robustness against software and hardware malfunctions, regular backups of all data in the central knowledge graph should be implemented. Further development could also be made to federate the SDLs, where each lab hosts its data and digital twins locally and only expose its capabilities in the central registry without revealing confidential information. An authentication and authorisation mechanism should be added to control access to the equipment and grant permission for federated learning.  
+
+34 Looking forward, achieving a globally collaborative research network requires collective efforts. The knowledge graph reflects a communal understanding of the field, and involving different stakeholders early on can accelerate collaboration and increase the chance of success. Industrial partners are encouraged to work together and provide a unified API for interacting with their proprietary software. This can be facilitated by efforts such as OPC UA [54] and SiLA [27], which aim to establish interoperability standards. Recent studies have shown the successful exchange of HPLC methods between vendors in the Chromatography Data System (CDS), demonstrating the potential for the ontology- based approach [55]. Collaboration between scientists and industry is also important at various stages of research and development [56].  
+
+44 Overall, we believe the dynamic knowledge graph approach demonstrated in this work provides the first evidence of its potential to establish a network of globally distributed SDLs.
+
+<--- Page Split --->
+
+1 Although we focus on flow chemistry in this study, the principles are generic. The same approach can be applied to DMTA cycles for other domains should relevant ontologies and agents be made available, for example, to support research in deep space [17]. We anticipate that this work will serve as an initial step towards the creation of a digital research scientist and eventually move us towards the realisation of "AI Scientists" [36].  
+
+## 4 Methods  
+
+### 4.1 The World Avatar knowledge graph  
+
+This work follows the best practices in the World Avatar project. All ontologies and agents are version- controlled on GitHub. We provide our thought process during the development below. The same principles can be followed for self- optimisation applications in other domains.  
+
+### 4.1.1 Ontology development  
+
+Developing ontologies is often an iterative process and it is not a goal in and of itself [57]. As suggested in [30, 34, 35], we follow the steps from specifying target deliverables to conceptualising relevant concepts and finally implementing codes for queries. Aimed at capturing data and material flow in distributed SDLs, the relevant concepts range from the reaction experiment to the hardware employed to conduct it. In the World Avatar, ontologies are typically developed to be digested by software agents which mimic the human way of conducting different tasks [58]. Therefore, the development draws inspiration from relevant software tools [49, 59, 60] and existing reaction database schemas [46, 47]. Views of the domain experts [61- 63] are also consulted to better align with the communal understanding of the subject. During iterations, competency questions are used to test if the ontologies meet case study requirements. The answers to these questions are provided in the form of SPARQL queries that are executed by the agents during their operations. Another essential aspect to consider is data instantiation, where we adopted pydantic to simplify the querying and processing of data from the knowledge graph. Overall, the ontology development process starts as easily as drawing concepts and their relationships on a whiteboard and then gradually materialising them in code.  
+
+### 4.1.2 Agent development  
+
+Following the development of ontologies, agents are defined as executables that process inputs and generate outputs. Their I/O signatures are represented following OntoAgent [64]. At the implementation level, all agents inherit the DerivationAgent template in python provided by the derived information framework [40]. Specifically, agents utilise the asynchronous communication mode when interacting with the knowledge graph as conducting experiments is inherently a time- consuming process. Each of the agents monitors the jobs assigned to itself and records the progress of execution in the knowledge graph. The derived information framework does most of the work behind the scenes, leaving the developer with the only task of implementing each agent's internal logic. As agents modify the knowledge graph and subsequently actuate the real world autonomously once active, it is important to
+
+<--- Page Split --->
+
+1 make sure they behave as expected. In this regard, unit and integration tests are provided to 2 help with responsible development.  
+
+### 3 4.1.3 Distributed deployment  
+
+4 Taking inspiration from remote control practices in lab automation [65- 67], the knowledge 5 graph is designed to span across the internet. It follows deployment practices commonly used 6 by cloud- native applications and is implemented through docker containers. The triplestore 7 and file server containing the knowledge statements are deployed at internet- resolvable 8 locations. Depending on capabilities, agents are located at different host machines. Those 9 who monitor and control the hardware are deployed in the corresponding laboratory for 10 security reasons. They transmit data collected from the hardware to the knowledge graph and 11 in reverse configure and actuate the equipment when a new experiment arises. At start- up, 12 agents register their OntoAgent instances in the knowledge graph, then act autonomously 13 should tasks be assigned to them. Altogether, these agents form a distributed network 14 of "digital scientists" that pass on information within the knowledge graph and bridge 15 cyberspace and the physical world.  
+
+## 4.2 Flow chemistry platforms  
+
+17 This work connects two similar automated flow chemistry platforms located in Cambridge 18 and Singapore. The method of sourcing input chemicals differs, with a liquid handler 19 employed in Cambridge and reagent bottles utilised in Singapore. We provide below brief 20 descriptions of the experimental setup. All chemicals were used as received.  
+
+## 4.2.1 Cambridge lab  
+
+22 On the Cambridge side, the experimental setup consists of two Vapourtec R2 pump modules, 23 one Vapourtec R4 reactor module, one Gilson GX- 271 liquid handler, one four- way VICI 24 switching valve (CI4W.06/.5 injector), and Shimadzu CBM- 20A HPLC analytical equipment 25 equipped with Eclipse XDB- C18 column (Agilent part number: 993967- 902). To initiate the 26 reaction, the liquid handler dispenses a \(2\mathrm{mL}\) solution of \(0.5\mathrm{M}\) benzaldehyde 1 dissolved in 27 acetonitrile (with \(0.06\mathrm{M}\) biphenyl as an internal standard) into the sample loop of pump A. 28 Acetone 2 ( \(50\%\) v/v in acetonitrile) and \(0.1\mathrm{M}\) NaOH 3 in ethanol are similarly loaded into 29 sample loops for pump B and C. After being transferred by the switching valve, the product 30 (benzylideneacetone 4) is analysed using online HPLC. The HPLC analysis lasts \(17\mathrm{min}\) 31 with a mobile phase consisting of an \(80:20\) (v/v) binary mixture of water and acetonitrile 32 running at a rate of \(2\mathrm{mL}\mathrm{min}^{- 1}\) . All compounds are detected at an absorption wavelength of 33 \(254\mathrm{nm}\)  
+
+## 4.2.2 Singapore lab  
+
+35 On the Singapore side, the experimental setup consists of two Vapourtec R2 pump modules, 36 one Vapourtec R4 reactor module, one 6- port 2- position VICI switch valve equipped with 37 \(60\mathrm{nL}\) sampling rotor, and an Agilent 1260 Infinity II system equipped with a G1311B 38 quaternary pump, Eclipse XDB- C18 column (Agilent product number: 961967- 302), and
+
+<--- Page Split --->
+
+1 G1314F variable wavelength detector (VWD). The input chemical for the reaction is sourced from three reagent bottles that are directly attached to the Vapourtec pumps: pump A contains \(0.5\mathrm{M}\) benzaldehyde 1 in acetonitrile (with \(0.05\mathrm{M}\) naphthalene as an internal standard), pump B contains \(6.73\mathrm{M}\) acetone 2 in acetonitrile ( \(50\%\) v/v in acetonitrile), and pump C contains \(0.1\mathrm{M}\) NaOH 3 in ethanol. The HPLC quaternary pump method for online HPLC described by Jeraal et al. [49] is used. However, the VWD wavelength was changed differently over the 8 min analysis time as follows: the absorption wavelength is \(248\mathrm{nm}\) for the initial \(6.05\mathrm{min}\) and then switched to \(228\mathrm{nm}\) until the end of acquisition.  
+
+## 9 Research data  
+
+10 Research data supporting this publication is available in the University of Cambridge data repository (doi:10.17863/CAM.97058).  
+
+11 All the codes developed are available on The World Avatar GitHub repository https://github.com/cambridge- cares/TheWorldAvatar. The docker images of agents are available at GitHub's public registry located at ghcr.io/cambridge- cares:/doe_agent:1.2.0,vapourtec_schedule_agent:1.2.0, vapourtec_agent:1.2.0, hplc_agent:1.2.0, hplc_postpro_agent:1.2.0, rxn_opt_goal_iter_agent:1.2.0, and rxn_opt_goal_agent:1.0.0. The deployment instructions can be found in folder TheWorldAvatar/Deploy/pips.  
+
+## 10 Acknowledgements  
+
+20 This research was supported by the National Research Foundation, Prime Minister's Office, 21 Singapore, under its Campus for Research Excellence and Technological Enterprise (CRE- 22 ATE) programme, and Pharma Innovation Platform Singapore (PIPS) via grant to CARES 23 Ltd "Data2Knowledge, C12". This project was cofunded by European Regional Develop- 24 ment Fund via the project "Innovation Centre in Digital Molecular Technologies", UKRI 25 via project EP/S024220/1 "EPSRC Centre for Doctoral Training in Automated Chemical 26 Synthesis Enabled by Digital Molecular Technologies". Part of this work was also supported 27 by Towards Turing 2.0 under the EPSRC Grant EP/W037211/1. The authors thank Dr. 28 Andrew C. Breeson for his helpful suggestions on graphical design. J. Bai acknowledges 29 financial support provided by CSC Cambridge International Scholarship from Cambridge 30 Trust and China Scholarship Council. C.J. Taylor is a Sustaining Innovation Postdoctoral 31 Research Associate at Astex Pharmaceuticals and thanks Astex Pharmaceuticals for funding, 32 as well as his Astex colleagues Mark Wade, Gianni Chessari, and David Rees for their 33 support. S.D. Rihm acknowledges financial support from Fitzwilliam College, Cambridge, 34 and the Cambridge Trust. M. Kraft gratefully acknowledges the support of the Alexander 35 von Humboldt Foundation.  
+
+36 For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising.  
+
+37 Author contributions: M.K., A.A.L., J.B., and S.M. conceived the project. J.B., S.M. and M.K. designed the ontological representation and agent workflow. J.B. implemented the ontologies/agents and deployed the knowledge graph under the advisement of S.M.
+
+<--- Page Split --->
+
+1 and K.F.L. The chemistry and HPLC method were developed by C.J.T. (Cambridge) and 2 D.K. (Singapore). D.K. and C.J.T. validated the calculation of objective functions. The 3 self- optimisation campaign involving two labs was set up by C.J.T. (hardware and chemicals 4 in Cambridge), D.K. (hardware and chemicals in Singapore) and J.B. (software on both sides 5 and goal request). M.K. and A.A.L. acquired funding and administrated the project. J.B. 6 drafted the body of this manuscript and SI with inputs from S.M., J.A., S.D.R. and C.J.T. All 7 authors provided feedback on the manuscript.  
+
+8 Competing interests: The authors declare no competing interests.
+
+<--- Page Split --->
+
+## References  
+
+[1] Florian Hase, Loic M Roch, and Alan Aspuru- Guzik. Next- Generation Experimentation with Self- Driving Laboratories. Trends Chem., 1(3):282- 291, 2019. doi:10.1016/j.trechm.2019.02.007.  [2] Milad Abolhasani and Eugenia Kumacheva. The Rise of Self- Driving Labs in Chemical and Materials Sciences. Nat. Synth., 2023. doi:10.1038/s44160- 022- 00231- 0.  [3] Robert B Merrifield, John Morrow Stewart, and Nils Jernberg. Instrument for Automated Synthesis of Peptides. Anal. Chem., 38(13):1905- 1914, 1966. doi:10.1021/ac50155a057.  [4] Connor W Coley, Dale A Thomas, Justin A M Lummis, Jonathan N Jaworski, Christopher P Breen, Victor Schultz, Travis Hart, Joshua S Fishman, Luke Rogers, Hanyu Gao, Robert W. Hicklin, Pieter P. Plehiers, Joshua Byington, John S. Piotti, William H. Green, A. John Hart, Timothy F. Jamison, and Klavs F. Jensen. A Robotic Platform for Flow Synthesis of Organic Compounds Informed by AI Planning. Science, 365(6453): eaax1566, 2019. doi:10.1126/science.aax1566.  [5] Sebastian Steiner, Jakob Wolf, Stefan Glatzel, Anna Andreou, Jaroslaw M Granda, Graham Keenan, Trevor Hinkley, Gerardo Aragon- Camarasa, Philip J Kitson, Davide Angelone, and Leroy Cronin. Organic Synthesis in a Modular Robotic System Driven by a Chemical Programming Language. Science, 363(6423):eaav2211, 2019. doi:10.1126/science.aav2211.  [6] Benjamin Burger, Phillip M Maffettone, Vladimir V Gusev, Catherine M Aitchison, Yang Bai, Xiaoyan Wang, Xiaobo Li, Ben M Alston, Buyi Li, Rob Clowes, Nicola Rankin, Brandon Harris, Reiner Sebastian Sprick, and Andrew I. Cooper. A Mobile Robotic Chemist. Nature, 583(7815):237- 241, 2020. doi:10.1038/s41586- 020- 2442- 2.  [7] Sourav Chatterjee, Mara Guidi, Peter H Seeberger, and Kerry Gilmore. Automated Radial Synthesis of Organic Molecules. Nature, 579(7799):379- 384, 2020. doi:10.1038/s41586- 020- 2083- 5.  [8] Daniel P Tabor, Loic M Roch, Semion K Saikin, Christoph Kreisbeck, Dennis Sheberla, Joseph H Montoya, Shyam Dwaraknath, Muratahan Aykol, Carlos Ortiz, Hermann Tribukait, Carlos Amador- Bedolla, Christoph J. Brabec, Benji Maruyama, Kristin A. Persson, and Alan Aspuru- Guzik. Accelerating the Discovery of Materials for Clean Energy in the Era of Smart Automation. Nat. Rev. Mater., 3(5):5- 20, 2018. doi:10.1038/s41578- 018- 0005- z.  [9] Qing Zhu, Fei Zhang, Yan Huang, Hengyu Xiao, LuYuan Zhao, XuChun Zhang, Tao Song, XinSheng Tang, Xiang Li, Guo He, BaoChen Chong, JunYi Zhou, YiHan Zhang, Baicheng Zhang, JiaQi Cao, Man Luo, Song Wang, GuiLin Ye, WanJun Zhang, Xin Chen, Shuang Cong, Donglai Zhou, Huirong Li, Jialei Li, Gang Zou, WeiWei Shang, Jun Jiang, and Yi Luo. An All- Round AI- Chemist with a Scientific Mind. Natl. Sci. Rev., 9(10), 2022. doi:10.1093/nsr/nwac190.  [10] Ross D King, Jem Rowland, Stephen G Oliver, Michael Young, Wayne Aubrey, Emma Byrne, Maria Liakata, Magdalena Markham, Pinar Pir, Larisa N Soldatova, Andrew
+
+<--- Page Split --->
+
+1 Sparkes, Kenneth E. Whelan, and Amanda Clare. The Automation of Science. Science, 324(5923):85- 89, 2009. doi:10.1126/science.1165620.
+
+<--- Page Split --->
+
+1 by Artificial Intelligence. Nat. Mach. Intell., 5(3):208- 219, 2023. doi:10.1038/s42256- 023- 00618- 4. 3 [18] Edward O Pyzer- Knapp, Jed W Pitera, Peter WJ Staar, Seiji Takeda, Teodoro Laino, Daniel P Sanders, James Sexton, John R Smith, and Alessandro Curioni. Accelerating Materials Discovery Using Artificial Intelligence, High Performance Computing and Robotics. npj Comput. Mater., 8(1):84, 2022. doi:10.1038/s41524- 022- 00765- z. 7 [19] Connor W Coley, Natalie S Eyke, and Klavs F Jensen. Autonomous Discovery in the Chemical Sciences Part II: Outlook. Angew. Chem., Int. Ed., 59(52):23414- 23436, 2020. doi:10.1002/anie.201909989. 10 [20] Liam Wilbraham, S Hessam M Mehr, and Leroy Cronin. Digitizing Chemistry Using the Chemical Processing Unit: From Synthesis to Discovery. Acc. Chem. Res., 54(2): 253- 262, 2021. doi:10.1021/acs.accounts.0c00674. 13 [21] Mark D. Wilkinson, Michel Dumontier, IJsbrand Jan Aalbersberg, Gabrielle Appleton, Myles Axton, Arie Baak, Niklas Blomberg, Jan- Willem Boiten, Luiz Bonino da Silva Santos, Philip E. Bourne, Jildau Bouwman, Anthony J. Brookes, Tim Clark, Merce Crosas, Ingrid Dillo, Olivier Dumon, Scott Edmunds, Chris T. Evelo, Richard Finkers, Alejandra Gonzalez- Beltran, Alasdair J.G. Gray, Paul Groth, Carole Goble, Jeffrey S. Grethe, Jaap Heringa, Peter A.C 't Hoen, Rob Hooft, Tobias Kuhn, Ruben Kok, Joost Kok, Scott J. Lusher, Maryann E. Martone, Albert Mons, Abel L. Packer, Bengt Persson, Philippe Rocca- Serra, Marco Roos, Rene van Schaik, Susanna- Assunta Sansone, Erik Schultes, Thierry Sengstag, Ted Slater, George Strawn, Morris A. Swertz, Mark Thompson, Johan van der Lei, Erik van Mulligen, Jan Velterop, Andra Waagmeester, Peter Wittenburg, Katherine Wolstencroft, Jun Zhao, and Barend Mons. The FAIR Guiding Principles for Scientific Data Management and Stewardship. Sci. Data, 3:160018, 2016. doi:10.1038/sdata.2016.18. 26 [22] Loic M Roch, Florian Hase, Christoph Kreisbeck, Teresa Tamayo- Mendoza, Lars P E Yunker, Jason E Hein, and Alan Aspuru- Guzik. ChemOS: Orchestrating Autonomous Experimentation. Sci. Robot., 3(19):eaat5559, 2018. doi:10.1126/sicrobotics.aat5559. 29 [23] Ian M Pendleton, Gary Cattabriga, Zhi Li, Mansoor Ani Najeeb, Sorelle A Friedler, Alexander J Norquist, Emory M Chan, and Joshua Schrier. Experiment Specification, Capture and Laboratory Automation Technology (ESCALATE): A Software Pipeline for Automated Chemical Experimentation and Data Management. MRS Commun., 9 (3):846- 859, 2019. doi:10.1557/mrc.2019.72. 34 [24] Fuzhan Rahmanian, Jackson Flowers, Dan Guevarra, Matthias Richter, Maximilian Fichtner, Phillip Donnely, John M Gregoire, and Helge S Stein. Enabling Modular Autonomous Feedback- Loops in Materials Science through Hierarchical Experimental Laboratory Automation and Orchestration. Adv. Mater. Interfaces, 9(8):2101987, 2022. doi:10.1002/admi.202101987. 39 [25] S Hessam M Mehr, Matthew Craven, Artem I Leonov, Graham Keenan, and Leroy Cronin. A Universal System for Digitization and Automatic Execution of the Chemical Synthesis Literature. Science, 370(6512):101- 108, 2020. doi:10.1126/science.abc2986.
+
+<--- Page Split --->
+
+[26] Jason Hein, Robert Rauschen, Mason Guy, and Leroy Cronin. Universal Chemical Programming Language for Robotic Synthesis Reproducibility, 2023. Research Square Platform LLC. doi:10.21203/rs.3. rs- 2761997/v1.  
+
+[27] Burkhard Schäfer. Data Exchange in the Laboratory of the Future – A Glimpse at AnIML and SiLA, 2018. URL https://analyticalscience.wiley.com/do/10.1002/gitlab.17270/full/. Accessed 30 May 2023.  
+
+[28] Sonia Natalie Mitchell, Andrew Lahiff, Nathan Cummings, Jonathan Hollocombe, Bram Boskamp, Ryan Field, Dennis Reddyhoff, Kristian Zarebski, Antony Wilson, Bruno Viola, Martin Burke, Blair Archibald, Paul Bessell, Richard Blackwell, Lisa A. Boden, Alys Brett, Sam Brett, Ruth Dundas, Jessica Enright, Alejandra N. Gonzalez- Beltran, Claire Harris, Ian Hinder, Christopher David Hughes, Martin Knight, Vino Mano, Ciaran McMonagle, Dominic Mellor, Sibylle Mohr, Glenn Marion, Louise Matthews, Iain J. McKendrick, Christopher Mark Pooley, Thibaud Porphyre, Aaron Reeves, Edward Townsend, Robert Turner, Jeremy Walton, and Richard Reeve. FAIR Data Pipeline: Provenance- Driven Data Management for Traceable Scientific Workflows. Philos. Trans. R. Soc., A, 380(2233):20210300, 2022. doi:10.1098/rsta.2021.0300.  
+
+[29] Angiras Menon, Nenad B Krdzavac, and Markus Kraft. From Database to Knowledge Graph—Using Data in Chemistry. Curr. Opin. Chem. Eng., 26:33- 37, 2019. doi:10.1016/j.coche.2019.08.004.  
+
+[30] Jiaru Bai, Liwei Cao, Sebastian Mosbach, Jethro Akroyd, Alexei A Lapkin, and Markus Kraft. From Platform to Knowledge Graph: Evolution of Laboratory Automation. JACS Au, 2(2):292- 309, 2022. doi:10.1021/jacsau.1c00438.  
+
+[31] Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia D'Amato, Gerard De Melo, Claudio Gutierrez, Sabrina Kirrane, José Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, Axel- Cyrille Ngonga Ngomo, Axel Polleres, Sabbir M. Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequeda, Steffen Staab, and Antoine Zimmermann. Knowledge Graphs. ACM Comput. Surv., 54(4):1- 37, 2022. doi:10.1145/3447772.  
+
+[32] Jethro Akroyd, Sebastian Mosbach, Amit Bhave, and Markus Kraft. Universal Digital Twin – A Dynamic Knowledge Graph. Data- Centric Engineering, 2:e14, 2021. doi:10.1017/dce.2021.10.  
+
+[33] Mei Qi Lim, Xiaonan Wang, Oliver Inderwildi, and Markus Kraft. The World Avatar – A World Model for Facilitating Interoperability. In Intelligent Decarbonisation: Can Artificial Intelligence and Cyber- Physical Systems Help Achieve Climate Mitigation Targets?, pages 39- 53. Springer International Publishing, 2022. doi:10.1007/978- 3- 030- 86215- 2_4.  
+
+[34] Simon D. Rihm, Jiaru Bai, Aleksandar Kondinski, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. A Comprehensive Laboratory Digital Twin: Top- Down Automation of a Chemical Research Laboratory, 2023. In preparation.  
+
+[35] Aleksandar Kondinski, Jiaru Bai, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. Knowledge Engineering in Chemistry: From Expert Systems to Agents of Creation. Acc. Chem. Res., 56(2):128- 139, 2023. doi:10.1021/acs.accounts.2c00617.
+
+<--- Page Split --->
+
+[36] Hiroaki Kitano. Nobel Turing Challenge: Creating the Engine for Scientific Discovery. npj Syst. Biol. Appl., 7(1):1- 12, 2021. doi:10.1038/s41540- 021- 00189- 3. 
+
+[37] Connor W Coley, Natalie S Eyke, and Klavs F Jensen. Autonomous Discovery in the Chemical Sciences Part I: Progress. Angew. Chem., Int. Ed., 59(51):22858- 22893, 2020. doi:10.1002/anie.201909987. 
+
+[38] Connor J Taylor, Alexander Pomberger, Kobi C Felton, Rachel Grainger, Magda Barecka, Thomas W Chamberlain, Richard A Bourne, Christopher N Johnson, and Alexei A Lapkin. A Brief Introduction to Chemical Reaction Optimization. Chem. Rev., 123(6):3089- 3126, 2023. doi:10.1021/acs.chemrev.2c00798. 
+
+[39] Andy B. Yoo, Morris A. Jette, and Mark Grondona. SLURM: Simple Linux Utility for Resource Management. In Job Scheduling Strategies for Parallel Processing, pages 44- 60. Springer Berlin Heidelberg, 2003. doi:10.1007/10968987_3. 
+
+[40] Jiaru Bai, Kok Foong Lee, Markus Hofmeister, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. A Derived Information Framework for a Dynamic Knowledge Graph and its Application to Smart Cities, 2023. Submitted for publication. Preprint available from https://como.ceb.cam.ac.uk/preprints/302/. 
+
+[41] Ian Holland and Jamie A. Davies. Automation in the Life Science Research Laboratory. Front. Bioeng. Biotechnol., 8:571777, 2020. doi:10.3389/fbioe.2020.571777. 
+
+[42] Anand S. Rao and Michael P. Georgeff. Modeling Rational Agents within a BDI- Architecture. In Proceedings of the Second International Conference on Principles of Knowledge Representation and Reasoning, KR'91, page 473- 484, San Francisco, CA, USA, 1991. Morgan Kaufmann Publishers Inc. ISBN 1558601651. doi:10.5555/3087158.3087205. 
+
+[43] Helge S Stein and John M Gregoire. Progress and Prospects for Accelerating Materials Science with Automated and Autonomous Workflows. Chem. Sci., 10(42):9640- 9649, 2019. doi:10.1039/C9SC03766G. 
+
+[44] Michel Krämer, Hendrik M Würz, and Christian Altenhofen. Executing Cyclic Scientific Workflows in the Cloud. J. Cloud Comput., 10(1):1- 26, 2021. doi:10.1186/s13677- 021- 00229- 7. 
+
+[45] Jan Morbach, Aidong Yang, and Wolfgang Marquardt. OntoCAPE - A Large-Scale Ontology for Chemical Process Engineering. Eng. Appl. Artif. Intell., 20(2):147- 161, 2007. doi:10.1016/j.engappai.2006.06.010. 
+
+[46] Steven M Kearnes, Michael R Maser, Michael Wleklinski, Anton Kast, Abigail G Doyle, Spencer D Dreher, Joel M Hawkins, Klavs F Jensen, and Connor W Coley. The Open Reaction Database. J. Am. Chem. Soc., 143(45):18820- 18826, 2021. doi:10.1021/jacs.1c09820. 
+
+[47] Pistoia Alliance. Unified Data Model, 2020. URL https://github.com/Pistoia Alliance/UDM. Accessed 30 May 2023. 
+
+[48] Laura Daniele, Raúl Garcia-Castro, Maxime Lefrançois, and Maria Poveda-Villalon. SAREF: The Smart Applications REference ontology, 2020. URL https://saref. etsi.org/core/v3.1.1/. Accessed 21 Feb 2023.
+
+<--- Page Split --->
+
+[49] Mohammed I Jeraal, Simon Sung, and Alexei A Lapkin. A Machine Learning- Enabled Autonomous Flow Chemistry Platform for Process Optimization of Multiple Reaction Metrics. Chem. Methods, 1(1):71- 77, 2021. doi:10.1002/cmtd.202000044. [50] Jean Soon Park, Ping Hu, Yakang Lin, and Lori Ann Reinsalu. Composition and Method for Preventing, Reducing, Alleviating or Treating Idiopathic Vomiting, February 4 2020. URL https://testpubchem.ncbi.nlm.nih.gov/patent/US- 10548935- B2. US Patent 10,548,935. Accessed 11 Feb 2023. [51] Laura Pascazio, Simon D. Rihm, Ali Naseri, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. A Chemical Species Ontology for Data Integration and Knowledge Discovery, 2023. Submitted for publication. Preprint available from https://como .ceb.cam.ac.uk/preprints/306/. [52] Eric Bradford, Artur M Schweidtmann, and Alexei Lapkin. Efficient Multiobjective Optimization Employing Gaussian Processes, Spectral Sampling and A Genetic Algorithm. J. Glob. Optim., 71(2):407- 438, 2018. doi:10.1007/s10898- 018- 0609- 2. [53] Laura von Rueden, Sebastian Mayer, Katharina Beckh, Bogdan Georgiev, Sven Gieselbach, Raoul Heese, Birgit Kirsch, Michal Walczak, Julius Pfrommer, Annika Pick, Rajkumar Ramamurthy, Jochen Garcke, Christian Bauckhage, and Jannis Schuecker. Informed Machine Learning - A Taxonomy and Survey of Integrating Prior Knowledge into Learning Systems. IEEE Trans. Knowl. Data Eng., 35(1):614- 633, 2021. doi:10.1109/TKDE.2021.3079836. [54] OPC Foundation. Unified Architecture, 2023. URL https://opcfoundation.org/g/about/opc- technologies/opc- ua/. Accessed 8 March 2023. [55] Pistoia Alliance. Update from the Pistoia Alliance's Methods Hub Project, 2022. URL https://www.pistoiaalliance.org/methods/april- 2022- methods- database- hplc- uv- methods/. Accessed 10 Feb 2023. [56] Melodie Christensen, Lars PE Yunker, Parisa Shiri, Tara Zepel, Paloma L Prieto, Shad Grunert, Finn Bork, and Jason E Hein. Automation isn't Automatic. Chem. Sci., 12 (47):15473- 15490, 2021. doi:10.1039/D1SC04588A. [57] Steffen Staab and Rudi Studer. Handbook on Ontologies. Springer Science & Business Media, 2010. [58] Aleksandar Kondinski, Angiras Menon, Daniel Nurkowski, Feroz Farazi, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. Automated Rational Design of Metal- Organic Polyhedra. J. Am. Chem. Soc., 144(26):11713- 11728, 2022. doi:10.1021/jacs.2c03402. [59] Jiang Guo, A Santiago Ibanez- Lopez, Hanyu Gao, Victor Quach, Connor W Coley, Klavs F Jensen, and Regina Barzilay. Automated Chemical Reaction Extraction from Scientific Literature. J. Chem. Inf. Model., 2021. doi:10.1021/acs.jcim.1c00284. [60] Kobi C Felton, Jan G Rittig, and Alexei A Lapkin. Summit: Benchmarking Machine Learning Methods for Reaction Optimisation. Chemistry- Methods, 1(2):116- 122, 2021. doi:10.1002/cmtd.202000051.
+
+<--- Page Split --->
+
+[61] Matthew B Plutschack, Bartholomäus Pieber, Kerry Gilmore, and Peter H Seeberger. The Hitchhiker's Guide to Flow Chemistry. Chem. Rev., 117(18):11796- 11893, 2017. doi:10.1021/acs.chemrev.7b00183. [62] Sushant S Garud, Iftekhar A Karimi, and Markus Kraft. Design of Computer Experiments: A Review. Comput. Chem. Eng., 106:71- 95, 2017. doi:10.1016/j.compchemeng.2017.05.010. [63] Adam D Clayton, Jamie A Manson, Connor J Taylor, Thomas W Chamberlain, Brian A Taylor, Graeme Clemens, and Richard A Bourne. Algorithms for the SelfOptimisation of Chemical Reactions. React. Chem. Eng., 4(9):1545- 1554, 2019. doi:10.1039/C9RE00209J. [64] Xiaochi Zhou, Andreas Eibeck, Mei Qi Lim, Nenad B Krdzavac, and Markus Kraft. An Agent Composition Framework for the J- Park Simulator - A Knowledge Graph for the Process Industry. Comput. Chem. Eng., 130:106577, 2019. doi:10.1016/j.compchemeng.2019.106577. [65] Daniel E Fitzpatrick, Timothé Maujean, Amanda C Evans, and Steven V Ley. Across- the- World Automated Optimization and Continuous- Flow Synthesis of Pharmaceutical Agents Operating through a Cloud- Based Server. Angew. Chem., Int. Ed., 57(46):15128- 15132, 2018. doi:10.1002/anie.201809080. [66] Juan Ramírez, David Soto, Sebastian López, Jethro Akroyd, Daniel Nurkowski, Maria L. Botero, Nicola Bianco, George Brownbridge, Markus Kraft, and Alejandro Molina. A Virtual Laboratory to Support Chemical Reaction Engineering Courses Using Real- Life Problems and Industrial Software. Educ. Chem. Eng., 33:36- 44, 2020. doi:10.1016/j.ece.2020.07.002. [67] CMU Cloud Lab, 2023. URL https://cloudlab.cmu.edu/. Accessed 12 May 2023.
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+suppinfo.pdf  suppvideo.html
+
+<--- Page Split --->

preprint/preprint__2b4640cbbd24fef518b88dcfeb69d338b82c53d043af3c97eb48d85e522c80ce/preprint__2b4640cbbd24fef518b88dcfeb69d338b82c53d043af3c97eb48d85e522c80ce_det.mmd

@@ -0,0 +1,445 @@
+<|ref|>title<|/ref|><|det|>[[44, 107, 872, 175]]<|/det|>
+# From Platform to Knowledge Graph: Distributed Self-Driving Laboratories  
+
+<|ref|>text<|/ref|><|det|>[[44, 194, 625, 238]]<|/det|>
+Markus Kraft (mk306@cam.ac.uk) University of Cambridge https://orcid.org/0000- 0002- 4293- 8924  
+
+<|ref|>text<|/ref|><|det|>[[44, 243, 625, 308]]<|/det|>
+Jiaru Bai Sebastian Mosbach University of Cambridge https://orcid.org/0000- 0001- 7018- 9433  
+
+<|ref|>text<|/ref|><|det|>[[44, 312, 260, 352]]<|/det|>
+Connor Taylor Astex Pharmaceuticals  
+
+<|ref|>text<|/ref|><|det|>[[44, 358, 667, 401]]<|/det|>
+Dogancan Karan Cambridge Centre for Advanced, Research and Education in Singapore  
+
+<|ref|>text<|/ref|><|det|>[[44, 405, 667, 448]]<|/det|>
+Kok Foong Lee Cambridge Centre for Advanced, Research and Education in Singapore  
+
+<|ref|>text<|/ref|><|det|>[[44, 452, 667, 495]]<|/det|>
+Simon Rihm Cambridge Centre for Advanced, Research and Education in Singapore  
+
+<|ref|>text<|/ref|><|det|>[[44, 499, 267, 540]]<|/det|>
+Jethro Akroyd University of Cambridge  
+
+<|ref|>text<|/ref|><|det|>[[44, 545, 267, 586]]<|/det|>
+Alexei Lapkin University of Cambridge  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 626, 101, 644]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 664, 930, 707]]<|/det|>
+Keywords: Dynamic knowledge graph, laboratory automation, digital twin, distributed laboratory, multiagent system, goal- driven self- optimisation  
+
+<|ref|>text<|/ref|><|det|>[[44, 724, 296, 743]]<|/det|>
+Posted Date: July 25th, 2023  
+
+<|ref|>text<|/ref|><|det|>[[44, 761, 473, 781]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3. rs- 3141873/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 799, 910, 842]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[44, 860, 530, 880]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 915, 940, 958]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on January 23rd, 2024. See the published version at https://doi.org/10.1038/s41467- 023- 44599- 9.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[0, 0, 997, 997]]<|/det|>
+# 1. 1. 1. 1. 1. 1. 1. 2. 2. 2. 2. 2. 2. 2.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[179, 90, 817, 144]]<|/det|>
+# From Platform to Knowledge Graph: Distributed Self-Driving Laboratories  
+
+<|ref|>text<|/ref|><|det|>[[192, 159, 808, 210]]<|/det|>
+Jiaru Bai \(^{1}\) , Sebastian Mosbach \(^{1,2}\) , Connor J. Taylor \(^{3,4}\) , Dogancan Karan \(^{2}\) , Kok Foong Lee \(^{5}\) , Simon D. Rihm \(^{1,2}\) , Jethro Akroyd \(^{1,2}\) , Alexei A. Lapkin \(^{1,2,4}\) ,  
+
+<|ref|>text<|/ref|><|det|>[[417, 220, 578, 236]]<|/det|>
+Markus Kraft \(^{1,2,6,7,*}\)  
+
+<|ref|>text<|/ref|><|det|>[[365, 260, 631, 274]]<|/det|>
+\*Corresponding author: mk306@cam.ac.uk  
+
+<|ref|>text<|/ref|><|det|>[[163, 297, 835, 472]]<|/det|>
+\(^{1}\) Department of Chemical Engineering and Biotechnology, University of Cambridge, Philippa Fawcett Drive, Cambridge CB3 0AS, United Kingdom \(^{2}\) CARES, Cambridge Centre for Advanced, Research and Education in Singapore, 1 Create Way, CREATE Tower, #05- 05, 138602 Singapore \(^{3}\) Astex Pharmaceuticals, 436 Cambridge Science Park Milton Road, Cambridge CB4 0QA, United Kingdom \(^{4}\) Innovation Centre in Digital Molecular Technologies, Yusuf Hamied Department of Chemistry, University of Cambridge, Lensfield Road, Cambridge CB2 1EW, United Kingdom \(^{5}\) CMCL Innovations, Sheraton House, Cambridge CB3 0AX, United Kingdom \(^{6}\) School of Chemical and Biomedical Engineering, Nanyang Technological University, 62 Nanyang Drive, 637459 Singapore \(^{7}\) The Alan Turing Institute, London NW1 2DB, United Kingdom  
+
+<|ref|>sub_title<|/ref|><|det|>[[466, 508, 530, 522]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[210, 534, 787, 762]]<|/det|>
+The ability to integrate resources and share knowledge across organisations empowers scientists to expedite the scientific discovery process. This is especially crucial in addressing emerging global challenges that require global solutions. In this work, we develop an architecture to enable distributed self- driving laboratories as part of The World Avatar project, which seeks to demonstrate how to create an all- encompassing digital twin based on a dynamic knowledge graph. Our approach utilises ontologies to capture the data and material flows involved in design- make- test- analyse cycles, and employs autonomous agents as executable knowledge components to carry out the experimentation workflow. All data provenance is recorded following the FAIR principles, ensuring its accessibility and interoperability. We demonstrate the practical application of our framework by linking two robots in Cambridge and Singapore to achieve a collaborative closed- loop optimisation for a pharmaceutically- relevant aldol condensation reaction in real time. The knowledge graph evolves autonomously while progressing towards the research goals set by the scientist. The two robots effectively produced a Pareto front for the cost- yield optimisation problem over the course of three days of operation.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[140, 90, 808, 360]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[207, 378, 787, 408]]<|/det|>
+<center>Keywords: Dynamic knowledge graph, laboratory automation, digital twin, distributed laboratory, multi-agent system, goal-driven self-optimisation </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[150, 92, 334, 112]]<|/det|>
+## 1 Introduction  
+
+<|ref|>text<|/ref|><|det|>[[140, 130, 833, 293]]<|/det|>
+2 The concept of laboratory automation, recently reinterpreted as self- driving laboratories (SDLs) [1, 2], has been in existence since the 1960s, when Merrifield et al. [3] introduced the first automated chemistry hardware. Since then, SDLs have gained widespread adoption in chemistry [4- 7], materials science [8, 9], biotechnology [10, 11] and robotics [12], resulting in accelerated scientific discovery and societal development. However, the implementation of SDLs can be challenging and typically requires a highly specialised team of researchers with expertise in chemistry, engineering, and computer science. Consequently, studies are often conducted by large research groups within a single organisation. Even in cases where collaborations occur between research groups, the SDL is usually centralised within the same laboratory.  
+
+<|ref|>text<|/ref|><|det|>[[140, 299, 833, 508]]<|/det|>
+In response to the pressing global challenges of today, there is a growing consensus within the scientific community that a paradigm shift towards a globally collaborative research network is necessary [13- 15]. This shift requires decentralising SDLs to integrate different research groups to contribute their expertise towards solving emerging problems [16]. Such decentralisation also holds great potential in supporting human exploration in deep space [17]. Achieving this vision is not an easy task and entails three major challenges. The first challenge is efficiently orchestrating heterogeneous resources [18], which includes hardware from different vendors and diverse computing environments. The second challenge is sharing data across organisations [19], which requires standardising language in which the research is communicated [20]. During this process, the source and metadata of the research need to be tracked to facilitate reproducibility, which leads to the third challenge of data provenance recording following FAIR principles - Findable, Accessible, Interoperable and Reusable [21].  
+
+<|ref|>text<|/ref|><|det|>[[140, 514, 833, 660]]<|/det|>
+Many attempts have been made to tackle each of these challenges separately. For resource orchestration, middleware such as ChemOS [22], ESCALATE [23], and HELAO [24] exist to glue different components within an SDL and abstract the hardware resources. For data sharing, XDL [25, 26] and AnIML [27] are examples of standard protocols developed for synthesis and analysis respectively. For data provenance, Mitchell et al. [28] proposed a data pipeline to support the modelling of the COVID pandemic. Although these studies provide insights into building a collaborative research environment, they are developed in isolation with customised data interfaces. Enhancing interoperability between these systems is essential to establish a truly connected research network.  
+
+<|ref|>text<|/ref|><|det|>[[140, 666, 833, 859]]<|/det|>
+As discussed in our previous work [29, 30], semantic web technologies such as knowledge graphs [31] offer a viable path forward. Ontologies abstract both resources and data using the same notion, allowing for a common language between participants when allocating tasks and sharing results. The World Avatar [32, 33] is such a knowledge graph that aims to encompass all aspects of scientific research laboratories as shown in Fig. 1 in their entirety: The experiment itself, including its physical setup and underlying chemistry; moving handlers that can be of human or robotic nature; and the laboratory providing necessary infrastructure and resources [34]. The World Avatar goes beyond static knowledge representation by encoding software agents as executable knowledge components, enabling dynamicity and continuous incorporation of new concepts and data while preserving connections to existing information. As the knowledge graph expands, this characteristic allows for capturing data provenance from experimental processes as knowledge statements, effectively acting as a
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 94, 834, 247]]<|/det|>
+1 living copy of the real world. By facilitating immediate dissemination of data between SDLs at its creation, the dynamic knowledge graph presents a promising holistic solution to the challenges aforementioned [30, 35] and supports the realisation of "AI Scientists" [34, 36]. 4 The purpose of this paper is to demonstrate a proof- of- concept for a distributed network of SDLs enabled by a dynamic knowledge graph. This signifies the first step towards digital research scientists (as shown in Fig. 1) collaborating autonomously. To illustrate the effectiveness of this approach, we present a demonstration using two robots in Cambridge and Singapore collaborating on a multi- objective closed- loop optimisation problem in response to a goal request from scientists.  
+
+<|ref|>image<|/ref|><|det|>[[330, 263, 664, 491]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[163, 505, 834, 570]]<|/det|>
+<center>Figure 1: An overview of intertwined aspects of a chemical research laboratory that need to be represented by a connected lab digital twin, adapted from [34]. This paper focuses on the automation of chemical reaction optimisation, a task that can be viewed as part of the daily work of a research scientist. </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 608, 278, 627]]<|/det|>
+## 2 Results  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 647, 495, 664]]<|/det|>
+### 2.1 Architecture of distributed SDLs  
+
+<|ref|>text<|/ref|><|det|>[[140, 679, 834, 793]]<|/det|>
+Closed- loop optimisation in SDLs is a dynamic process that revolves around design- make- test- analyse (DMTA) cycles [37, 38]. Compared to machine learning systems and scientific workflows that only capture data flows, SDLs offer an integrated approach by orchestrating both computational and physical resources. This involves the integration of data and material flows, as well as the interface that bridges the gap between the virtual and physical worlds. To this end, we propose a conceptual architecture of distributed SDLs that effectively incorporates all three flows, as illustrated in Fig. 2(a).  
+
+<|ref|>text<|/ref|><|det|>[[140, 800, 834, 864]]<|/det|>
+The proposed architecture presents a framework to enable scientists to set research goals and resource restrictions for a particular chemical reaction and have them trigger a closed- loop process in cyberspace. The process is initiated by the monitoring component, which parses the research goals and requests the iterations needed to achieve the objectives. The
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[174, 179, 833, 325]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[164, 331, 839, 372]]<|/det|>
+(a) Conceptual framework of components used to build a network of distributed SDLs for closed-loop optimisation. Information and materials from a chemical reaction flow autonomously across cyber and physical space until they accomplish the research goals set by scientists or the allocated resources have been consumed.  
+
+<|ref|>image<|/ref|><|det|>[[175, 384, 837, 703]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[275, 766, 719, 782]]<|/det|>
+<center>Figure 2: An illustration of a distributed SDLs architecture. </center>  
+
+<|ref|>text<|/ref|><|det|>[[164, 710, 838, 750]]<|/det|>
+(b) Dynamic knowledge graph approach. The overall system comprises three layers: the real world where the hardware is located, the dynamic knowledge graph that hosts all information in cyberspace, and the layer of active agents that manages the knowledge graph.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 94, 835, 290]]<|/det|>
+iterating component collects prior information about the design space and passes it on to the component that designs the next experiment. The algorithm employed, as well as the availability of prior data, determines the combination of design variables to be proposed within the domain provided by the scientist. Similar to the scheduling of high- performance computing (HPC) jobs [39], the proposed physical experimentation is scheduled for execution in one of the available laboratories. The suggested conditions are translated to the machine- actionable recipe that enables the control of hardware for reaction and characterisation. In the physical world, this is reflected in the material flow between the two pieces of equipment. The data processing component is then responsible for computing the objectives by analysing the complete job information and raw data. If the resources are still available, a comparison of these objectives with the research goals determines whether the system should proceed to the next iteration.  
+
+<|ref|>text<|/ref|><|det|>[[140, 295, 835, 376]]<|/det|>
+This architecture liberates the scientists from routine work, however, it also poses challenges in the implementation in terms of ensuring robustness, scalability, maintainability, safety, and ethics. Ideally, the system should enable seamless integration of new devices, resources, and algorithms without disrupting the system's overall functioning. It is also critical to allow for dynamic adaption to changes in research goals and resource restrictions.  
+
+<|ref|>text<|/ref|><|det|>[[140, 381, 835, 560]]<|/det|>
+We believe dynamic knowledge graph technology can help with realising this design [30]. Specifically, as illustrated in Fig. 2(b), this technology represents the software components as agents that receive inputs and produce outputs, while the data flow between components is marked up as the agents' messages. Physical entities can be virtualised as digital twins in cyberspace, enabling real- time control and eliminating geospatial boundaries when multiple labs are involved. This reformulation of the closed- loop optimisation problem as information travelling through the knowledge graph and reflecting their changes in the real world offers a powerful framework for achieving true distributed SDLs. In this way, we can think of an occurrence of physical experimentation as a sequence of actions that dynamically generates information about a reaction experiment as it progresses in time, analogous to computational workflows [40].  
+
+<|ref|>text<|/ref|><|det|>[[140, 566, 835, 775]]<|/det|>
+This work is part of a series of papers introducing a holistic approach to lab automation by including all aspects of research laboratories (see Fig. 1) in an all- encompassing digital twin [34]. By employing dynamic knowledge graphs embedded with knowledge models across different domains, we can address the challenges related to interoperability and adaptability encountered in platform- based approaches [30]. The goal- driven top- down architecture enables reasoning across the knowledge base and ensures alignment. Consequently, only high- level, abstract goals need to be defined by humans, which are subsequently decomposed into concrete sub- goals and tasks. In this framework, humans play a dual role, acting both as handlers to execute and intervene in experiments. They can be represented in the knowledge graph in a similar manner to robots and are provided with instructions in a human- readable format. This enables the realisation of a hybrid and evolving digital laboratory, bridging potential "interim technology gaps" [41]. It is important to note that the work described in this paper relies fully on robotic handling only.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[144, 94, 533, 111]]<|/det|>
+### 2.2 Chemical ontologies and digital twins  
+
+<|ref|>text<|/ref|><|det|>[[140, 125, 834, 272]]<|/det|>
+2.2 Chemical ontologies and digital twins2 The realisation of SDLs requires a connection between abstract chemistry knowledge and concrete hardware for execution [20]. This calls for a set of connected ontologies, as identified in our previous analysis on the gaps in current semantic representations for chemical digitalisation [30]. Figure 3 presents a selection of concepts and relationships as an effort to address these gaps. These concepts span various abstraction levels, ranging from the high- level research goals to the conceptual level of chemical reactions to the mathematical level of design of experiments to the physical realisation of reaction experiments and the laboratory digital twin. We describe below ontologies' cross- domain characteristics, for technical details on each ontology please see Supplementary Information section A.1.  
+
+<|ref|>text<|/ref|><|det|>[[140, 278, 834, 423]]<|/det|>
+For closed- loop optimisation in SDLs, we draw parallels between the pursuit of optimal objectives and the reasoning cycles involved in pursuing a goal [42, 43]. The multi- objective problem can be formulated as a GoalSet which comprises individual Goals. Each goal is associated with specific dimensional quantities that can be achieved by a Plan, which consists of multiple Steps to be carried out by corresponding agents. From the implementation perspective, this is akin to research in the scientific workflow community which seeks to manage workflow expressed as a directed cyclic graph (DCG) [44]. In this regard, we adopt the derived information framework [40], a knowledge- graph- native approach, to manage the iterative workflow.  
+
+<|ref|>text<|/ref|><|det|>[[137, 429, 834, 656]]<|/det|>
+In developing chemical ontologies for SDLs, we draw upon the lessons learnt in creating ontologies for chemical plants. One prominent example is the OntoCAPE material and chemical process system [45] ontology, which describes materials from three aspects: the ChemicalSpecies that reflects the intrinsic characteristics, Material as part of the phase system which describes macroscopic thermodynamic behaviour, and MaterialAmount that refers to a concrete occurrence of an amount of matter in the physical world. Building on this foundation, we introduce OntoReaction, an ontology that captures knowledge in wet- lab reaction experiments, and OntoDoE, an ontology for the design of experiments (DoE) in optimisation campaigns. As an effort to align with existing data, OntoReaction draws inspiration from established schemas used in chemical reaction databases like ORD [46] and UDM [47]. ReactionExperiment is a concrete realisation of a ChemicalReaction that is sampled at a set of ReactionConditions and measures certain PerformanceIndicators. When grouped together, they can form HistoricalData that are utilised by a DesignOfExperiment study to propose new experiments.  
+
+<|ref|>text<|/ref|><|det|>[[137, 661, 833, 805]]<|/det|>
+In the development of our hardware ontologies, we have utilised the Smart Applications REFerence (SAREF) ontology [48] wherever feasible, which is widely adopted in the field of the Internet of Things (IoT). We introduce OntoLab to represent the digital twin of a laboratory, comprising a group of LabEquipment and ChemicalContainers that contain ChemicalAmount. Furthermore, we create OntoVapourtec and OntoHPLC as ontologies for the equipment involved in this work, linking them to the concrete realisation aspect of OntoCAPE. We establish the link between abstract chemical knowledge and hardware by translating ReactionCondition to ParameterSetting, which can be combined to form EquipmentSettings for configuration.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[167, 220, 832, 692]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[164, 706, 835, 754]]<|/det|>
+<center>Figure 3: A selection of concepts and relationships capturing different aspects in SDLs. Their namespaces correspond to the colour coding. For complete knowledge representation and namespace definitions see Supplementary Information. </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[144, 94, 518, 111]]<|/det|>
+### 2.3 Contextualised reaction informatics  
+
+<|ref|>text<|/ref|><|det|>[[140, 125, 833, 256]]<|/det|>
+2.3 Contextualised reaction informatics2 By utilising ontologies as blueprints, it becomes possible to instantiate reaction information while preserving connections with contextual recordings. The reaction we choose for demonstration is an aldol condensation reaction between benzaldehyde 1 (bold numbers for reference) and acetone 2, catalysed by sodium hydroxide 3 to yield the target product benzylideneacetone 4 [49]. There are also the reported side products dibenzylideneacetone 5 and further condensation products from acetone polymerisation. The target product is pharmaceutically relevant and can be used as an NK- 1 receptor inhibitor for the treatment of idiopathic vomiting [50].  
+
+<|ref|>text<|/ref|><|det|>[[140, 261, 833, 471]]<|/det|>
+Figure 4 provides an illustrative representation of the chosen reaction in the knowledge graph as viewed through various roles within a laboratory, each with its unique perspective on the same chemical. Taking the starting material benzaldehyde as an example, a chemist, more interested in conceptual description, focuses on its role as a reactant and seeks relevant species information. A data scientist, on the other hand, looks into its concentration to determine the appropriate usage of other chemicals when designing conditions for a particular reaction experiment. Meanwhile, a lab manager inspects via the 3D digital twin and ensures the availability of an internal standard that can be mixed with the physical existence of benzaldehyde to enable characterisation during the actual execution of the experiment. Lastly, the knowledge engineer seeks to represent and access this information in a machine- readable format following the proposed ontologies. In practice, the same individual might play several roles, and the emphasis here is on the cross- domain interoperability enabled by the knowledge graph.  
+
+<|ref|>text<|/ref|><|det|>[[139, 477, 833, 606]]<|/det|>
+The integration of chemical knowledge from PubChem, represented by OntoSpecies for unique species identification [51], serves as a critical link between these facets of chemicals. It enables the identification of potential input chemicals based on the reactant and solvent during DoE and allows for the selection of appropriate sources of starting materials from multiple chemical containers (see Supplementary Information section A.2). By combining various aspects into a unified representation, the knowledge graph encompasses information that is pertinent to diverse users while keeping humans in the loop [34]. For concrete examples of ontology instantiation see Supplementary Information section A.1.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[175, 160, 822, 730]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[163, 740, 835, 806]]<|/det|>
+<center>Figure 4: A snapshot of reaction views from different perspectives. The knowledge graph representation puts chemical informatics into context, allowing for queries and answers from different layers of abstraction. The colour coding corresponds to the ontological expression. </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[144, 94, 499, 112]]<|/det|>
+### 2.4 Goal-driven knowledge dynamics  
+
+<|ref|>text<|/ref|><|det|>[[140, 125, 834, 305]]<|/det|>
+2 Figure 5 presents a high- level overview of the goal- driven self- evolution of the knowledge graph during closed- loop optimisation. The dynamicity of the knowledge graph is enabled by the presence of software agents that realise each component of the distributed architecture and facilitate the flow of information within the graph. The process begins with goal derivation when the scientist initiates a goal request. The Reaction Optimisation Goal (ROG) Agent is responsible for translating the goal request into a machine- readable statement that captures the scientist's intention. To accommodate all objectives, a goal set is formulated that considers each objective as a reward function for the agents' operations. For each participating laboratory, a Goal Iteration Derivation instance is created using the derived information framework [40] and requested for execution by the Reaction Optimisation Goal Iteration (ROGI) Agent.  
+
+<|ref|>text<|/ref|><|det|>[[140, 310, 834, 503]]<|/det|>
+13 The goal iteration stage plays a central role in the self- evolution of the dynamic knowledge graph. It involves the ROGI Agent initiating the flow of information among the participating agents towards achieving the goal set. This process begins with ROGI Agent creating tasks for the corresponding agents according to the DMTTA cycle, including the DoE Agent, Schedule Agent, and Post- Processing Agent. The DoE Agent perceives the knowledge graph to retrieve prior data and chemical stock available for experiments and then proposes a new experiment. The Schedule Agent evaluates the hardware available in the specified laboratory according to the proposed conditions and subsequently selects the most appropriate hardware to execute the experiment. This is accomplished by generating tasks for the agents responsible for managing the selected digital twin. These agents actuate the equipment to perform reaction and characterisation in the physical world. When the HPLC report is generated, the Post- Processing Agent analyses the chromatogram data to calculate the objectives.  
+
+<|ref|>text<|/ref|><|det|>[[140, 509, 834, 623]]<|/det|>
+25 During the third stage, the ROG Agent utilises the obtained results to determine whether the next iteration should be pursued. To do so, it checks if the Pareto front of the multi- objective fulfils the pre- defined goals and if the resources are still available. The reaction experiment performed in the current iteration is then considered historical data to be included as input for the succeeding round of the Goal Iteration Derivation across all participating SDLs. Afterwards, a new request will be made to the ROGI Agent to start a new iteration, forming a self- evolving feedback loop.  
+
+<|ref|>text<|/ref|><|det|>[[140, 629, 834, 709]]<|/det|>
+32 To ensure correct data dependencies and task execution, we employed the derived information framework [40] to manage the iterative workflow. We implemented each software agent using the derivation agent template provided by the framework. Once deployed, these agents autonomously update the knowledge graph to actively reflect and influence the state of the world.  
+
+<|ref|>text<|/ref|><|det|>[[140, 716, 834, 829]]<|/det|>
+37 This approach enables flexibility and extensibility in the system. As the digital twin of each lab is represented as a node in the knowledge graph, new hardware can be added or removed during the optimisation campaign by simply modifying the list of participating laboratories. The experimental allowance can also be updated when more chemicals become available. The system also supports data sharing across organisations at the very moment the data are generated. Details on the internal logic and technical aspects of the agents in the knowledge graph implementation are available in the Supplementary Information section A.2.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[170, 199, 866, 780]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[150, 797, 880, 821]]<|/det|>
+<center>Figure 5: Autonomous workflow triggered in response to goal requests from scientists as information travelling within the knowledge graph. </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[144, 94, 541, 111]]<|/det|>
+## 2.5 Collaborative closed-loop optimisation  
+
+<|ref|>text<|/ref|><|det|>[[140, 125, 834, 288]]<|/det|>
+2 To demonstrate the scalability and modularity, the knowledge graph approach was applied to a real- time collaborative closed- loop optimisation distributed over two SDLs in Cambridge and Singapore. The objectives selected are run material cost and yield that were sampled for a search space of molar equivalents (relative to benzaldehyde 1) of acetone 2, NaOH 3, residence time and reaction temperature. The research goals and restrictions were populated in the knowledge graph via a web front end. As no prior experimental data was provided, the agents start experiments with random conditions and gradually update their beliefs using TSEMO algorithm [52]. Before running the optimisation, two control conditions were performed to validate the reproducibility across the two setups. For experimental details see Supplementary Information section A.3.  
+
+<|ref|>text<|/ref|><|det|>[[140, 293, 834, 457]]<|/det|>
+Figure 6(a) presents the cost- yield objectives consisting of 65 data points collected during the self- optimisation. Throughout the operation, two SDLs share the results with each other when proposing new experimental conditions. The real- time collaboration demonstrated faster advances in the Pareto front with the highest yield of \(93\%\) . The chemicals used in this study were obtained from different vendors compared to Jeraal et al. [49], the cost is therefore not directly comparable due to different prices. Although not considered in the optimisation, the environment factor and space- time yield were found to be highly correlated to the yield objective. The best values obtained are 26.17 and \(258.175 \mathrm{g L}^{- 1} \mathrm{h}^{- 1}\) when scaled to the same benzaldehyde injection volume (5 mL), both outperformed the previous study [49].  
+
+<|ref|>text<|/ref|><|det|>[[140, 462, 834, 606]]<|/det|>
+Figures 6(b) and 6(c) illustrate the influence of the continuous variables on the cost and yield objectives respectively. The cost is calculated to count for the molar amount of input chemicals sourced from the pumps for the reaction. Therefore, it increases linearly with the molar equivalents of the starting materials. Similarly as identified by Jeraal et al. [49], reaction temperature has a positive correlation with the yield of reaction, whereas the residence time shows a poor correlation. Upon examination of the molar equivalent of acetone 2, it can be observed that its further increase after 30 results in a reduction in yield. This decrease can be attributed to the formation of more side product 5 and other further condensation products of acetone and benzaldehyde.  
+
+<|ref|>text<|/ref|><|det|>[[140, 612, 834, 840]]<|/det|>
+Notably, the Singapore setup encountered an HPLC failure after running for approximately 10 hours. This caused peak shifting of the internal standard which resulted in a wrongly identified peak that gives more than \(3500\%\) yield. This point is considered abnormal by the agents and therefore not utilised in the following DoE. An email notification was sent to the developer for maintenance which took the hardware out of the campaign. The asynchronous and distributed design enabled the Cambridge side to further advance the Pareto front for the cost- yield trade- offs. It is also notable that the product peak was missed for one run at the Cambridge side due to a small shift of the peak which gives a yield of \(0\%\) . This point was taken into consideration in the DoE, but fortunately, it did not affect the final Pareto front as the corrected yield is still Pareto- dominated. The optimisation campaign was stopped since no more significant improvement was observed in terms of hypervolume, and also due to requests for repurposing the equipment for other projects. The complete provenance records (knowledge graph triples) are provided as Supplementary Data. An interactive animation of the optimisation progress is provided as a Supplementary Video.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[330, 167, 660, 401]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[327, 408, 669, 448]]<|/det|>
+<center>(a) Pareto front plot of the yield and cost objectives for the aldol condensation reaction collaboratively optimised by two distributed SDLs. </center>  
+
+<|ref|>image<|/ref|><|det|>[[186, 460, 480, 670]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[175, 678, 485, 731]]<|/det|>
+<center>(b) Three-dimensional plot of the four sampled design variables colour coded for run material cost during the closed-loop optimisation. The size of the dots denotes the molar equivalents of 3 in each run. </center>  
+
+<|ref|>image<|/ref|><|det|>[[515, 460, 811, 670]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[512, 678, 822, 730]]<|/det|>
+<center>(c) Three-dimensional plot of the four sampled design variables colour coded for yield during the closed-loop optimisation. The size of the dots denotes the molar equivalents of 3 in each run. </center>  
+
+<|ref|>text<|/ref|><|det|>[[163, 746, 835, 794]]<|/det|>
+Figure 6: Plots for objectives and design variables of experiments conducted in the distributed closed-loop optimisation campaign. Each dot refers to a single run. The animation of the optimisation progress is available as Supplementary Video.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[150, 92, 312, 111]]<|/det|>
+## 3 Discussion  
+
+<|ref|>text<|/ref|><|det|>[[139, 130, 834, 308]]<|/det|>
+2 In this contribution, we presented a dynamic knowledge graph approach to realise a conceptual architecture for distributed SDLs. We developed ontologies to represent various aspects of chemical knowledge and hardware digital twins involved in a closed- loop optimisation campaign. By employing autonomous agents as executable knowledge components to update and restructure the knowledge graph, we have enabled collaborative management of data and material flow across SDLs. Our approach allows scientists to initiate the autonomous workflow by setting up a goal request, which triggers the flow of information through the knowledge graph as the experimentation workflow progresses. Compared to contemporary designs, where a central orchestrator is responsible for all the data transfer, our approach emphasises information flows in the data layer, reducing the traffic stress on a single point in the network.  
+
+<|ref|>text<|/ref|><|det|>[[139, 315, 833, 427]]<|/det|>
+13 As a proof- of- concept demonstration, we applied the system to an aldol condensation reaction using two setups across different parts of the globe. Despite the differences in configurations, the reaction data produced by both machines were interoperable owing to the layered knowledge abstraction. Throughout the experiment, the system recorded all data provenance as the knowledge graph evolved autonomously, providing opportunities for informed machine learning [53]. Our collaborative approach resulted in faster data generation and advanced the Pareto front while exhibiting resilience to hardware failure.  
+
+<|ref|>text<|/ref|><|det|>[[137, 433, 834, 660]]<|/det|>
+20 The implementation of this work has provided valuable insights and identified areas for future improvement. In terms of orchestration, it is crucial for the system to be robust to network disruption, since it is distributed over the internet. We have implemented measures to ensure that agents deployed in the lab can handle internet cut- offs and resume operations once back online. To minimise downtime during reconnection, future developments could provide on- demand deployment of local agents to continue operation. For efficient optimisation and data quality, it is critical to have control conditions in place when adding new setups to the network, and only those generated results within the tolerance should be approved. To increase the system's robustness against software and hardware malfunctions, regular backups of all data in the central knowledge graph should be implemented. Further development could also be made to federate the SDLs, where each lab hosts its data and digital twins locally and only expose its capabilities in the central registry without revealing confidential information. An authentication and authorisation mechanism should be added to control access to the equipment and grant permission for federated learning.  
+
+<|ref|>text<|/ref|><|det|>[[137, 666, 833, 826]]<|/det|>
+34 Looking forward, achieving a globally collaborative research network requires collective efforts. The knowledge graph reflects a communal understanding of the field, and involving different stakeholders early on can accelerate collaboration and increase the chance of success. Industrial partners are encouraged to work together and provide a unified API for interacting with their proprietary software. This can be facilitated by efforts such as OPC UA [54] and SiLA [27], which aim to establish interoperability standards. Recent studies have shown the successful exchange of HPLC methods between vendors in the Chromatography Data System (CDS), demonstrating the potential for the ontology- based approach [55]. Collaboration between scientists and industry is also important at various stages of research and development [56].  
+
+<|ref|>text<|/ref|><|det|>[[137, 833, 833, 865]]<|/det|>
+44 Overall, we believe the dynamic knowledge graph approach demonstrated in this work provides the first evidence of its potential to establish a network of globally distributed SDLs.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 94, 833, 177]]<|/det|>
+1 Although we focus on flow chemistry in this study, the principles are generic. The same approach can be applied to DMTA cycles for other domains should relevant ontologies and agents be made available, for example, to support research in deep space [17]. We anticipate that this work will serve as an initial step towards the creation of a digital research scientist and eventually move us towards the realisation of "AI Scientists" [36].  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 204, 293, 223]]<|/det|>
+## 4 Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 243, 520, 261]]<|/det|>
+### 4.1 The World Avatar knowledge graph  
+
+<|ref|>text<|/ref|><|det|>[[140, 275, 833, 341]]<|/det|>
+This work follows the best practices in the World Avatar project. All ontologies and agents are version- controlled on GitHub. We provide our thought process during the development below. The same principles can be followed for self- optimisation applications in other domains.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 365, 397, 381]]<|/det|>
+### 4.1.1 Ontology development  
+
+<|ref|>text<|/ref|><|det|>[[138, 396, 833, 655]]<|/det|>
+Developing ontologies is often an iterative process and it is not a goal in and of itself [57]. As suggested in [30, 34, 35], we follow the steps from specifying target deliverables to conceptualising relevant concepts and finally implementing codes for queries. Aimed at capturing data and material flow in distributed SDLs, the relevant concepts range from the reaction experiment to the hardware employed to conduct it. In the World Avatar, ontologies are typically developed to be digested by software agents which mimic the human way of conducting different tasks [58]. Therefore, the development draws inspiration from relevant software tools [49, 59, 60] and existing reaction database schemas [46, 47]. Views of the domain experts [61- 63] are also consulted to better align with the communal understanding of the subject. During iterations, competency questions are used to test if the ontologies meet case study requirements. The answers to these questions are provided in the form of SPARQL queries that are executed by the agents during their operations. Another essential aspect to consider is data instantiation, where we adopted pydantic to simplify the querying and processing of data from the knowledge graph. Overall, the ontology development process starts as easily as drawing concepts and their relationships on a whiteboard and then gradually materialising them in code.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 680, 372, 696]]<|/det|>
+### 4.1.2 Agent development  
+
+<|ref|>text<|/ref|><|det|>[[137, 711, 833, 872]]<|/det|>
+Following the development of ontologies, agents are defined as executables that process inputs and generate outputs. Their I/O signatures are represented following OntoAgent [64]. At the implementation level, all agents inherit the DerivationAgent template in python provided by the derived information framework [40]. Specifically, agents utilise the asynchronous communication mode when interacting with the knowledge graph as conducting experiments is inherently a time- consuming process. Each of the agents monitors the jobs assigned to itself and records the progress of execution in the knowledge graph. The derived information framework does most of the work behind the scenes, leaving the developer with the only task of implementing each agent's internal logic. As agents modify the knowledge graph and subsequently actuate the real world autonomously once active, it is important to
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 95, 833, 128]]<|/det|>
+1 make sure they behave as expected. In this regard, unit and integration tests are provided to 2 help with responsible development.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 152, 407, 169]]<|/det|>
+### 3 4.1.3 Distributed deployment  
+
+<|ref|>text<|/ref|><|det|>[[140, 184, 834, 378]]<|/det|>
+4 Taking inspiration from remote control practices in lab automation [65- 67], the knowledge 5 graph is designed to span across the internet. It follows deployment practices commonly used 6 by cloud- native applications and is implemented through docker containers. The triplestore 7 and file server containing the knowledge statements are deployed at internet- resolvable 8 locations. Depending on capabilities, agents are located at different host machines. Those 9 who monitor and control the hardware are deployed in the corresponding laboratory for 10 security reasons. They transmit data collected from the hardware to the knowledge graph and 11 in reverse configure and actuate the equipment when a new experiment arises. At start- up, 12 agents register their OntoAgent instances in the knowledge graph, then act autonomously 13 should tasks be assigned to them. Altogether, these agents form a distributed network 14 of "digital scientists" that pass on information within the knowledge graph and bridge 15 cyberspace and the physical world.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 402, 432, 420]]<|/det|>
+## 4.2 Flow chemistry platforms  
+
+<|ref|>text<|/ref|><|det|>[[140, 434, 833, 499]]<|/det|>
+17 This work connects two similar automated flow chemistry platforms located in Cambridge 18 and Singapore. The method of sourcing input chemicals differs, with a liquid handler 19 employed in Cambridge and reagent bottles utilised in Singapore. We provide below brief 20 descriptions of the experimental setup. All chemicals were used as received.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 525, 339, 541]]<|/det|>
+## 4.2.1 Cambridge lab  
+
+<|ref|>text<|/ref|><|det|>[[138, 556, 834, 750]]<|/det|>
+22 On the Cambridge side, the experimental setup consists of two Vapourtec R2 pump modules, 23 one Vapourtec R4 reactor module, one Gilson GX- 271 liquid handler, one four- way VICI 24 switching valve (CI4W.06/.5 injector), and Shimadzu CBM- 20A HPLC analytical equipment 25 equipped with Eclipse XDB- C18 column (Agilent part number: 993967- 902). To initiate the 26 reaction, the liquid handler dispenses a \(2\mathrm{mL}\) solution of \(0.5\mathrm{M}\) benzaldehyde 1 dissolved in 27 acetonitrile (with \(0.06\mathrm{M}\) biphenyl as an internal standard) into the sample loop of pump A. 28 Acetone 2 ( \(50\%\) v/v in acetonitrile) and \(0.1\mathrm{M}\) NaOH 3 in ethanol are similarly loaded into 29 sample loops for pump B and C. After being transferred by the switching valve, the product 30 (benzylideneacetone 4) is analysed using online HPLC. The HPLC analysis lasts \(17\mathrm{min}\) 31 with a mobile phase consisting of an \(80:20\) (v/v) binary mixture of water and acetonitrile 32 running at a rate of \(2\mathrm{mL}\mathrm{min}^{- 1}\) . All compounds are detected at an absorption wavelength of 33 \(254\mathrm{nm}\)  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 775, 330, 791]]<|/det|>
+## 4.2.2 Singapore lab  
+
+<|ref|>text<|/ref|><|det|>[[138, 806, 833, 872]]<|/det|>
+35 On the Singapore side, the experimental setup consists of two Vapourtec R2 pump modules, 36 one Vapourtec R4 reactor module, one 6- port 2- position VICI switch valve equipped with 37 \(60\mathrm{nL}\) sampling rotor, and an Agilent 1260 Infinity II system equipped with a G1311B 38 quaternary pump, Eclipse XDB- C18 column (Agilent product number: 961967- 302), and
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 94, 833, 225]]<|/det|>
+1 G1314F variable wavelength detector (VWD). The input chemical for the reaction is sourced from three reagent bottles that are directly attached to the Vapourtec pumps: pump A contains \(0.5\mathrm{M}\) benzaldehyde 1 in acetonitrile (with \(0.05\mathrm{M}\) naphthalene as an internal standard), pump B contains \(6.73\mathrm{M}\) acetone 2 in acetonitrile ( \(50\%\) v/v in acetonitrile), and pump C contains \(0.1\mathrm{M}\) NaOH 3 in ethanol. The HPLC quaternary pump method for online HPLC described by Jeraal et al. [49] is used. However, the VWD wavelength was changed differently over the 8 min analysis time as follows: the absorption wavelength is \(248\mathrm{nm}\) for the initial \(6.05\mathrm{min}\) and then switched to \(228\mathrm{nm}\) until the end of acquisition.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 254, 312, 272]]<|/det|>
+## 9 Research data  
+
+<|ref|>text<|/ref|><|det|>[[140, 292, 833, 325]]<|/det|>
+10 Research data supporting this publication is available in the University of Cambridge data repository (doi:10.17863/CAM.97058).  
+
+<|ref|>text<|/ref|><|det|>[[140, 330, 833, 444]]<|/det|>
+11 All the codes developed are available on The World Avatar GitHub repository https://github.com/cambridge- cares/TheWorldAvatar. The docker images of agents are available at GitHub's public registry located at ghcr.io/cambridge- cares:/doe_agent:1.2.0,vapourtec_schedule_agent:1.2.0, vapourtec_agent:1.2.0, hplc_agent:1.2.0, hplc_postpro_agent:1.2.0, rxn_opt_goal_iter_agent:1.2.0, and rxn_opt_goal_agent:1.0.0. The deployment instructions can be found in folder TheWorldAvatar/Deploy/pips.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 473, 366, 492]]<|/det|>
+## 10 Acknowledgements  
+
+<|ref|>text<|/ref|><|det|>[[137, 511, 833, 770]]<|/det|>
+20 This research was supported by the National Research Foundation, Prime Minister's Office, 21 Singapore, under its Campus for Research Excellence and Technological Enterprise (CRE- 22 ATE) programme, and Pharma Innovation Platform Singapore (PIPS) via grant to CARES 23 Ltd "Data2Knowledge, C12". This project was cofunded by European Regional Develop- 24 ment Fund via the project "Innovation Centre in Digital Molecular Technologies", UKRI 25 via project EP/S024220/1 "EPSRC Centre for Doctoral Training in Automated Chemical 26 Synthesis Enabled by Digital Molecular Technologies". Part of this work was also supported 27 by Towards Turing 2.0 under the EPSRC Grant EP/W037211/1. The authors thank Dr. 28 Andrew C. Breeson for his helpful suggestions on graphical design. J. Bai acknowledges 29 financial support provided by CSC Cambridge International Scholarship from Cambridge 30 Trust and China Scholarship Council. C.J. Taylor is a Sustaining Innovation Postdoctoral 31 Research Associate at Astex Pharmaceuticals and thanks Astex Pharmaceuticals for funding, 32 as well as his Astex colleagues Mark Wade, Gianni Chessari, and David Rees for their 33 support. S.D. Rihm acknowledges financial support from Fitzwilliam College, Cambridge, 34 and the Cambridge Trust. M. Kraft gratefully acknowledges the support of the Alexander 35 von Humboldt Foundation.  
+
+<|ref|>text<|/ref|><|det|>[[137, 776, 833, 810]]<|/det|>
+36 For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising.  
+
+<|ref|>text<|/ref|><|det|>[[137, 814, 833, 863]]<|/det|>
+37 Author contributions: M.K., A.A.L., J.B., and S.M. conceived the project. J.B., S.M. and M.K. designed the ontological representation and agent workflow. J.B. implemented the ontologies/agents and deployed the knowledge graph under the advisement of S.M.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 94, 835, 214]]<|/det|>
+1 and K.F.L. The chemistry and HPLC method were developed by C.J.T. (Cambridge) and 2 D.K. (Singapore). D.K. and C.J.T. validated the calculation of objective functions. The 3 self- optimisation campaign involving two labs was set up by C.J.T. (hardware and chemicals 4 in Cambridge), D.K. (hardware and chemicals in Singapore) and J.B. (software on both sides 5 and goal request). M.K. and A.A.L. acquired funding and administrated the project. J.B. 6 drafted the body of this manuscript and SI with inputs from S.M., J.A., S.D.R. and C.J.T. All 7 authors provided feedback on the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[140, 216, 655, 232]]<|/det|>
+8 Competing interests: The authors declare no competing interests.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[150, 92, 279, 110]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[135, 120, 838, 870]]<|/det|>
+[1] Florian Hase, Loic M Roch, and Alan Aspuru- Guzik. Next- Generation Experimentation with Self- Driving Laboratories. Trends Chem., 1(3):282- 291, 2019. doi:10.1016/j.trechm.2019.02.007.  [2] Milad Abolhasani and Eugenia Kumacheva. The Rise of Self- Driving Labs in Chemical and Materials Sciences. Nat. Synth., 2023. doi:10.1038/s44160- 022- 00231- 0.  [3] Robert B Merrifield, John Morrow Stewart, and Nils Jernberg. Instrument for Automated Synthesis of Peptides. Anal. Chem., 38(13):1905- 1914, 1966. doi:10.1021/ac50155a057.  [4] Connor W Coley, Dale A Thomas, Justin A M Lummis, Jonathan N Jaworski, Christopher P Breen, Victor Schultz, Travis Hart, Joshua S Fishman, Luke Rogers, Hanyu Gao, Robert W. Hicklin, Pieter P. Plehiers, Joshua Byington, John S. Piotti, William H. Green, A. John Hart, Timothy F. Jamison, and Klavs F. Jensen. A Robotic Platform for Flow Synthesis of Organic Compounds Informed by AI Planning. Science, 365(6453): eaax1566, 2019. doi:10.1126/science.aax1566.  [5] Sebastian Steiner, Jakob Wolf, Stefan Glatzel, Anna Andreou, Jaroslaw M Granda, Graham Keenan, Trevor Hinkley, Gerardo Aragon- Camarasa, Philip J Kitson, Davide Angelone, and Leroy Cronin. Organic Synthesis in a Modular Robotic System Driven by a Chemical Programming Language. Science, 363(6423):eaav2211, 2019. doi:10.1126/science.aav2211.  [6] Benjamin Burger, Phillip M Maffettone, Vladimir V Gusev, Catherine M Aitchison, Yang Bai, Xiaoyan Wang, Xiaobo Li, Ben M Alston, Buyi Li, Rob Clowes, Nicola Rankin, Brandon Harris, Reiner Sebastian Sprick, and Andrew I. Cooper. A Mobile Robotic Chemist. Nature, 583(7815):237- 241, 2020. doi:10.1038/s41586- 020- 2442- 2.  [7] Sourav Chatterjee, Mara Guidi, Peter H Seeberger, and Kerry Gilmore. Automated Radial Synthesis of Organic Molecules. Nature, 579(7799):379- 384, 2020. doi:10.1038/s41586- 020- 2083- 5.  [8] Daniel P Tabor, Loic M Roch, Semion K Saikin, Christoph Kreisbeck, Dennis Sheberla, Joseph H Montoya, Shyam Dwaraknath, Muratahan Aykol, Carlos Ortiz, Hermann Tribukait, Carlos Amador- Bedolla, Christoph J. Brabec, Benji Maruyama, Kristin A. Persson, and Alan Aspuru- Guzik. Accelerating the Discovery of Materials for Clean Energy in the Era of Smart Automation. Nat. Rev. Mater., 3(5):5- 20, 2018. doi:10.1038/s41578- 018- 0005- z.  [9] Qing Zhu, Fei Zhang, Yan Huang, Hengyu Xiao, LuYuan Zhao, XuChun Zhang, Tao Song, XinSheng Tang, Xiang Li, Guo He, BaoChen Chong, JunYi Zhou, YiHan Zhang, Baicheng Zhang, JiaQi Cao, Man Luo, Song Wang, GuiLin Ye, WanJun Zhang, Xin Chen, Shuang Cong, Donglai Zhou, Huirong Li, Jialei Li, Gang Zou, WeiWei Shang, Jun Jiang, and Yi Luo. An All- Round AI- Chemist with a Scientific Mind. Natl. Sci. Rev., 9(10), 2022. doi:10.1093/nsr/nwac190.  [10] Ross D King, Jem Rowland, Stephen G Oliver, Michael Young, Wayne Aubrey, Emma Byrne, Maria Liakata, Magdalena Markham, Pinar Pir, Larisa N Soldatova, Andrew
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 95, 836, 129]]<|/det|>
+1 Sparkes, Kenneth E. Whelan, and Amanda Clare. The Automation of Science. Science, 324(5923):85- 89, 2009. doi:10.1126/science.1165620.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[135, 92, 836, 842]]<|/det|>
+1 by Artificial Intelligence. Nat. Mach. Intell., 5(3):208- 219, 2023. doi:10.1038/s42256- 023- 00618- 4. 3 [18] Edward O Pyzer- Knapp, Jed W Pitera, Peter WJ Staar, Seiji Takeda, Teodoro Laino, Daniel P Sanders, James Sexton, John R Smith, and Alessandro Curioni. Accelerating Materials Discovery Using Artificial Intelligence, High Performance Computing and Robotics. npj Comput. Mater., 8(1):84, 2022. doi:10.1038/s41524- 022- 00765- z. 7 [19] Connor W Coley, Natalie S Eyke, and Klavs F Jensen. Autonomous Discovery in the Chemical Sciences Part II: Outlook. Angew. Chem., Int. Ed., 59(52):23414- 23436, 2020. doi:10.1002/anie.201909989. 10 [20] Liam Wilbraham, S Hessam M Mehr, and Leroy Cronin. Digitizing Chemistry Using the Chemical Processing Unit: From Synthesis to Discovery. Acc. Chem. Res., 54(2): 253- 262, 2021. doi:10.1021/acs.accounts.0c00674. 13 [21] Mark D. Wilkinson, Michel Dumontier, IJsbrand Jan Aalbersberg, Gabrielle Appleton, Myles Axton, Arie Baak, Niklas Blomberg, Jan- Willem Boiten, Luiz Bonino da Silva Santos, Philip E. Bourne, Jildau Bouwman, Anthony J. Brookes, Tim Clark, Merce Crosas, Ingrid Dillo, Olivier Dumon, Scott Edmunds, Chris T. Evelo, Richard Finkers, Alejandra Gonzalez- Beltran, Alasdair J.G. Gray, Paul Groth, Carole Goble, Jeffrey S. Grethe, Jaap Heringa, Peter A.C 't Hoen, Rob Hooft, Tobias Kuhn, Ruben Kok, Joost Kok, Scott J. Lusher, Maryann E. Martone, Albert Mons, Abel L. Packer, Bengt Persson, Philippe Rocca- Serra, Marco Roos, Rene van Schaik, Susanna- Assunta Sansone, Erik Schultes, Thierry Sengstag, Ted Slater, George Strawn, Morris A. Swertz, Mark Thompson, Johan van der Lei, Erik van Mulligen, Jan Velterop, Andra Waagmeester, Peter Wittenburg, Katherine Wolstencroft, Jun Zhao, and Barend Mons. The FAIR Guiding Principles for Scientific Data Management and Stewardship. Sci. Data, 3:160018, 2016. doi:10.1038/sdata.2016.18. 26 [22] Loic M Roch, Florian Hase, Christoph Kreisbeck, Teresa Tamayo- Mendoza, Lars P E Yunker, Jason E Hein, and Alan Aspuru- Guzik. ChemOS: Orchestrating Autonomous Experimentation. Sci. Robot., 3(19):eaat5559, 2018. doi:10.1126/sicrobotics.aat5559. 29 [23] Ian M Pendleton, Gary Cattabriga, Zhi Li, Mansoor Ani Najeeb, Sorelle A Friedler, Alexander J Norquist, Emory M Chan, and Joshua Schrier. Experiment Specification, Capture and Laboratory Automation Technology (ESCALATE): A Software Pipeline for Automated Chemical Experimentation and Data Management. MRS Commun., 9 (3):846- 859, 2019. doi:10.1557/mrc.2019.72. 34 [24] Fuzhan Rahmanian, Jackson Flowers, Dan Guevarra, Matthias Richter, Maximilian Fichtner, Phillip Donnely, John M Gregoire, and Helge S Stein. Enabling Modular Autonomous Feedback- Loops in Materials Science through Hierarchical Experimental Laboratory Automation and Orchestration. Adv. Mater. Interfaces, 9(8):2101987, 2022. doi:10.1002/admi.202101987. 39 [25] S Hessam M Mehr, Matthew Craven, Artem I Leonov, Graham Keenan, and Leroy Cronin. A Universal System for Digitization and Automatic Execution of the Chemical Synthesis Literature. Science, 370(6512):101- 108, 2020. doi:10.1126/science.abc2986.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 95, 835, 144]]<|/det|>
+[26] Jason Hein, Robert Rauschen, Mason Guy, and Leroy Cronin. Universal Chemical Programming Language for Robotic Synthesis Reproducibility, 2023. Research Square Platform LLC. doi:10.21203/rs.3. rs- 2761997/v1.  
+
+<|ref|>text<|/ref|><|det|>[[140, 154, 835, 202]]<|/det|>
+[27] Burkhard Schäfer. Data Exchange in the Laboratory of the Future – A Glimpse at AnIML and SiLA, 2018. URL https://analyticalscience.wiley.com/do/10.1002/gitlab.17270/full/. Accessed 30 May 2023.  
+
+<|ref|>text<|/ref|><|det|>[[140, 212, 835, 390]]<|/det|>
+[28] Sonia Natalie Mitchell, Andrew Lahiff, Nathan Cummings, Jonathan Hollocombe, Bram Boskamp, Ryan Field, Dennis Reddyhoff, Kristian Zarebski, Antony Wilson, Bruno Viola, Martin Burke, Blair Archibald, Paul Bessell, Richard Blackwell, Lisa A. Boden, Alys Brett, Sam Brett, Ruth Dundas, Jessica Enright, Alejandra N. Gonzalez- Beltran, Claire Harris, Ian Hinder, Christopher David Hughes, Martin Knight, Vino Mano, Ciaran McMonagle, Dominic Mellor, Sibylle Mohr, Glenn Marion, Louise Matthews, Iain J. McKendrick, Christopher Mark Pooley, Thibaud Porphyre, Aaron Reeves, Edward Townsend, Robert Turner, Jeremy Walton, and Richard Reeve. FAIR Data Pipeline: Provenance- Driven Data Management for Traceable Scientific Workflows. Philos. Trans. R. Soc., A, 380(2233):20210300, 2022. doi:10.1098/rsta.2021.0300.  
+
+<|ref|>text<|/ref|><|det|>[[140, 400, 835, 450]]<|/det|>
+[29] Angiras Menon, Nenad B Krdzavac, and Markus Kraft. From Database to Knowledge Graph—Using Data in Chemistry. Curr. Opin. Chem. Eng., 26:33- 37, 2019. doi:10.1016/j.coche.2019.08.004.  
+
+<|ref|>text<|/ref|><|det|>[[140, 459, 835, 508]]<|/det|>
+[30] Jiaru Bai, Liwei Cao, Sebastian Mosbach, Jethro Akroyd, Alexei A Lapkin, and Markus Kraft. From Platform to Knowledge Graph: Evolution of Laboratory Automation. JACS Au, 2(2):292- 309, 2022. doi:10.1021/jacsau.1c00438.  
+
+<|ref|>text<|/ref|><|det|>[[140, 518, 835, 599]]<|/det|>
+[31] Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia D'Amato, Gerard De Melo, Claudio Gutierrez, Sabrina Kirrane, José Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, Axel- Cyrille Ngonga Ngomo, Axel Polleres, Sabbir M. Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequeda, Steffen Staab, and Antoine Zimmermann. Knowledge Graphs. ACM Comput. Surv., 54(4):1- 37, 2022. doi:10.1145/3447772.  
+
+<|ref|>text<|/ref|><|det|>[[140, 609, 835, 658]]<|/det|>
+[32] Jethro Akroyd, Sebastian Mosbach, Amit Bhave, and Markus Kraft. Universal Digital Twin – A Dynamic Knowledge Graph. Data- Centric Engineering, 2:e14, 2021. doi:10.1017/dce.2021.10.  
+
+<|ref|>text<|/ref|><|det|>[[140, 668, 835, 749]]<|/det|>
+[33] Mei Qi Lim, Xiaonan Wang, Oliver Inderwildi, and Markus Kraft. The World Avatar – A World Model for Facilitating Interoperability. In Intelligent Decarbonisation: Can Artificial Intelligence and Cyber- Physical Systems Help Achieve Climate Mitigation Targets?, pages 39- 53. Springer International Publishing, 2022. doi:10.1007/978- 3- 030- 86215- 2_4.  
+
+<|ref|>text<|/ref|><|det|>[[140, 759, 835, 808]]<|/det|>
+[34] Simon D. Rihm, Jiaru Bai, Aleksandar Kondinski, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. A Comprehensive Laboratory Digital Twin: Top- Down Automation of a Chemical Research Laboratory, 2023. In preparation.  
+
+<|ref|>text<|/ref|><|det|>[[140, 818, 835, 865]]<|/det|>
+[35] Aleksandar Kondinski, Jiaru Bai, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. Knowledge Engineering in Chemistry: From Expert Systems to Agents of Creation. Acc. Chem. Res., 56(2):128- 139, 2023. doi:10.1021/acs.accounts.2c00617.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[135, 92, 837, 125]]<|/det|>
+[36] Hiroaki Kitano. Nobel Turing Challenge: Creating the Engine for Scientific Discovery. npj Syst. Biol. Appl., 7(1):1- 12, 2021. doi:10.1038/s41540- 021- 00189- 3. 
+
+<|ref|>text<|/ref|><|det|>[[135, 137, 835, 184]]<|/det|>
+[37] Connor W Coley, Natalie S Eyke, and Klavs F Jensen. Autonomous Discovery in the Chemical Sciences Part I: Progress. Angew. Chem., Int. Ed., 59(51):22858- 22893, 2020. doi:10.1002/anie.201909987. 
+
+<|ref|>text<|/ref|><|det|>[[135, 195, 835, 259]]<|/det|>
+[38] Connor J Taylor, Alexander Pomberger, Kobi C Felton, Rachel Grainger, Magda Barecka, Thomas W Chamberlain, Richard A Bourne, Christopher N Johnson, and Alexei A Lapkin. A Brief Introduction to Chemical Reaction Optimization. Chem. Rev., 123(6):3089- 3126, 2023. doi:10.1021/acs.chemrev.2c00798. 
+
+<|ref|>text<|/ref|><|det|>[[135, 270, 835, 318]]<|/det|>
+[39] Andy B. Yoo, Morris A. Jette, and Mark Grondona. SLURM: Simple Linux Utility for Resource Management. In Job Scheduling Strategies for Parallel Processing, pages 44- 60. Springer Berlin Heidelberg, 2003. doi:10.1007/10968987_3. 
+
+<|ref|>text<|/ref|><|det|>[[135, 329, 835, 392]]<|/det|>
+[40] Jiaru Bai, Kok Foong Lee, Markus Hofmeister, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. A Derived Information Framework for a Dynamic Knowledge Graph and its Application to Smart Cities, 2023. Submitted for publication. Preprint available from https://como.ceb.cam.ac.uk/preprints/302/. 
+
+<|ref|>text<|/ref|><|det|>[[135, 403, 835, 435]]<|/det|>
+[41] Ian Holland and Jamie A. Davies. Automation in the Life Science Research Laboratory. Front. Bioeng. Biotechnol., 8:571777, 2020. doi:10.3389/fbioe.2020.571777. 
+
+<|ref|>text<|/ref|><|det|>[[135, 446, 835, 523]]<|/det|>
+[42] Anand S. Rao and Michael P. Georgeff. Modeling Rational Agents within a BDI- Architecture. In Proceedings of the Second International Conference on Principles of Knowledge Representation and Reasoning, KR'91, page 473- 484, San Francisco, CA, USA, 1991. Morgan Kaufmann Publishers Inc. ISBN 1558601651. doi:10.5555/3087158.3087205. 
+
+<|ref|>text<|/ref|><|det|>[[135, 535, 835, 582]]<|/det|>
+[43] Helge S Stein and John M Gregoire. Progress and Prospects for Accelerating Materials Science with Automated and Autonomous Workflows. Chem. Sci., 10(42):9640- 9649, 2019. doi:10.1039/C9SC03766G. 
+
+<|ref|>text<|/ref|><|det|>[[135, 593, 835, 640]]<|/det|>
+[44] Michel Krämer, Hendrik M Würz, and Christian Altenhofen. Executing Cyclic Scientific Workflows in the Cloud. J. Cloud Comput., 10(1):1- 26, 2021. doi:10.1186/s13677- 021- 00229- 7. 
+
+<|ref|>text<|/ref|><|det|>[[135, 652, 835, 699]]<|/det|>
+[45] Jan Morbach, Aidong Yang, and Wolfgang Marquardt. OntoCAPE - A Large-Scale Ontology for Chemical Process Engineering. Eng. Appl. Artif. Intell., 20(2):147- 161, 2007. doi:10.1016/j.engappai.2006.06.010. 
+
+<|ref|>text<|/ref|><|det|>[[135, 710, 835, 773]]<|/det|>
+[46] Steven M Kearnes, Michael R Maser, Michael Wleklinski, Anton Kast, Abigail G Doyle, Spencer D Dreher, Joel M Hawkins, Klavs F Jensen, and Connor W Coley. The Open Reaction Database. J. Am. Chem. Soc., 143(45):18820- 18826, 2021. doi:10.1021/jacs.1c09820. 
+
+<|ref|>text<|/ref|><|det|>[[135, 784, 833, 816]]<|/det|>
+[47] Pistoia Alliance. Unified Data Model, 2020. URL https://github.com/Pistoia Alliance/UDM. Accessed 30 May 2023. 
+
+<|ref|>text<|/ref|><|det|>[[135, 827, 835, 874]]<|/det|>
+[48] Laura Daniele, Raúl Garcia-Castro, Maxime Lefrançois, and Maria Poveda-Villalon. SAREF: The Smart Applications REference ontology, 2020. URL https://saref. etsi.org/core/v3.1.1/. Accessed 21 Feb 2023.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[135, 92, 836, 905]]<|/det|>
+[49] Mohammed I Jeraal, Simon Sung, and Alexei A Lapkin. A Machine Learning- Enabled Autonomous Flow Chemistry Platform for Process Optimization of Multiple Reaction Metrics. Chem. Methods, 1(1):71- 77, 2021. doi:10.1002/cmtd.202000044. [50] Jean Soon Park, Ping Hu, Yakang Lin, and Lori Ann Reinsalu. Composition and Method for Preventing, Reducing, Alleviating or Treating Idiopathic Vomiting, February 4 2020. URL https://testpubchem.ncbi.nlm.nih.gov/patent/US- 10548935- B2. US Patent 10,548,935. Accessed 11 Feb 2023. [51] Laura Pascazio, Simon D. Rihm, Ali Naseri, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. A Chemical Species Ontology for Data Integration and Knowledge Discovery, 2023. Submitted for publication. Preprint available from https://como .ceb.cam.ac.uk/preprints/306/. [52] Eric Bradford, Artur M Schweidtmann, and Alexei Lapkin. Efficient Multiobjective Optimization Employing Gaussian Processes, Spectral Sampling and A Genetic Algorithm. J. Glob. Optim., 71(2):407- 438, 2018. doi:10.1007/s10898- 018- 0609- 2. [53] Laura von Rueden, Sebastian Mayer, Katharina Beckh, Bogdan Georgiev, Sven Gieselbach, Raoul Heese, Birgit Kirsch, Michal Walczak, Julius Pfrommer, Annika Pick, Rajkumar Ramamurthy, Jochen Garcke, Christian Bauckhage, and Jannis Schuecker. Informed Machine Learning - A Taxonomy and Survey of Integrating Prior Knowledge into Learning Systems. IEEE Trans. Knowl. Data Eng., 35(1):614- 633, 2021. doi:10.1109/TKDE.2021.3079836. [54] OPC Foundation. Unified Architecture, 2023. URL https://opcfoundation.org/g/about/opc- technologies/opc- ua/. Accessed 8 March 2023. [55] Pistoia Alliance. Update from the Pistoia Alliance's Methods Hub Project, 2022. URL https://www.pistoiaalliance.org/methods/april- 2022- methods- database- hplc- uv- methods/. Accessed 10 Feb 2023. [56] Melodie Christensen, Lars PE Yunker, Parisa Shiri, Tara Zepel, Paloma L Prieto, Shad Grunert, Finn Bork, and Jason E Hein. Automation isn't Automatic. Chem. Sci., 12 (47):15473- 15490, 2021. doi:10.1039/D1SC04588A. [57] Steffen Staab and Rudi Studer. Handbook on Ontologies. Springer Science & Business Media, 2010. [58] Aleksandar Kondinski, Angiras Menon, Daniel Nurkowski, Feroz Farazi, Sebastian Mosbach, Jethro Akroyd, and Markus Kraft. Automated Rational Design of Metal- Organic Polyhedra. J. Am. Chem. Soc., 144(26):11713- 11728, 2022. doi:10.1021/jacs.2c03402. [59] Jiang Guo, A Santiago Ibanez- Lopez, Hanyu Gao, Victor Quach, Connor W Coley, Klavs F Jensen, and Regina Barzilay. Automated Chemical Reaction Extraction from Scientific Literature. J. Chem. Inf. Model., 2021. doi:10.1021/acs.jcim.1c00284. [60] Kobi C Felton, Jan G Rittig, and Alexei A Lapkin. Summit: Benchmarking Machine Learning Methods for Reaction Optimisation. Chemistry- Methods, 1(2):116- 122, 2021. doi:10.1002/cmtd.202000051.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[135, 95, 837, 560]]<|/det|>
+[61] Matthew B Plutschack, Bartholomäus Pieber, Kerry Gilmore, and Peter H Seeberger. The Hitchhiker's Guide to Flow Chemistry. Chem. Rev., 117(18):11796- 11893, 2017. doi:10.1021/acs.chemrev.7b00183. [62] Sushant S Garud, Iftekhar A Karimi, and Markus Kraft. Design of Computer Experiments: A Review. Comput. Chem. Eng., 106:71- 95, 2017. doi:10.1016/j.compchemeng.2017.05.010. [63] Adam D Clayton, Jamie A Manson, Connor J Taylor, Thomas W Chamberlain, Brian A Taylor, Graeme Clemens, and Richard A Bourne. Algorithms for the SelfOptimisation of Chemical Reactions. React. Chem. Eng., 4(9):1545- 1554, 2019. doi:10.1039/C9RE00209J. [64] Xiaochi Zhou, Andreas Eibeck, Mei Qi Lim, Nenad B Krdzavac, and Markus Kraft. An Agent Composition Framework for the J- Park Simulator - A Knowledge Graph for the Process Industry. Comput. Chem. Eng., 130:106577, 2019. doi:10.1016/j.compchemeng.2019.106577. [65] Daniel E Fitzpatrick, Timothé Maujean, Amanda C Evans, and Steven V Ley. Across- the- World Automated Optimization and Continuous- Flow Synthesis of Pharmaceutical Agents Operating through a Cloud- Based Server. Angew. Chem., Int. Ed., 57(46):15128- 15132, 2018. doi:10.1002/anie.201809080. [66] Juan Ramírez, David Soto, Sebastian López, Jethro Akroyd, Daniel Nurkowski, Maria L. Botero, Nicola Bianco, George Brownbridge, Markus Kraft, and Alejandro Molina. A Virtual Laboratory to Support Chemical Reaction Engineering Courses Using Real- Life Problems and Industrial Software. Educ. Chem. Eng., 33:36- 44, 2020. doi:10.1016/j.ece.2020.07.002. [67] CMU Cloud Lab, 2023. URL https://cloudlab.cmu.edu/. Accessed 12 May 2023.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[44, 93, 765, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 131, 220, 177]]<|/det|>
+suppinfo.pdf  suppvideo.html
+
+<--- Page Split --->

preprint/preprint__2b5449656c9a8636b77273dc85ce96cbe00a6238ddd32058f9565878862bd45c/images_list.json

@@ -0,0 +1,77 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1: Structure of the microcavity and steady-state spectra. a Schematic illustration of the mixed (6,5)/(7,5) CNT microcavity investigated in this work. b Linear absorption spectra of the uncoupled CNTs measured in 1,2-dichlorobenzene, each normalized to the \\(\\mathrm{S}_{11}\\) peak maximum. The two peaks labeled with asterisks denote minor impurities in the (7,5) CNTs from residual (7,6) \\((*,1.09\\mathrm{eV})\\) and (6,5) \\((^{**},1.23\\mathrm{eV})\\) chiralities \\(^{32}\\) . c Angle-dependent reflectance \\((R)\\) spectra of the microcavity shown in a. Dashed lines illustrate the uncoupled CNT state energies as shown in b, where the energy levels of the two most intense transitions, i.e., those that couple most strongly to light, are highlighted in purple ((6,5)) and green ((7,5)). Black solid lines are coupled oscillator model fits to the measured spectra, and the yellow solid line shows the cavity dispersion profile determined from the model (see Supplementary Note 3).",
+    "footnote": [],
+    "bbox": [
+      [
+        268,
+        153,
+        730,
+        560
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2: 2DWL spectra and energy relaxation dynamics in the microcavities. a–c Cartoon illustrations of the (6,5)- (a), (7,5)- (b) and mixed (6,5)/(7,5) microcavities (c). d–f 2DWL spectra of the microcavities at waiting times \\((T)\\) of 100 fs (top) and 500 fs (bottom). The spectra are normalized to the maximum magnitude of the 2D signal in the \\(T = 100\\) fs spectrum for each sample. Ground-state bleach/stimulated emission and excited-state absorption are plotted in negative and positive signs, respectively. g–i Normalized waiting time traces (colored solid lines) generated at the peak positions labeled with open squares in d–f. Black dashed lines show the fit curves for the UP diagonal and UP/LP peaks. All traces are plotted in absolute values, i.e., traces for negative peaks are plotted with signs flipped. See also Supplementary Figs. 16–21 for additional data.",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        135,
+        884,
+        603
+      ]
+    ],
+    "page_idx": 7
+  },
+  {
+    "type": "image",
+    "img_path": "images/Supplementary_Figure_13.jpg",
+    "caption": "Figure 3: Simulated transient reflection spectra and Rabi contraction dynamics. Transient reflection spectra for the (6,5)- (a), (7,5)- (b), and mixed (6,5)/(7,5) microcavities (c) simulated using the transfer matrix method. Measured transient reflection spectra at \\(T = 100\\) fs are overlaid in light blue for comparison. The dark-to-light color gradient in each panel illustrates the gradual decrease of the bleached population over time (see Supplementary Fig. 13 for the exact values and also Supplementary Note 2.2). d–f Measured kinetic traces (light blue solid lines) of the LP population overlaid with the pseudo-time trace generated from the simulated spectra (black dashed lines). The measured traces are obtained from transient reflection spectra for d, e, and from 2DWL spectra for f to separately plot the diagonal (light blue) and cross peak (green) contributions. The simulated pseudo-time traces are obtained by taking slices at the wavelengths labeled with black arrows in a–c. All traces are plotted in absolute values, i.e., traces for negative peaks are plotted with signs flipped.",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        151,
+        881,
+        521
+      ]
+    ],
+    "page_idx": 10
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4: Calculated spectra, energy levels, and dynamics of the microcavities. a–c Calculated linear \\(1 - R\\) spectra of the (6,5)- (a), (7,5)- (b) and mixed (6,5)/(7,5) (c) microcavities. In c, the diagonal slice from the measured 2DWL spectrum in Fig. 2f, top (light blue line) is overlaid with the calculated spectrum (black line) with \\(J = -10 \\mathrm{meV}\\) and \\(g = 42.9 \\mathrm{meV}\\) , which yielded the best match. These \\(J\\) and \\(g\\) values were then applied to generate the spectra for the single band-gap microcavities shown as black lines in a, b. The vertical bars plot the individual eigenstates with heights proportional to the transition dipole strength, color coded to illustrate the composition of each eigenstate. d–f Energy level diagrams. Bright states are shown with color codes as in a–c, and dark states (states with transition dipoles smaller than \\(0.01\\%\\) of the maximum dipole strength in each case) are shown in gray. g–i Calculated LP population dynamics upon photoexcitation of the UP band. In i, the two traces plot the dynamics upon pumping states with high- (yellow) and low (purple) photon content, also illustrated as colored arrows in c. For comparison, the experimental UP/LP cross peak trace is overlaid as black open circles.",
+    "footnote": [],
+    "bbox": [
+      [
+        150,
+        90,
+        840,
+        576
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Supplementary_Figure_14.jpg",
+    "caption": "Figure 5: System parameters affecting the energy landscape and population dynamics. a–c Energy level diagrams of the mixed (6,5)/(7,5) microcavity in the absence of inter-tube coupling \\((J)\\) (a), when both \\(J\\) and light-matter coupling \\((g)\\) are present with comparable amounts (b), and when \\(g\\) is significantly larger than \\(J\\) (c). Bright states are color coded to illustrate the contribution from the cavity photon and the CNTs, and dark states (states with transition dipoles smaller than \\(0.01\\%\\) of the maximum dipole strength in each case) are shown in gray. The specific \\(J\\) and \\(g\\) values used for each case are displayed on top. See also Supplementary Fig. 14 for the simulated spectra. b is a replicate of the energy level diagram in Fig. 4c plotted on a different energy axis. d–f show the corresponding relaxation dynamics of the LP population upon photoexcitation into the UP band. The gray and purple traces plot the contribution from the dark states and the UP-to-LP transfer, i.e., population that would appear as a UP/LP cross peak in the experiment, respectively. The purple trace in e is a replicate of the purple trace shown in Fig. 4i, and is overlaid with the experimental UP/LP cross peak trace (open circles).",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        140,
+        880,
+        504
+      ]
+    ],
+    "page_idx": 16
+  }
+]

preprint/preprint__2b5449656c9a8636b77273dc85ce96cbe00a6238ddd32058f9565878862bd45c/preprint__2b5449656c9a8636b77273dc85ce96cbe00a6238ddd32058f9565878862bd45c.mmd

@@ -0,0 +1,355 @@
+
+# Energy cascades in polaritonic systems with energetic disorder observed by ultrafast two-dimensional white-light spectroscopy  
+
+Minjung Son University of Wisconsin- Madison https://orcid.org/0000- 0002- 8385- 062X  
+
+Zachary Armstrong University of Wisconsin- Madison  
+
+Ryan Allen University of Wisconsin- Madison  
+
+Abitha Dhavamani University of Wisconsin- Madison  
+
+Michael Arnold University of Wisconsin- Madison https://orcid.org/0000- 0002- 2044- 7032  
+
+Martin Zanni ( zanni@chem.wisc.edu )  
+
+University of Wisconsin- Madison https://orcid.org/0000- 0001- 7191- 9768  
+
+Article  
+
+Keywords:  
+
+Posted Date: May 4th, 2022  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 1582812/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: Yes there is potential Competing Interest. Martin T. Zanni is a co- owner of PhaseTech Spectroscopy, which sells ultrafast pulse shapers and multidimensional spectrometers.  
+
+Version of Record: A version of this preprint was published at Nature Communications on November 27th, 2022. See the published version at https://doi.org/10.1038/s41467- 022- 35046- 2.
+
+<--- Page Split --->
+
+# Energy cascades in polaritonic systems with energetic disorder observed by ultrafast two-dimensional white-light spectroscopy  
+
+3 Minjung Son \(^{1,\dagger}\) , Zachary T. Armstrong \(^{1,\dagger}\) , Ryan T. Allen \(^{1}\) , Abitha Dhavamani \(^{2}\) , Michael S. Arnold \(^{2}\) & Martin T. Zanni \(^{1,*}\)  
+
+5 \(^{1}\) Department of Chemistry, University of Wisconsin- Madison, 1101 University Ave, Madison, WI 6 53706, USA 7 \(^{2}\) Department of Materials Science and Engineering, University of Wisconsin- Madison, 1509 Uni- 8 versity Ave, Madison, WI 53706, USA  
+
+## Abstract  
+
+10 Exciton-polaritons are hybrid states formed when molecular excitons are strongly coupled to 11 photons trapped in an optical cavity. These systems have many attractive properties, including 12 large delocalization lengths, but questions regarding the role of energetic disorder remain 13 unanswered. Here, we fabricate microcavities with two different layers of semiconducting 14 carbon nanotubes as a way of controlling the energetic disorder and exploring its impact 15 on energy transfer. Using ultrafast two- dimensional white- light spectroscopy, we observe a 16 delayed growth of a cross peak between the upper- and lower- polariton states. Using Redfield 17 theory, we assign the growth to cascading energy transfer down a manifold of new electronic 18 states created by energetic disorder that is of comparable magnitude to the light- matter 19 coupling. These results broaden our understanding of energy transfer dynamics in exciton
+
+<--- Page Split --->
+
+20 polariton systems beyond the Rabi contraction picture and enable control over how energy is transported in polaritonic systems.  
+
+## 22 Introduction  
+
+22 IntroductionMicrocavity exciton- polaritons are quasiparticles formed when molecular transitions resonantly exchange energy with light trapped in an optical cavity \(^{1 - 4}\) . The marriage of short- lived cavity photons and longer- lived molecular states, two very disparate entities, gives rise to unique properties that are unobserved in purely photonic or purely molecular systems \(^{2,5 - 8}\) . Notably, the photon character causes the wavefunction of the system to be spatially delocalized across the cavity, spanning many molecules simultaneously. The delocalized nature of polariton states has been shown to impact a wide range of photophysical and chemical processes \(^{9 - 20}\) .  
+
+20 Polariton states have wavefunctions that are a linear combination of molecular and photon states. A set of bright eigenstates emerge at energies distinct from those of the molecular states with an energy separation termed the "Rabi splitting", whose magnitude reports on the strength of the light- matter coupling. When more than one chromophore is coupled to the cavity photon, a series of dark states also appear in between the energy of the bright polariton states \(^{21}\) . Measuring energy flow in strongly- coupled polaritonic systems is difficult, because the bright states are short- lived, since they collapse when the light leaves the cavity, and the dark states cannot be directly probed, since they are not optically active \(^{22 - 26}\) . Characterization of energy flow in polaritons will help the understanding of how they impact photophysical and chemical processes \(^{4,27}\) .
+
+<--- Page Split --->
+
+It is well- known from studies on photosynthetic light- harvesting proteins that the energetic disorder of the individual chromophores and the couplings among them are important factors for determining the nature of the bright and dark eigenstates as well as their corresponding dynamics. Indeed, the couplings are often comparable to the energetic disorder, creating energetically overlapping eigenstates that are not spectroscopically resolved but have very different wavefunctions. In contrast, very large light- matter couplings are often used for polaritons to maximize the light- matter hybridization and to make the bright states energetically well- separated from the molecular states. As a result, polaritons have mostly been studied in the limit where the role of the molecular parameters specific to the chromophores was obscured.  
+
+We have created a polaritonic microcavity suitable for studying the interplay of light- matter coupling and energetic disorder. Our system consists of two thin- film layers of semiconducting single- walled carbon nanotubes (CNTs) with different band- gap energies. The differences in band gap and the coupling among the CNTs introduce energetic disorder into the microcavity. The two layers are spatially separated by an insulating polymer barrier to limit inter- tube coupling to those of the same band gap and ensure that energy transfer can only occur due to polaritonic effects. Using two- dimensional white- light (2DWL) spectroscopy, we observe a 200- fs growth of a cross peak between the upper- (UP) and lower- polariton (LP) energies. The UP and LP relax with distinct lifetimes, which is unobserved in the absence of energetic disorder where the dynamics are dominated by Rabi contraction. Through Redfield theory calculations, we find that energetic disorder comparable to the light- matter coupling creates a manifold of eigenstates with varying photon character, and hence, lifetimes. Energy cascades down this manifold depending
+
+<--- Page Split --->
+
+60 on the overlap of the eigenstates, their energy gaps, and the bath frequencies. Our findings shed 61 light on the previously underappreciated role of intermolecular coupling and energetic disorder in 62 exciton-polaritons and help explain the seemingly counterintuitive long lifetimes often observed in 63 them.  
+
+## 64 Results  
+
+65 2DWL spectroscopy of CNT microcavities. Our microcavity contains semiconducting singlewalled CNTs of two different chiralities, known as (6,5) and (7,5), held between a pair of partially reflective gold mirrors (Fig. 1a and Supplementary Figs. 1, 2). The (6,5)- and (7,5) CNTs have different diameters and so exhibit different \(\mathrm{S}_{11}\) band- gap energies, at 1.23 and 1.18 eV, respectively 33, thereby introducing 50 meV of static energetic disorder into the microcavity (Fig. 1b). The weaker bands at 1.43 eV and 1.38 eV are phonon sidebands (PSBs) of the (6,5)- and (7,5) CNTs, respectively 34- 36. The overall thickness of our microcavity was designed such that the cavity mode energy falls between the two \(\mathrm{S}_{11}\) transitions (1.16 - 1.29 eV at normal incidence estimated by a transfer matrix method simulation; see Supplementary Fig. 3 and Methods). We also fabricated microcavities that are made solely from the (6,5)- or (7,5) CNTs, and so have much less energetic disorder than the mixed cavity made from both (6,5) and (7,5) (Fig. 2a, b).  
+
+Angle- dependent reflectance spectra were measured to characterize the light- matter coupling (Fig. 1c and Supplementary Fig. 4). Peaks are observed at energies distinct from those of the uncoupled CNT states and shift in position as a function of the incident angle as expected for
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1: Structure of the microcavity and steady-state spectra. a Schematic illustration of the mixed (6,5)/(7,5) CNT microcavity investigated in this work. b Linear absorption spectra of the uncoupled CNTs measured in 1,2-dichlorobenzene, each normalized to the \(\mathrm{S}_{11}\) peak maximum. The two peaks labeled with asterisks denote minor impurities in the (7,5) CNTs from residual (7,6) \((*,1.09\mathrm{eV})\) and (6,5) \((^{**},1.23\mathrm{eV})\) chiralities \(^{32}\) . c Angle-dependent reflectance \((R)\) spectra of the microcavity shown in a. Dashed lines illustrate the uncoupled CNT state energies as shown in b, where the energy levels of the two most intense transitions, i.e., those that couple most strongly to light, are highlighted in purple ((6,5)) and green ((7,5)). Black solid lines are coupled oscillator model fits to the measured spectra, and the yellow solid line shows the cavity dispersion profile determined from the model (see Supplementary Note 3). </center>
+
+<--- Page Split --->
+
+polaritons. The highest- and lowest- energy eigenstates, labeled PSB and (7,6) in Fig. 1c, are mostly caused by coupling of the PSB- and (7,6) transitions to the cavity photon. In between, there are two transitions that are the focus of this manuscript, which we label Upper Polariton (UP) and Lower Polariton (LP) even though they are not the outermost states. They result, predominantly, from the coupling of the \(\mathrm{S}_{11}\) states of the (6,5)- and (7,5) CNTs to the cavity photon that interest us most. A coupled oscillator model was employed to determine the polariton eigenstate energies as well as the cavity dispersion profile, the result of which is shown as black and yellow solid lines in Fig. 1c, Supplementary Figs. 5, 6 and discussed further in Supplementary Note 3.  
+
+The excited- state dynamics of the microcavities (Fig. 2a- c) were characterized by 2DWL spectroscopy, a technique that maps out energy transfer through cross peaks with ultrafast time resolution \(^{37}\) . Fig. 2d, e shows representative 2DWL spectra of the single band- gap microcavities at \(T = 100\) and 500 fs. The spectra exhibit derivative lineshapes with positive and negative features along the probe energy axis. Based on the peak positions identified in the angle- dependent reflectance spectra, we assign the diagonal features in each of the spectra to the UP and LP (for the (7,5) microcavity, the UP diagonal peak falls outside our detection range).  
+
+The samples made solely from single band- gap CNTs exhibit similar kinetics across the spectral range measured (Fig. 2g, h). All features show a growth within the time resolution of our spectrometer (70 fs) followed by a slow biexponential return to the baseline. To within the error bars, the return to baseline has the same kinetics in each single band- gap microcavity, with time constants of \(1.1 \pm 0.2\) and \(8.5 \pm 0.9\) ps for the (6,5)- and \(0.8 \pm 0.2\) ps and \(7.9 \pm 0.8\) ps for
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure 2: 2DWL spectra and energy relaxation dynamics in the microcavities. a–c Cartoon illustrations of the (6,5)- (a), (7,5)- (b) and mixed (6,5)/(7,5) microcavities (c). d–f 2DWL spectra of the microcavities at waiting times \((T)\) of 100 fs (top) and 500 fs (bottom). The spectra are normalized to the maximum magnitude of the 2D signal in the \(T = 100\) fs spectrum for each sample. Ground-state bleach/stimulated emission and excited-state absorption are plotted in negative and positive signs, respectively. g–i Normalized waiting time traces (colored solid lines) generated at the peak positions labeled with open squares in d–f. Black dashed lines show the fit curves for the UP diagonal and UP/LP peaks. All traces are plotted in absolute values, i.e., traces for negative peaks are plotted with signs flipped. See also Supplementary Figs. 16–21 for additional data. </center>
+
+<--- Page Split --->
+
+the (7,5) microcavity. Thus, to a first approximation, all of the features in the single band- gap samples have uniform kinetics. As we show below, the derivative lineshapes and the uniformity of kinetics originate from a Rabi contraction, where the depletion of ground- state population upon photoexcitation by the pump pulse results in a reduction in Rabi splitting \(^{38,39}\) .  
+
+The 2DWL spectra of the mixed (6,5)/(7,5) microcavity (Fig. 2c) are shown in Fig. 2f. In the \(T = 100\) fs spectrum, we identify three diagonal peaks at energies of 1.27, 1.14, and 1.09 eV, labeled UP, LP, and (7,6), respectively. The peaks are analogous to those in the other two samples; the LP and (7,6) peaks resemble those in the (7,5) microcavity, and the UP peak is consistent with that of the (6,5) microcavity. The Rabi splitting between the UP and LP, identified from their peak positions, is 140 meV. While the spectral features are analogous in all three microcavities, the kinetics are distinct between the single band- gap and mixed microcavities. In the mixed (6,5)/(7,5) sample, the UP and LP diagonal peaks decay on distinctly different timescales. The LP goes back to baseline with time constants of \(1.0 \pm 0.2\) ps and \(8.4 \pm 0.9\) ps (gray trace, Fig. 2i), whereas the UP decays much more rapidly with a time constant of \(0.21 \pm 0.04\) ps, followed by a small- amplitude slow component of \(8.2 \pm 1.3\) ps (purple trace, Fig. 2i). Also different is the appearance of a cross peak with a delayed rise, labeled UP/LP, unlike any other feature in the single band- gap microcavities. This cross peak grows in on a \(200 \pm 20\) fs timescale (green trace, Fig. 2i). Thus, unlike the microcavities made from single band- gap CNTs that show uniform dynamics regardless of the feature, the mixed (6,5)/(7,5) microcavity exhibits complex dynamics with different kinetics for nearly every feature.
+
+<--- Page Split --->
+
+Cross peaks typically measure energy transfer \(^{40}\) . To investigate the origin of the delayed growth of the UP/LP cross peak in the mixed (6,5)/(7,5) microcavity, we performed control experiments in a sample where the cavity mirrors are replaced with plain quartz substrates. No cross peaks are observed, indicating that the cavity mode is required to create this energy transfer pathway (Supplementary Figs. 7, 8). We know that the UP is mostly composed of the (6,5) excitons and the LP mostly of the (7,5) excitons, since the cavity energy sits midway between the band gaps of the two types of CNTs (Fig. 1c). Thus, we conclude that the cavity enables long- range energy transfer across the insulating barrier as measured by the growth of the UP/LP cross peak. These empirical conclusions are confirmed by classical and quantum mechanical simulations below.  
+
+Simulations. Modeling the spectra using Rabi contractions and the transfer matrix method. Absorption/reflection spectra of polaritons are often modeled with classical electrodynamics, using the transfer matrix method \(^{39,41}\) . We simulated the ground- state (i.e., pump- off) and excited- state (i.e., pump- on) reflection spectra of the microcavities as described in Supplementary Note 2.2. The difference of the ground- and excited- state spectra gives the transient reflection spectra as shown in Fig. 3a–c. These transient spectra are analogous to the 2DWL spectra integrated over all pump energies measured and contain the derivative lineshapes characteristic of Rabi contraction.  
+
+As shown in Fig. 3a–c, the model yields exceptionally good agreement with the measured transient reflection spectra (see also Supplementary Figs. 16, 17). Because the linewidths are much larger than the frequency shift caused by the Rabi contraction, the percentage of excited population does not appreciably alter the spectral profile. Instead, regardless of how much photoexcitation
+
+<--- Page Split --->
+![](images/Supplementary_Figure_13.jpg)
+
+<center>Figure 3: Simulated transient reflection spectra and Rabi contraction dynamics. Transient reflection spectra for the (6,5)- (a), (7,5)- (b), and mixed (6,5)/(7,5) microcavities (c) simulated using the transfer matrix method. Measured transient reflection spectra at \(T = 100\) fs are overlaid in light blue for comparison. The dark-to-light color gradient in each panel illustrates the gradual decrease of the bleached population over time (see Supplementary Fig. 13 for the exact values and also Supplementary Note 2.2). d–f Measured kinetic traces (light blue solid lines) of the LP population overlaid with the pseudo-time trace generated from the simulated spectra (black dashed lines). The measured traces are obtained from transient reflection spectra for d, e, and from 2DWL spectra for f to separately plot the diagonal (light blue) and cross peak (green) contributions. The simulated pseudo-time traces are obtained by taking slices at the wavelengths labeled with black arrows in a–c. All traces are plotted in absolute values, i.e., traces for negative peaks are plotted with signs flipped. </center>
+
+<--- Page Split --->
+
+occurs or how much energy transfers from the UP to the LP, the effect is mostly a change in intensity. Thus, we predict that a Rabi contraction model will display uniform, wavelength- independent kinetics, which is indeed what is observed in the experimental data for the single band- gap samples (Fig. 3d, e). In contrast, neither the non- monotonic, wavelength- dependent kinetics nor the delayed rise of the UP/LP cross peak observed in the mixed (6,5)/(7,5) sample can be reproduced by this model alone, which only considers energy transfer from the UP to the dark states (Fig. 3f). To explain the kinetics of the mixed (6,5)/(7,5) microcavity, additional states and energy transfer pathways need to be considered. We also note that the cross peak in the mixed cavity does not have derivative lineshapes and so is not spectroscopically consistent with Rabi contraction.  
+
+Modeling energy transfer using Redfield theory. To simulate the spectra and kinetics we have also employed system- bath quantum dynamics calculations using Redfield theory, which considers all possible energy transfer pathways between eigenstates \(^{42 - 46}\) . We modeled our system with a Holstein- Tavis- Cummings Hamiltonian, analogous to that employed by Herrera and Spano \(^{47 - 50}\) :  
+
+\[\hat{\mathcal{H}} = \sum_{i}\hbar \omega_{i}|i\rangle \langle i| + \sum_{i}\sum_{j}J|i\rangle \langle j| + \hbar \omega_{c}a^{\dagger}a + \sum_{i}g_{i}(|G\rangle \langle i|a^{\dagger} + |i\rangle \langle G|a) \quad (1)\]  
+
+where \(\omega_{i}\) is the transition frequency of the \(i^{t h}\) chromophore, \(J\) is the nearest- neighbor inter- tube coupling, \(\omega_{c}\) is the cavity mode energy, \(a^{\dagger}\) and \(a\) are photon creation and annihilation operators, and \(g_{i}\) is the light- matter coupling given by:  
+
+\[g_{i} = \mu_{i}\sqrt{\frac{N_{i}\hbar\omega_{c}}{2V\epsilon_{o}}} \quad (2)\]  
+
+where \(\mu_{i}\) is the transition dipole of the \(i^{t h}\) chromophore, \(N_{i}\) is the number of chromophores of that
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Figure 4: Calculated spectra, energy levels, and dynamics of the microcavities. a–c Calculated linear \(1 - R\) spectra of the (6,5)- (a), (7,5)- (b) and mixed (6,5)/(7,5) (c) microcavities. In c, the diagonal slice from the measured 2DWL spectrum in Fig. 2f, top (light blue line) is overlaid with the calculated spectrum (black line) with \(J = -10 \mathrm{meV}\) and \(g = 42.9 \mathrm{meV}\) , which yielded the best match. These \(J\) and \(g\) values were then applied to generate the spectra for the single band-gap microcavities shown as black lines in a, b. The vertical bars plot the individual eigenstates with heights proportional to the transition dipole strength, color coded to illustrate the composition of each eigenstate. d–f Energy level diagrams. Bright states are shown with color codes as in a–c, and dark states (states with transition dipoles smaller than \(0.01\%\) of the maximum dipole strength in each case) are shown in gray. g–i Calculated LP population dynamics upon photoexcitation of the UP band. In i, the two traces plot the dynamics upon pumping states with high- (yellow) and low (purple) photon content, also illustrated as colored arrows in c. For comparison, the experimental UP/LP cross peak trace is overlaid as black open circles. </center>
+
+<--- Page Split --->
+
+type, and \(V\) is the mode volume. The CNTs are modeled as hexagonally packed bundles of 36 nanotubes. The nearest- neighbor electronic coupling is introduced only among CNTs of the same band gaps \((J = - 10 \mathrm{meV})^{51 - 53}\) , but not between the (6,5)- and (7,5) CNTs due to the presence of the insulating polymer barrier (see Supplementary Notes 5.1, 5.2 and Supplementary Fig. 9 for details of the procedure and associated data). By fitting the calculated linear spectrum to a diagonal slice through the experimental 2DWL spectrum, we determine that the best match occurs when \(\omega_{c} = 1.217 \mathrm{eV}\) and \(g = 42.9 \mathrm{meV}\) (Fig. 4c).  
+
+The simulation of the mixed (6,5)/(7,5) system predicts a myriad of eigenstates with various oscillator strengths (Fig. 4c, f). The spectrum can be divided into three bands: eigenstates that form the UP, the LP, and those that fall in between. The eigenstates in both the UP and LP bands have character of both the (6,5)- and (7,5) CNTs, with more (6,5) in the UP and more (7,5) in the LP, as expected from the cavity dispersion profile shown above (Fig. 1c). Many states are dark, with no photon character, while others have small amount of photon character with the maximum photon character in a single eigenstate of \(26\%\) . Many of the bright middle states between the UP and LP are predominantly molecular in character with near- zero photon content. The eigenstates are so closely spaced that many cannot be resolved spectroscopically. Thus, pumping the UP band will result in simultaneous excitation of several eigenstates within this band with varying CNT- and photon character.  
+
+The eigenstates of the single band- gap samples can also be grouped into UP, LP and middle states, but with differences (Fig. 4a, b, d, e). The states that fall in between the UP and LP have
+
+<--- Page Split --->
+
+170 far weaker transition dipoles than the bright “molecular” states in the mixed (6,5)/(7,5) cavity. 171 Furthermore, the spectra are much simpler than the mixed (6,5)/(7,5) spectrum, because they lack 172 the energetic disorder created by having more than one band gaps present.  
+
+We simulated the dynamics by evolving the reduced density matrix in time using Redfield 173 theory 44- 46 (see Supplementary Notes 5.1, 5.3, 5.4 and Supplementary Figs. 10, 11 for additional 174 details and data). To mimic the experimental conditions, the dynamics are initiated with all the 175 population residing in the UP band. For the single band- gap systems, the simulations predict rapid 176 relaxation of the UP to lower- lying states (Supplementary Fig. 12). The LP state relaxes within 177 tens of femtoseconds because it has \(14 - 34\%\) photon character (Fig. 4g, h), unlike the picosecond 178 LP dynamics observed in the experiment (Fig. 2g, h). The mismatch is caused by the measured 179 dynamics being dominated by the dark state population as discussed earlier.  
+
+Simulated dynamics of the mixed (6,5)/(7,5) microcavity, in contrast, recover the initial 180 growth of the LP population observed in the experimental cross peak under certain conditions. 181 When an eigenstate in the UP band with high photon content is initially pumped, the dynamics 182 are similar to those for the single band- gap microcavities. The LP instantaneously grows in due 183 to rapid relaxation from the UP and mostly decays within the first 500 fs (Fig. 4i, yellow trace). 184 On the other hand, initial pumping of an eigenstate with low photon content results in strikingly 185 different dynamics (Fig. 4i, purple trace). There is a slow buildup in the LP population over the 186 first 500 fs, followed by a several picosecond decay to the ground state. Kinetics associated with 187 low photon content shows excellent agreement with the measured UP/LP cross peak dynamics (Fig.
+
+<--- Page Split --->
+
+4i, circles).  
+
+Using the kinetics from Redfield theory, we return to the classical Rabi contraction model discussed earlier. If the dark state populations obtained from Redfield theory (Supplementary Fig. 13) are used to scale the bleach terms in the transfer matrix method simulations, then the measured LP dynamics of the two single band- gap microcavities are recovered nearly perfectly (Fig. 3d, e) as are the LP diagonal dynamics of the mixed (6,5)/(7,5) cavity (Fig. 3f, light blue trace). However, no features from this classical model mimic the 200- fs growth of the UP/LP cross peak (Fig. 3f, green trace). Thus, we conclude that transient reflection spectra and the diagonal features in the 2DWL spectra mostly reflect Rabi contractions and, hence, exhibit dynamics that mirror the dark state populations, whereas the cross peak in the 2DWL spectra directly measures energy transfer between states.  
+
+## Discussion  
+
+Redfield theory allows us to understand how energetic disorder creates a manifold of optically bright states through which energy cascades. We find that cascading occurs when there is a source of energetic disorder that is comparable in magnitude to the light- matter coupling. The sources of disorder are the two band- gap energies of the CNTs and the splitting caused by inter- tube couplings (J). To illustrate this point, we adjust the best- fit parameters of the Hamiltonian in Eq. 1 to explore two different limits, one in which we decrease the energetic disorder (Fig. 5a) and the other in which we increase the light- matter coupling (Fig. 5c).
+
+<--- Page Split --->
+![](images/Supplementary_Figure_14.jpg)
+
+<center>Figure 5: System parameters affecting the energy landscape and population dynamics. a–c Energy level diagrams of the mixed (6,5)/(7,5) microcavity in the absence of inter-tube coupling \((J)\) (a), when both \(J\) and light-matter coupling \((g)\) are present with comparable amounts (b), and when \(g\) is significantly larger than \(J\) (c). Bright states are color coded to illustrate the contribution from the cavity photon and the CNTs, and dark states (states with transition dipoles smaller than \(0.01\%\) of the maximum dipole strength in each case) are shown in gray. The specific \(J\) and \(g\) values used for each case are displayed on top. See also Supplementary Fig. 14 for the simulated spectra. b is a replicate of the energy level diagram in Fig. 4c plotted on a different energy axis. d–f show the corresponding relaxation dynamics of the LP population upon photoexcitation into the UP band. The gray and purple traces plot the contribution from the dark states and the UP-to-LP transfer, i.e., population that would appear as a UP/LP cross peak in the experiment, respectively. The purple trace in e is a replicate of the purple trace shown in Fig. 4i, and is overlaid with the experimental UP/LP cross peak trace (open circles). </center>
+
+<--- Page Split --->
+
+Fig. 5a illustrates the case where no inter- tube coupling is present ( \(J = 0 \mathrm{meV}\) ). The energy landscape converges into only three well- separated bright states (UP, MP (middle polariton), LP) and degenerate dark states, which is the conventional simple picture used to describe the energetics of exciton- polaritons. Photoexcitation of the UP results in rapid energy transfer to all lower- lying states, i.e., the dark states, MP, LP, and the ground state, due to its sizeable (41%) photon character. Energy that does make it into the dark states within the UP lifetime creates a Rabi contraction (Fig. 5d, gray trace), but in this limit the populations that create cross peaks are extremely short- lived (Fig. 5d, purple trace). Fig. 5c illustrates the case where the light- matter coupling is much larger ( \(g = 100 \mathrm{meV}\) ) than both the inter- tube coupling and difference in the band- gap energies. In this limit, the UP and LP bands appear with sizeable photon character ( \(> 31\%\) ), in between which a manifold of states with little to no photon character appears. The UP and LP bands are energetically well- separated from the manifold, similarly to the limit shown in Fig. 5a. Once again, the UP decays rapidly, which does not permit sufficient time for energy to flow into the manifold of molecular states (Fig. 5f, purple trace).  
+
+On the contrary, when the energetic disorder is of comparable magnitude with the light- matter coupling (Fig. 5b), the UP and LP bands are no longer well- separated from the manifold, and the photon content is redistributed over many states within this manifold. This allows for sufficient time for the energy to enter and cascade down the manifold when a UP state with low photon content is initially excited, creating the delayed rise of the cross peak in the experiment. Within the linewidth of the UP band, one cannot resolve eigenstates with large versus small photon character, and so the experimental dynamics reflect a convolution of contributions from states with varying photon
+
+<--- Page Split --->
+
+230 character.  
+
+## Conclusion  
+
+We have studied the role that energetic disorder plays in the formation and dynamics of exciton- polaritons. Energetic disorder, created by intermolecular couplings or differences in band gaps, creates manifolds of closely spaced, but non- degenerate eigenstates, many of which are bright states with varying photon character. Energy can cascade through these states from the UP to the LP, which are themselves bands of states, creating a cross peak in the 2DWL spectra that exhibits a delayed rise with a \(\sim 200\) fs time constant. The effects of disorder are observed in the kinetics more so than in the spectra. Kinetics from the transient reflection spectra and the diagonal peaks in the 2DWL spectra are dominated by Rabi contractions, while the cross peak in the 2DWL spectra helps resolve specific transfer pathways. In biology, energetic disorder plays a central role in long- range energy transfer. We note that the kinetics of the mixed (6,5)/(7,5) microcavity reported here is accompanied by energy transfer across the 150- nm insulating barrier, which is significantly longer than the typical exciton diffusion length of \(\sim 10\) nm in most organic semiconductors \(^{54}\) . Thus, it may be advantageous to engineer disorder into exciton- polaritons and thereby manipulate both their dynamics and spatial energy flow as often occurs in biological systems.  
+
+## Methods  
+
+Purification of (6,5)- and (7,5)- enriched semiconducting single- walled CNTs. Semiconducting singled- walled CNTs enriched in (6,5)- and (7,5) chiralities were isolated from as- produced het
+
+<--- Page Split --->
+
+ergoneous single- walled CNTs (CoMoCAT SG65i, Sigma Aldrich) following previously reported protocols with minor modifications \(^{55 - 57}\) . Briefly, polyfluorene derivatives with different functional groups were employed as wrapping polymers to selectively wrap and isolate the (6,5)- and (7,5) CNTs from the heterogeneous mixture (see Supplementary Note 1 for details of the purification procedure). The concentration of the final (6,5)- and (7,5) CNT suspensions (in 1,2- dichlorobenzene) used for fabrication of all microcavity samples in this work was \(50 \mu \mathrm{g / mL}\) .  
+
+Fabrication of the microcavities. The 10- nm and 60- nm thick gold mirrors were prepared by thermal evaporation of Au pellets (Kurt J. Lesker) at 0.02 nm/s on glass substrates (Electron Microscopy Sciences; 15 mm diameter, 0.15 mm thickness). Prior to evaporation, the glass substrates were cleaned by successive sonication in acetone for 1 h and soaking in 2- propanol at \(80^{\circ} \mathrm{C}\) for 40 min. For each sample, a pair of half- cavities was fabricated by drop- casting \(60 \mu \mathrm{L}\) of either (6,5)- or (7,5) CNT suspension on the 10- nm/60- nm gold- deposited substrates. The solvent was evaporated by drying the half- cavities overnight. Aggregates of the wrapping polymer and CNT bundles were removed by successive soaking in tetrahydrofuran and then 2- propanol ( \(80^{\circ} \mathrm{C}\) , 2 h each). Poly(vinyl acetate) (PVA) polymer beads (Sigma Aldrich) were added to chlorobenzene to a concentration of \(40 \mathrm{mg / mL}\) and dissolved by stirring overnight at \(60^{\circ} \mathrm{C}\) . Immediately before lamination of each microcavity sample, the PVA suspension was spin- cast (2,000 rpm, 2 min) on one of the two half- cavities, on top of the drop- cast and washed CNT layer. The two half- cavities were assembled in a home- built compressor tool and laminated in an oven at \(110^{\circ} \mathrm{C}\) for \(3 \mathrm{h}^{36}\) . The cavity- less control sample (Supplementary Fig. 7a) was fabricated following the same cleaning, drop- casting, and lamination procedures but with a pair of plain glass substrates. Surface roughness
+
+<--- Page Split --->
+
+and thickness of each component layer were characterized on an atomic force microscope (Bruker, Icon; see Supplementary Figs. 1, 2). The thickness of the CNT layers and the PVA layer was characterized to be \(50 - 70 \mathrm{nm}\) (CNT) and \(150 \mathrm{nm}\) (PVA). The \(20 - \mathrm{nm}\) variation in the thickness of the CNT layers ( \(50 - 70 \mathrm{nm}\) ) originates from the drop- casting method employed in fabrication, which creates a radial gradient of layer thickness, as reported and discussed in our previous work. The microcavity design was identical in all three microcavities with CNT ( \(50 - 70 \mathrm{nm}\) )- PVA ( \(150 \mathrm{nm}\) )- CNT ( \(50 - 70 \mathrm{nm}\) ) layers, where for the (6,5)- and (7,5) microcavities both CNT layers were of the same chirality. The thickness of each layer, and thus the total cavity thickness, was maintained across the three systems so that the cavity mode dispersion profile was nearly identical for all three microcavities (Supplementary Fig. 4).  
+
+Angle- dependent reflectance spectroscopy. Steady- state angle- dependent reflectance spectra were measured on a V- VASE variable- angle ellipsometer (J. A. Woollam). White light with transverse electric (TE) polarization was focused into a \(0.2\mathrm{- mm}\) spot on the sample position, and the light reflected from the \(10\mathrm{- nm}\) gold mirror was collected. The angle of incidence was varied from \(20^{\circ}\) to \(70^{\circ}\) in \(5^{\circ}\) steps with a motorized goniometer.  
+
+2DWL spectroscopy. Detailed description of the 2DWL apparatus and data processing can be found elsewhere. 2DWL was implemented in a pump- probe geometry in an all- parallel polarization scheme. Briefly, the pump and probe white light was generated by splitting the output of a Ti:Sapphire regenerative amplifier (Spectra Physics, Spitfire; \(800 \mathrm{nm}\) center wavelength, \(1 \mathrm{kHz}\) , \(150 \mathrm{fs}\) pulse duration) into pump and probe arms and then focusing each onto a \(4\mathrm{- mm}\) thick yttrium aluminum garnet (YAG) crystal (NewLight Photonics). Following collimation, the pump
+
+<--- Page Split --->
+
+white light was sent into a prism compressor for removal of unwanted wavelengths including the residual \(800~\mathrm{nm}\) fundamental as well as compression of the pulse. The temporal resolution of our 2DWL measurements was \(\sim 70\) fs as characterized by fitting the rise time of a non- resonant transient signal. Two pairs of birefringent wedges mounted on a motorized translational stage (Newport) generated the pump pulse pair and controlled the time delay between the pump pulses, the coherence time \(^{58}\) . The coherence time was sampled in 0.4- fs steps in the range of \(0 - 200\) fs. The waiting time \((T)\) , the time delay between the pump and the probe pulses, was controlled by delaying the probe beam with a retroreflector mounted on a motorized translational stage (Thorlabs). \(T\) was incremented in steps of 50 fs for \(T = - 50 - 500\) fs, 100 fs for \(T = 500 - 1,000\) fs, 1,000 fs for \(T = 1,000 - 2,000\) fs, and 2,000 fs for \(T = 2,000 - 10,000\) fs. The pump and the probe beams were focused onto the sample with a \(90^{\circ}\) off- axis parabolic mirror, and were temporally and spatially overlapped. The angle of incidence on the sample plane was \(\sim 5^{\circ}\) for the pump and \(\sim 30^{\circ}\) for the probe. The probe beam reflected off the sample was picked off, collimated, and directed into a spectrograph (Princeton Instruments, SpectraPro 2150i). The pump- on and pump- off probe spectra were recorded shot- to- shot at \(1\mathrm{kHz}\) by chopping the pump beam at \(0.5\mathrm{kHz}\) . The 2DWL spectrum at each \(T\) was obtained by Fourier transforming the raw data, which were spectral interferograms collected in the probe frequency domain and coherence time domain, along the coherence time axis. Transient reflection spectra were also collected by setting the coherence time to zero. A pump fluence of \(2.90\times 10^{13}\) photons per pulse per \(\mathrm{cm}^2\) was employed for all 2DWL measurements, which was previously reported to be in the single exciton regime \(^{59,60}\) and yielded transient reflection signal intensity linearly dependent on the pump fluence (Supplementary Fig. 17b). Given the 70 fs
+
+<--- Page Split --->
+
+time resolution mentioned above, we only analyze spectra measured at \(T \geq 100\) fs in this work.  
+
+## References  
+
+1. Lidzey, D. G. et al. Strong exciton-photon coupling in an organic semiconductor microcavity. Nature 395, 53–55 (1998).  
+
+2. Ebbesen, T. W. Hybrid light–matter states in a molecular and material science perspective. Acc. Chem. Res. 49, 2403–2412 (2016).  
+
+3. Hertzog, M., Wang, M., Mony, J. & Börjesson, K. Strong light–matter interactions: A new direction within chemistry. Chem. Soc. Rev. 48, 937–961 (2019).  
+
+4. Ribeiro, R. F., Martínez-Martínez, L. A., Du, M., Campos-Gonzalez-Angulo, J. & Yuen-Zhou, J. Polariton chemistry: Controlling molecular dynamics with optical cavities. Chem. Sci. 9, 6325–6339 (2018).  
+
+5. Byrnes, T., Kim, N. Y. & Yamamoto, Y. Exciton–polariton condensates. Nat. Phys. 10, 803–813 (2014).  
+
+6. Flick, J., Ruggenthaler, M., Appel, H. & Rubio, A. Atoms and molecules in cavities, from weak to strong coupling in quantum-electrodynamics (QED) chemistry. Proc. Natl. Acad. Sci. U. S. A 114, 3026–3034 (2017).  
+
+7. Frisk Kockum, A., Miranowicz, A., De Liberato, S., Savasta, S. & Nori, F. Ultrastrong coupling between light and matter. Nat. Rev. Phys. 1, 19–40 (2019).
+
+<--- Page Split --->
+
+8. Keeling, J. & Kéna-Cohen, S. Bose-Einstein condensation of exciton-polaritons in organic microcavities. Annu. Rev. Phys. Chem. 71, 435-459 (2020).  
+
+9. Hutchison, J. A., Schwartz, T., Genet, C., Devaux, E. & Ebbesen, T. W. Modifying chemical landscapes by coupling to vacuum fields. Angew. Chem. Int. Ed. 51, 1592-1596 (2012).  
+
+10. Avramenko, A. G. & Rury, A. S. Quantum control of ultrafast internal conversion using nanoconfined virtual photons. J. Phys. Chem. Lett. 11, 1013-1021 (2020).  
+
+11. Stranius, K., Hertzog, M. & Börjesson, K. Selective manipulation of electronically excited states through strong light-matter interactions. Nat. Commun. 9, 2273 (2018).  
+
+12. Eizner, E., Martínez-Martínez, L. A., Yuen-Zhou, J. & Kéna-Cohen, S. Inverting singlet and triplet excited states using strong light-matter coupling. Sci. Adv. 5, eaax4482 (2019).  
+
+13. Polak, D. et al. Manipulating molecules with strong coupling: Harvesting triplet excitons in organic exciton microcavities. Chem. Sci. 11, 343-354 (2020).  
+
+14. Coles, D. M. et al. Polariton-mediated energy transfer between organic dyes in a strongly coupled optical microcavity. Nat. Mater. 13, 712-719 (2014).  
+
+15. Zhong, X. et al. Non-radiative energy transfer mediated by hybrid light-matter states. Angew. Chem. Int. Ed. 55, 6202-6206 (2016).  
+
+16. Zhong, X. et al. Energy transfer between spatially separated entangled molecules. Angew. Chem. Int. Ed. 56, 9034-9038 (2017).
+
+<--- Page Split --->
+
+17. DelPo, C. A. et al. Polariton decay in donor-acceptor cavity systems. J. Phys. Chem. Lett. 12, 9774-9782 (2021).  
+
+18. Dovzhenko, D. et al. Polariton-assisted manipulation of energy relaxation pathways: Donor-acceptor role reversal in a tuneable microcavity. Chem. Sci. 12, 12794-12805 (2021).  
+
+19. Jiang, Z., Ren, A., Yan, Y., Yao, J. & Zhao, Y. S. Exciton-polaritons and their Bose-Einstein condensates in organic semiconductor microcavities. Adv. Mater. 34, 2106095 (2022).  
+
+20. Du, M. et al. Theory for polariton-assisted remote energy transfer. Chem. Sci. 9, 6659-6669 (2018).  
+
+21. Tavis, M. & Cummings, F. W. Exact solution for an \(N\) -molecule—radiation-field hamiltonian. Phys. Rev. 170, 379 (1968).  
+
+22. Schwartz, T. et al. Polariton dynamics under strong light–molecule coupling. ChemPhysChem 14, 125-131 (2013).  
+
+23. Wang, S. et al. Quantum yield of polariton emission from hybrid light-matter states. J. Phys. Chem. Lett. 5, 1433-1439 (2014).  
+
+24. Mony, J., Hertzog, M., Kushwaha, K. & Börjesson, K. Angle-independent polariton emission lifetime shown by perylene hybridized to the vacuum field inside a Fabry–Pérot cavity. J. Phys. Chem. C 122, 24917-24923 (2018).  
+
+25. Virgili, T. et al. Ultrafast polariton relaxation dynamics in an organic semiconductor microcavity. Phys. Rev. B 83, 245309 (2011).
+
+<--- Page Split --->
+
+26. Renken, S. et al. Untargeted effects in organic exciton-polariton transient spectroscopy: A cautionary tale. J. Chem. Phys. 155, 154701 (2021).  
+
+27. Herrera, F. & Owrutsky, J. Molecular polaritons for controlling chemistry with quantum optics. J. Chem. Phys. 152, 100902 (2020).  
+
+28. Fassioli, F., Dinshaw, R., Arpin, P. C. & Scholes, G. D. Photosynthetic light harvesting: Excitons and coherence. J. R. Soc. Interface 11, 20130901 (2014).  
+
+29. Mirkovic, T., Ostroumov, E. E., Anna, J. M., Van Grondelle, R. & Scholes, G. D. Light absorption and energy transfer in the antenna complexes of photosynthetic organisms. Chem. Rev. 117, 249-293 (2017).  
+
+30. Pavel, M. & van Grondelle, R. Interplay of disorder and delocalization in photosynthetic light harvesting. Curr. Opin. Chem. Biol. 47, 1-6 (2018).  
+
+31. Georgiou, K., Jayaprakash, R., Othonos, A. & Lidzey, D. Ultralong-range polariton-assisted energy transfer in organic microcavities. Angew. Chem. Int. Ed. 60, 6202-6206 (2021).  
+
+32. Flach, J. T., Wang, J., Arnold, M. S. & Zanni, M. T. Providing time to transfer: Longer lifetimes lead to improved energy transfer in films of semiconducting carbon nanotubes. J. Phys. Chem. Lett. 11, 6016-6024 (2020).  
+
+33. Bachilo, S. M. et al. Structure-assigned optical spectra of single-walled carbon nanotubes. Science 298, 2361-2366 (2002).
+
+<--- Page Split --->
+
+34. Torrens, O. N., Zheng, M. & Kikkawa, J. M. Energy of \(K\) -momentum dark excitons in carbon nanotubes by optical spectroscopy. Phys. Rev. Lett. 101, 157401 (2008).  
+
+35. Blackburn, J. L., Holt, J. M., Irurzun, V. M., Resasco, D. E. & Rumbles, G. Confirmation of \(K\) -momentum dark exciton vibronic sidebands using \(^{13}\mathrm{C}\) -labeled, highly enriched (6, 5) single-walled carbon nanotubes. Nano Lett. 12, 1398–1403 (2012).  
+
+36. Dhavamani, A., Haeberlé, L., Wang, J., Kéna-Cohen, S. & Arnold, M. S. Cavity-mediated hybridization of bright and dark excitons in an ultrastrongly coupled carbon nanotube microcavity. ACS Photonics 8, 2375–2383 (2021).  
+
+37. Mehlenbacher, R. D. et al. Energy transfer pathways in semiconducting carbon nanotubes revealed using two-dimensional white-light spectroscopy. Nat. Commun. 6, 6732 (2015).  
+
+38. Xiang, B. et al. Manipulating optical nonlinearities of molecular polaritons by delocalization. Sci. Adv. 5, eaax5196 (2019).  
+
+39. Grafton, A. B. et al. Excited-state vibration-polariton transitions and dynamics in nitroprusside. Nat. Commun. 12, 214 (2021).  
+
+40. Ginsberg, N. S., Cheng, Y.-C. & Fleming, G. R. Two-dimensional electronic spectroscopy of molecular aggregates. Acc. Chem. Res. 42, 1352–1363 (2009).  
+
+41. Dunkelberger, A. D., Davidson, R. B., Ahn, W., Simpkins, B. S. & Owrutsky, J. C. Ultrafast transmission modulation and recovery via vibrational strong coupling. J. Phys. Chem. A 122, 965–971 (2018).
+
+<--- Page Split --->
+
+42. Ishizaki, A., Calhoun, T. R., Schlau-Cohen, G. S. & Fleming, G. R. Quantum coherence and its interplay with protein environments in photosynthetic electronic energy transfer. Phys. Chem. Chem. Phys. 12, 7319–7337 (2010).  
+
+43. Hwang-Fu, Y.-H., Chen, W. & Cheng, Y.-C. A coherent modified Redfield theory for excitation energy transfer in molecular aggregates. Chem. Phys. 447, 46–53 (2015).  
+
+44. May, V. & Kühn, O. Charge and Energy Transfer Dynamics in Molecular Systems (John Wiley & Sons, 2008).  
+
+45. Berkelbach, T. C., Hybertsen, M. S. & Reichman, D. R. Microscopic theory of singlet exciton fission. I. General formulation. J. Chem. Phys. 138, 114102 (2013).  
+
+46. Tempelaar, R. & Reichman, D. R. Vibronic exciton theory of singlet fission. III. How vibronic coupling and thermodynamics promote rapid triplet generation in pentacene crystals. J. Chem. Phys. 148, 244701 (2018).  
+
+47. Spano, F. C. Optical microcavities enhance the exciton coherence length and eliminate vibronic coupling in J-aggregates. J. Chem. Phys. 142, 184707 (2015).  
+
+48. Herrera, F. & Spano, F. C. Dark vibronic polaritons and the spectroscopy of organic microcavities. Phys. Rev. Lett. 118, 223601 (2017).  
+
+49. Herrera, F. & Spano, F. C. Theory of nanoscale organic cavities: The essential role of vibration-photon dressed states. ACS Photonics 5, 65–79 (2018).
+
+<--- Page Split --->
+
+50. Michetti, P. & La Rocca, G. C. Exciton-phonon scattering and photoexcitation dynamics in J-aggregate microcavities. Phys. Rev. B 79, 035325 (2009).  
+
+51. Mehlenbacher, R. D. et al. Ultrafast exciton hopping observed in bare semiconducting carbon nanotube thin films with two-dimensional white-light spectroscopy. J. Phys. Chem. Lett. 7, 2024-2031 (2016).  
+
+52. Wong, C. Y., Curutchet, C., Tretiak, S. & Scholes, G. D. Ideal dipole approximation fails to predict electronic coupling and energy transfer between semiconducting single-wall carbon nanotubes. J. Chem. Phys. 130, 081104 (2009).  
+
+53. Crochet, J. J., Sau, J. D., Duque, J. G., Doorn, S. K. & Cohen, M. L. Electrodynamic and excitonic intertube interactions in semiconducting carbon nanotube aggregates. ACS Nano 5, 2611-2618 (2011).  
+
+54. Menke, S. M. & Holmes, R. J. Exciton diffusion in organic photovoltaic cells. Energy Environ. Sci. 7, 499-512 (2014).  
+
+55. Ozawa, H., Ide, N., Fujigaya, T., Niidome, Y. & Nakashima, N. One-pot separation of highly enriched (6,5)-single-walled carbon nanotubes using a fluorene-based copolymer. Chem. Lett. 40, 239-241 (2011).  
+
+56. Nish, A., Hwang, J.-Y., Doig, J. & Nicholas, R. J. Highly selective dispersion of single-walled carbon nanotubes using aromatic polymers. Nat. Nanotechnol. 2, 640-646 (2007).  
+
+57. Graf, A. et al. Large scale, selective dispersion of long single-walled carbon nanotubes with high photoluminescence quantum yield by shear force mixing. Carbon 105, 593-599 (2016).
+
+<--- Page Split --->
+
+58. Brida, D., Manzoni, C. & Cerullo, G. Phase-locked pulses for two-dimensional spectroscopy by a birefringent delay line. Opt. Lett. 37, 3027–3029 (2012).  
+
+59. Harrah, D. M. et al. Intensity-dependent exciton dynamics of (6,5) single-walled carbon nanotubes: Momentum selection rules, diffusion, and nonlinear interactions. ACS Nano 5, 9898–9906 (2011).  
+
+60. Park, J., Deria, P., Olivier, J.-H. & Therien, M. J. Fluence-dependent singlet exciton dynamics in length-sorted chirality-enriched single-walled carbon nanotubes. Nano Lett. 14, 504–511 (2014).  
+
+Acknowledgements The authors gratefully acknowledge the financial support from the Air Force Office of Scientific Research awarded to M.T.Z. and M.S.A. under grant no. FA9550- 19- 1- 0093. We thank Louis Haeberlé and Prof. Stéphane Kéna-Cohen (Polytechnique Montréal) for their help with measuring the optical constants for the transfer matrix method simulations.  
+
+Author Contributions M.S. fabricated and characterized the microcavity samples with assistance from A.D. M.S. and R.T.A. performed the 2DWL experiments and analyzed the data. M.S. and Z.T.A. performed the modeling. M.S., Z.T.A., and M.T.Z. co-wrote the paper. All authors discussed the results and commented on the manuscript. † M.S. and Z.T.A. contributed equally to this work.  
+
+Competing Interests M.T.Z. is a co- owner of PhaseTech Spectroscopy, which sells ultrafast pulse shapers and multidimensional spectrometers.  
+
+Data Availability The data presented in this work are available from the corresponding author upon
+
+<--- Page Split --->
+
+462 reasonable request.  
+
+463 Correspondence Correspondence and requests for materials should be addressed to M.T.Z. (zanni@chem.wisc.edu).
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SupplementaryInformation.pdf
+
+<--- Page Split --->

preprint/preprint__2b5449656c9a8636b77273dc85ce96cbe00a6238ddd32058f9565878862bd45c/preprint__2b5449656c9a8636b77273dc85ce96cbe00a6238ddd32058f9565878862bd45c_det.mmd

@@ -0,0 +1,466 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 826, 210]]<|/det|>
+# Energy cascades in polaritonic systems with energetic disorder observed by ultrafast two-dimensional white-light spectroscopy  
+
+<|ref|>text<|/ref|><|det|>[[44, 230, 700, 270]]<|/det|>
+Minjung Son University of Wisconsin- Madison https://orcid.org/0000- 0002- 8385- 062X  
+
+<|ref|>text<|/ref|><|det|>[[44, 277, 344, 315]]<|/det|>
+Zachary Armstrong University of Wisconsin- Madison  
+
+<|ref|>text<|/ref|><|det|>[[44, 323, 344, 362]]<|/det|>
+Ryan Allen University of Wisconsin- Madison  
+
+<|ref|>text<|/ref|><|det|>[[44, 369, 344, 408]]<|/det|>
+Abitha Dhavamani University of Wisconsin- Madison  
+
+<|ref|>text<|/ref|><|det|>[[44, 415, 700, 455]]<|/det|>
+Michael Arnold University of Wisconsin- Madison https://orcid.org/0000- 0002- 2044- 7032  
+
+<|ref|>text<|/ref|><|det|>[[44, 460, 410, 479]]<|/det|>
+Martin Zanni ( zanni@chem.wisc.edu )  
+
+<|ref|>text<|/ref|><|det|>[[50, 483, 700, 501]]<|/det|>
+University of Wisconsin- Madison https://orcid.org/0000- 0001- 7191- 9768  
+
+<|ref|>text<|/ref|><|det|>[[44, 544, 102, 561]]<|/det|>
+Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 582, 136, 600]]<|/det|>
+Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[44, 620, 286, 639]]<|/det|>
+Posted Date: May 4th, 2022  
+
+<|ref|>text<|/ref|><|det|>[[44, 658, 474, 678]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 1582812/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 695, 909, 737]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[44, 756, 896, 799]]<|/det|>
+Additional Declarations: Yes there is potential Competing Interest. Martin T. Zanni is a co- owner of PhaseTech Spectroscopy, which sells ultrafast pulse shapers and multidimensional spectrometers.  
+
+<|ref|>text<|/ref|><|det|>[[44, 834, 956, 877]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on November 27th, 2022. See the published version at https://doi.org/10.1038/s41467- 022- 35046- 2.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[92, 85, 883, 139]]<|/det|>
+# Energy cascades in polaritonic systems with energetic disorder observed by ultrafast two-dimensional white-light spectroscopy  
+
+<|ref|>text<|/ref|><|det|>[[92, 166, 883, 225]]<|/det|>
+3 Minjung Son \(^{1,\dagger}\) , Zachary T. Armstrong \(^{1,\dagger}\) , Ryan T. Allen \(^{1}\) , Abitha Dhavamani \(^{2}\) , Michael S. Arnold \(^{2}\) & Martin T. Zanni \(^{1,*}\)  
+
+<|ref|>text<|/ref|><|det|>[[92, 267, 884, 397]]<|/det|>
+5 \(^{1}\) Department of Chemistry, University of Wisconsin- Madison, 1101 University Ave, Madison, WI 6 53706, USA 7 \(^{2}\) Department of Materials Science and Engineering, University of Wisconsin- Madison, 1509 Uni- 8 versity Ave, Madison, WI 53706, USA  
+
+<|ref|>sub_title<|/ref|><|det|>[[92, 464, 191, 481]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[87, 520, 886, 874]]<|/det|>
+10 Exciton-polaritons are hybrid states formed when molecular excitons are strongly coupled to 11 photons trapped in an optical cavity. These systems have many attractive properties, including 12 large delocalization lengths, but questions regarding the role of energetic disorder remain 13 unanswered. Here, we fabricate microcavities with two different layers of semiconducting 14 carbon nanotubes as a way of controlling the energetic disorder and exploring its impact 15 on energy transfer. Using ultrafast two- dimensional white- light spectroscopy, we observe a 16 delayed growth of a cross peak between the upper- and lower- polariton states. Using Redfield 17 theory, we assign the growth to cascading energy transfer down a manifold of new electronic 18 states created by energetic disorder that is of comparable magnitude to the light- matter 19 coupling. These results broaden our understanding of energy transfer dynamics in exciton
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[86, 90, 884, 148]]<|/det|>
+20 polariton systems beyond the Rabi contraction picture and enable control over how energy is transported in polaritonic systems.  
+
+<|ref|>sub_title<|/ref|><|det|>[[87, 194, 223, 211]]<|/det|>
+## 22 Introduction  
+
+<|ref|>text<|/ref|><|det|>[[86, 252, 884, 494]]<|/det|>
+22 IntroductionMicrocavity exciton- polaritons are quasiparticles formed when molecular transitions resonantly exchange energy with light trapped in an optical cavity \(^{1 - 4}\) . The marriage of short- lived cavity photons and longer- lived molecular states, two very disparate entities, gives rise to unique properties that are unobserved in purely photonic or purely molecular systems \(^{2,5 - 8}\) . Notably, the photon character causes the wavefunction of the system to be spatially delocalized across the cavity, spanning many molecules simultaneously. The delocalized nature of polariton states has been shown to impact a wide range of photophysical and chemical processes \(^{9 - 20}\) .  
+
+<|ref|>text<|/ref|><|det|>[[86, 531, 885, 844]]<|/det|>
+20 Polariton states have wavefunctions that are a linear combination of molecular and photon states. A set of bright eigenstates emerge at energies distinct from those of the molecular states with an energy separation termed the "Rabi splitting", whose magnitude reports on the strength of the light- matter coupling. When more than one chromophore is coupled to the cavity photon, a series of dark states also appear in between the energy of the bright polariton states \(^{21}\) . Measuring energy flow in strongly- coupled polaritonic systems is difficult, because the bright states are short- lived, since they collapse when the light leaves the cavity, and the dark states cannot be directly probed, since they are not optically active \(^{22 - 26}\) . Characterization of energy flow in polaritons will help the understanding of how they impact photophysical and chemical processes \(^{4,27}\) .
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[87, 88, 886, 404]]<|/det|>
+It is well- known from studies on photosynthetic light- harvesting proteins that the energetic disorder of the individual chromophores and the couplings among them are important factors for determining the nature of the bright and dark eigenstates as well as their corresponding dynamics. Indeed, the couplings are often comparable to the energetic disorder, creating energetically overlapping eigenstates that are not spectroscopically resolved but have very different wavefunctions. In contrast, very large light- matter couplings are often used for polaritons to maximize the light- matter hybridization and to make the bright states energetically well- separated from the molecular states. As a result, polaritons have mostly been studied in the limit where the role of the molecular parameters specific to the chromophores was obscured.  
+
+<|ref|>text<|/ref|><|det|>[[87, 440, 886, 863]]<|/det|>
+We have created a polaritonic microcavity suitable for studying the interplay of light- matter coupling and energetic disorder. Our system consists of two thin- film layers of semiconducting single- walled carbon nanotubes (CNTs) with different band- gap energies. The differences in band gap and the coupling among the CNTs introduce energetic disorder into the microcavity. The two layers are spatially separated by an insulating polymer barrier to limit inter- tube coupling to those of the same band gap and ensure that energy transfer can only occur due to polaritonic effects. Using two- dimensional white- light (2DWL) spectroscopy, we observe a 200- fs growth of a cross peak between the upper- (UP) and lower- polariton (LP) energies. The UP and LP relax with distinct lifetimes, which is unobserved in the absence of energetic disorder where the dynamics are dominated by Rabi contraction. Through Redfield theory calculations, we find that energetic disorder comparable to the light- matter coupling creates a manifold of eigenstates with varying photon character, and hence, lifetimes. Energy cascades down this manifold depending
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[85, 88, 884, 220]]<|/det|>
+60 on the overlap of the eigenstates, their energy gaps, and the bath frequencies. Our findings shed 61 light on the previously underappreciated role of intermolecular coupling and energetic disorder in 62 exciton-polaritons and help explain the seemingly counterintuitive long lifetimes often observed in 63 them.  
+
+<|ref|>sub_title<|/ref|><|det|>[[86, 268, 179, 285]]<|/det|>
+## 64 Results  
+
+<|ref|>text<|/ref|><|det|>[[85, 325, 886, 712]]<|/det|>
+65 2DWL spectroscopy of CNT microcavities. Our microcavity contains semiconducting singlewalled CNTs of two different chiralities, known as (6,5) and (7,5), held between a pair of partially reflective gold mirrors (Fig. 1a and Supplementary Figs. 1, 2). The (6,5)- and (7,5) CNTs have different diameters and so exhibit different \(\mathrm{S}_{11}\) band- gap energies, at 1.23 and 1.18 eV, respectively 33, thereby introducing 50 meV of static energetic disorder into the microcavity (Fig. 1b). The weaker bands at 1.43 eV and 1.38 eV are phonon sidebands (PSBs) of the (6,5)- and (7,5) CNTs, respectively 34- 36. The overall thickness of our microcavity was designed such that the cavity mode energy falls between the two \(\mathrm{S}_{11}\) transitions (1.16 - 1.29 eV at normal incidence estimated by a transfer matrix method simulation; see Supplementary Fig. 3 and Methods). We also fabricated microcavities that are made solely from the (6,5)- or (7,5) CNTs, and so have much less energetic disorder than the mixed cavity made from both (6,5) and (7,5) (Fig. 2a, b).  
+
+<|ref|>text<|/ref|><|det|>[[86, 750, 884, 842]]<|/det|>
+Angle- dependent reflectance spectra were measured to characterize the light- matter coupling (Fig. 1c and Supplementary Fig. 4). Peaks are observed at energies distinct from those of the uncoupled CNT states and shift in position as a function of the incident angle as expected for
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[268, 153, 730, 560]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 579, 884, 805]]<|/det|>
+<center>Figure 1: Structure of the microcavity and steady-state spectra. a Schematic illustration of the mixed (6,5)/(7,5) CNT microcavity investigated in this work. b Linear absorption spectra of the uncoupled CNTs measured in 1,2-dichlorobenzene, each normalized to the \(\mathrm{S}_{11}\) peak maximum. The two peaks labeled with asterisks denote minor impurities in the (7,5) CNTs from residual (7,6) \((*,1.09\mathrm{eV})\) and (6,5) \((^{**},1.23\mathrm{eV})\) chiralities \(^{32}\) . c Angle-dependent reflectance \((R)\) spectra of the microcavity shown in a. Dashed lines illustrate the uncoupled CNT state energies as shown in b, where the energy levels of the two most intense transitions, i.e., those that couple most strongly to light, are highlighted in purple ((6,5)) and green ((7,5)). Black solid lines are coupled oscillator model fits to the measured spectra, and the yellow solid line shows the cavity dispersion profile determined from the model (see Supplementary Note 3). </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[85, 88, 886, 365]]<|/det|>
+polaritons. The highest- and lowest- energy eigenstates, labeled PSB and (7,6) in Fig. 1c, are mostly caused by coupling of the PSB- and (7,6) transitions to the cavity photon. In between, there are two transitions that are the focus of this manuscript, which we label Upper Polariton (UP) and Lower Polariton (LP) even though they are not the outermost states. They result, predominantly, from the coupling of the \(\mathrm{S}_{11}\) states of the (6,5)- and (7,5) CNTs to the cavity photon that interest us most. A coupled oscillator model was employed to determine the polariton eigenstate energies as well as the cavity dispersion profile, the result of which is shown as black and yellow solid lines in Fig. 1c, Supplementary Figs. 5, 6 and discussed further in Supplementary Note 3.  
+
+<|ref|>text<|/ref|><|det|>[[85, 404, 886, 644]]<|/det|>
+The excited- state dynamics of the microcavities (Fig. 2a- c) were characterized by 2DWL spectroscopy, a technique that maps out energy transfer through cross peaks with ultrafast time resolution \(^{37}\) . Fig. 2d, e shows representative 2DWL spectra of the single band- gap microcavities at \(T = 100\) and 500 fs. The spectra exhibit derivative lineshapes with positive and negative features along the probe energy axis. Based on the peak positions identified in the angle- dependent reflectance spectra, we assign the diagonal features in each of the spectra to the UP and LP (for the (7,5) microcavity, the UP diagonal peak falls outside our detection range).  
+
+<|ref|>text<|/ref|><|det|>[[85, 682, 886, 848]]<|/det|>
+The samples made solely from single band- gap CNTs exhibit similar kinetics across the spectral range measured (Fig. 2g, h). All features show a growth within the time resolution of our spectrometer (70 fs) followed by a slow biexponential return to the baseline. To within the error bars, the return to baseline has the same kinetics in each single band- gap microcavity, with time constants of \(1.1 \pm 0.2\) and \(8.5 \pm 0.9\) ps for the (6,5)- and \(0.8 \pm 0.2\) ps and \(7.9 \pm 0.8\) ps for
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[112, 135, 884, 603]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 622, 884, 824]]<|/det|>
+<center>Figure 2: 2DWL spectra and energy relaxation dynamics in the microcavities. a–c Cartoon illustrations of the (6,5)- (a), (7,5)- (b) and mixed (6,5)/(7,5) microcavities (c). d–f 2DWL spectra of the microcavities at waiting times \((T)\) of 100 fs (top) and 500 fs (bottom). The spectra are normalized to the maximum magnitude of the 2D signal in the \(T = 100\) fs spectrum for each sample. Ground-state bleach/stimulated emission and excited-state absorption are plotted in negative and positive signs, respectively. g–i Normalized waiting time traces (colored solid lines) generated at the peak positions labeled with open squares in d–f. Black dashed lines show the fit curves for the UP diagonal and UP/LP peaks. All traces are plotted in absolute values, i.e., traces for negative peaks are plotted with signs flipped. See also Supplementary Figs. 16–21 for additional data. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[88, 88, 884, 220]]<|/det|>
+the (7,5) microcavity. Thus, to a first approximation, all of the features in the single band- gap samples have uniform kinetics. As we show below, the derivative lineshapes and the uniformity of kinetics originate from a Rabi contraction, where the depletion of ground- state population upon photoexcitation by the pump pulse results in a reduction in Rabi splitting \(^{38,39}\) .  
+
+<|ref|>text<|/ref|><|det|>[[88, 255, 886, 828]]<|/det|>
+The 2DWL spectra of the mixed (6,5)/(7,5) microcavity (Fig. 2c) are shown in Fig. 2f. In the \(T = 100\) fs spectrum, we identify three diagonal peaks at energies of 1.27, 1.14, and 1.09 eV, labeled UP, LP, and (7,6), respectively. The peaks are analogous to those in the other two samples; the LP and (7,6) peaks resemble those in the (7,5) microcavity, and the UP peak is consistent with that of the (6,5) microcavity. The Rabi splitting between the UP and LP, identified from their peak positions, is 140 meV. While the spectral features are analogous in all three microcavities, the kinetics are distinct between the single band- gap and mixed microcavities. In the mixed (6,5)/(7,5) sample, the UP and LP diagonal peaks decay on distinctly different timescales. The LP goes back to baseline with time constants of \(1.0 \pm 0.2\) ps and \(8.4 \pm 0.9\) ps (gray trace, Fig. 2i), whereas the UP decays much more rapidly with a time constant of \(0.21 \pm 0.04\) ps, followed by a small- amplitude slow component of \(8.2 \pm 1.3\) ps (purple trace, Fig. 2i). Also different is the appearance of a cross peak with a delayed rise, labeled UP/LP, unlike any other feature in the single band- gap microcavities. This cross peak grows in on a \(200 \pm 20\) fs timescale (green trace, Fig. 2i). Thus, unlike the microcavities made from single band- gap CNTs that show uniform dynamics regardless of the feature, the mixed (6,5)/(7,5) microcavity exhibits complex dynamics with different kinetics for nearly every feature.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 884, 404]]<|/det|>
+Cross peaks typically measure energy transfer \(^{40}\) . To investigate the origin of the delayed growth of the UP/LP cross peak in the mixed (6,5)/(7,5) microcavity, we performed control experiments in a sample where the cavity mirrors are replaced with plain quartz substrates. No cross peaks are observed, indicating that the cavity mode is required to create this energy transfer pathway (Supplementary Figs. 7, 8). We know that the UP is mostly composed of the (6,5) excitons and the LP mostly of the (7,5) excitons, since the cavity energy sits midway between the band gaps of the two types of CNTs (Fig. 1c). Thus, we conclude that the cavity enables long- range energy transfer across the insulating barrier as measured by the growth of the UP/LP cross peak. These empirical conclusions are confirmed by classical and quantum mechanical simulations below.  
+
+<|ref|>text<|/ref|><|det|>[[111, 440, 884, 680]]<|/det|>
+Simulations. Modeling the spectra using Rabi contractions and the transfer matrix method. Absorption/reflection spectra of polaritons are often modeled with classical electrodynamics, using the transfer matrix method \(^{39,41}\) . We simulated the ground- state (i.e., pump- off) and excited- state (i.e., pump- on) reflection spectra of the microcavities as described in Supplementary Note 2.2. The difference of the ground- and excited- state spectra gives the transient reflection spectra as shown in Fig. 3a–c. These transient spectra are analogous to the 2DWL spectra integrated over all pump energies measured and contain the derivative lineshapes characteristic of Rabi contraction.  
+
+<|ref|>text<|/ref|><|det|>[[111, 718, 884, 848]]<|/det|>
+As shown in Fig. 3a–c, the model yields exceptionally good agreement with the measured transient reflection spectra (see also Supplementary Figs. 16, 17). Because the linewidths are much larger than the frequency shift caused by the Rabi contraction, the percentage of excited population does not appreciably alter the spectral profile. Instead, regardless of how much photoexcitation
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[112, 151, 881, 521]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 539, 884, 812]]<|/det|>
+<center>Figure 3: Simulated transient reflection spectra and Rabi contraction dynamics. Transient reflection spectra for the (6,5)- (a), (7,5)- (b), and mixed (6,5)/(7,5) microcavities (c) simulated using the transfer matrix method. Measured transient reflection spectra at \(T = 100\) fs are overlaid in light blue for comparison. The dark-to-light color gradient in each panel illustrates the gradual decrease of the bleached population over time (see Supplementary Fig. 13 for the exact values and also Supplementary Note 2.2). d–f Measured kinetic traces (light blue solid lines) of the LP population overlaid with the pseudo-time trace generated from the simulated spectra (black dashed lines). The measured traces are obtained from transient reflection spectra for d, e, and from 2DWL spectra for f to separately plot the diagonal (light blue) and cross peak (green) contributions. The simulated pseudo-time traces are obtained by taking slices at the wavelengths labeled with black arrows in a–c. All traces are plotted in absolute values, i.e., traces for negative peaks are plotted with signs flipped. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[81, 87, 886, 440]]<|/det|>
+occurs or how much energy transfers from the UP to the LP, the effect is mostly a change in intensity. Thus, we predict that a Rabi contraction model will display uniform, wavelength- independent kinetics, which is indeed what is observed in the experimental data for the single band- gap samples (Fig. 3d, e). In contrast, neither the non- monotonic, wavelength- dependent kinetics nor the delayed rise of the UP/LP cross peak observed in the mixed (6,5)/(7,5) sample can be reproduced by this model alone, which only considers energy transfer from the UP to the dark states (Fig. 3f). To explain the kinetics of the mixed (6,5)/(7,5) microcavity, additional states and energy transfer pathways need to be considered. We also note that the cross peak in the mixed cavity does not have derivative lineshapes and so is not spectroscopically consistent with Rabi contraction.  
+
+<|ref|>text<|/ref|><|det|>[[112, 475, 886, 607]]<|/det|>
+Modeling energy transfer using Redfield theory. To simulate the spectra and kinetics we have also employed system- bath quantum dynamics calculations using Redfield theory, which considers all possible energy transfer pathways between eigenstates \(^{42 - 46}\) . We modeled our system with a Holstein- Tavis- Cummings Hamiltonian, analogous to that employed by Herrera and Spano \(^{47 - 50}\) :  
+
+<|ref|>equation<|/ref|><|det|>[[208, 628, 880, 675]]<|/det|>
+\[\hat{\mathcal{H}} = \sum_{i}\hbar \omega_{i}|i\rangle \langle i| + \sum_{i}\sum_{j}J|i\rangle \langle j| + \hbar \omega_{c}a^{\dagger}a + \sum_{i}g_{i}(|G\rangle \langle i|a^{\dagger} + |i\rangle \langle G|a) \quad (1)\]  
+
+<|ref|>text<|/ref|><|det|>[[112, 682, 884, 778]]<|/det|>
+where \(\omega_{i}\) is the transition frequency of the \(i^{t h}\) chromophore, \(J\) is the nearest- neighbor inter- tube coupling, \(\omega_{c}\) is the cavity mode energy, \(a^{\dagger}\) and \(a\) are photon creation and annihilation operators, and \(g_{i}\) is the light- matter coupling given by:  
+
+<|ref|>equation<|/ref|><|det|>[[431, 787, 880, 838]]<|/det|>
+\[g_{i} = \mu_{i}\sqrt{\frac{N_{i}\hbar\omega_{c}}{2V\epsilon_{o}}} \quad (2)\]  
+
+<|ref|>text<|/ref|><|det|>[[84, 853, 884, 876]]<|/det|>
+where \(\mu_{i}\) is the transition dipole of the \(i^{t h}\) chromophore, \(N_{i}\) is the number of chromophores of that
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[150, 90, 840, 576]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 594, 884, 876]]<|/det|>
+<center>Figure 4: Calculated spectra, energy levels, and dynamics of the microcavities. a–c Calculated linear \(1 - R\) spectra of the (6,5)- (a), (7,5)- (b) and mixed (6,5)/(7,5) (c) microcavities. In c, the diagonal slice from the measured 2DWL spectrum in Fig. 2f, top (light blue line) is overlaid with the calculated spectrum (black line) with \(J = -10 \mathrm{meV}\) and \(g = 42.9 \mathrm{meV}\) , which yielded the best match. These \(J\) and \(g\) values were then applied to generate the spectra for the single band-gap microcavities shown as black lines in a, b. The vertical bars plot the individual eigenstates with heights proportional to the transition dipole strength, color coded to illustrate the composition of each eigenstate. d–f Energy level diagrams. Bright states are shown with color codes as in a–c, and dark states (states with transition dipoles smaller than \(0.01\%\) of the maximum dipole strength in each case) are shown in gray. g–i Calculated LP population dynamics upon photoexcitation of the UP band. In i, the two traces plot the dynamics upon pumping states with high- (yellow) and low (purple) photon content, also illustrated as colored arrows in c. For comparison, the experimental UP/LP cross peak trace is overlaid as black open circles. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[112, 88, 884, 330]]<|/det|>
+type, and \(V\) is the mode volume. The CNTs are modeled as hexagonally packed bundles of 36 nanotubes. The nearest- neighbor electronic coupling is introduced only among CNTs of the same band gaps \((J = - 10 \mathrm{meV})^{51 - 53}\) , but not between the (6,5)- and (7,5) CNTs due to the presence of the insulating polymer barrier (see Supplementary Notes 5.1, 5.2 and Supplementary Fig. 9 for details of the procedure and associated data). By fitting the calculated linear spectrum to a diagonal slice through the experimental 2DWL spectrum, we determine that the best match occurs when \(\omega_{c} = 1.217 \mathrm{eV}\) and \(g = 42.9 \mathrm{meV}\) (Fig. 4c).  
+
+<|ref|>text<|/ref|><|det|>[[112, 367, 884, 752]]<|/det|>
+The simulation of the mixed (6,5)/(7,5) system predicts a myriad of eigenstates with various oscillator strengths (Fig. 4c, f). The spectrum can be divided into three bands: eigenstates that form the UP, the LP, and those that fall in between. The eigenstates in both the UP and LP bands have character of both the (6,5)- and (7,5) CNTs, with more (6,5) in the UP and more (7,5) in the LP, as expected from the cavity dispersion profile shown above (Fig. 1c). Many states are dark, with no photon character, while others have small amount of photon character with the maximum photon character in a single eigenstate of \(26\%\) . Many of the bright middle states between the UP and LP are predominantly molecular in character with near- zero photon content. The eigenstates are so closely spaced that many cannot be resolved spectroscopically. Thus, pumping the UP band will result in simultaneous excitation of several eigenstates within this band with varying CNT- and photon character.  
+
+<|ref|>text<|/ref|><|det|>[[112, 791, 883, 847]]<|/det|>
+The eigenstates of the single band- gap samples can also be grouped into UP, LP and middle states, but with differences (Fig. 4a, b, d, e). The states that fall in between the UP and LP have
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[83, 88, 884, 184]]<|/det|>
+170 far weaker transition dipoles than the bright “molecular” states in the mixed (6,5)/(7,5) cavity. 171 Furthermore, the spectra are much simpler than the mixed (6,5)/(7,5) spectrum, because they lack 172 the energetic disorder created by having more than one band gaps present.  
+
+<|ref|>text<|/ref|><|det|>[[111, 220, 884, 500]]<|/det|>
+We simulated the dynamics by evolving the reduced density matrix in time using Redfield 173 theory 44- 46 (see Supplementary Notes 5.1, 5.3, 5.4 and Supplementary Figs. 10, 11 for additional 174 details and data). To mimic the experimental conditions, the dynamics are initiated with all the 175 population residing in the UP band. For the single band- gap systems, the simulations predict rapid 176 relaxation of the UP to lower- lying states (Supplementary Fig. 12). The LP state relaxes within 177 tens of femtoseconds because it has \(14 - 34\%\) photon character (Fig. 4g, h), unlike the picosecond 178 LP dynamics observed in the experiment (Fig. 2g, h). The mismatch is caused by the measured 179 dynamics being dominated by the dark state population as discussed earlier.  
+
+<|ref|>text<|/ref|><|det|>[[111, 536, 884, 851]]<|/det|>
+Simulated dynamics of the mixed (6,5)/(7,5) microcavity, in contrast, recover the initial 180 growth of the LP population observed in the experimental cross peak under certain conditions. 181 When an eigenstate in the UP band with high photon content is initially pumped, the dynamics 182 are similar to those for the single band- gap microcavities. The LP instantaneously grows in due 183 to rapid relaxation from the UP and mostly decays within the first 500 fs (Fig. 4i, yellow trace). 184 On the other hand, initial pumping of an eigenstate with low photon content results in strikingly 185 different dynamics (Fig. 4i, purple trace). There is a slow buildup in the LP population over the 186 first 500 fs, followed by a several picosecond decay to the ground state. Kinetics associated with 187 low photon content shows excellent agreement with the measured UP/LP cross peak dynamics (Fig.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 92, 205, 108]]<|/det|>
+4i, circles).  
+
+<|ref|>text<|/ref|><|det|>[[111, 149, 886, 499]]<|/det|>
+Using the kinetics from Redfield theory, we return to the classical Rabi contraction model discussed earlier. If the dark state populations obtained from Redfield theory (Supplementary Fig. 13) are used to scale the bleach terms in the transfer matrix method simulations, then the measured LP dynamics of the two single band- gap microcavities are recovered nearly perfectly (Fig. 3d, e) as are the LP diagonal dynamics of the mixed (6,5)/(7,5) cavity (Fig. 3f, light blue trace). However, no features from this classical model mimic the 200- fs growth of the UP/LP cross peak (Fig. 3f, green trace). Thus, we conclude that transient reflection spectra and the diagonal features in the 2DWL spectra mostly reflect Rabi contractions and, hence, exhibit dynamics that mirror the dark state populations, whereas the cross peak in the 2DWL spectra directly measures energy transfer between states.  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 544, 206, 561]]<|/det|>
+## Discussion  
+
+<|ref|>text<|/ref|><|det|>[[111, 602, 886, 844]]<|/det|>
+Redfield theory allows us to understand how energetic disorder creates a manifold of optically bright states through which energy cascades. We find that cascading occurs when there is a source of energetic disorder that is comparable in magnitude to the light- matter coupling. The sources of disorder are the two band- gap energies of the CNTs and the splitting caused by inter- tube couplings (J). To illustrate this point, we adjust the best- fit parameters of the Hamiltonian in Eq. 1 to explore two different limits, one in which we decrease the energetic disorder (Fig. 5a) and the other in which we increase the light- matter coupling (Fig. 5c).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[112, 140, 880, 504]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 525, 884, 819]]<|/det|>
+<center>Figure 5: System parameters affecting the energy landscape and population dynamics. a–c Energy level diagrams of the mixed (6,5)/(7,5) microcavity in the absence of inter-tube coupling \((J)\) (a), when both \(J\) and light-matter coupling \((g)\) are present with comparable amounts (b), and when \(g\) is significantly larger than \(J\) (c). Bright states are color coded to illustrate the contribution from the cavity photon and the CNTs, and dark states (states with transition dipoles smaller than \(0.01\%\) of the maximum dipole strength in each case) are shown in gray. The specific \(J\) and \(g\) values used for each case are displayed on top. See also Supplementary Fig. 14 for the simulated spectra. b is a replicate of the energy level diagram in Fig. 4c plotted on a different energy axis. d–f show the corresponding relaxation dynamics of the LP population upon photoexcitation into the UP band. The gray and purple traces plot the contribution from the dark states and the UP-to-LP transfer, i.e., population that would appear as a UP/LP cross peak in the experiment, respectively. The purple trace in e is a replicate of the purple trace shown in Fig. 4i, and is overlaid with the experimental UP/LP cross peak trace (open circles). </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[112, 90, 884, 585]]<|/det|>
+Fig. 5a illustrates the case where no inter- tube coupling is present ( \(J = 0 \mathrm{meV}\) ). The energy landscape converges into only three well- separated bright states (UP, MP (middle polariton), LP) and degenerate dark states, which is the conventional simple picture used to describe the energetics of exciton- polaritons. Photoexcitation of the UP results in rapid energy transfer to all lower- lying states, i.e., the dark states, MP, LP, and the ground state, due to its sizeable (41%) photon character. Energy that does make it into the dark states within the UP lifetime creates a Rabi contraction (Fig. 5d, gray trace), but in this limit the populations that create cross peaks are extremely short- lived (Fig. 5d, purple trace). Fig. 5c illustrates the case where the light- matter coupling is much larger ( \(g = 100 \mathrm{meV}\) ) than both the inter- tube coupling and difference in the band- gap energies. In this limit, the UP and LP bands appear with sizeable photon character ( \(> 31\%\) ), in between which a manifold of states with little to no photon character appears. The UP and LP bands are energetically well- separated from the manifold, similarly to the limit shown in Fig. 5a. Once again, the UP decays rapidly, which does not permit sufficient time for energy to flow into the manifold of molecular states (Fig. 5f, purple trace).  
+
+<|ref|>text<|/ref|><|det|>[[112, 623, 884, 862]]<|/det|>
+On the contrary, when the energetic disorder is of comparable magnitude with the light- matter coupling (Fig. 5b), the UP and LP bands are no longer well- separated from the manifold, and the photon content is redistributed over many states within this manifold. This allows for sufficient time for the energy to enter and cascade down the manifold when a UP state with low photon content is initially excited, creating the delayed rise of the cross peak in the experiment. Within the linewidth of the UP band, one cannot resolve eigenstates with large versus small photon character, and so the experimental dynamics reflect a convolution of contributions from states with varying photon
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[77, 92, 192, 108]]<|/det|>
+230 character.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 158, 211, 176]]<|/det|>
+## Conclusion  
+
+<|ref|>text<|/ref|><|det|>[[110, 214, 886, 714]]<|/det|>
+We have studied the role that energetic disorder plays in the formation and dynamics of exciton- polaritons. Energetic disorder, created by intermolecular couplings or differences in band gaps, creates manifolds of closely spaced, but non- degenerate eigenstates, many of which are bright states with varying photon character. Energy can cascade through these states from the UP to the LP, which are themselves bands of states, creating a cross peak in the 2DWL spectra that exhibits a delayed rise with a \(\sim 200\) fs time constant. The effects of disorder are observed in the kinetics more so than in the spectra. Kinetics from the transient reflection spectra and the diagonal peaks in the 2DWL spectra are dominated by Rabi contractions, while the cross peak in the 2DWL spectra helps resolve specific transfer pathways. In biology, energetic disorder plays a central role in long- range energy transfer. We note that the kinetics of the mixed (6,5)/(7,5) microcavity reported here is accompanied by energy transfer across the 150- nm insulating barrier, which is significantly longer than the typical exciton diffusion length of \(\sim 10\) nm in most organic semiconductors \(^{54}\) . Thus, it may be advantageous to engineer disorder into exciton- polaritons and thereby manipulate both their dynamics and spatial energy flow as often occurs in biological systems.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 760, 191, 777]]<|/det|>
+## Methods  
+
+<|ref|>text<|/ref|><|det|>[[115, 810, 884, 864]]<|/det|>
+Purification of (6,5)- and (7,5)- enriched semiconducting single- walled CNTs. Semiconducting singled- walled CNTs enriched in (6,5)- and (7,5) chiralities were isolated from as- produced het
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 884, 293]]<|/det|>
+ergoneous single- walled CNTs (CoMoCAT SG65i, Sigma Aldrich) following previously reported protocols with minor modifications \(^{55 - 57}\) . Briefly, polyfluorene derivatives with different functional groups were employed as wrapping polymers to selectively wrap and isolate the (6,5)- and (7,5) CNTs from the heterogeneous mixture (see Supplementary Note 1 for details of the purification procedure). The concentration of the final (6,5)- and (7,5) CNT suspensions (in 1,2- dichlorobenzene) used for fabrication of all microcavity samples in this work was \(50 \mu \mathrm{g / mL}\) .  
+
+<|ref|>text<|/ref|><|det|>[[111, 321, 886, 860]]<|/det|>
+Fabrication of the microcavities. The 10- nm and 60- nm thick gold mirrors were prepared by thermal evaporation of Au pellets (Kurt J. Lesker) at 0.02 nm/s on glass substrates (Electron Microscopy Sciences; 15 mm diameter, 0.15 mm thickness). Prior to evaporation, the glass substrates were cleaned by successive sonication in acetone for 1 h and soaking in 2- propanol at \(80^{\circ} \mathrm{C}\) for 40 min. For each sample, a pair of half- cavities was fabricated by drop- casting \(60 \mu \mathrm{L}\) of either (6,5)- or (7,5) CNT suspension on the 10- nm/60- nm gold- deposited substrates. The solvent was evaporated by drying the half- cavities overnight. Aggregates of the wrapping polymer and CNT bundles were removed by successive soaking in tetrahydrofuran and then 2- propanol ( \(80^{\circ} \mathrm{C}\) , 2 h each). Poly(vinyl acetate) (PVA) polymer beads (Sigma Aldrich) were added to chlorobenzene to a concentration of \(40 \mathrm{mg / mL}\) and dissolved by stirring overnight at \(60^{\circ} \mathrm{C}\) . Immediately before lamination of each microcavity sample, the PVA suspension was spin- cast (2,000 rpm, 2 min) on one of the two half- cavities, on top of the drop- cast and washed CNT layer. The two half- cavities were assembled in a home- built compressor tool and laminated in an oven at \(110^{\circ} \mathrm{C}\) for \(3 \mathrm{h}^{36}\) . The cavity- less control sample (Supplementary Fig. 7a) was fabricated following the same cleaning, drop- casting, and lamination procedures but with a pair of plain glass substrates. Surface roughness
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[80, 88, 886, 440]]<|/det|>
+and thickness of each component layer were characterized on an atomic force microscope (Bruker, Icon; see Supplementary Figs. 1, 2). The thickness of the CNT layers and the PVA layer was characterized to be \(50 - 70 \mathrm{nm}\) (CNT) and \(150 \mathrm{nm}\) (PVA). The \(20 - \mathrm{nm}\) variation in the thickness of the CNT layers ( \(50 - 70 \mathrm{nm}\) ) originates from the drop- casting method employed in fabrication, which creates a radial gradient of layer thickness, as reported and discussed in our previous work. The microcavity design was identical in all three microcavities with CNT ( \(50 - 70 \mathrm{nm}\) )- PVA ( \(150 \mathrm{nm}\) )- CNT ( \(50 - 70 \mathrm{nm}\) ) layers, where for the (6,5)- and (7,5) microcavities both CNT layers were of the same chirality. The thickness of each layer, and thus the total cavity thickness, was maintained across the three systems so that the cavity mode dispersion profile was nearly identical for all three microcavities (Supplementary Fig. 4).  
+
+<|ref|>text<|/ref|><|det|>[[80, 470, 886, 635]]<|/det|>
+Angle- dependent reflectance spectroscopy. Steady- state angle- dependent reflectance spectra were measured on a V- VASE variable- angle ellipsometer (J. A. Woollam). White light with transverse electric (TE) polarization was focused into a \(0.2\mathrm{- mm}\) spot on the sample position, and the light reflected from the \(10\mathrm{- nm}\) gold mirror was collected. The angle of incidence was varied from \(20^{\circ}\) to \(70^{\circ}\) in \(5^{\circ}\) steps with a motorized goniometer.  
+
+<|ref|>text<|/ref|><|det|>[[80, 666, 886, 870]]<|/det|>
+2DWL spectroscopy. Detailed description of the 2DWL apparatus and data processing can be found elsewhere. 2DWL was implemented in a pump- probe geometry in an all- parallel polarization scheme. Briefly, the pump and probe white light was generated by splitting the output of a Ti:Sapphire regenerative amplifier (Spectra Physics, Spitfire; \(800 \mathrm{nm}\) center wavelength, \(1 \mathrm{kHz}\) , \(150 \mathrm{fs}\) pulse duration) into pump and probe arms and then focusing each onto a \(4\mathrm{- mm}\) thick yttrium aluminum garnet (YAG) crystal (NewLight Photonics). Following collimation, the pump
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 886, 844]]<|/det|>
+white light was sent into a prism compressor for removal of unwanted wavelengths including the residual \(800~\mathrm{nm}\) fundamental as well as compression of the pulse. The temporal resolution of our 2DWL measurements was \(\sim 70\) fs as characterized by fitting the rise time of a non- resonant transient signal. Two pairs of birefringent wedges mounted on a motorized translational stage (Newport) generated the pump pulse pair and controlled the time delay between the pump pulses, the coherence time \(^{58}\) . The coherence time was sampled in 0.4- fs steps in the range of \(0 - 200\) fs. The waiting time \((T)\) , the time delay between the pump and the probe pulses, was controlled by delaying the probe beam with a retroreflector mounted on a motorized translational stage (Thorlabs). \(T\) was incremented in steps of 50 fs for \(T = - 50 - 500\) fs, 100 fs for \(T = 500 - 1,000\) fs, 1,000 fs for \(T = 1,000 - 2,000\) fs, and 2,000 fs for \(T = 2,000 - 10,000\) fs. The pump and the probe beams were focused onto the sample with a \(90^{\circ}\) off- axis parabolic mirror, and were temporally and spatially overlapped. The angle of incidence on the sample plane was \(\sim 5^{\circ}\) for the pump and \(\sim 30^{\circ}\) for the probe. The probe beam reflected off the sample was picked off, collimated, and directed into a spectrograph (Princeton Instruments, SpectraPro 2150i). The pump- on and pump- off probe spectra were recorded shot- to- shot at \(1\mathrm{kHz}\) by chopping the pump beam at \(0.5\mathrm{kHz}\) . The 2DWL spectrum at each \(T\) was obtained by Fourier transforming the raw data, which were spectral interferograms collected in the probe frequency domain and coherence time domain, along the coherence time axis. Transient reflection spectra were also collected by setting the coherence time to zero. A pump fluence of \(2.90\times 10^{13}\) photons per pulse per \(\mathrm{cm}^2\) was employed for all 2DWL measurements, which was previously reported to be in the single exciton regime \(^{59,60}\) and yielded transient reflection signal intensity linearly dependent on the pump fluence (Supplementary Fig. 17b). Given the 70 fs
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[78, 90, 860, 110]]<|/det|>
+time resolution mentioned above, we only analyze spectra measured at \(T \geq 100\) fs in this work.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 157, 208, 175]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[115, 214, 882, 272]]<|/det|>
+1. Lidzey, D. G. et al. Strong exciton-photon coupling in an organic semiconductor microcavity. Nature 395, 53–55 (1998).  
+
+<|ref|>text<|/ref|><|det|>[[115, 301, 882, 358]]<|/det|>
+2. Ebbesen, T. W. Hybrid light–matter states in a molecular and material science perspective. Acc. Chem. Res. 49, 2403–2412 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[115, 386, 882, 444]]<|/det|>
+3. Hertzog, M., Wang, M., Mony, J. & Börjesson, K. Strong light–matter interactions: A new direction within chemistry. Chem. Soc. Rev. 48, 937–961 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[115, 472, 883, 565]]<|/det|>
+4. Ribeiro, R. F., Martínez-Martínez, L. A., Du, M., Campos-Gonzalez-Angulo, J. & Yuen-Zhou, J. Polariton chemistry: Controlling molecular dynamics with optical cavities. Chem. Sci. 9, 6325–6339 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[115, 594, 882, 651]]<|/det|>
+5. Byrnes, T., Kim, N. Y. & Yamamoto, Y. Exciton–polariton condensates. Nat. Phys. 10, 803–813 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[115, 680, 883, 775]]<|/det|>
+6. Flick, J., Ruggenthaler, M., Appel, H. & Rubio, A. Atoms and molecules in cavities, from weak to strong coupling in quantum-electrodynamics (QED) chemistry. Proc. Natl. Acad. Sci. U. S. A 114, 3026–3034 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[115, 803, 883, 859]]<|/det|>
+7. Frisk Kockum, A., Miranowicz, A., De Liberato, S., Savasta, S. & Nori, F. Ultrastrong coupling between light and matter. Nat. Rev. Phys. 1, 19–40 (2019).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[71, 88, 886, 149]]<|/det|>
+8. Keeling, J. & Kéna-Cohen, S. Bose-Einstein condensation of exciton-polaritons in organic microcavities. Annu. Rev. Phys. Chem. 71, 435-459 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[71, 175, 886, 234]]<|/det|>
+9. Hutchison, J. A., Schwartz, T., Genet, C., Devaux, E. & Ebbesen, T. W. Modifying chemical landscapes by coupling to vacuum fields. Angew. Chem. Int. Ed. 51, 1592-1596 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[71, 260, 886, 319]]<|/det|>
+10. Avramenko, A. G. & Rury, A. S. Quantum control of ultrafast internal conversion using nanoconfined virtual photons. J. Phys. Chem. Lett. 11, 1013-1021 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[71, 344, 886, 404]]<|/det|>
+11. Stranius, K., Hertzog, M. & Börjesson, K. Selective manipulation of electronically excited states through strong light-matter interactions. Nat. Commun. 9, 2273 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[71, 430, 886, 490]]<|/det|>
+12. Eizner, E., Martínez-Martínez, L. A., Yuen-Zhou, J. & Kéna-Cohen, S. Inverting singlet and triplet excited states using strong light-matter coupling. Sci. Adv. 5, eaax4482 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[71, 515, 886, 575]]<|/det|>
+13. Polak, D. et al. Manipulating molecules with strong coupling: Harvesting triplet excitons in organic exciton microcavities. Chem. Sci. 11, 343-354 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[71, 601, 886, 660]]<|/det|>
+14. Coles, D. M. et al. Polariton-mediated energy transfer between organic dyes in a strongly coupled optical microcavity. Nat. Mater. 13, 712-719 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[71, 686, 884, 745]]<|/det|>
+15. Zhong, X. et al. Non-radiative energy transfer mediated by hybrid light-matter states. Angew. Chem. Int. Ed. 55, 6202-6206 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[71, 772, 884, 831]]<|/det|>
+16. Zhong, X. et al. Energy transfer between spatially separated entangled molecules. Angew. Chem. Int. Ed. 56, 9034-9038 (2017).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[81, 88, 884, 147]]<|/det|>
+17. DelPo, C. A. et al. Polariton decay in donor-acceptor cavity systems. J. Phys. Chem. Lett. 12, 9774-9782 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[81, 175, 884, 234]]<|/det|>
+18. Dovzhenko, D. et al. Polariton-assisted manipulation of energy relaxation pathways: Donor-acceptor role reversal in a tuneable microcavity. Chem. Sci. 12, 12794-12805 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[81, 260, 884, 319]]<|/det|>
+19. Jiang, Z., Ren, A., Yan, Y., Yao, J. & Zhao, Y. S. Exciton-polaritons and their Bose-Einstein condensates in organic semiconductor microcavities. Adv. Mater. 34, 2106095 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[81, 345, 884, 404]]<|/det|>
+20. Du, M. et al. Theory for polariton-assisted remote energy transfer. Chem. Sci. 9, 6659-6669 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[81, 430, 884, 489]]<|/det|>
+21. Tavis, M. & Cummings, F. W. Exact solution for an \(N\) -molecule—radiation-field hamiltonian. Phys. Rev. 170, 379 (1968).  
+
+<|ref|>text<|/ref|><|det|>[[81, 515, 884, 574]]<|/det|>
+22. Schwartz, T. et al. Polariton dynamics under strong light–molecule coupling. ChemPhysChem 14, 125-131 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[81, 600, 884, 660]]<|/det|>
+23. Wang, S. et al. Quantum yield of polariton emission from hybrid light-matter states. J. Phys. Chem. Lett. 5, 1433-1439 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[81, 686, 884, 781]]<|/det|>
+24. Mony, J., Hertzog, M., Kushwaha, K. & Börjesson, K. Angle-independent polariton emission lifetime shown by perylene hybridized to the vacuum field inside a Fabry–Pérot cavity. J. Phys. Chem. C 122, 24917-24923 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[81, 808, 884, 867]]<|/det|>
+25. Virgili, T. et al. Ultrafast polariton relaxation dynamics in an organic semiconductor microcavity. Phys. Rev. B 83, 245309 (2011).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[80, 88, 884, 147]]<|/det|>
+26. Renken, S. et al. Untargeted effects in organic exciton-polariton transient spectroscopy: A cautionary tale. J. Chem. Phys. 155, 154701 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[80, 174, 883, 232]]<|/det|>
+27. Herrera, F. & Owrutsky, J. Molecular polaritons for controlling chemistry with quantum optics. J. Chem. Phys. 152, 100902 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[80, 260, 883, 318]]<|/det|>
+28. Fassioli, F., Dinshaw, R., Arpin, P. C. & Scholes, G. D. Photosynthetic light harvesting: Excitons and coherence. J. R. Soc. Interface 11, 20130901 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[80, 345, 884, 440]]<|/det|>
+29. Mirkovic, T., Ostroumov, E. E., Anna, J. M., Van Grondelle, R. & Scholes, G. D. Light absorption and energy transfer in the antenna complexes of photosynthetic organisms. Chem. Rev. 117, 249-293 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[80, 467, 884, 525]]<|/det|>
+30. Pavel, M. & van Grondelle, R. Interplay of disorder and delocalization in photosynthetic light harvesting. Curr. Opin. Chem. Biol. 47, 1-6 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[80, 552, 884, 611]]<|/det|>
+31. Georgiou, K., Jayaprakash, R., Othonos, A. & Lidzey, D. Ultralong-range polariton-assisted energy transfer in organic microcavities. Angew. Chem. Int. Ed. 60, 6202-6206 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[80, 638, 884, 732]]<|/det|>
+32. Flach, J. T., Wang, J., Arnold, M. S. & Zanni, M. T. Providing time to transfer: Longer lifetimes lead to improved energy transfer in films of semiconducting carbon nanotubes. J. Phys. Chem. Lett. 11, 6016-6024 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[80, 760, 883, 818]]<|/det|>
+33. Bachilo, S. M. et al. Structure-assigned optical spectra of single-walled carbon nanotubes. Science 298, 2361-2366 (2002).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[80, 88, 884, 147]]<|/det|>
+34. Torrens, O. N., Zheng, M. & Kikkawa, J. M. Energy of \(K\) -momentum dark excitons in carbon nanotubes by optical spectroscopy. Phys. Rev. Lett. 101, 157401 (2008).  
+
+<|ref|>text<|/ref|><|det|>[[80, 174, 884, 270]]<|/det|>
+35. Blackburn, J. L., Holt, J. M., Irurzun, V. M., Resasco, D. E. & Rumbles, G. Confirmation of \(K\) -momentum dark exciton vibronic sidebands using \(^{13}\mathrm{C}\) -labeled, highly enriched (6, 5) single-walled carbon nanotubes. Nano Lett. 12, 1398–1403 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[80, 296, 884, 392]]<|/det|>
+36. Dhavamani, A., Haeberlé, L., Wang, J., Kéna-Cohen, S. & Arnold, M. S. Cavity-mediated hybridization of bright and dark excitons in an ultrastrongly coupled carbon nanotube microcavity. ACS Photonics 8, 2375–2383 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[80, 419, 884, 478]]<|/det|>
+37. Mehlenbacher, R. D. et al. Energy transfer pathways in semiconducting carbon nanotubes revealed using two-dimensional white-light spectroscopy. Nat. Commun. 6, 6732 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[80, 504, 884, 562]]<|/det|>
+38. Xiang, B. et al. Manipulating optical nonlinearities of molecular polaritons by delocalization. Sci. Adv. 5, eaax5196 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[80, 589, 884, 648]]<|/det|>
+39. Grafton, A. B. et al. Excited-state vibration-polariton transitions and dynamics in nitroprusside. Nat. Commun. 12, 214 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[80, 674, 884, 733]]<|/det|>
+40. Ginsberg, N. S., Cheng, Y.-C. & Fleming, G. R. Two-dimensional electronic spectroscopy of molecular aggregates. Acc. Chem. Res. 42, 1352–1363 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[80, 760, 884, 856]]<|/det|>
+41. Dunkelberger, A. D., Davidson, R. B., Ahn, W., Simpkins, B. S. & Owrutsky, J. C. Ultrafast transmission modulation and recovery via vibrational strong coupling. J. Phys. Chem. A 122, 965–971 (2018).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[81, 88, 884, 183]]<|/det|>
+42. Ishizaki, A., Calhoun, T. R., Schlau-Cohen, G. S. & Fleming, G. R. Quantum coherence and its interplay with protein environments in photosynthetic electronic energy transfer. Phys. Chem. Chem. Phys. 12, 7319–7337 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[81, 210, 884, 269]]<|/det|>
+43. Hwang-Fu, Y.-H., Chen, W. & Cheng, Y.-C. A coherent modified Redfield theory for excitation energy transfer in molecular aggregates. Chem. Phys. 447, 46–53 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[81, 295, 884, 354]]<|/det|>
+44. May, V. & Kühn, O. Charge and Energy Transfer Dynamics in Molecular Systems (John Wiley & Sons, 2008).  
+
+<|ref|>text<|/ref|><|det|>[[81, 381, 884, 440]]<|/det|>
+45. Berkelbach, T. C., Hybertsen, M. S. & Reichman, D. R. Microscopic theory of singlet exciton fission. I. General formulation. J. Chem. Phys. 138, 114102 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[81, 467, 884, 562]]<|/det|>
+46. Tempelaar, R. & Reichman, D. R. Vibronic exciton theory of singlet fission. III. How vibronic coupling and thermodynamics promote rapid triplet generation in pentacene crystals. J. Chem. Phys. 148, 244701 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[81, 589, 884, 648]]<|/det|>
+47. Spano, F. C. Optical microcavities enhance the exciton coherence length and eliminate vibronic coupling in J-aggregates. J. Chem. Phys. 142, 184707 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[81, 674, 884, 733]]<|/det|>
+48. Herrera, F. & Spano, F. C. Dark vibronic polaritons and the spectroscopy of organic microcavities. Phys. Rev. Lett. 118, 223601 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[81, 759, 884, 818]]<|/det|>
+49. Herrera, F. & Spano, F. C. Theory of nanoscale organic cavities: The essential role of vibration-photon dressed states. ACS Photonics 5, 65–79 (2018).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[81, 88, 884, 149]]<|/det|>
+50. Michetti, P. & La Rocca, G. C. Exciton-phonon scattering and photoexcitation dynamics in J-aggregate microcavities. Phys. Rev. B 79, 035325 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[81, 173, 884, 268]]<|/det|>
+51. Mehlenbacher, R. D. et al. Ultrafast exciton hopping observed in bare semiconducting carbon nanotube thin films with two-dimensional white-light spectroscopy. J. Phys. Chem. Lett. 7, 2024-2031 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[81, 291, 884, 387]]<|/det|>
+52. Wong, C. Y., Curutchet, C., Tretiak, S. & Scholes, G. D. Ideal dipole approximation fails to predict electronic coupling and energy transfer between semiconducting single-wall carbon nanotubes. J. Chem. Phys. 130, 081104 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[81, 410, 884, 507]]<|/det|>
+53. Crochet, J. J., Sau, J. D., Duque, J. G., Doorn, S. K. & Cohen, M. L. Electrodynamic and excitonic intertube interactions in semiconducting carbon nanotube aggregates. ACS Nano 5, 2611-2618 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[81, 530, 883, 590]]<|/det|>
+54. Menke, S. M. & Holmes, R. J. Exciton diffusion in organic photovoltaic cells. Energy Environ. Sci. 7, 499-512 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[81, 613, 884, 707]]<|/det|>
+55. Ozawa, H., Ide, N., Fujigaya, T., Niidome, Y. & Nakashima, N. One-pot separation of highly enriched (6,5)-single-walled carbon nanotubes using a fluorene-based copolymer. Chem. Lett. 40, 239-241 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[81, 732, 884, 792]]<|/det|>
+56. Nish, A., Hwang, J.-Y., Doig, J. & Nicholas, R. J. Highly selective dispersion of single-walled carbon nanotubes using aromatic polymers. Nat. Nanotechnol. 2, 640-646 (2007).  
+
+<|ref|>text<|/ref|><|det|>[[81, 816, 884, 876]]<|/det|>
+57. Graf, A. et al. Large scale, selective dispersion of long single-walled carbon nanotubes with high photoluminescence quantum yield by shear force mixing. Carbon 105, 593-599 (2016).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[81, 88, 884, 149]]<|/det|>
+58. Brida, D., Manzoni, C. & Cerullo, G. Phase-locked pulses for two-dimensional spectroscopy by a birefringent delay line. Opt. Lett. 37, 3027–3029 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[81, 174, 884, 270]]<|/det|>
+59. Harrah, D. M. et al. Intensity-dependent exciton dynamics of (6,5) single-walled carbon nanotubes: Momentum selection rules, diffusion, and nonlinear interactions. ACS Nano 5, 9898–9906 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[81, 296, 884, 393]]<|/det|>
+60. Park, J., Deria, P., Olivier, J.-H. & Therien, M. J. Fluence-dependent singlet exciton dynamics in length-sorted chirality-enriched single-walled carbon nanotubes. Nano Lett. 14, 504–511 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[81, 442, 884, 567]]<|/det|>
+Acknowledgements The authors gratefully acknowledge the financial support from the Air Force Office of Scientific Research awarded to M.T.Z. and M.S.A. under grant no. FA9550- 19- 1- 0093. We thank Louis Haeberlé and Prof. Stéphane Kéna-Cohen (Polytechnique Montréal) for their help with measuring the optical constants for the transfer matrix method simulations.  
+
+<|ref|>text<|/ref|><|det|>[[81, 600, 884, 724]]<|/det|>
+Author Contributions M.S. fabricated and characterized the microcavity samples with assistance from A.D. M.S. and R.T.A. performed the 2DWL experiments and analyzed the data. M.S. and Z.T.A. performed the modeling. M.S., Z.T.A., and M.T.Z. co-wrote the paper. All authors discussed the results and commented on the manuscript. † M.S. and Z.T.A. contributed equally to this work.  
+
+<|ref|>text<|/ref|><|det|>[[81, 758, 884, 812]]<|/det|>
+Competing Interests M.T.Z. is a co- owner of PhaseTech Spectroscopy, which sells ultrafast pulse shapers and multidimensional spectrometers.  
+
+<|ref|>text<|/ref|><|det|>[[81, 846, 884, 867]]<|/det|>
+Data Availability The data presented in this work are available from the corresponding author upon
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[84, 94, 252, 108]]<|/det|>
+462 reasonable request.  
+
+<|ref|>text<|/ref|><|det|>[[84, 146, 945, 164]]<|/det|>
+463 Correspondence Correspondence and requests for materials should be addressed to M.T.Z. (zanni@chem.wisc.edu).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 42, 311, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[44, 93, 765, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 353, 150]]<|/det|>
+SupplementaryInformation.pdf
+
+<--- Page Split --->

preprint/preprint__2b80d2125fb4e34dd3cd9f801c4387ddb7d0c5c4f3aaa3b81839750ee75cc826/images_list.json

@@ -0,0 +1,182 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1. Schematic of the WTCL for real-time and in situ glaucoma diagnosis and therapy in a closed-loop manner. (a) Schematic of the WTCL for wireless glaucoma diagnosis and therapy. (b) Photograph of WTCL worn on the eyes of a live rabbit. (c) Schematic of wireless operation for the purpose of IOP monitoring and on-demand medicines administration in a closed-loop manner. The soft device, engineered as a double layer contact lens structure, was integrated with an LCR and a WPT receiver circuit. These modules were wirelessly connected to external integrated antenna that could record the IOP signal and trigger iontophoresis for drug delivery if needed. Insert figures respectively highlight critical IOP sensing and drug delivery unit. (d) Structure of the WTCL in an exploded view. (e) Optical image of the WTCL. (f) Schematic showing the structure of the cantilever capacitive sensor, which could be highly sensitive to pressure, allowing drug delivery circuits to integrate in limited space.",
+    "footnote": [],
+    "bbox": [
+      [
+        160,
+        88,
+        848,
+        435
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2. Schematic illustration of the WTCL's design and fabrication process. (a) The snowflake-shaped layout design and the photograph of the sensing circuit. (b) The microscopic image of the reference plate, coils, and sensing plate deployed on sensing circuit. The photograph of (c) the folded sensing circuit and (d) the upper layer of contact lens. (e) Image of the integrated antenna. (f)(I) The flower-shaped layout design, (II) back surface (III) front surface images, and (IV) microscopic image of the drug delivery circuit. (g) The photograph of bottom layer lens integrated with drug delivery circuit. Illustration of the fabrication process of (h) IOP monitoring circuit, (i) drug delivery circuit and the device integration. The fabrication of the sensing and delivery modulus employed a printed circuit process coupled with cast-molding method.",
+    "footnote": [],
+    "bbox": [
+      [
+        150,
+        82,
+        848,
+        528
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3. IOP sensing performance of the WTCL. (a-I) Schematic and (II) experimental set-up of the wireless IOP sensing experiments. (b) The reflection spectra of six representative WTCL devices worn on porcine eyeball at different IOP. (c) The results of the S11 values at different frequencies and IOP conditions in (b) were plotted as heatmap diagram, where the value of S11 exhibited a linear pattern in the frequency-IOP heatmap. (d) Linear regression of resonant frequency versus IOP value of each WTCL device. (e) The averaged linear regression of resonant frequency versus IOP value of the six WTCL devices in (b). (f) Error grid analysis and statical analysis of the IOP sensing accuracy via WTCL. Region A, B, C and D referred to errors \\(< 10\\%\\) , 10-20%, 20-40% and \\(>40\\%\\) , respectively. (g) Heatmap plot of the reflection coefficients recorded during the continuous recording of IOP via WTCL. (h) Continuous IOP signals monitored by WTCL on in-vitro porcine eyeball. Calibration point using reference IOP was marked with blue asterisks. The black arrow referred to the time",
+    "footnote": [],
+    "bbox": [
+      [
+        145,
+        84,
+        852,
+        546
+      ]
+    ],
+    "page_idx": 14
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4. WPT performance of the WTCL. (a-I) Schematic and (II) experimental set-up of the WPT experiments. (b) Reflection coefficient spectra (S11 and S21) recorded from four receivers at different radiation distance. (c) The S21 recorded by the four receivers at \\(850\\mathrm{kHz}\\) were plotted as a function of radiation distance. (d) Alternating voltage signals collected from four receivers wirelessly under different",
+    "footnote": [],
+    "bbox": [
+      [
+        147,
+        226,
+        852,
+        753
+      ]
+    ],
+    "page_idx": 15
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5. (a) The mutual inductance and (b) Power transfer efficiencies were theoretically calculated according to the circuit design of four receivers and the radiation distance. (c) Rader chart summarized performance (S21, Vpp, M, and \\(\\eta\\) ) of etch WPT group (Rec#2, Rec#5, Rec#9 or Rec#17 linked to transmitter with \\(6\\mathrm{mm}\\) radiation distance and \\(850\\mathrm{kHz}\\) ) (d) Schematic showing the experiments studying the cross-coupling between IOP monitoring and WPT module. Red arrow denoted the interference generated by the radiation of WPT transmitter to sensing module. Blue arrow denotes the cross-coupling between IOP reading coil and the WPT receiver. All examinations were performed with the radiation distance of \\(6\\mathrm{mm}\\) . (e) The IOP reading coil and WPT transmitter were coupled with the WPT receiver (using Rec#17), respectively, and the S21 indicating the coupling efficiency and the generated voltages on receiver radiated at \\(850\\mathrm{kHz}\\) were separately measured. (f) The WTCL was placed on porcine eye at different IOP, and the reading signals (the resonance frequency and the S11) were recorded with or without the presence of radiation from WPT transmitter. (g) The (top) 3D COMSOL model and (bottom) the simulated distribution profile of electric potential and electric field through the anterior region under the condition of iontophoresis at \\(3\\mathrm{V}\\) for \\(\\mathrm{t} = 30\\mathrm{min}\\) . (h) The time slots of drug concentration profile delivered by WTCL at various applied voltages (0, 1, 2, and 3 V). (i) The delivered",
+    "footnote": [],
+    "bbox": [
+      [
+        147,
+        90,
+        854,
+        414
+      ]
+    ],
+    "page_idx": 20
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6. Sensing and therapeutic performance of the integrated WTCL. (a) Quantitative analysis of in vitro rhodamine B released from WTCL at different alternating voltages for 30 min. N=6 measurements. (b) Rhodamine B was utilized as the medicines analog in ex vivo experiments on porcine eyes, to examine the influence of iontophoresis on drug delivery across cornea. CB, ciliary body; IS, iris; CA, Cornea. Scale bar is \\(500\\mu \\mathrm{m}\\) . After delivery, fluorescence visualization in the anterior tissue was",
+    "footnote": [],
+    "bbox": [
+      [
+        147,
+        85,
+        848,
+        730
+      ]
+    ],
+    "page_idx": 23
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1",
+    "footnote": [],
+    "bbox": [
+      [
+        42,
+        95,
+        951,
+        600
+      ]
+    ],
+    "page_idx": 43
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2",
+    "footnote": [],
+    "bbox": [
+      [
+        42,
+        40,
+        953,
+        680
+      ]
+    ],
+    "page_idx": 44
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3",
+    "footnote": [],
+    "bbox": [
+      [
+        42,
+        42,
+        950,
+        707
+      ]
+    ],
+    "page_idx": 45
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4",
+    "footnote": [],
+    "bbox": [
+      [
+        40,
+        175,
+        949,
+        920
+      ]
+    ],
+    "page_idx": 46
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5",
+    "footnote": [],
+    "bbox": [
+      [
+        39,
+        293,
+        960,
+        752
+      ]
+    ],
+    "page_idx": 47
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6",
+    "footnote": [],
+    "bbox": [
+      [
+        42,
+        37,
+        780,
+        789
+      ]
+    ],
+    "page_idx": 49
+  }
+]

preprint/preprint__2b80d2125fb4e34dd3cd9f801c4387ddb7d0c5c4f3aaa3b81839750ee75cc826/preprint__2b80d2125fb4e34dd3cd9f801c4387ddb7d0c5c4f3aaa3b81839750ee75cc826.mmd

@@ -0,0 +1,448 @@
+
+# Intelligent wireless theranostic contact lens for close-loop electrical sensing and regulation of glaucoma  
+
+Cheng Yang State Key Laboratory of Optoelectronic Materials and Technologies, Sun Yat- Sen University  
+
+Qianni Wu Sun Yat- Sen University  
+
+Junqing Liu Department of Cardiology, the First Affiliated Hospital of Jinan University, Guangzhou 510630, China  
+
+Jingshan Mo State Key Laboratory of Optoelectronic Materials and Technologies, Sun Yat- Sen University  
+
+Xiangling Li State Key Laboratory of Optoelectronic Materials and Technologies, Sun Yat- Sen University  
+
+Chengduan Yang State Key Laboratory of Optoelectronic Materials and Technologies, Sun Yat- Sen University  
+
+Ziqi Liu State Key Laboratory of Optoelectronic Materials and Technologies, Sun Yat- Sen University  
+
+Jingbo Yang State Key Laboratory of Optoelectronic Materials and Technologies, Sun Yat- Sen University  
+
+Lelun Jiang School of Biomedical Engineering, Sun Yat- Sen University  
+
+Weirong Chen  
+
+State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat- Sen University  
+
+Hui- juan Chen  
+
+State Key Laboratory of Optoelectronic Materials and Technologies, Sun Yat- Sen University  
+
+Ji Wang  
+
+The First Affiliated Hospital of Sun Yat- Sen University, Sun Yat- Sen University  
+
+Xi Xie ( \(\boxed{\times}\) xiexi27@mail.sysu.edu.cn) State Key Laboratory of Optoelectronic Materials and Technologies, Sun Yat- Sen University https://orcid.org/0000- 0001- 7406- 8444
+
+<--- Page Split --->
+
+**Keywords:** Wireless theranostic system, Smart contact lens, Glaucoma monitoring and therapy, Closed-loop system, Wireless sensing
+
+**Posted Date:** November 5th, 2021
+
+**DOI:** https://doi.org/10.21203/rs.3.rs-1020390/v1
+
+**License:** © This work is licensed under a Creative Commons Attribution 4.0 International License.Read Full License
+
+**Version of Record:** A version of this preprint was published at Nature Communications on May 17th,2022. See the published version at https://doi.org/10.1038/s41467-022-29860-x.
+
+<--- Page Split --->
+
+# Intelligent wireless theranostic contact lens for close-loop electrical sensing and regulation of glaucoma  
+
+Cheng Yang \(^{1,3}\) , Qianni Wu \(^{2}\) , Junqing Liu \(^{4}\) , Jingshan Mo \(^{1}\) , Xiangling Li \(^{1,5}\) , Chengduan Yang \(^{1,3}\) , Ziqi Liu \(^{1}\) , Jingbo Yang \(^{1,5}\) , Lelun Jiang \(^{5}\) , Weirong Chen \(^{2}\) , Hui- jiuan Chen \(^{1}\) , Ji Wang \(^{3}\) , Xi Xie \(^{1,3*}\)  
+
+1. State Key Laboratory of Optoelectronic Materials and Technologies, Guangdong Province Key Laboratory of Display Material and Technology, School of Electronics and Information Technology, Sun Yat-Sen University, Guangzhou, 510006, China  
+2. State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat-Sen University, Guangzhou, 510006, China  
+3. The First Affiliated Hospital of Sun Yat-Sen University, Sun Yat-Sen University, Guangzhou, 510006, China  
+4. Department of Cardiology, the First Affiliated Hospital of Jinan University, Guangzhou 510630, China  
+5. School of Biomedical Engineering, Sun Yat-Sen University, Guangzhou, 510006, China  
+
+Email: xiexi27@mail.sysu.edu.cn
+
+<--- Page Split --->
+
+**Keywords:** Wireless theranostic system, Smart contact lens, Glaucoma monitoring and therapy, Closed-loop system, Wireless sensing.  
+
+## Abstract:  
+
+Engineered closed- loop devices that can wirelessly track intraocular pressure (IOP) and offer feedback- medicine administrations are highly desirable for glaucoma treatments, yet remain difficult to develop. Integrated theranostic systems based on contact lens still confront several challenges, including size limits, requirements of wireless operations, and cross- coupling between multiple functional modulus. Here, for the first time to our knowledge, an integrated wireless theranostic contact lens (WTCL) for in situ electrical sensing and on- demand drug delivery of glaucoma was developed. The WTCL utilized a highly compact circuitry and structural design, which enabled high- degree integration of IOP sensing and electrically controlled delivery modulus on the curved and limited surface of contact lens. The wireless IOP sensing modulus could ultra- sensitively detect IOP fluctuations, due to the unique cantilever configuration design of LCR circuit with ultra- soft air dielectric film sandwiched between each capacitive sensing plate. The drug delivery modulus employed a highly efficient wireless power transfer circuit, to trigger delivery of anti- glaucoma drug into aqueous chamber via iontophoresis to enhance drug permeation across cornea. The specialized design of frequency separation enabled individual operations of different modules without cross- coupling. The minimally invasive, smart, wireless and closed- loop theranostic features endowed the WTCL as a highly promising system for glaucoma treatments.
+
+<--- Page Split --->
+
+## Introduction  
+
+Intelligent point- of- care electrical platforms that could provide real- time health assessment and medical intervention would greatly relieve many acute and stubborn diseases \(^{1 - 6}\) . Among the diseases, glaucoma and its combined ophthalmic diseases can cause irreversible vision loss of patients \(^{7}\) , which is often deteriorated by the elevation of intraocular pressure (IOP) due to abnormal circulation of aqueous humor \(^{7 - 9}\) . Since IOP varies associated with human activities and circadian rhythm \(^{10}\) , it needs long- term and continuous tracking to analyze the critical IOP fluctuations for identifying optimal therapeutic conditions \(^{11}\) . At present, many types of ophthalmotonometers (e.g. indentation tonometry, applanation tonometry, rebound tonometry, and dynamic contour tonometry) have provided snapshot measurements of IOP for glaucoma diagnosis in hospitals \(^{12}\) , yet the operations generally required trained clinicians, and failed to collect many critical IOP fluctuation \(^{13}\) . On the other hand, clinically medicine administrations for glaucoma treatments have been relying on topical drug delivery via eyedrops to reduce IOP for suspend the deterioration of vision that glaucoma caused \(^{8,12,14}\) . However, conventional drug deliveries into the anterior chamber remain challenging (low intraocular bioavailability, inevitable side- effects, and poor patient adherence) due to the diffusion barriers of cornea \(^{15}\) , and lack the possibility of integration with smart biodevices for on- demand drug delivery. Especially for acute angle- closure glaucoma featured with sudden rise of IOP \(^{16}\) , it is usually accompanied by headache, nausea and vomiting that hinders manually self- administrations by patients \(^{8}\) , while the delayed reduction of IOP will inevitably causes ischemic infarcts and damage optic nerve \(^{12}\) .  
+
+Contact lens, an ideal platform contacted with human eye intimately \(^{17,18}\) , has been exploited as wearable devices for physiological measurements \(^{2,19 - 22}\) . In recent decades, contact lens- based IOP sensors integrated with resonant circuits, microfluidic chips, piezoresistive and photonic crystal technologies have emerged \(^{13,19,20,23 - 26}\) . For example, Park et al. developed colorimetric contact lens for IOP reading based on photonic crystal sensor coupled with microhydraulic strategy to amplify sensitivity \(^{20}\) . To acquire electrical signals of IOP, Kim et al. demonstrated graphene/Ag nanowires and silicon
+
+<--- Page Split --->
+
+strain sensors- based contact lens that could detect IOP with high sensitivities<sup>19,23,27</sup>. Besides, controlled ocular drug deliveries mediated via contact lens devices have employed versatile strategies including thermal- responsive, enzyme triggering, and hydrogel layer- controlled drug release<sup>25,26,28,29</sup>. To reduce burst release of drug from devices, Cakmak et al. fabricated a multi- diffusion layers- based contact lens that could achieve stable ophthalmic drug administration with constant rate<sup>29</sup>. However, medicines permeabilities into aqueous chamber by these passive diffusion methodologies are generally compromised due to the physiological barriers of eye, especially the frequent tear clearance and the tightly packed corneal epithelium cells<sup>30</sup>. While most of the existing strategies for glaucoma applications focus on either sensing or delivery separately, integrated wireless electrical systems for closed- loop IOP monitor and regulation are highly desirable to treat glaucoma, yet rarely developed due to challenges (Supplementary Materials S2).  
+
+Closed- loop theranostic systems on flexible patches have recently been developed to automatically monitor biomarkers, and respond rapidly to treat these complications<sup>1,3,4,31</sup>. However, in contrast to patch devices worn on skin, closed- loop theranostic systems based on contact lens confront several complicating challenges due to its nature of limited size and requirement of wireless operations. First, contact lens is flexible, lightweight, curved and ultrathin device with extremely limited area<sup>32,33</sup>, which is highly challenging to install intricate theranostic system composited by multi- modules, and less compatible with standard 2D micro/nano- fabrication routes. Second, contact lens devices need to operate wirelessly to promote patients' comforts<sup>34</sup>, yet the potential cross- coupling between wireless sensing and delivery modulus on limited device area would interfere to their individual operations. Third, simultaneous satisfactions of detection sensitivity and on- demand drug delivery on a single device are also difficult, since the limited space of contact lens would restrict the sizes of sensor or delivery module to achieve effective operations.
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1. Schematic of the WTCL for real-time and in situ glaucoma diagnosis and therapy in a closed-loop manner. (a) Schematic of the WTCL for wireless glaucoma diagnosis and therapy. (b) Photograph of WTCL worn on the eyes of a live rabbit. (c) Schematic of wireless operation for the purpose of IOP monitoring and on-demand medicines administration in a closed-loop manner. The soft device, engineered as a double layer contact lens structure, was integrated with an LCR and a WPT receiver circuit. These modules were wirelessly connected to external integrated antenna that could record the IOP signal and trigger iontophoresis for drug delivery if needed. Insert figures respectively highlight critical IOP sensing and drug delivery unit. (d) Structure of the WTCL in an exploded view. (e) Optical image of the WTCL. (f) Schematic showing the structure of the cantilever capacitive sensor, which could be highly sensitive to pressure, allowing drug delivery circuits to integrate in limited space. </center>  
+
+In this work, for the first time to our knowledge, an integrated wireless theranostic contact lens (WTCL) was developed, for in situ glaucoma monitoring and electrically triggered drug administration at high- risk IOP conditions (Figure 1a). The WTCL employed a highly compact structural design and circuits layout, which enabled high
+
+<--- Page Split --->
+
+degraded integration of IOP sensing and on- demand delivery modulus on the curved and limited surface of contact lens without vision blockage (Figure 1b). The IOP sensing modulus possessed a unique cantilever configuration of LCR circuit, where each capacitive sensing plate sandwiching ultra- soft air dielectric film could ultra- sensitively respond to the IOP changes, producing detectable resonant frequency signals for wireless recording. The drug delivery modulus utilized a highly efficient wireless power transfer (WPT) circuit to drive anti- glaucoma drugs coated on electrode surface to migrate into the aqueous chamber via iontophoresis, which offered an electrical switch for drug delivery and enhancement of drug permeation across cornea (Figure 1c). The specialized design of wireless sensor and WPT receiver enabled channel separation via different operational frequencies without cross- coupling, ensuring the individual functions of modulus in an integrated closed- loop system. The minimally invasive, smart, wireless and closed- loop theranostic features of the WTCL endowed this platform as a highly promising tool for facilitating glaucoma treatments and preventing acute symptoms.  
+
+## Results  
+
+## System design and fabrications of the WTCL  
+
+Soft contact lens conformally interface with the cornea could effectively deform to transduce the expansion of the corneal limbus to the integrated sensor circuit when IOP increases, and localize external stimulations (e.g. electricity or chemicals) exerted on the cornea. Double layer- lens structure, a typical design acceptable for contact lens devices<sup>34,35</sup>, was adopted for fabricating WTCL to benefit integration of LCR circuit and drug administration module (Figure 1d) on the extremely limited space of contact lens (Figure 1e). The air film sandwiched between two layers of lens combined with LCR circuit characterized by cantilever structure formed the IOP transducer that could detect pressure fluctuations and transmit it wirelessly<sup>23,36</sup>. At high- risk IOP condition (IOP>21 mmHg), WPT triggered iontophoresis enables in situ drug administration effectively. Key advantages of this device contain: 1) soft, lightweight, re- usable, and minimally invasive features as well as wireless operations are highly compatible with contact lens platform. 2) Compact structural design and circuits layout enable high-
+
+<--- Page Split --->
+
+degraded integration of IOP monitoring and on- demand delivery modulus in limited area without vision blockage. 3) Rational circuit designs enabled sensitive IOP monitoring by unique cantilever sensor structure, on- demand and effective ocular drug delivery via iontophoresis, individually controlled channel without cross- coupling via frequency separation. 4) the cantilever capacitive sensor is highly sensitive to pressure, which allows drug delivery circuits to integrate in limited space, while maintaining sensitive IOP monitoring performance. Any slight distance displacement or angle displacement of the capacitive plates could easily induce significant electrical signals. Due to the cantilever design that ultra- soft air layer is present between the capacitive plates, the displacement of capacitive plates could be highly responsive to the pressure even the pressure could be partially buffered by the delivery coils on top of the sensor (Figure 1f). 5) The closed- loop system with entirely electrical interface is beneficial for signal collection, processing, feedback, and transmission, as well as programmable on- demand drug administration. 6) Fabrication of the device is compatible with existing large- scale and cost- effective manufacturing process, emphasizing its potential for widespread applications.  
+
+The IOP monitoring circuit employed a unique snowflake- shaped layout design (Figure 2a), where each capacitive sensing plate (totally 6 plates) was then aligned with reference plate (Figure 2b) by folding to form a cantilever configuration (Figure 2c). The reference plates and 5 coils of inductance were embedded in the upper lens (Figure 2d), while dangling sensing plates contacting the front surface of the lower lens, with a dielectric air film between the reference and sensing plates forming a variable capacitor (Figure S3.1). The capacitance combined with inductance coil formed a LCR circuit for wirelessly IOP monitoring. The deformation of corneal curvature caused by increased IOP compresses the thickness of air dielectric layer (Δd), leading to the rise of capacitance (C<sub>SR</sub>) and reduction of resonant frequency of LCR circuit that could be recorded by reading coil of integrated antenna (Figure 2e) wirelessly<sup>23,36</sup>. Due to the ultra- soft (ultra- low elastic modulus and zero viscoelasticity) feature of the sandwiched air film, the variable capacitors formed by cantilever configuration can ultra- sensitively respond according to the change of pressure (Figure S3.2). The cantilever design
+
+<--- Page Split --->
+
+effectively avoids the issues of redundant serial capacitors, and complicated device fabrication process especially the wire bonding step that have been encountered in previous reported strategies<sup>36,37</sup>. On the other hand, the drug delivery circuit utilized a flower- shaped layout design (Figure 2f) that enables robust interlocking mechanically between the flexible circuit and the lower layer of contact lens (Figure 2g)<sup>38</sup>. The front side of circuit embedded in the lower lens possessed coils (Figure 2f- IV and Figure S3.1) connected with chip capacitor for wireless power harvest, while the drugs- coated iontophoretic electrodes on the bottom side of delivery circuit were exposed and would be in contact with the cornea. Anti- glaucoma drugs, brimonidine, was loaded in a hydrogel layer coated on the iontophoretic electrode, which could be delivered into the aqueous humor via wirelessly iontophoresis to reduce IOP. The iontophoresis not only offered non- mechanical switch for drug delivery in a low- power consummation manner, but also facilitate drug penetration across cornea via electrophoresis effects<sup>39</sup>. The rational design of wireless sensor and WPT receiver at different operational frequencies (\~3.8 GHz and \~850 kHz, respectively) enabled channel separation for individual function. The double layer lens design enabled a compact structure to accommodate multiple electronic modulus positioned in the rim region of the contact lens, hence possessed an open vision window larger than the pupil’s size without blocking the views of wearers.  
+
+The fabrication of the sensing and delivery modulus employed a printed circuit process coupled with cast- molding method. For the sensing module, \(\mathrm{Cu} (\sim 100 \mu \mathrm{m})\) was electro- deposited on a flexible polyimide (PI) substrate and patterned via photolithography and wet etching, followed by covering with Ni/Au to improve biocompatibility of electrodes. The flexible substrate was cut by laser according to the snowflake- shaped circuit design, folded, and embedded into the upper polydimethylsiloxane (PDMS) lens via cast- molding technique (Figure S3.3), where each folded sensing plate was detached from the contact lens to form a cantilever configuration (Figure 2h). For the delivery module, similar to the sensing module, Cu was electro- deposited on PI substrate and patterned as coils features via photolithography and wet etching to form
+
+<--- Page Split --->
+
+the WPT receiver. A second layer of PI was spinning coated on top of the coils, and another Cu layer was prepared according to the iontophoretic electrode pattern, which was connected to the coils at the bottom side by through- holes. The electrodes were further covered with Ni/Au, and capacitors were soldered onto the front side of circuit to tune the WPT operation frequency, followed by embedding the circuit into the lower PDMS lens via cast- molding. The delivery electrode of lower lens was coated with a thin layer of drugs- loaded hydrogel, and assembled with the upper lens to form the final WTCL (Figure 2i). The compact layout and double layer lens design enables sensors and WPT receiver to be embedded inside contact lens, avoiding direct contact of these components to the ocular surface that might cause potential irritations to eye. The WTCL fabrication is compatible with commercial printed circuit board process, indicating the potentials for large- scale manufactures of this biomedical device. For experiments, an integrated antenna (Figure 2e) consisting of concentrically aligned IOP reading coils and WPT transmitter soldered on a matching circuit board was fabricated, which could collect the output signals from the wireless sensor and transfer power to the WPT receiver of the WTCL.
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure 2. Schematic illustration of the WTCL's design and fabrication process. (a) The snowflake-shaped layout design and the photograph of the sensing circuit. (b) The microscopic image of the reference plate, coils, and sensing plate deployed on sensing circuit. The photograph of (c) the folded sensing circuit and (d) the upper layer of contact lens. (e) Image of the integrated antenna. (f)(I) The flower-shaped layout design, (II) back surface (III) front surface images, and (IV) microscopic image of the drug delivery circuit. (g) The photograph of bottom layer lens integrated with drug delivery circuit. Illustration of the fabrication process of (h) IOP monitoring circuit, (i) drug delivery circuit and the device integration. The fabrication of the sensing and delivery modulus employed a printed circuit process coupled with cast-molding method. </center>  
+
+## In-vitro performance of wireless IOP monitoring  
+
+The sensing performance of the WTCL was tested in vitro using porcine eyeballs, where
+
+<--- Page Split --->
+
+porcine eyeballs' features similar with human eyeball have been widely employed in many physiological experiments. The IOP in porcine eye was tuned by controlled infusion of saline solution into the anterior chamber via microinfusion pump, with a pressure gauge to monitor the reference IOP. The IOP reading coil (diameter: \(17\mathrm{mm}\) , turns: 1) of the integrated antenna connected to a network analyzer was positioned on top of the WTCL to monitoring the resonance frequency (Figure 3a). The static sensing performance was conducted by stepwise increase of IOP, while the resonance frequency of WTCL at each IOP condition was recorded. The reflection spectra of six representative WTCL devices worn on the porcine eyeball at different IOP (5- \(50\mathrm{mmHg}\) ) were recorded and analyzed (Figure 3b and Figure S4.1a), where the resonant frequency of IOP monitoring module was found to shift to the lower frequency at higher IOP.  
+
+The return loss (S11) values at different frequencies and IOP conditions were plotted as heatmap diagrams, where the S11 value exhibited a linear pattern in the frequency- IOP heatmap (Figure 3c). The relation of resonance frequency and IOP of each device was analyzed (Figure 3d and Figure S4.2), which was revealed to be in an inversely linear profile (average R- Square \(= 0.976 \pm 0.015\) ). Theoretically the resonant frequency is in inversely related to the capacitance according to the LCR circuit equation \(\mathrm{f} = \left(2\pi \sqrt{\mathrm{LC}}\right)^{- 1}\) . Our results were consistent with the theoretical prediction in that the increase of IOP would reduce the distance between the capacitance electrodes, hence led to the elevation of capacitance value (Figure S4.1b) and reduction of resonant frequency. Noted that the linear relations of all the six devices overlapped well (normalized slope variation \(< 15\%\) ) with each other (Figure S4.3), indicating the reliability and repeatability of the fabricated device using our design.  
+
+A universal standard curve between resonant frequency and IOP was established by averaging all the six linear curves obtained from the measurements on the six representative devices (Figure 3e), where this standard curve could be employed to calculate detected IOP based on measured resonant frequency. The results suggested that the IOP sensors in the WTCL possessed sufficient sensitivity of \(1.28 \pm 0.09\)
+
+<--- Page Split --->
+
+MHz/mmHg, which was superior or comparable to other wireless IOP sensors (Table S2). This is likely due to the specific cantilever design of sensor, where the ultra- soft air film sandwiched between the sensing capacitive plates is highly mobile, so that the variable capacitors formed by cantilever configuration can respond to the change of pressure in a highly sensitive manner. The linear range of WTCL was wider than 5- 50 mmHg, which were desirable for glaucoma monitoring applications. The measured IOP values by the six WTCL devices were derived from the recorded values of resonant frequency, and compared to the reference IOP measured by pressure gauge for analyzing the sensor's static accuracy via error grid analysis (Figure 3f). The percentage of data points at different error range was quantified, where \(>50\%\) recording was found to be within error \(< 10\%\) , and \(>75\%\) recording was found to be within error \(< 20\%\) . The continuous recording of IOP via WTCL was also examined by measuring the resonant frequency and reference IOP via pressure gauge, respectively, where saline solution was injected into anterior chamber at \(t = 0\) s and \(883\) s intending to induce IOP spikes (Figure 3g). The measured resonant frequency (Figure S4.4) was calculated into IOP according to the WTCL's averaged standard curve, and the results were calibrated (Figure S4.5). Considering the possible batch variations of devices and porcine eyeballs in experiments, the detected IOP results via WTCL were calibrated by a pressure gauge- measured data point at \(t = 0\) s (indicated with blue star in Figure 3h). The injections of saline induced rising of IOP, followed by slight IOP declines potentially due to the gradual leakage of solution from eyeball, which were all consistently recorded by both WTCL and pressure gauge. The dynamic recording accuracy of WTCL at each time point was analyzed (Figure 3i) and plotted via an error grid analysis (Figure 3j), and the average error was found to be \(16.49 \pm 7.58\%\) with all the errors below \(30\%\) . The above results demonstrated the WTCL possessed sufficient sensitivity, linear region and reliability that were desirable for glaucoma monitoring.
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Figure 3. IOP sensing performance of the WTCL. (a-I) Schematic and (II) experimental set-up of the wireless IOP sensing experiments. (b) The reflection spectra of six representative WTCL devices worn on porcine eyeball at different IOP. (c) The results of the S11 values at different frequencies and IOP conditions in (b) were plotted as heatmap diagram, where the value of S11 exhibited a linear pattern in the frequency-IOP heatmap. (d) Linear regression of resonant frequency versus IOP value of each WTCL device. (e) The averaged linear regression of resonant frequency versus IOP value of the six WTCL devices in (b). (f) Error grid analysis and statical analysis of the IOP sensing accuracy via WTCL. Region A, B, C and D referred to errors \(< 10\%\) , 10-20%, 20-40% and \(>40\%\) , respectively. (g) Heatmap plot of the reflection coefficients recorded during the continuous recording of IOP via WTCL. (h) Continuous IOP signals monitored by WTCL on in-vitro porcine eyeball. Calibration point using reference IOP was marked with blue asterisks. The black arrow referred to the time </center>
+
+<--- Page Split --->
+
+point of saline injections. (i) Statistical analysis of detection errors via WTCL compared to commercial pressure gauge at different time points. Calibration point was marked with blue asterisks. (j) Error grid analysis of the continuous IOP sensing via WTCL. Region A+B, C and D referred to errors \(< 20\%\) , \(20 - 40\%\) and \(>40\%\) , respectively.  
+
+![](images/Figure_4.jpg)
+
+<center>Figure 4. WPT performance of the WTCL. (a-I) Schematic and (II) experimental set-up of the WPT experiments. (b) Reflection coefficient spectra (S11 and S21) recorded from four receivers at different radiation distance. (c) The S21 recorded by the four receivers at \(850\mathrm{kHz}\) were plotted as a function of radiation distance. (d) Alternating voltage signals collected from four receivers wirelessly under different </center>
+
+<--- Page Split --->
+
+frequency radiation at 20 Vpp applied on transmitter, and (e) Wirelessly transferred alternating voltage waveforms of four receivers under different radiation distance at 20 Vpp applied on transmitter, and (f) the Vpp were plotted as a function of frequency and (g) as a function of distance. (h) The wirelessly transferred voltage signals collected by Rec#17 activated by SquWave or SinWave voltages at different frequencies, and the Vpp were plotted as a function of (i)frequency or (j) distance. (k) Heatmap plot summarized the Vpp recorded from four receive circuits under different voltage transfer conditions, including the coupling frequency, the radiation distance, and the waveforms.  
+
+## The studies on WPT, cross-coupling and drug delivery  
+
+Magnetic resonance coupling- based WPT has been a competing technique for wireless bioelectronics due to its relatively high power transfer efficiency and resistance to environmental inference<sup>40</sup>. To achieve optimal coupling performance, four types of WPT receivers with 2, 5, 9, and 17 coils- design (namely Rec#2, Rec#5, Rec#9, and Rec#17, respectively) were designed with other circuit parameters accordingly modified (Figure S5.1 and Table S3). In order to evaluate the power transfer performance, the optimal coupling frequency and acceptable radiation distance between WPT receiver and transmitter were examined. During experiments, the WPT transmitter of the integrated antenna connected to a waveform generator and a network analyzer was aligned over the WTCL with identical axis (Figure 4a), while the WPT receivers were connected to an oscilloscope to monitored the generated voltages. The Reflection coefficient spectra from four receivers at different radiation distance were recorded (Figure 4b and Figure S5.2), where the resonant frequency of transmitter and all receivers were observed to be at \(\sim 850 \mathrm{kHz}\) according to the circuit designs. The channel separation between IOP monitoring ( \(\sim 3.8 \mathrm{GHz}\) ) and WPT ( \(\sim 850 \mathrm{kHz}\) ) was sufficiently large to avoid cross- coupling, which might prevent unexpected activation of a nontargeted wireless channel in the closed- loop glaucoma diagnosis and treatment<sup>41</sup>. The return loss (S11) reveals that most of the energy carried by electromagnetic wave could be radiated rather than dissipated in the frequency range of \(837.38 \mathrm{kHz}\) to \(867.45 \mathrm{kHz}\) , with a bandwidth of transmitter about \(30 \mathrm{kHz}^{42}\) . The S21
+
+<--- Page Split --->
+
+under \(850\mathrm{kHz}\) of all receivers decreased linearly with the increase of radiation distance (Figure 4c). Considering that certain distance between transmitting coils and contact lens is required to avoid interference to human eyes in practical applications, \(6\mathrm{mm}\) was chosen as optimal distance between transmitting coils and WTCL in experiments. Sequentially, a series of square wave (20 Vpp) with different frequencies (500 kHz to \(1200\mathrm{kH}\) , \(50\mathrm{kHz}\) step) or different distances ( \(0\mathrm{mm}\) to \(15\mathrm{mm}\) , \(1\mathrm{mm}\) step) were wirelessly exerted on the transmitter, to further verify the optimized coupling frequency and distance. The generated sinusoidal voltages on the receivers were recorded by oscilloscope (Figure 4d, Figure 4e and Figure S5.3- 5.4), and the relations between peak to peak (Vpp) values and frequencies were analyzed. At the set distance of \(6\mathrm{mm}\) , the Vpp increased sharply from \(500\) to \(850\mathrm{kHz}\) , and dramatically decreased from \(850\mathrm{kHz}\) to \(1.2\mathrm{MHz}\) , where the coupling at \(850\mathrm{kHz}\) displayed maximum Vpp of \(\sim 6\mathrm{V}\) (Figure 4f), consisting with the previous results of resonant frequency at \(\sim 850\mathrm{kHz}\) . On the other hand, at the set frequency of \(850\mathrm{kHz}\) , the Vpp of all receivers decreased with the increase of radiation distance (Figure 4g). The insert loss (S21) and Vpp of Rec#17 were both significantly higher than those of other receivers at identical conditions, suggesting that the Rec#17 possessed better matching with the WPT transmitter. Square wave (SquWave) and sine wave (SinWave), representing common voltage signals in analog electronics, possess distinct characteristics in rising and falling edges. In our experiments, SquWave and SinWave voltage signals featured with \(20\mathrm{Vpp}\) and different frequencies ( \(500\mathrm{kHz}\) to \(1200\mathrm{kHz}\) , with step of \(50\mathrm{kHz}\) ) were exerted on WPT transmitter (Figure 4h and Figure S5.5- 5.6) at different set distances ( \(0\mathrm{mm}\) to \(15\mathrm{mm}\) , \(1\mathrm{mm}\) step) to the Rec#17 which was chosen as the optimized receiver design. The correspondingly collected Vpp of receiver showed that the voltage transfer behaviors at SinWave voltage were similar to that at SquWave, where \(850\mathrm{kHz}\) was close to the optimal frequency. Moreover, the Vpp induced by the SquWave voltage wave was slightly higher than the that by SinWave (Figure 4i and Figure 4j), likely due to the fact that SquWave signals with more steeper edges created more rapidly changed magnetic field that is more favorable for WPT performance, compared to the SinWave at identical conditions.
+
+<--- Page Split --->
+
+To comprehensively evaluate the optimal conditions of WPT, receiver designs and the voltage transfer conditions (the coupling frequency, the radiation distance, and the waveforms) were systematically analyzed (Figure S5.3- 5.8) and summarized in two heatmap diagram (Figure 4k). Although the WPT efficiency was higher at shorter radiation distance, \(6\mathrm{mm}\) was selected as optimal distance between transmitting coils and WTCL since the contact lens needs certain separation from the transmitting coils in practical applications. The maximum transferred Vpp was observed on the optimal receiver Rec#17 at the applied SquWave with frequency of \(850\mathrm{kHz}\) , which were identified as the optimal conditions for the final WTCL.  
+
+The WPT performance of the WTCL was further theoretically analyzed, where the mutual inductance (M), power transfer efficiency \((\eta)\) (Figure S6.1), and the skin effects (Figure S6.2) were calculated based on the circuit design. The Mutual inductance (M), a key factor in the technology of WPT, determines voltage in the coil of receiver circuits, were derived from the magnetic coupling coefficient according to the circuit design of four receivers<sup>43</sup>. The mutual inductance was inversely proportional to the radiation distance between transmitter and receiver circuits, where the Rec#17 group (transmitter and Rec#17) exhibited the highest value of mutual inductance than other groups (Figure 5a). The power transfer efficiency of Rec#17 was further calculated to be \(48.4\%\) at the resonate frequency of \(850\mathrm{kHz}\) and \(6\mathrm{mm}\) radiation distance, significantly higher than other three receiver designs (Figure 5b), also consisting with the experimental results. Rader chart (Figure 5c) visually summarized performance (S21, Vpp, M, and \(\eta\) ) of etch WPT group (Rec#2, Rec#5, Rec#9 or Rec#17 linked to transmitter with \(6\mathrm{mm}\) radiation distance and \(850\mathrm{kHz}\) ), where Rec#17 group showing grater area in the chart compared to other alternatives. It demonstrated that Rec#17 group could be served as optimal power transfer platform for further iontophoretic drug administration in this work.  
+
+The cross- coupling between multiple wireless channels is a significant concern since it may disturb the independent control over the in situ sensing and delivery modulus (Figure 5d). Conventional strategy to spatially avoid cross- coupling is less compatible
+
+<--- Page Split --->
+
+with contact lens devices due to their limited space for spatially separation of channels<sup>44</sup>. Here we employed a specialized technique of radio frequency separation to solve the cross- coupling issue, based on a compact design of device to accommodate distinguished wireless circuits on the limited area of contact lens. Firstly, the IOP reading coil and WPT transmitter were coupled with the WPT receiver (using Rec#17) at a set distance of \(6\mathrm{mm}\) , respectively, and the S21 indicating the coupling efficiency and the generated voltages on receiver were separately measured. The coupling between WPT transmitter and receiver exhibited S21 higher than \(- 40\mathrm{dB}\) and reached its maximum value (- 15.6 dB) at around \(850\mathrm{kHz}\) , and apparent sinusoidal voltage waveform with \(6\mathrm{Vpp}\) was recorded. The coupling between reading coil and WPT receiver displayed ultra- low S21 ( \(< - 60\mathrm{dB}\) ), and generated negligible voltage that was close to blank group of uncoupled receiver, suggested cross- coupling between IOP reading coil and WPT receiver rarely occurred.  
+
+The WTCL was placed on porcine eye at different IOP (0- 50 mmHg), and the reading signals were recorded with or without the presence of radiation from WPT transmitter (Figure S7.1). The S11- frequency spectra appeared to be overlapping well disregard of the presence of WPT radiation, where a typical example at \(30\mathrm{mmHg}\) IOP was shown in Figure 5f- I. The resonance frequency and the peak S11 at different IOP were quantitatively analyzed (Figure 5f- II), where the radiation of WPT transmitter did not significantly influent the IOP monitoring, indicating cross- coupling between IOP sensor and WPT transmitter was negligible.
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Figure 5. (a) The mutual inductance and (b) Power transfer efficiencies were theoretically calculated according to the circuit design of four receivers and the radiation distance. (c) Rader chart summarized performance (S21, Vpp, M, and \(\eta\) ) of etch WPT group (Rec#2, Rec#5, Rec#9 or Rec#17 linked to transmitter with \(6\mathrm{mm}\) radiation distance and \(850\mathrm{kHz}\) ) (d) Schematic showing the experiments studying the cross-coupling between IOP monitoring and WPT module. Red arrow denoted the interference generated by the radiation of WPT transmitter to sensing module. Blue arrow denotes the cross-coupling between IOP reading coil and the WPT receiver. All examinations were performed with the radiation distance of \(6\mathrm{mm}\) . (e) The IOP reading coil and WPT transmitter were coupled with the WPT receiver (using Rec#17), respectively, and the S21 indicating the coupling efficiency and the generated voltages on receiver radiated at \(850\mathrm{kHz}\) were separately measured. (f) The WTCL was placed on porcine eye at different IOP, and the reading signals (the resonance frequency and the S11) were recorded with or without the presence of radiation from WPT transmitter. (g) The (top) 3D COMSOL model and (bottom) the simulated distribution profile of electric potential and electric field through the anterior region under the condition of iontophoresis at \(3\mathrm{V}\) for \(\mathrm{t} = 30\mathrm{min}\) . (h) The time slots of drug concentration profile delivered by WTCL at various applied voltages (0, 1, 2, and 3 V). (i) The delivered </center>
+
+<--- Page Split --->
+
+amounts of drugs at different conditions, including the (I) applied voltages, (II) iontophoretic duration and (III) assumed diffusivities, were systematically evaluated.  
+
+Brimonidine, maintaining positive charge, is a medication that has been topically delivered in aqueous humor or ciliary body clinically to treat glaucoma by increasing uveoscleral outflow and reducing aqueous fluid production<sup>8,14</sup>. However, corneal barrier comprised of tightly packed epithelium and hydrophilic- hydrophobic interfaces could significantly hinder the passive diffusion of drug molecules from ocular surface into anterior chamber<sup>39,45</sup>. Iontophoresis that drives the migration of charged species via electric field have been successful on transdermal pharmaceuticals delivery<sup>46,47</sup>, and the delivery dose can be tuned by the iontophoretic strength or duration<sup>48</sup>. To achieve effective and controllable ocular drug delivery, iontophoresis is coupled to the WTCL to enhance the transport of bromonidine across the cornea layer in an on- demand manner. Mixture solution of HEMA monomers, crosslinker, photoinitiator and brimonidine tartrate was drop- casted on the iontophoretic electrode surface, and was irradiated with UVB light to form a brimonidine- loaded pHEMA hydrogel layer<sup>49</sup>. When wirelessly powered by the transmitter, the WPT receiver generated alternating voltages on the electrode, which electrically drove the brimonidine into anterior chamber<sup>39</sup>.  
+
+To understand the process of iontophoretic delivery, a simplified 3D model (top of Figure 5g and Figure S8.1) imitating the actual scenario of WTCL worn on an eye was established using COMSOL Multiphysics 5.5, where the electric currents interface and the transport of diluted species interface were employed to calculate the electrically- driven drug diffusion profile. The anterior was modeled as a cornea layer (consisting of epithelial cell layer, stroma and epithelial cell layer) covered on anterior chamber, with corresponding electric conductivities and mass diffusivities. A working electrode coated with a thin layer of drug- loaded hydrogel and a counter electrode were conformally placed on the eye, and constant voltage was applied instead of alternating voltage in order to simplify the dynamic simulation process. Detailed boundary condition and parameters were shown in Figure S8.2, Figure S8.3 and Table S6 in
+
+<--- Page Split --->
+
+supplementary materials. The distribution of electric potential and electric field through the anterior region under iontophoresis at \(3\mathrm{V}\) at \(\mathrm{t} = 30\) min were shown in bottom of Figure 5g, and the time slots of drug concentration profile at various applied voltages (0, 1, 2, and 3 V) were shown in Figure 5h and Figure S8.4. Under iontophoresis, positively charged drug molecule migrated into the anterior chamber, where the drug delivery via iontophoresis was more effective ( \(\sim 3\) to 5- folds higher) than the passive diffusion, due to difficulty of drug diffusion across cornea. The delivered amounts of drugs at different conditions, including the applied voltages, iontophoretic duration and assumed diffusivities, were systematically evaluated (Figure 5i), and the results indicated the increase of these parameters would effectively enhance the drug delivery efficiency.  
+
+In vivo performance of the integrated WTCL
+
+<--- Page Split --->
+![](images/Figure_6.jpg)
+
+<center>Figure 6. Sensing and therapeutic performance of the integrated WTCL. (a) Quantitative analysis of in vitro rhodamine B released from WTCL at different alternating voltages for 30 min. N=6 measurements. (b) Rhodamine B was utilized as the medicines analog in ex vivo experiments on porcine eyes, to examine the influence of iontophoresis on drug delivery across cornea. CB, ciliary body; IS, iris; CA, Cornea. Scale bar is \(500\mu \mathrm{m}\) . After delivery, fluorescence visualization in the anterior tissue was </center>
+
+<--- Page Split --->
+
+observed via microscope and (c) quantitatively analyzed. N=4 sites per group. (d- I) Schematic and (II) experimental set- up of the in vivo WTCL experiments. The rabbits were anesthetized and worn with WTCL on their eyes, while the signals recording of WTCL and WPT operation were conducted by the integrated antenna. The rabbits' IOP were monitored with Tonopen, and brimonidine deliveries were performed via (e) Wirelessly powered iontophoresis or (f) passive free diffusion based on WTCL. (g) Simultaneous IOP sensing and drug delivery using a single WTCL device. The rabbit's IOP were wirelessly monitored with the WTCL, and brimonidine delivery via wireless iontophoresis on the same WTCL was conducted. The green asterisk indicated the IOP measurements via Tonopen for calibration or accuracy comparison. (h) Rabbit's eye was treated with eye drops of brimonidine, and the IOP was measured via Tonopen. N=2 rabbits in (e)- (g) and N=1 rabbit in (h). In (e)- (h), at each time point, 8 Topopen measurements were conducted, and the purple, blue, gray and white regions indicated the periods of prior to delivery, during delivery, 0.5 h after delivery and 2 h after delivery, respectively. (i) The IOP reductions effects after 0.5 and 2 h of drug delivery via iontophoresis, free diffusion and eye drops were summarized. Data were presented as mean ± s.d. Significance was evaluated by one- way analysis of variance. \*P < 0.05. (j) Monitoring of the thermal effects generated during WPT operation via infrared thermal camera, and (k) the temperate on cornea, WTCL, and transmitter coils during WPT process were analyzed. N=4 measurements per group. Data were presented as mean ± s.d. Significance was evaluated by one- way analysis of variance. \*P < 0.05.  
+
+The in vitro release of brimonidine (Mw 292.1) from pHEMA hydrogel- coated electrode surface at different iontophoretic voltages (alternating voltages with 0- 6 Vpp) was performed using a red fluorescent dye, rhodamine B (Mw 479.0), as the medicines analog to facilitate quantifications via optical measurements. The results (Figure 6a) showed that the dye molecule was continuously released at a higher rate when higher voltages were applied, likely due to the fact that the electric field facilitated the diffusion of dye out of the hydrogel layer. Ex vivo experiments on porcine eyes were performed to examine the influence of iontophoresis on delivery across cornea, where
+
+<--- Page Split --->
+
+rhodamine B was utilized as the medicines analog to facilitate visualization of distribution in tissue. The WTCL was worn on porcine eyes, and voltages with 6 Vpp at \(850\mathrm{kHz}\) (the determined optimal WPT operation frequency) was applied to facilitate dye delivery via iontophoresis, and the anterior tissue was then fixed and sectioned for fluorescence visualization via microscope. Iontophoresis at other frequencies (650 and \(1\mathrm{MHz}\) ) or passive free diffusion were tested to optimize the iontophoresis conditions, and the fluorescence intensity and distribution area in the anterior tissues were analyzed. Red fluorescence was clearly observed in the tissues of ciliary body and anterior chamber angle for all the samples treated via iontophoresis (Figure 6b), while the group of free diffusion exhibited significantly ( \(>3\) -folds) lower fluorescence intensity and less ( \(>3\) -folds) fluorescence distribution compared to the iontophoresis groups (Figure 6c). Of note, the ciliary body and anterior chamber angle have been proven to be the target sites for suppressing IOP by brimonidine through reducing aqueous humor production and increasing uveoscleral outflow. These results suggested the coupled iontophoresis could facilitate the delivery of drug analog molecules into anterior segment, and effectively work at the determined optimal WPT operation frequency of \(850\mathrm{kHz}\) .  
+
+Next, in vivo experiments were conducted on rabbits, while the size of WTCL was proportionally scaled down to fit the rabbits' eyes. The WTCL was worn on the anesthetized rabbits' eyes, while the signals recording of WTCL and WPT operation were conducted by the integrated antenna (Figure 6d). The rabbits' IOP were monitored with either WTCL or commercial tonometry as a standard reference, and brimonidine delivery via wirelessly powered iontophoresis of WTCL was performed to reduce the IOP and compared to that via eyedrop. The initial IOP of rabbits exhibited slight fluctuation within the range of \(10 - 15\mathrm{mmHg}\) as measured by Tonopen (Figure 6e), which rapidly ( \(< 0.5\mathrm{h}\) ) dropped by \(39.2\pm 10.3\%\) (Figure S9.1) after brimonidine delivery via wirelessly powered iontophoresis (at \(6\mathrm{Vpp}\) , \(850\mathrm{kHz}\) , for \(30\mathrm{min}\) ), and the IOP reduction remained above \(20\%\) for the prolonged period ( \(\sim 2\mathrm{h}\) ) after delivery (Figure S9.2). In contrast, brimonidine delivery via free diffusion (for \(30\mathrm{min}\) ) from WTCL only slightly reduced IOP by \(12.4\pm 14.3\%\) within \(0.5\mathrm{h}\) after delivery (Figure 6f
+
+<--- Page Split --->
+
+and Figure S9.1), and produced negligible effects \((6.85 + 14.7\%)\) within \(2\mathrm{~h}\) (Figure S9.3). These results suggested that the slow diffusion of brimonidine from WTCL might form a basal delivery to stabilize the IOP, while iontophoresis was able to facilitate a bolus delivery to more effectively reduce IOP spikes. Simultaneous IOP sensing and drug delivery using a single WTCL device were next performed (Figure 6g). The rabbit's IOP were wirelessly monitored with the WTCL for the first hour, then in situ brimonidine delivery via wireless iontophoresis on the same WTCL was conducted to reduce IOP, which were still continuously monitored by the WTCL (Figure 9.4 and Figure S9.5). Considering the variations between in vitro and vivo sensing and the differences between rabbit' porcine' eyes, the rabbits' IOP were measured with Tonopen before experiments (at \(\mathrm{t} = 0\mathrm{h}\) ) to calibrate the WTCL's sensing results (Figure 9.6 and Figure S9.7). The last data points of IOP recorded by WTCL were compared to the reference IOP measured via Tonopen after experiments, and the results showed that the sensing error of WTCL was \(< 42\%\) . The IOP was observed to gradually drop by \(32.5 \pm 35.9\%\) (Figure S9.1) within 0.5 hour, and remained reduction of \(43.2 \pm 38.8\%\) (Figure S9.3) for the prolonged period. As control, eye drops of brimonidine (1 mg/ml, 50 uL) were instilled into rabbit's eye, and the IOP measured via Tonopen showed a reduction of \(30.9 \pm 14.4\%\) during a short period ( \(< 30\) min) (Figure S9.1), followed by rebounding rapidly to the initial IOP state. The IOP reductions effects via iontophoresis, free diffusion and eye drops were summarized in Figure 6e-i and Figure S9.1-9.3, and the results confirmed that the iontophoresis via WTCL rapidly reduce the IOP with pronounced and prolonged effects that was desirable for regulating glaucoma. At the end, since WPT operation at high frequency is likely to produce thermal effects that are harmful to animal eyes, the temperatures of rabbits' eye surface (cornea) and WTCL were monitored via infrared thermal camera during the process of WPT operation (Figure 6j and Figure S10.1). The temperature of the and cornea was not increased, while the temperature of WTCL was observed to increase only by \(< 3^{\circ}\mathrm{C}\) (Figure 6k), respectively, during WPT for 30 min, suggesting negligible thermal effects produced by WTCL.
+
+<--- Page Split --->
+
+## Conclusion  
+
+ConclusionIn this work, a soft, minimally invasive and battery- free WTCL system for in situ IOP tracking and on- demand medicines administration was developed. The delicate design for structure, circuits layout of the device enabled highly integration on limited area and curved surface without causing vision blockage as well as potential irritations. The compact lens was exploited as a platform for deploying wireless bioelectronics and intimate contacting with human cornea, while the fabrication is compatible with high- throughput standard manufacturing process. The specialized design of frequency separation enabled individual operations of sensing and delivery modules without cross- coupling. Due to the unique cantilever configuration design of LCR circuit, the embedded wireless IOP sensor could ultra- sensitively detect IOP fluctuation, while the drug delivery modulus coupled with iontophoresis enabled highly efficient release of drug permeating across cornea. Systematic characterizations of IOP sensing, WPT, cross- coupling between individual sub- systems, iontophoretic medicines administration, and in vivo experiments all demonstrated the feasibility and promise of this WTCL platform for real- time monitoring and wireless controlled medical intervention. This smart system provides promising methodologies that could be expanded to other ophthalmic disease, which would positively promote the emerging of new generation of theranostic system for personalized health management.  
+
+## Methods  
+
+Theoretical analysis of mutual inductance (M), power transfer efficiency \((\eta)\) , and the skin effects. Mutual inductance \((M)\) , a key factor in the technology of WPT, determines voltage in a secondary coil of receiver circuits. The critical parameter could be expressed as \(^{50}\) :  
+
+\[M = k\sqrt{L_1L_2}\]  
+
+where \(L_{1}\) , \(L_{2}\) represent the inductance value of coil integrated in WPT transmitter and receiver circuit, respectively. \(k\) , denotes the magnetic coupling coefficient, means the link of magnetic flux between the WPT transmitter and receiver side \(^{43}\) . The parameter
+
+<--- Page Split --->
+
+is approximately equal to the following equation when the radiation distance is comparable to coils dimension \(^{43}\) .  
+
+\[k = \frac{1}{\left[1 + 2^{\frac{2}{3}}\left(\frac{d}{\sqrt{r_1r_2}}\right)^2\right]^{\frac{2}{3}}}\]  
+
+where \(d\) refers to distance between WPT transmitter and receiver circuit. Furthermore, \(\mathbf{r}_1, \mathbf{r}_2\) denote radius of inductance coil of transmitter and receiver circuit.  
+
+According to these two equations mentioned above, the \(M\) was inversely proportional to the radiation distance for these two coaxial coils of transmitter and receiver circuit.  
+
+According to Kirchhoff's voltage law, the equation of the WPT system could be described as  
+
+\[\left[ \begin{array}{c}U_{s}\\ 0 \end{array} \right] = \left[ \begin{array}{c}R_{PT} + j\omega L_{PT} - j\frac{1}{\omega C_{PT}} -j\omega M\\ -j\omega M R_{PR} + j\omega L_{PR} + \frac{Z_{L}}{1 - j\omega C_{PR}Z_{L}} \end{array} \right]\left[ \begin{array}{c}I_{PT}\\ I_{PR} \end{array} \right] \quad (1)\]  
+
+Where \(U_{s}, R_{PT}, L_{PT}, C_{PT}, I_{PT}\) refer to the alternating voltage supplied for the transmitter, parasitic resistance, inductor, capacitor and alternating current in transmitter. Correspondingly, \(R_{PR}, L_{PR}, C_{PR}, R_{L}\) denote the parasitic resistance, inductor, capacitor and electric load in receiver circuit. \(I_{PR}\) represents the total alternating current in receiver circuit. \(I_{L}\) is the alternating current flow through electric load.  
+
+To simplify the matrix, \(Z_{PT}\) and \(Z_{PR}\) were introduced as the impedance of transmitter and receiver circuit and expressed as  
+
+\[Z_{PT} = R_{PT} + j\omega L_{PT} - j\frac{1}{\omega C_{PT}} \quad (2)\]  
+
+\[Z_{PR} = R_{PR} + j\omega L_{PR} + \frac{Z_{L}}{1 - j\omega C_{PR}Z_{L}} \quad (3)\]  
+
+Therefore, equation (1) could be transformed as  
+
+\[\left[ \begin{array}{c}I_{PT}\\ I_{PR} \end{array} \right] = \frac{U_{S}}{Z_{PT}Z_{PR} + \omega^{2}M^{2}}\left[ \begin{array}{c}Z_{PR}\\ j\omega M \end{array} \right]\]  
+
+And the current consumed by load illustrated as  
+
+\[I_{L} = \frac{j\omega M U_{S}}{(1 - j\omega C_{PR}Z_{L})(Z_{PT}Z_{PR} + \omega^{2}M^{2})}\]  
+
+The power transfer efficiency \(\eta\) was regarded to be the ratio of the real power dissipated in the load impedance \(P_{out}\) to the power supplied from the source side \(P_{in}\) ,
+
+<--- Page Split --->
+
+\[\eta = \frac{P_{out}}{P_{in}} = \frac{\omega^2 M^2 Z_L}{Z_{PR}(1 + \omega^2 C_{PR}^2 Z_L^2)(Z_{PT}Z_{PR} + \omega^2 M^2)}\]  
+
+As regards high frequency circuit, alternating high- frequency currents tend to distributed toward the surface of conductor. This phenomenon, known as skin effect, will increase the resistance of the conductor and reduce the effective electric power exerted on load. The effective cross- section of the conductor for alternating currents was defined as skin depth that could be expressed by the following equation:  
+
+\[\delta = (\pi f\mu_r\mu_0\sigma)^{-\frac{1}{2}}\]  
+
+where \(\delta\) , \(f\) , \(\mu_r\) , \(\mu_0\) , and \(\sigma\) represent skin depth in meters, frequency of the alternating current in Hz, relative magnetic permeability of the conductive matter, permeability of free space \((4\pi \times 10^{- 7} \mathrm{H / m})\) , and conductivity of conductor. Detailed parameters (relative magnetic permeability and conductivity of the conductive matter) were listed in Table S4.  
+
+Fabrication of IOP monitoring circuits. The sensing and delivery were designed, and the fabricating process including Copper (Cu) film deposition, photolithography, etching, and laser cutting were performed by Shenzhen Gaoyue Electronics Co. Ltd, China. Cu film deposition process (including sputtering and electrical plating) was performed to establish electric film on the surface of PI substrate. After that, Cu film was patterned by photolithography and development process. Extra Cu film was etched by to form Cu electrodes, which were then covered with nickel (Ni) and gold (Au) to improve biocompatibility for the flexible circuits. Subsequently, ultraviolet beam excited by high energy YAG laser was utilized to cut the PI substrate to form the snowflake- shaped layout for the flexible IOP sensing circuits.  
+
+Fabrication of upper lens. Each capacitive sensing plate (totally 6 plates) of the flexible IOP sensing circuit was aligned with reference plate and folded manually. Then the folded IOP sensing circuit was positioned into the metal mold for contact lens. Polydimethylsiloxane (PDMS, Sylgard 184, Dow Corning) and curing agent were
+
+<--- Page Split --->
+
+prepared according to the ration of 10:1 and then stirred sufficiently. The transparent PDMS solution was placed in vacuum with pressure of 10 Pa for 30 min to remove bubbles, and injected into the metal mold. After vacuum treatment (10 Pa, 30 min), the upper mold and bottom mold were assembled and placed into oven (80 °C, 1.5 h). Afterwards, the contact lens embedded with IOP monitoring circuit was disassembled from mold carefully. Finally, sensing plates were detached from the upper contact lens manually. While reference plates and 5 coils of inductance were kept inside the upper lens. The dangling sensing plates aligned with reference plates served as cantilever configuration.  
+
+Fabrication, soldering of drug delivery circuits. The fabricating process contained film deposition, electrical plating, chemical plating, photolithography, etching, drilling, and laser cutting (by Shenzhen Gaoyue Electronics Co. Ltd, China). Cu film deposition process (including chemical and electrical plating) was performed to establish electric film on the surface of PI substrate. After that, Cu film was patterned by photolithography and development process and etched to form Cu electrodes, and the surface except the iontophoretic electrodes were insulated by a thin layer of PI. Ultraviolet laser with high energy was adopted to fabricate through- hole on PI substrate. Chemical and electrical plating were conducted to deposit Cu film on the surface of PI substrate and through- hole, which enables electric connections between electrodes on the top and bottom layers. Sequentially, photolithography, development and wet etching process were performed to form patterned Cu electrodes that were then covered with Ni and Au layer to improve biocompatibility for the flexible circuits. High energy laser was utilized to cut the PI substrate to define the flower- shaped layout for drug delivery circuits. Subsequently, ceramic chip capacitors (1 mm length, 0.5 mm width, 0.5 mm thickness) were attached to their respective sites on flexible circuits using low temperature solder by electrical soldering iron.
+
+<--- Page Split --->
+
+Fabrication of bottom lens. Drug delivery circuit was positioned in metal mold for contact lens. PDMS solution was placed in vacuum with pressure of \(10\mathrm{Pa}\) for \(30\mathrm{min}\) to remove bubbles, and injected into the mold. After vacuum treatment ( \(10\mathrm{Pa}\) , \(30\mathrm{min}\) ), the top and bottom molds were assembled and placed into oven ( \(80^{\circ}\mathrm{C}\) , \(1.5\mathrm{h}\) ). Afterwards, the PDMS contact lens integrated with drug delivery circuit was disassembled from mold carefully. Finally, extra PDMS film covered on iontophoretic electrodes (including delivery and counter electrodes) was removed manually.  
+
+WTCL integration and drug loading. The upper lens integrated with IOP monitoring circuit and bottom lens embedded with drug delivery circuit were assembled by liquid PDMS glue in oven ( \(60^{\circ}\mathrm{C}\) , \(3\mathrm{h}\) ). The materials for preparing pHEMA hydrogel preparation included HEMA monomer ( \(1.45\mathrm{ml}\) ), EGDMA ( \(5\mu \mathrm{l}\) ) as crosslinker, DI water ( \(0.5\mathrm{ml}\) ), and brimonidine tartrate ( \(10\mathrm{mg}\) ). Darocur ( \(6\mathrm{mg}\) ), a photoinitiator was mixed into monomer mixture and sonicated. The mixture solution of pHEMA hydrogel ( \(10\mu \mathrm{l}\) ) loaded with brimonidine tartrate ( \(5\mathrm{mg / ml}\) ) was drop- casted onto the drug delivery electrode. The solution was irradiated with UVB light ( \(365\mathrm{nm}\) ) for \(20\mathrm{min}\) for the hydrogel polymerization, and kept at room temperature overnight.  
+
+Characterization of circuits. Microscopic images of IOP sensing and drug delivery circuits were captured by inverted fluorescence microscope (MF52- N, Guangzhou Micro- shot Technology Co., Ltd, China)  
+
+In- vitro performance of wireless IOP monitoring. The porcine eyes were placed in the lab with room temperature ( \(27\pm 3^{\circ}\mathrm{C}\) ) and humidity ( \(50\pm 10\%\) ) to avoid the shape changes of eyeball induced by water loss. Before the deployment of wearable smart contact lens, physiological saline solution ( \(200\mu \mathrm{l}\) ) was drip on the surface of cornea to build a water film. The layer could be used to simulate tear film to avoid bubble between cornea and contact lens. IOPs ranging from 5 to \(50\mathrm{mmHg}\) were achieved inside the eye by injecting saline solution into the anterior chamber via a disposable
+
+<--- Page Split --->
+
+intravenous infusion needles \((0.45 \times 13.5 \mathrm{mm})\) controlled by syringe pump (PHD ULTRA, Harvard Apparatus, Inc., U.S.A.) During the experiments, a pressure gauge (GM511, Shenzhen Jumaoyuan Science And Technology Co., Ltd, China) was connected to the anterior chamber by a disposable intravenous infusion needle to independently track the value of IOP. The resonance frequency of the IOP monitoring module was record wirelessly by IOP reading coil (diameter: \(17 \mathrm{mm}\) , turns: 1) of the integrated antenna connected to a network analyzer (E5063A, Agilent Technologies Inc., Santa Clara, CA, USA).  
+
+Dynamical responson of wireless IOP monitoring. Physiological saline solution was injected into the anterior chamber of porcine eye to elevate the IOP from \(4.5 \mathrm{mmHg}\) to \(30 \mathrm{mmHg}\) . Then, the pressure decreases down to \(13 \mathrm{mmHg}\) with the continuous leaking of solution from eyeball. Sequentially, saline was filled into in vitro eyeball again to raise the pressure again. During this process, resonant frequency changes of the IOP monitoring module in WTCL were recorded wirelessly by network analyzer, and the pressure data in anterior chamber was validated using a commercial pressure gauge via a disposable intravenous infusion needle.  
+
+WPT performance characterization of the WTCL system. During the measurements, integrated antenna was immobilized above WTCL deployed on porcine eye. The in vitro organ that was posited on cystosepiment board held by a multi- axis stages to adjust the distance between WTCL and integrated antenna. During the collection of scattering parameter, two ports of vector network analyzer (E5063A, Keysight Technologies, USA) was connected to the WPT transmitter of the integrated antenna. Four different receivers (Rec#2, Rec#5, Rec#9, and Rec#17 integrated in four WTCLs) were connected to network analyzer successively to conduct data collections. During the measurements of wireless voltage transfer performance, waveform generator (DG1022, Beijing RIGOL Technology Co., Ltd., China) was adopted as power source for operating the WPT transmitter of the integrated antenna. Oscilloscope (TDS2014C,
+
+<--- Page Split --->
+
+Tektronix, USA) was connected to WPT receiver circuit of WTCL for recording the voltage signal collected wirelessly.  
+
+Characterization of disturbance generated by the radiation of WPT transmitter to IOP sensing module. Saline solution were injected into the anterior chamber of in vitro porcine eyes via a infusion needles \((0.45 \times 13.5 \mathrm{mm})\) controlled by syringe pump to achieve IOPs ranging from 5 to \(50 \mathrm{mmHg}\) . Network analyzer was connected to the IOP reading coil of the integrated antenna to monitor the physiological pressure transduced by WTCL. The distance between integrated antenna and WTCL was set as \(6 \mathrm{mm}\) . A commercial pressure gauge was exploited to validate the shifts of pressure inside of the anterior chamber through a infusion needles \((0.45 \times 13.5 \mathrm{mm})\) . Moreover, waveform generator was connected to WPT transmitter of the integrated antenna. During the process of IOP sensing, the power of waveform generator to support WPT transmitter was turned on and off to observe and record the S11 response of the IOP monitoring module.  
+
+Characterization of WPT performance of Rec#17 integrated in WTCL under the radiation of IOP reading coil. During the measurements, integrated antenna was immobilized above WTCL deployed on porcine eye. The distance between integrated antenna and WTCL was set as \(6 \mathrm{mm}\) . The IOP reading coil and Rec#17 integrated in WTCL was connected to two ports of network analyzer for the recording of the S21 parameters. As a control group, IOP reading coil was replaced by WPT transmitter that was coupled with Rec#17 to record S21 parameters. Similarly, The IOP reading coil of integrated antenna was connected to waveform generator. While, the Rec#17 circuit integrated in WTCL was connected to oscilloscope for the collecting of voltage signal. Correspondingly, WPT transmitter was adopted to replace the IOP sensing coil. Moreover, integrated antenna (including IOP reading coil and WPT transmitter) was disconnected from waveform generator, while Rec#17 was still connected with oscilloscope. The collected data serve as blank group.
+
+<--- Page Split --->
+
+Theoretical simulations of iontophoretic medicines administration via COMSOL Multiphysics 5.5. The theoretical simulations of iontophoretic medicines administration were performed with COMSOL Multiphysics software using the AC/DC module and Chemical Species Transport module. To visualize the effect of active agents administration and spatial distributions of electric potential, the process of cargo delivery were simulated with a 3D model, where the components and geometric layouts mimicked the actual experimental setup. Anterior segment of the eye was modeled as aqueous humor, cornea including epithelial cell, stroma, and endothelial cell layer. Counter electrode and pHEMA hydrogel that served as reservoir to load bio- active compounds were attached on cornea. Furthermore, the back surface of hydrogel was set as a drug delivery electrode to generate electric potential. Correspondingly, the counter electrode was labeled with ground in electric field. Under the action of constant electrical voltages, the working electrode combined with counter electrode form electric field through the tissue of cornea. For drug delivery, the drug concentration was set as Cg0 in hydrogel, and gradually diffused into aqueous humor through corneal barriers facilitated by electric field. After that, the average compounds concentration in the anterior chamber (aqueous humor) was calculated to evaluate the drug delivery efficiency. The bio- active molecules concentration was then normalized by comparing to the initial cargo's concentration loaded in hydrogel. Critical factors in this simulation work involves: 1) the drug diffusivities and the electrical conductivities in the pHEMA hydrogel, corneal layers, and aqueous humor; 2) Electrical charge of drugs. Detailed physic setting of cargo administrations, related parameters were demonstrated in table S5 in this supporting information file.  
+
+AC/DC module was exploited to simulate the steady electric filed distribution, which was performed by electric currents interface, following the theoretical equation:  
+
+\[\nabla \cdot J = 0\] \[J = \sigma E\] \[E = -\nabla V\]
+
+<--- Page Split --->
+
+Where V denotes potential, E refers to the intensity of electric field, J represents current density, \(\sigma\) is the material conductivity, \(\nabla\) refers to Hamiltonian. These equations mentioned above contributed to a Laplace equation that could be adopted to calculate electric potential and electric field in this model:  
+
+\[\nabla (\sigma \cdot \nabla V) = 0\]  
+
+On the top boundary of drug delivery electrode, a boundary voltage terminal was used to simulate the constant voltage source:  
+
+\[V = V_{0}\]  
+
+Where \(V_{0}\) refers to constant voltage.  
+
+Then, dynamic iontophoresis process was simulated by Chemical Species Transport module according to the electric filed distribution. The theoretical equation could be expressed as:  
+
+\[\frac{\partial c}{\partial t} +\nabla \cdot J_{t d s} = 0\] \[J_{t d s} = -D_{e}\nabla_{C} - z u_{m e}F c\nabla V\]  
+
+Where \(J_{t d s}\) refers to diffusion flux vector, c denotes the concentration, z represents the charge number, F refers to Faraday constant, V is electric potential, \(D_{e}\) corresponding to the effective diffusion coefficient, \(u_{m e}\) denotes the effective mobility. Therefore, the relationship of \(D_{e}\) and \(u_{m e}\) can be demonstrated by Nernst- Einstein equation:  
+
+\[u_{m e} = \frac{D e}{R T}\]  
+
+Where R represents Moore gas constant, T is temperature.  
+
+Quantitative analysis of rhodamine B released from WTCL. In this work, rhodamine B was exploited as the medicines analog to visualize the distribution of medicines in bio- tissue. Rhodamine B in PBS solution with the concentration of 0.5 to 30 ug/ml was prepared, and the standard curve of absorption value and solution's concentration was established. Then HEMA monomer (1.45 ml), EGDMA (5 μl) as crosslinker, DI water (0.5 ml), photoinitiator Darocur (9 mg) were mixed. The mixture solution of pHEMA hydrogel (10 μl) loaded with rhodamine B (0.4 mg/ml) was drop
+
+<--- Page Split --->
+
+casted onto the drug delivery electrode of WTCL. After that, the solution coated on WTCL was irradiated with UVB light (365 nm) for 20 min for the hydrogel polymerization, and kept at room temperature overnight. The WTCL connected with oscilloscope, and powered by the WPT transmitter of the integrated antenna underneath the WTCL (distance: 6 mm). \(200~\mu \mathrm{l}\) PBS solution was placed on the WTCL to allow dye diffusion, and \(100~\mu \mathrm{l}\) solution was withdrawn every 5 minutes for analysis of the diffusion rate. Accordingly, fresh PBS with equal volume was re- supplemented into WTCL. These absorbance values of solutions collected at each time point was measured by microplate reader. According to these absorbance values, the accumulated concentration of released rhodamine B could be calculated according to the equation mentioned above.  
+
+Ex vivo delivery of Rhodamine B in to porcine eyes by WTCL. The WTCL was worn on in vitro porcine eye, and the integrated antenna was deployed on top of WTCL with the distance of \(6\mathrm{mm}\) . Waveform generator was connected to the WPT transmitter of integrated antenna as power source. Oscilloscope was connected to the WPT Rec#17 circuit in WTCL to monitor the voltage adopted for iontophoresis and ensure that the received alternating voltage stabilized around \(6\mathrm{Vpp}\) with the frequency of 650 kHz, \(850\mathrm{kHz}\) , and \(1\mathrm{MHz}\) , respectively. Moreover, drug delivery in the manner of free diffusion was performed as a control group. A through- hole (diameter: \(4\mathrm{mm}\) ) was created in the central area of WTCL to allow dropping of PBS solution to the eye surface for maintaining humidity. During experiments, PBS was instilled into the central hole of WTCL with the speed of \(30~\mu \mathrm{L}\) every \(30\mathrm{s}\) to form a thin solution film on the corneal surface. The liquid film has been regarded as simulant of tears to ensure reliable connection between drug delivery and cornea electrically, and also prevent drying of ocular surface tissue<sup>39</sup>. After the completion of examinations, the corneal surface of porcine eye was irrigated by PBS. Then extra tissues (muscle, fat) outside of eyeball were removed by dissecting scissors. The whole eyeball was fixed in a paraformaldehyde solution (Fixative Solution, \(4\%\) formaldehyde, methanol- free,
+
+<--- Page Split --->
+
+Biosharp Co., Ltd, China), sectioned, stained by Wuhan Servicebio Technology Co., Ltd. Sequentially, fluorescent microscopic images including bio- tissue of ciliary body, iris, cornea (site 1- 4) in each experiment condition (free diffusion, and 6 Vpp with the frequency of \(650\mathrm{kHz}\) , \(850\mathrm{kHz}\) , \(1\mathrm{MHz}\) ) were taken and processed by Image J program to quantify the fluorescence intensity and distribution area in the anterior tissues. The mean distribution area and integrated density of rhodamine B in the sample treated by iontophoretic drug administration with \(20\mathrm{Vpp}\) at \(850\mathrm{kHz}\) were set to be base reference of 1 for normalization. Correspondingly, normalized values of distribution area and integrated density of rhodamine B in the sample treated by other drug delivery conditions could be quantified.  
+
+Animal experiments. Male New Zealand white rabbits weighing about \(3\mathrm{kg}\) (Animal Center, Sun Yat- sen University, Guangzhou, China), adopted for in vivo experiments, were maintained in climate- controlled independent room with \(12\mathrm{h} / 12\mathrm{h}\) light/dark cycle separately. All in vivo experiments in this study were reviewed, permitted, and supervised by the Institutional Animal Care and Use Committee of the Sun Yat- sen University. For all in vivo experiment process, the rabbits were deeply anesthetized with pentobarbital sodium solution ( \(0.8\mathrm{ml / kg}\) body weight). To minimize side effects, the administration of anesthetic solution was separately into three times through ear venous and twice intramuscular injection every ten minutes successively. Moreover, Isoflurane and oxygen were supplied through gas anaesthesia machine for rabbit to obtain prolonged anesthesia effects. Propivacaine hydrochloride eye drops (S. A. ALCON- COUVREUR N.V. Belgium) were dropped onto the rabbit cornea surface for topical anesthesia to avoid ocular movement including blink, facilitate WTCL wearing and IOP measurement by commercial ophthalmotonometer.
+
+<--- Page Split --->
+
+In vivo experiments of WTCL performance. Pentobarbital sodium (Nembutal, Ovation Pharmaceuticals Inc. Deerfield, USA) solution in saline (0.3 wt%) was prepared. New Zealand white rabbit (3 kg) was initially anesthetized with an appropriate dose of pentobarbital sodium solution (0.8 ml/kg body weight), and continuously anesthetized with Isoflurane anesthesia machine. To avoid unexpected situations, the administration of anesthetic solution was divided into three times through ear venous and twice intramuscular injection every ten minutes successively. After anesthesia, the rabbits were covered with blanket to maintain body temperature. Propivacaine hydrochloride eye drops were dropped onto the rabbit cornea surface for local anesthesia to further avoid of ocular movement including blink. Commercial applanation tonometer (Tono- Pen Avia; Reichert, Inc., Depew, NY) was applied to acquire IOP measurement as reference. WTCL was worn on rabbit's eye, and oscilloscope was connected to the WTCL to monitor the Vpp between delivery and counter electrode during WPT process. Integrated antenna connected to network analyzer and waveform generator was posited above WTCL with the distance of 6 mm. Sequentially, measurements of return loss by IOP reading coil was collected by network analyzer to wirelessly detect IOP. After one hour, square voltage with 20 Vpp at 850 kHz produced from waveform generator was exerted on WPT transmitter to trigger iontophoretic delivery wirelessly. Meanwhile, wireless IOP monitoring was continuously performed until the end of experiments.  
+
+Thermal characterization. New Zealand white rabbit (3 kg) was anesthetized with an appropriate dose of pentobarbital sodium solution (0.8 ml/kg body weight) through ear venous injections. After general anesthesia, the rabbits were covered with blanket to maintain body temperature. Then anesthesia machine was further adopted to supply isoflurane and oxygen via facemask for the rabbit, which could obtain prolonged anesthesia effects. WTCL was worn on rabbit's eye, and integrated antenna connected to waveform generator was posited above WTCL with the distance of 6 mm. Square voltage with 20 Vpp at 850 kHz produced from waveform generator was exerted on
+
+<--- Page Split --->
+
+WPT transmitter. Infrared camera (T650sc, FLIR Systems, Wilsonille, OR, USA) was exploited to monitor thermal changes of ocular surface tissue, WTCL, and integrated antenna during the experimental process.  
+
+## References  
+
+1 Lee, H. et al. Wearable/disposable sweat- based glucose monitoring device with multistage transdermal drug delivery module. Sci. Adv. 3, e1601314, doi:10.1126/sciadv.1601314 (2017).  2 Keum, D. H. et al. Wireless smart contact lens for diabetic diagnosis and therapy. Sci. Adv. 6, eaba3252, doi:10.1126/sciadv.aba3252 (2020).  3 Lee, H. et al. A graphene- based electrochemical device with thermoresponsive microneedles for diabetes monitoring and therapy. Nat. Nanotechnol. 11, 566- 572, doi:10.1038/nnano.2016.38 (2016).  4 Son, D. et al. Multifunctional wearable devices for diagnosis and therapy of movement disorders. Nat. Nanotechnol. 9, 397- 404, doi:10.1038/nnano.2014.38 (2014).  5 Han, M. et al. Catheter- integrated soft multilayer electronic arrays for multiplexed sensing and actuation during cardiac surgery. Nat. Biomed. Eng. 4, 997- 1009, doi:10.1038/s41551- 020- 00604- w (2020).  6 Mickle, A. D. et al. A wireless closed- loop system for optogenetic peripheral neuromodulation. Nature 565, 361- 365, doi:10.1038/s41586- 018- 0823- 6 (2019).  7 Weinreb, R. N. et al. Primary open- angle glaucoma. Nat. Rev. Dis. Primers 2, 16067, doi:10.1038/nrdp.2016.67 (2016).  8 Jonas, J. B. et al. Glaucoma. Lancet 390, 2183- 2193, doi:10.1016/s0140- 6736(17)31469- 1 (2017).  9 Wang, N., Chintala, S. K., Fini, M. E. & Schuman, J. S. Activation of a tissue- specific stress response in the aqueous outflow pathway of the eye defines the glaucoma disease phenotype. Nat. Med. 7, 304- 309, doi:10.1038/85446 (2001).  10 Hughes, E., Spry, P. & Diamond, J. 24- hour monitoring of intraocular pressure in glaucoma management: A retrospective review. J. Glaucoma 12, 232- 236, doi:10.1097/00061198- 200306000- 00009 (2003).  11 Taylor, S. A., Galbraith, S. M. & Mills, R. P. Causes of non- compliance with drug regimens in glaucoma patients: A qualitative study. J. Ocul. Pharmacol. Ther. 18, 401- 409, doi:10.1089/10807680260362687 (2002).  12 Grehn, F. & Stamper, R. Glaucoma: [progress III]. (Springer, 2006).  13 Agaoglu, S. et al. Ultra- sensitive microfluidic wearable strain sensor for intraocular pressure monitoring. Lab Chip 18, 3471- 3483, doi:10.1039/c8lc00758f (2018).  14 Morrison, J. C. & Pollack, I. P. Glaucoma: Science and Practice. (Thieme, 2003).  15 Novack, G. D. Ophthalmic Drug Delivery: Development and Regulatory Considerations. Clin. Pharmacol. Ther. 85, 539- 543, doi:10.1038/clpt.2008.297 (2009).  16 Richa, S. & Yazbek, J. C. Ocular Adverse Effects of Common Psychotropic Agents A Review. CNS Drugs 24, 501- 526, doi:10.2165/11533180- 000000000- 00000 (2010).  17 Farandos, N. M., Yetisen, A. K., Monteiro, M. J., Lowe, C. R. & Yun, S. H. Contact Lens Sensors in Ocular Diagnostics. Adv. Healthc. Mater. 4, 792- 810, doi:10.1002/adhm.201400504
+
+<--- Page Split --->
+
+(2015).18 Tseng, R. C., Chen, C. C., Hsu, S. M. & Chuang, H. S. Contact-Lens Biosensors. Sensors 18, 2651, doi:10.3390/s18082651 (2018).19 Kim, J. et al. Intraocular Pressure Monitoring Following Islet Transplantation to the Anterior Chamber of the Eye. Nano Lett. 20, 1517-1525, doi:10.1021/acs.nanolett.9b03605 (2020).20 Maeng, B., Chang, H.-k. & Park, J. Photonic crystal-based smart contact lens for continuous intraocular pressure monitoring. Lab Chip 20, 1740-1750, doi:10.1039/c9lc01268k (2020).21 Park, J. et al. Soft, smart contact lenses with integrations of wireless circuits, glucose sensors, and displays. Sci. Adv. 4, eaap9841, doi:10.1126/sciadv.aap9841 (2018).22 Elsherif, M., Hassan, M. U., Yetisen, A. K. & Butt, H. Glucose Sensing with Phenylboronic Acid Functionalized Hydrogel-Based Optical Diffusers. Acs Nano 12, 2283-2291, doi:10.1021/acsnano.7b07082 (2018).23 Kim, J. et al. Wearable smart sensor systems integrated on soft contact lenses for wireless ocular diagnostics. Nat. Commun. 8, 14997, doi:10.1038/ncomms14997 (2017).24 medGadget. Smart, Continuous Monitoring of Intra-Ocular Pressure with Triggerfish Contact Lens: Q&A with René Goedkoop, CMO of Sensimed, <https://www.medgadget.com/2013/07/smart-continuous-monitoring-of-the-intra-ocular-pressure-with-the-triggerfish-contact-lens-qa-with-rene-goedkoop-cmo-of-sensimed.html> (2013).25 Lee, S.-H., Shin, K.-S., Kim, J.-W., Kang, J.-Y. & Kim, J.-K. Stimulus-Responsive Contact Lens for IOP Measurement or Temperature-Triggered Drug Release. Transl. Vis. Sci. Technol. 9, 1-1, doi:10.1167/tvst.9.4.1 (2020).26 Song, C., Ben-Shlomo, G. & Que, L. A Multifunctional Smart Soft Contact Lens Device Enabled by Nanopore Thin Film for Glaucoma Diagnostics and in situ Drug Delivery. J. Microelectromech. Syst. 28, 810-816, doi:10.1109/JMEMS.2019.2927232 (2019).27 Kim, J. et al. A soft and transparent contact lens for the wireless quantitative monitoring of intraocular pressure. Nat. Biomed. Eng 5, 772-782, doi:10.1038/s41551-021-00719-8 (2021).28 Kim, H.-J., Zhang, K., Moore, L. & Ho, D. Diamond Nanogel-Embedded Contact Lenses Mediate Lysozyme-Dependent Therapeutic Release. Acs Nano 8, 2998-3005, doi:10.1021/nn5002968 (2014).29 Guzman, G., Es-haghi, S. S., Nugay, T. & Cakmak, M. Zero-Order Antibiotic Release from Multilayer Contact Lenses: Nonuniform Drug and Diffusivity Distributions Produce Constant-Rate Drug Delivery. Adv. Healthc. Mater. 6, 1600775, doi:10.1002/adhm.201600775 (2017).30 Janagam, D. R., Wu, L. & Lowe, T. L. Nanoparticles for drug delivery to the anterior segment of the eye. Adv. Drug Deliv. Rev. 122, 31-64, doi:10.1016/j.addr.2017.04.001 (2017).31 Zhao, Y. et al. Skin-Inspired Antibacterial Conductive Hydrogels for Epidermal Sensors and Diabetic Foot Wound Dressings. Adv. Funct. Mater. 29, 1901474 doi:10.1002/adfm.201901474 (2019).32 Sim, K. et al. Three-dimensional curvy electronics created using conformal additive stamp printing. Nature Electronics 2, 471-479, doi:10.1038/s41928-019-0304-4 (2019).33 Guo, S. et al. Integrated contact lens sensor system based on multifunctional ultrathin MoS2 transistors. Matter 4, 969-985, doi:10.1016/j.matt.2020.12.002 (2021).34 Park, J. et al. Printing of wirelessly rechargeable solid-state supercapacitors for soft, smart contact lenses with continuous operations. Sci. Adv. 5, eaay0764, doi:10.1126/sciadv.aay0764
+
+<--- Page Split --->
+
+(2019). 35 Yuan, M. et al. Electronic Contact Lens: A Platform for Wireless Health Monitoring Applications. Adv. Intell. Syst. 2, 1900190, doi:10.1002/aisy.201900190 (2020). 36 Boutry, C. M. et al. Biodegradable and flexible arterial-pulse sensor for the wireless monitoring of blood flow. Nat. Biomed. Eng 3, 47- 57, doi:10.1038/s41551- 018- 0336- 5 (2019). 37 Chen, G.- Z., Chan, I.- S. & Lam, D. C. C. Capacitive contact lens sensor for continuous noninvasive intraocular pressure monitoring. Sens. Actuator A- Phys. 203, 112- 118, doi:10.1016/j.sna.2013.08.029 (2013). 38 Qi, D. et al. Highly Stretchable, Compliant, Polymeric Microelectrode Arrays for In Vivo Electrophysiological Interfacing. Adv. Mater. 29, 1702800, doi:10.1002/adma.201702800 (2017). 39 Christopher, K. & Chauhan, A. Delivery of ionic molecules to anterior chamber by iontophoretic contact lenses. Eur. J. Pharm. Biopharm. 140, 40- 49, doi:10.1016/j.ejpb.2019.04.016 (2019). 40 Kurs, A. et al. Wireless power transfer via strongly coupled magnetic resonances. Science 317, 83- 86, doi:10.1126/science.1143254 (2007). 41 Noh, K. N. et al. Miniaturized, Battery- Free Optofluidic Systems with Potential for Wireless Pharmacology and Optogenetics. Small 14, 1702479, doi:10.1002/smll.201702479 (2018). 42 Bansal, R. Antenna theory; analysis and design. Proceedings of the IEEE 72, 989- 990, doi:10.1109/PROC.1984.12959 (1984). 43 Imura;, T. in Wireless Power Transfer Using Magnetic and Electric Resonance Coupling Techniques 40- 41 (Springer, 2020). 44 Sempionatto, J. R. et al. An epidermal patch for the simultaneous monitoring of haemodynamic and metabolic biomarkers. Nat. Biomed. Eng 5, 737- 748, doi:10.1038/s41551- 021- 00685- 1 (2021). 45 Gaudana, R., Ananthula, H. K., Parenky, A. & Mitra, A. K. Ocular Drug Delivery. AAPS J. 12, 348- 360, doi:10.1208/s12248- 010- 9183- 3 (2010). 46 Wu, C. et al. Self- Powered Iontophoretic Transdermal Drug Delivery System Driven and Regulated by Biomechanical Motions. Adv. Funct. Mater. 30, 1907378, doi:10.1002/adfm.201907378 (2020). 47 Emaminejad, S. et al. Autonomous sweat extraction and analysis applied to cystic fibrosis and glucose monitoring using a fully integrated wearable platform. Proc. Natl. Acad. Sci. U. S. A. 114, 4625- 4630, doi:10.1073/pnas.1701740114 (2017). 48 Behar- Cohen, F. F. et al. Transscleral Coulomb- controlled iontophoresis of methyl prednisolone into the rabbit eye: Influence of duration of treatment, current intensity and drug concentration on ocular tissue and fluid levels. Exp. Eye Res. 74, 51- 59, doi:10.1006/exer.2001.1098 (2002). 49 Maulvi, F. A. et al. In vitro and in vivo evaluation of novel implantation technology in hydrogel contact lenses for controlled drug delivery. J. Control. Release 226, 47- 56, doi:10.1016/j.jconrel.2016.02.012 (2016). 50 Griffiths, D. J. Introduction to Electrodynamics. (Prentice Hall, 1999).  
+
+## Acknowledgements
+
+<--- Page Split --->
+
+The authors would like to acknowledge financial support from the National Natural Science Foundation of China (Grant No. 61771498, 61901535, 81970778) and Science and Technology Planning Project of Guangdong Province for Industrial Applications (Grant No. 2017B090917001), Guangdong Province Key Area R&D Program (Grant No.2018B030332001), Science and Technology Program of Guangzhou, China (Grant No. 202102080192) and Guangdong Basic and Applied Basic Research Foundation (Grant No. 2021A1515012261, 2019A1515012087, 2020A1515010987, 2020A1515110424), Key Program of Sun Yat- Sen University (20lgzd14).  
+
+## Author contributions  
+
+C.Y. and X.X. conceived the concept, designed the work, analysed data and wrote the manuscript. C.Y., Q.N.W., J.Q.L., J.S.M., X.L.L., C.D.Y., Z.Q.L., J.B.Y., L.L.J., W.R.C., H.J.C., J.W., and X.X. performed statistical analyses of datasets and aided in the preparation of displays communicating datasets. X.X. supervised the study. All authors discussed the results and assisted in the preparation of the manuscript.
+
+<--- Page Split --->
+
+## Figures  
+
+![](images/Figure_1.jpg)
+
+<center>Figure 1 </center>  
+
+Schematic of the WTCL for real- time and in situ glaucoma diagnosis and therapy in a closed- loop manner. (a) Schematic of the WTCL for wireless glaucoma diagnosis and therapy. (b) Photograph of WTCL worn on the eyes of a live rabbit. (c) Schematic of wireless operation for the purpose of IOP monitoring and on- demand medicines administration in a closed- loop manner. The soft device, engineered as a double layer contact lens structure, was integrated with an LCR and a WPT receiver circuit. These modules were wirelessly connected to external integrated antenna that could record the IOP signal and trigger iontophoresis for drug delivery if needed. Insert figures respectively highlight critical IOP sensing and drug delivery unit. (d) Structure of the WTCL in an exploded view. (e) Optical image of the WTCL. (f) Schematic showing the structure of the cantilever capacitive sensor, which could be highly sensitive to pressure, allowing drug delivery circuits to integrate in limited space.
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure 2 </center>  
+
+Schematic illustration of the WTCL's design and fabrication process. (a) The snowflake- shaped layout design and the photograph of the sensing circuit. (b) The microscopic image of the reference plate, coils, and sensing plate deployed on sensing circuit. The photograph of (c) the folded sensing circuit and (d) the upper layer of contact lens. (e) Image of the integrated antenna. (f)(l) The flower- shaped layout design, (ll) back surface (lll) front surface images, and (IV) microscopic image of the drug delivery circuit. (g) The photograph of bottom layer lens integrated with drug delivery circuit. Illustration of the fabrication process of (h) IOP monitoring circuit, (i) drug delivery circuit and the device integration. The fabrication of the sensing and delivery modulus employed a printed circuit process coupled with cast- molding method.
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Figure 3 </center>  
+
+IOP sensing performance of the WTCL. (a- I) Schematic and (II) experimental set- up of the wireless IOP sensing experiments. (b) The reflection spectra of six representative WTCL devices worn on porcine eyeball at different IOP. (c) The results of the S11 values at different frequencies and IOP conditions in (b) were plotted as heatmap diagram, where the value of S11 exhibited a linear pattern in the frequency- IOP heatmap. (d) Linear regression of resonant frequency versus IOP value of each WTCL device. (e) The averaged linear regression of resonant frequency versus IOP value of the six WTCL devices in (b). (f) Error grid analysis and statical analysis of the IOP sensing accuracy via WTCL. Region A, B, C and D referred to errors \(< 10\%\) , \(10 - 20\%\) , \(20 - 40\%\) and \(>40\%\) , respectively. (g) Heatmap plot of the reflection coefficients
+
+<--- Page Split --->
+
+recorded during the continuous recording of IOP via WTCL. (h) Continuous IOP signals monitored by WTCL on in- vitro porcine eyeball. Calibration point using reference IOP was marked with blue asterisks. The black arrow referred to the time point of saline injections. (i) Statistical analysis of detection errors via WTCL compared to commercial pressure gauge at different time points. Calibration point was marked with blue asterisks. (j) Error grid analysis of the continuous IOP sensing via WTCL. Region A+B, C and D referred to errors \(< 20\%\) , \(20 - 40\%\) and \(>40\%\) , respectively.  
+
+![](images/Figure_4.jpg)
+
+<center>Figure 4 </center>
+
+<--- Page Split --->
+
+WPT performance of the WTCL. (a- I) Schematic and (II) experimental set- up of the WPT experiments. (b) Reflection coefficient spectra (S11 and S21) recorded from four receivers at different radiation distance. (c) The S21 recorded by the four receivers at 850 kHz were plotted as a function of radiation distance. (d) Alternating voltage signals collected from four receivers wirelessly under different frequency radiation at 20 Vpp applied on transmitter, and (e) Wirelessly transferred alternating voltage waveforms of four receivers under different radiation distance at 20 Vpp applied on transmitter, and (f) the Vpp were plotted as a function of frequency and (g) as a function of distance. (h) The wirelessly transferred voltage signals collected by Rec#17 activated by SquWave or SinWave voltages at different frequencies, and the Vpp were plotted as a function of (i) frequency or (j) distance. (k) Heatmap plot summarized the Vpp recorded from four receive circuits under different voltage transfer conditions, including the coupling frequency, the radiation distance, and the waveforms.  
+
+![](images/Figure_5.jpg)
+
+<center>Figure 5 </center>  
+
+(a) The mutual inductance and (b) Power transfer efficiencies were theoretically calculated according to the circuit design of four receivers and the radiation distance. (c) Rader chart summarized performance (S21, Vpp, M, and n) of etch WPT group (Rec#2, Rec#5, Rec#9 or Rec#17 linked to transmitter with 6 mm radiation distance and 850 kHz) (d) Schematic showing the experiments studying the cross-coupling between IOP monitoring and WPT module. Red arrow denoted the interference generated by the radiation of WPT transmitter to sensing module. Blue arrow denotes the cross-coupling between IOP reading coil
+
+<--- Page Split --->
+
+and the WPT receiver. All examinations were performed with the radiation distance of \(6 \text{mm}\) . (e) The IOP reading coil and WPT transmitter were coupled with the WPT receiver (using Rec#17), respectively, and the S21 indicating the coupling efficiency and the generated voltages on receiver radiated at \(850 \text{kHz}\) were separately measured. (f) The WTCL was placed on porcine eye at different IOP, and the reading signals (the resonance frequency and the S11) were recorded with or without the presence of radiation from WPT transmitter. (g) The (top) 3D COMSOL model and (bottom) the simulated distribution profile of electric potential and electric field through the anterior region under the condition of iontophoresis at \(3 \text{V}\) for \(t = 30 \text{min}\) . (h) The time slots of drug concentration profile delivered by WTCL at various applied voltages (0, 1, 2, and 3 V). (i) The delivered amounts of drugs at different conditions, including the (I) applied voltages, (II) iontophoretic duration and (III) assumed diffusivities, were systematically evaluated.
+
+<--- Page Split --->
+![](images/Figure_6.jpg)
+
+<center>Figure 6 </center>  
+
+Sensing and therapeutic performance of the integrated WTCL. (a) Quantitative analysis of in vitro rhodamine B released from WTCL at different alternating voltages for 30 min. N=6 measurements. (b) Rhodamine B was utilized as the medicines analog in ex vivo experiments on porcine eyes, to examine the influence of iontophoresis on drug delivery across cornea. CB, ciliary body; IS, iris; CA, Cornea. Scale bar is 500 μm. After delivery, fluorescence visualization in the anterior tissue was observed via
+
+<--- Page Split --->
+
+microscope and (c) quantitatively analyzed. N=4 sites per group. (d- I) Schematic and (II) experimental setup of the in vivo WTCL experiments. The rabbits were anesthetized and worn with WTCL on their eyes, while the signals recording of WTCL and WPT operation were conducted by the integrated antenna. The rabbits' IOP were monitored with Tonopen, and brimonidine deliveries were performed via (e) Wirelessly powered iontophoresis or (f) passive free diffusion based on WTCL. (g) Simultaneous IOP sensing and drug delivery using a single WTCL device. The rabbit's IOP were wirelessly monitored with the WTCL, and brimonidine delivery via wireless iontophoresis on the same WTCL was conducted. The green asterisk indicated the IOP measurements via Tonopen for calibration or accuracy comparison. (h) Rabbit's eye was treated with eye drops of brimonidine, and the IOP was measured via Tonopen. N=2 rabbits in (e)–(g) and N=1 rabbit in (h). In (e)–(h), at each time point, 8 Topopen measurements were conducted, and the purple, blue, gray and white regions indicated the periods of prior to delivery, during delivery, 0.5 h after delivery and 2 h after delivery, respectively. (i) The IOP reductions effects after 0.5 and 2 h of drug delivery via iontophoresis, free diffusion and eye drops were summarized. Data were presented as mean ± s.d. Significance was evaluated by one- way analysis of variance. \*P < 0.05. (j) Monitoring of the thermal effects generated during WPT operation via infrared thermal camera, and (k) the temperate on cornea, WTCL, and transmitter coils during WPT process were analyzed. N=4 measurements per group. Data were presented as mean ± s.d. Significance was evaluated by one- way analysis of variance. \*P < 0.05.  
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+- WTCLSupplementaryMaterials.docx
+
+<--- Page Split --->

preprint/preprint__2b8c98fb229297a1262aa712f9c884323c7fbe1c918252cec11c83d0130a8dc7/images_list.json

@@ -0,0 +1,62 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1. Simulated and observed the total power outage and recovery process in Louisiana for Hurricane. a) Ida and b) Laura. The red curve shows median values, with 5% to 95% quantile range shown by shade and the blue curve shows the observation. The yellow curve in (a) shows the percent of customers impacted by heatwaves (value reads the right axis). c) Comparison of observed and simulated spatial distribution of power outage for Hurricane Ida.",
+    "footnote": [],
+    "bbox": [
+      [
+        128,
+        66,
+        770,
+        420
+      ]
+    ],
+    "page_idx": 16
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure. 2. The return period of hurricane events by various metrics for Louisiana. a) interruption hours under power outage, b) interruption hours under blackout-heatwave compound hazard. c) summary statistics for the return period and impact for Ida-level events under different climates. The red curve shows median values for SSP5 8.5 for the future climate, with the 5% to 95% quantile range shown by shade, the yellow curve represents SSP2 4.5 for the future, and the blue for the historical climate. The dashed lines highlight Hurricane Ida’s power outage and compound hazard return levels.",
+    "footnote": [],
+    "bbox": [
+      [
+        155,
+        72,
+        720,
+        747
+      ]
+    ],
+    "page_idx": 17
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure. 3. Estimated average interruption days of compound hazard. a) historical average interruption, b) future SSP 5 8.5 average interruption, and c) future SSP2 4.5 average interruption for each county in Louisiana for a compound hazard event with a 278-year return period. d) Distribution of Orleans Parish customers' compound hazard duration under a 278-year return period event. The solid lines show the percentage of residents affected by the compound hazard up to a certain length in the historical and future climates.",
+    "footnote": [],
+    "bbox": [
+      [
+        140,
+        70,
+        777,
+        560
+      ]
+    ],
+    "page_idx": 18
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure. 4. Relative impacts of climate change factors on Ida’s compound hazard return period. a). Relative impact of each climate change factor assuming a consistent TC frequency. b). Sensitivity to TC frequency change. Note that the combined impact of all climate factors on Ida’s compound hazard return period is highly non-linear and thus the sum of the relative impact of individual factors does not equal the total impact. The solid line below the points shows \\(\\pm 1\\sigma\\) of uncertainty level.",
+    "footnote": [],
+    "bbox": [
+      [
+        135,
+        60,
+        480,
+        321
+      ]
+    ],
+    "page_idx": 19
+  }
+]

preprint/preprint__2b8c98fb229297a1262aa712f9c884323c7fbe1c918252cec11c83d0130a8dc7/preprint__2b8c98fb229297a1262aa712f9c884323c7fbe1c918252cec11c83d0130a8dc7.mmd

@@ -0,0 +1,360 @@
+
+# Hurricane Ida's blackout-heatwave compound hazard risk in a changing climate  
+
+Ning Lin  
+
+nlin@princeton.edu  
+
+Princeton University https://orcid.org/0000- 0002- 5571- 1606  
+
+Kairui Feng  
+
+Princeton University https://orcid.org/0000- 0001- 8978- 2480  
+
+Avantika Gori  
+
+Princeton University  
+
+Dazhi Xi  
+
+Princeton University https://orcid.org/0000- 0002- 4096- 8441  
+
+Min Ouyang  
+
+Huazhong University of Science and Technology https://orcid.org/0000- 0002- 3190- 4390  
+
+Michael Oppenheimer  
+
+Princeton University https://orcid.org/0000- 0002- 9708- 5914  
+
+Article  
+
+Keywords:  
+
+Posted Date: March 29th, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4096843/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on May 15th, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 59737- 8.
+
+<--- Page Split --->
+
+## Title  
+
+Hurricane Ida's blackout- heatwave compound hazard risk in a changing climate  
+
+## Authors  
+
+Kairui Feng \(^{1,2}\) , Ning Lin \(^{1,*}\) , Avantika Gori \(^{1}\) , Dazhi Xi \(^{1,3}\) , Min Ouyang \(^{4}\) , Michael Oppenheimer \(^{2,3,5}\)  
+
+## Affiliations  
+
+\(^{1}\) Department of Civil and Environmental Engineering, Princeton University, New Jersey, U.S.  
+
+\(^{2}\) High Meadows Environmental Institute, Princeton University, New Jersey, U.S.  
+
+\(^{3}\) Program in Atmospheric and Oceanic Sciences, Princeton University, New Jersey, U.S.  
+
+\(^{4}\) School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, China.  
+
+\(^{5}\) Department of Geosciences, Princeton University, New Jersey, U.S.  
+
+\(^{*}\) nlin@princeton.edu  
+
+## Abstract  
+
+The emerging tropical cyclone (TC)- blackout- heatwave compound hazard under climate change are not well understood. In this study, we employ future projections of TCs, sea levels, and heatwaves, in conjunction with power system resilience modeling, to evaluate historical and future TC- blackout- heatwave compound hazard risks in Louisiana, US. We find that the return period for a compound hazard event comparable to Hurricane Ida (2021), with approximately 35 million customer hours of simultaneous power outage and heatwave exposure in Louisiana, is around 278 years in the historical climate (1980- 2005). Under the emissions scenario SSP5 8.5 (SSP2 4.5), this return period may decrease by a factor of \(\sim 17 \times (10x)\) to 16.2 (28.4) years in the future climate (2070- 2100). The significant increase in risk can be primarily attributed to projected escalations in heatwaves, which result in an approximate 5(2)- fold decrease in compound hazard return period, and in TC activity, which cause an estimated 2(1)- fold decrease in the return period. The findings contribute to our knowledge of and adaptation to compound climate hazards.
+
+<--- Page Split --->
+
+## Introduction  
+
+In August 2021, Hurricane Ida, a Category 4 storm, struck Louisiana with intense winds, heavy rainfall, and significant storm surges, resulting in widespread flooding and damage to the state's infrastructure systems. Subsequent to the hurricane's landfall, the state experienced a loss of approximately 200 million customer hours of electricity, affecting roughly 2.15 million customers for an average power outage duration of 96 hours. Data from the U.S. Department of Energy reveals that Hurricane Ida caused the most extensive power outage in Louisiana's history, largely surpassing Hurricane Katrina (Category 5; 2005) and Hurricane Laura (Category 4; 2020), which led to losses of approximately 140 million and 100 million customer hours of electricity, respectively (1,2,3).  
+
+Furthermore, a prolonged heatwave occurred in the aftermath of Hurricane Ida, particularly affecting households that lost power and thus had no air conditioning (3). Consequently, Louisiana residents experienced a total of 35 million hours of compound blackout- heatwave hazard risk (with a heat index surpassing \(37.8^{\circ}\mathrm{C} / 100^{\circ}\mathrm{F}\) , in accordance with Louisiana's heat advisory criteria, 4). Customers exposed to the compound hazard endured an average of approximately 98 hours of heatwave conditions (4). Prolonged heat exposure can cause hospitalization and mortality risks (5), especially among vulnerable populations, for example, elders (6). Understanding how often Ida- like compound blackout- heatwave events may occur is critical for the development of coastal risk mitigation strategies.  
+
+A "compound climate event" can result in significant impacts due to the combination of climate drivers and hazards such as floods, wildfires, heatwaves, and droughts (7). Traditional risk assessment methods typically consider one hazard at a time, potentially leading to an underestimation of risk, since the physical drivers causing extreme events may exhibit spatial and/or temporal dependencies and interact to exacerbate the overall impact. Hurricanes or intense tropical cyclones (TCs), as drivers of extreme wind, rainfall, and storm surge, inherently lead to compound impacts on coastal regions (8) and are responsible for nine of the ten largest power outages in the United States over the past two decades (9). While extreme winds are the primary source of damage to power systems, the presence of storm surges and heavy rainfall resulted in extensive flood inundation during Hurricane Ida, which caused additional physical damage and hindered power system resilience as repair crews were unable to access affected areas (2). In addition, sea- level rise (SLR) may intensify coastal flood inundation by extending and prolonging the flood coverage, further exacerbating power system damage and delaying recovery operations.  
+
+Due to the seasonal peak of intense heat being ahead of that of major TCs, TC- heatwave compound events have so far been rare worldwide (10). However, a previous study (11) found that for Harris County, TX (a portion of Houston situated on higher ground and affected by hurricane winds), the TC- blackout- heatwave compound hazard risk would increase by a factor as large as 23 over the course of \(21^{\mathrm{st}}\) century under the high emissions scenario RCP8.5. In recent years, TC- heatwave compound hazards have happened in the Gulf Coast region. Hurricane Ida may represent the first hurricane landfall on the mainland United States associated with a long- lasting (i.e., multi- day), large- scale blackout- heatwave compound hazard (during Hurricane Laura the state- average heat index was also high but did not reach the threshold of \(100^{\circ}\mathrm{F}\) ). During Hurricane Ida and Laura, at least eleven and eight Louisianans, respectively, died of heat- related illnesses (1,12). Investigating this emerging compound threat, possibly enhanced by climate change, will contribute to our knowledge of and adaptation to compound climate hazards.
+
+<--- Page Split --->
+
+In this study, we integrate hazard projection and power system analysis to examine TC- blackout- heatwave compound hazard risks for Louisiana over the 21st century under the combined influence of SLR and changes in heatwave and storm climatology. We highlight the change in the return period/recurrence interval of Ida- like compound events from the historical to future climates. We further quantify the relative importance of the change in various climatological variables (i.e., heat stress, sea level, storm frequency, storm intensity) in driving the changes in the compound hazard risk.  
+
+Our framework is an extension of the previous study (11) to incorporate multiple hazards, including storm surge, rainfall and SLR, in addition to wind and heatwave, to more comprehensively model TC- blackout- heatwave compound hazard risk, for the entire State of Louisiana under various climate conditions. Specifically, we combine projections of heatwaves (13), TC hazards (including wind, storm surge, and rainfall, 14), and SLR (15,16,17), driven by CMIP6 GCMs (13) under both the high (Shared Socioeconomic Pathway 5 8.5; SSP5 8.5) and moderate (SSP2 4.5) emission scenarios. We generate a large number of compound hazard events based on the combined projections for historical (1980- 2005) and future (2081- 2100) climates to estimate how hazard probabilities may change over the 21st century (Methods). Then we utilize a physics- based power outage and restoration model for Louisiana to simulate wind/surge/rainfall- induced power system failure and recovery for each hazard event, to estimate the probabilities of TC- blackout- heatwave hazards. We extends the existing wind- impact- only simulation method (county level; 11,18) to a wind- rainfall- surge coupled framework for power damage and recovery process modeling to consider a larger study area including coastal regions (state level; Methods). Considering the uncertainty surrounding the impact of climate change on the frequency of TCs making landfall along the Gulf Coast, we assume a constant TC frequency but also assess the sensitivity of the compound hazard risk to TC frequency projection. To focus on the impact of climate change, we assume that the power system, population distribution, and recovery plans in the study region will remain unchanged. However, we assume that the coastal levees will be elevated following a design based on the return period of storm tides, as this enhancement may be considered necessary to prevent the region from frequent inundation due to SLR (Methods).  
+
+## Results  
+
+## Historical Cases  
+
+We first examine power outage simulations of the historical cases of Hurricanes Ida and Laura, which are the two major events over the last decade that caused widespread power disruptions in Louisiana (1). Ida devastated the eastern half of Louisiana, which is more densely populated (including the city of New Orleans), whereas Laura grazed the western side. Ida destroyed 31,000 poles (reported by local utility company Entergy, 19) that carry lower- voltage distribution lines in the neighborhoods, twice as many as those in Hurricane Laura (14,000 poles) and Katrina (2005; 17,000 poles). As shown in Fig. 1a and b, the model's estimates for the overall impact of Hurricanes Ida and Laura on Louisiana compare relatively well with the observation. Hurricane Ida led to 47% (48% in simulation) of customers being out of power within the first 24 hours and it took \(\sim 10\) days (11 days in simulation) for 90% of customers to restore power. Meanwhile, up to 60% of Louisiana residents were under heatwave conditions within 6 days after Hurricane Ida's landfall. On average, 42% of customers experienced compound power outage- heatwave hazards for at least a day after the hurricane's landfall, based on the overlap of county- level power outage and heat index data. Hurricane Laura led to 27% (29% in simulation) of customers being out of electricity initially and it took \(\sim 6\) days (8 days in simulation) for 90% of customers to restore
+
+<--- Page Split --->
+
+power. To measure the overall severity of the blackout associated with each TC, we compute the cumulative interruption hours of customers throughout Louisiana (the total of all customers' power outage duration), a commonly used metric in evaluating power system reliability (20). The model estimates in total 189 (156- 242; \(\pm 3\sigma\) ) million power interruption hours for Ida, which is consistent with the observed 206 million hours, and 110 (77- 153) million power interruption hours for Laura, which compares relatively well with the observed 99 million interruption hours. As a comparison, Hurricane Katrina led to \(\sim 140\) million power interruption hours. The model estimation for the spatial and temporal distribution of power outage also correlates well with observations (the average relative error is \(< 10\%\) between the modeled and observed county- level power outage), as illustrated in Fig. 1c for the peak power outages within 24 hours and 5 days after landfall power outages at the county level for Hurricane Ida.  
+
+## Blackout and Compound Hazards  
+
+Integrating power outage and recovery modeling with projections of future TC, SLR, and heatwaves, we examine the risk of TC- induced blackout- heatwave compound hazards in Louisiana. We generate 10,000 simulations of synthetic hazard events for each of the historical (1980- 2005) and future (2081- 2100) SSP5 8.5 and SSP2 4.5 scenarios. Each stochastic simulation includes a continuous 20- year sequence of TC occurrences, along with the physical simulation of TC tracks, wind speeds, rainfall amounts, storm surge levels, and heatwaves. We track each customer's exposure (i.e., duration) to blackout or compound blackout- heatwave hazard in the power outage and recovery modeling process for each synthetic hazard event. Then, we integrate the customer- level results to obtain state- level statistics and estimate the return periods (i.e., reciprocal of annual exceedance probability) of event total interruption hours for the historical and future climates. As demonstrated by the substantial shift of the return period curves (Fig. 2), the power outage risk will increase dramatically from the historical to the future climate. Specifically, the historical return period of a power outage of 206 million customer hours, as in Hurricane Ida, is 64 years. Under the SSP5 8.5 (SSP2 4.5) scenario, the return period of Ida's total power outage is estimated to be 35.8 (38.2) years in the future. The power outage with Ida's return period of 64 years is estimated to be about 413 million (265 million) customer hours in the future. The return period of a TC- blackout- heatwave compound hazard of 35 million customer hours, as in Hurricane Ida, is 278 years in the historical climate. The return period of Ida's compound hazard of 35 million customer hours is estimated to be 16.2 (28.4) years in the future climate ( \(\sim 17x\) (10x) decrease in return period). The compound hazard with Ida's return period of 278 years is estimated to be about 435 million (138 million) customer hours in the future climate, which corresponds to an average 8.8 (2.8)- day blackout- heatwave compound hazard experience for every of Louisiana's 2.13 million customers.  
+
+Compared to that in SSP2 4.5, the power outage level is similar to that in SSP5 8.5 at Ida's return period or lower, although the power outage risk becomes significantly higher in SSP5 8.5 at higher return periods (Fig. 2a), due to higher extreme TC hazards and SLR in SSP5 8.5. The difference between the two emissions scenarios is larger for the compound hazard risks. Specifically, compared to that in SSP2 4.5, the compound hazard risk is slightly higher in SSP5 8.5 at Ida's return period or lower, and the compound hazard risk becomes dramatically higher in SSP5 8.5 at higher return periods (Fig. 2a), due to combined effects of larger increases in extreme heatwaves, TC hazards, SLR in SSP5 8.5. Specifically, the frequency of extremely severe compound events, such as those with triple the impact of Hurricane Ida (i.e., 100 million customer- hours of compound hazards), is expected to be 2.5 times higher under the high emission scenario (SSP5 8.5) compared to the moderate emission scenario (SSP2 4.5). However, for less severe events, such as those with a third of Ida's impact (i.e., 10 million customer- hours of
+
+<--- Page Split --->
+
+compound hazards), we do not find a statistically significant difference in event frequency between the two emission scenarios. These findings suggest that the combined effects of global warming and increasing hurricane intensity significantly amplify the risk of the most extreme compound events. Nonetheless, the moderate emission scenario may still lead to a similar level of compound hazard risk for events of Ida's magnitude as the high emission scenario.  
+
+## Spatial Pattern of Compound Hazard Risk  
+
+To investigate the spatial distribution of the compound hazard risk, we estimate the county- average compound hazard interruption days for each synthetic hazard event. Fig. 3 shows the compound hazard interruption days with Ida's return period of 278 years for each county in Louisiana in the historical and future climates. The coastal counties face a greater compound hazard risk than inland counties for both historical and future climates (Figs. 3a- 3c). For example, the counties with an average compound hazard impact larger than 20 days in the future climates are mostly coastal counties. Coastal counties often face a greater compound hazard risk since hurricane winds reach peak strength before the storm makes landfall, and storms tend to weaken significantly as they move inland, causing less damage to the inland power infrastructure. Moreover, the floods induced by storm surge and/or heavy rainfall can severely damage coastal power sectors. The flooding also hampers the recovery efforts of local contractors by submerging electrical components in water and obstructing local traffic and logistics with debris.  
+
+The general spatial disparities in compound hazard risks are also significant and will increase with climate change. For example, in the historical climate, the most impacted county may face on average a 1.8- day compound hazard with Ida's return period (278 years), and the least impacted county does not face any compound hazard risk (Fig. 3a). In the future under the SSP5 8.5 (SSP2 4.5) scenario, the county with the greatest impact may face the compound hazard of on average 12.7 (3.1) days with Ida's return period (278 years), and the least impacted area will face the compound hazard of on average 1.1 (0.1) days (Figs. 3b- 3c). To quantify the spatial disparity in compound hazard risks, we employ the Gini coefficient, which is a measure of statistical dispersion often used to represent income inequality, wealth inequality, or consumption inequality within a nation or a social group (21). It ranges from 0 to 1, where 0 represents perfect equality (i.e., every county has the same average compound hazard duration) and 1 represents full inequality (i.e., one county faces the compound hazard while others do not). In the historical climate, the Gini coefficient is around 0.312; however, it becomes 0.632 (0.411) in the future climate under SSP5 8.5 (SSP2 4.5). The results indicate that warming is likely to exacerbate the existing disparities and inequalities of TC- blackout- heatwave compound hazard risk in Louisiana.  
+
+We also investigate the distribution of compound hazard duration for residents within densely populated counties. In Fig. 3d, we show the distribution of compound hazard interruption days for affected customers in Orleans Parish for Ida's return period of 278 years. In the historical climate, only 5% of affected customers may face a \(>120\) - hour (5- day) compound hazard. However, in the future climate under the SSP5 8.5 (SSP2 4.5) scenario, over 70% (50%) of affected customers will confront a \(>120\) - hour (5- day) compound hazard, over 25% (3%) of affected customers will confront a \(>240\) - hour (10- day) compound hazard, and 9% (0%) of affected customers will encounter the compound hazard over 360 hours (15 days). Hence, climate change not only increases the average compound hazard impact but also intensifies the tail risk that vulnerable residents may encounter, especially under the high emission scenario.
+
+<--- Page Split --->
+
+## Drivers of Changes in Compound Hazard Risk  
+
+The change in the compound hazard risk is driven by the change in three climate factors: 1) heatwaves (heat index) 2) TC climatology and 3) the sea level. As we assume that TC frequency remains unchanged in the future, the changes in TC climatology include changes in TC characteristics, particularly intensity (which drives changes in wind, storm tide, and rainfall). To determine the relative effect of the changes in these factors, we estimate the changes in the compound hazard risk due to changes in temperature, SLR, and TC intensity, respectively, by adjusting each variable to its future value or distribution and calculating the resulting return period of Ida's compound hazard (i.e., 35 million customer hours of simultaneous power outage and heatwave impact), as shown in Fig. 4a. As discussed above, if all climate change factors are considered, Ida's return period would drop from 278 years to 16.2 (28.4) years from historical climate to future climate under SSP5 8.5 (SSP2 4.5). The change in heatwaves is the largest contributor to the change in Ida's return period from the historical climate to the future climate; due to solely the heatwave change, Ida's return period would drop from 278 years in the historical climate to 47.6 (84.3) years in the future climate, which is a \(\sim 5x\) (3x) return period decrease. This large impact is induced by the dramatic change in temperature and humidity: the annual number of heatwave days \((>37.8^{\circ}C)\) will increase by 8 times from the historical climate to the future climate (ensemble average of the six GCMs) for the study region. The contribution TC intensity change reduces Ida's return period to 167.8 (185.2) years in the future climate. The impact of SLR on Ida's return period is relatively small, reducing Ida's return period to 251.2 (263.3) years in the future climate. SLR appears to have a relatively low impact because we assume the levees along the coast will be elevated. Also, the impact of SLR is limited to coastal regions and it is averaged out when the compound hazard impact is calculated for the entire state. The contribution of the various climatological drivers to future compound hazard risk is consistent across the two different emission scenarios.  
+
+Considering that the projection of TC frequency is subject to significant uncertainty (22,23), we assumed a constant TC frequency in above analyses. Here we investigate the sensitivity of the estimated compound hazard risk to the projection of TC frequency change. In one scenario, we apply the TC frequency in ref. 14, which projects relatively high increases in TC frequency in the future climates under SSP5 8.5 (SSP2 4.5), and Ida's return period would drop to 7.9 (15.2) years, compared to 16.2 (28.4) years when accounting for all climate change factors except TC frequency change. In another scenario, we consider a \(30\%\) decrease in TC frequency, the lower bound of TC frequency projections ensembled in ref. 23. Ida's return period would become 23.1 (40.6) years, which is a similarly dramatic decrease from 278 years in the historical climate, compared to the case when TC frequency was hold constant. This sensitivity test indicates the relatively small impact of TC frequency change compared to the combined effects of other climate change factors.  
+
+## Discussion  
+
+This analysis highlights the substantial increase in the frequency of Ida- level extreme power outage- heatwave compound hazards over time, resulting from the combined effect of temperature increase, sea- level rise, and storm climatology changes under the SSP5 8.5 (SSP 2 4.5) climate change scenario. Linear interpolation reveals that the return period of Hurricane Ida has decreased from 278 years around 2000 to 225.6 (228.0) years in the 2020s, indicating a \(19\%\) reduction in the return period over the past 20 years. This real- life observation of an emerging climate compound
+
+<--- Page Split --->
+
+hazard motivates further research on projecting future compound climate hazard risks and developing strategies to mitigate climate risks for various regions around the world.  
+
+When examining the impact of various climate scenarios, such as the high emission scenario SSP5 8.5 and the moderate emission scenario SSP2 4.5, it appears that the risk associated with Ida- or smaller- scale compound hazard events may not exhibit substantial difference. This result indicates that utility companies urgently need to prepare for the compound events to prevent major impacts. On the other hand, for less frequent events, the impact of these compound hazards is expected to be significantly lower under the moderate emissions scenario. Moreover, the duration of interruptions caused by compound hazards will also be reduced with moderate emissions. This result highlights the importance of strengthening climate change mitigation policy to reduce the impact of extreme climate hazards.  
+
+We do not consider the potential change in the power grid or its operation in the future. In the future, localized solutions, including backup generators and solar panels, can provide temporary support to residents who lose power from the main grid, thus mitigating the impacts of compound hazards (24). These solutions can help reduce the exposure of vulnerable populations to the effects of power outages and extreme heat, thereby lessening the overall impact of compound hazard events. However, backup generators and solar panels may be cost- prohibitive for many middle and low- income communities, limiting their effectiveness in reducing heat stress. From the main power grid design perspective, adopting effective strategies like burying distribution networks (5) and developing distributed power systems (24) can bolster the resilience of power infrastructure against extreme weather events.  
+
+We also do not consider the potential changes in demographics and human habitat. As extreme climate events become more frequent, coastal megacities are also expected to develop rapidly (25,26). Encouraging climate- resilient urban design principles that prioritize green spaces, water management systems, and heat- resistant building materials can enhance cities' resilience against compound hazard risks (27). Furthermore, the implementation of advanced early warning systems and preparedness measures, combined with public awareness campaigns, can help minimize potential impacts on vulnerable communities (28,29). Moreover, changes in population patterns, such as urbanization in low- elevation coastal zones and the concentration of populations in areas vulnerable to climate hazards, can also influence the severity and duration of compound hazards, emphasizing the need to account for these demographic shifts when devising adaptation strategies (25,30).  
+
+Quantifying the reliability and resilience of infrastructure systems under the impact of future compound hazards is essential for climate change adaptation. Developing an integrated risk assessment framework that combines climatology, civil and electric engineering, urban planning, and social sciences is crucial for comprehensively understanding the interconnected nature of compound hazard risks and their societal impacts, and the formulation of effective mitigation strategies (31,32). For example, conventional statistical methods may fail to detect significant changes in compound hazard risk, especially for the most extremes. Our analysis shows that the intensity of relatively frequent hazards may not change significantly in the future, especially under moderate emissions scenarios. However, if such a conclusion for frequent, observable events is statistically extrapolated to that for extreme events, we may transform events like "Ida" or extreme, which could have been foreseen and prepared for, into "black swan" events—unpredictable extreme disasters with unimaginable losses. Only physics- based modeling integrating climate and hazards projection and infrastructure/social system analysis may provide reliable estimates of future risks. This multidisciplinary perspective is essential for capturing the
+
+<--- Page Split --->
+
+complex interactions between different hazards and their cascading effects on infrastructure systems and society as a whole, ultimately enabling the development of robust and resilient strategies to mitigate the impacts of compound hazards. Also, given various uncertainties in climate projection to social development, there is a need for continuous refinement and updating of risk analysis techniques as improved modeling approaches and new data become available. By adopting a comprehensive approach that integrates various disciplines and continuously enhances our understanding of compound hazard risks, we can work towards developing effective adaptation strategies for a sustainable future in the face of a changing climate.  
+
+## Materials and Methods  
+
+## Synthetic TCs, Storm Surge, Tide and Rainfall Modeling  
+
+We use the synthetic TC hazard dataset generated in ref. 14 for the North Atlantic basin and select the TC tracks passing within \(200\mathrm{km}\) of Louisiana. The dataset contains synthetic TC tracks generated with the statistical- deterministic TC model (33), which has been applied to TC hazard assessment (8,34,35). The synthetic TC tracks for the historical period (between 1980 and 2005) were generated based on the National Centers for Environmental Prediction (NCEP) reanalysis. The dataset also contains bias- corrected and weighted- average climate projections of TCs for the future period (2070 to 2100) under combined Shared Socioeconomic Pathway (SSP) emissions scenarios, SSP5 8.5 and SSP2 4.5, based on six CMIP6 climate models: CanESM5, CNRM- CM6- 1, UKESM1- 0- LL, EC- Earth3, IP- SL- CM6A- LR and MIROC6. The TC storm tides were modeled in ref. 13 using the Advanced Circulation (ADCIRC) hydrodynamic model (36,37). We extract peak storm tides at nodes ( \(\sim 1\mathrm{km}\) resolution) along the coastline of Louisiana for each TC and match these to the county level. The rain fields were simulated in ref. 14 for each synthetic TC using the physics- based Tropical Cyclone Rainfall (TCR) model (38). We apply area- averaged TCR estimates at the county level, and we employ the maximum 24- hour rainfall accumulation from each storm event, since the 24- hour storm duration is often utilized for rainfall risk assessment (38).  
+
+To perform sequential risk analysis, we generate 10,000 stochastic samples of storm sequences for both historical and future climate periods. These samples were derived by selecting storms (according to a Poisson process with a rate as the TC annual frequency) and their associated hazards from the TC hazard dataset (14) described above. Each stochastic sample consists of 20 consecutive years of TC activity. For the primary analysis in this study, we maintain a constant TC frequency in the future climate. For the sensitivity analysis, we consider the increased TC frequency projected by the statistical- deterministic TC model in ref. 14 and the decreased TC frequency by up to \(30\%\) projected by a range climate models ensembled in ref. 23.  
+
+For each sampled storm, we generate the spatial- temporal wind field, employing the classical Holland wind profile (39) and accounting for the effects of surface friction and large- scale background wind following ref. 40, and converting one- minute mean winds to 3- second wind gusts using gust factors (41). We estimate the coastal flood area by comparing the height of peak storm tide (if over levee height) to the ground surface elevation specified by the USGS thirty- meter DEM (42; Fig. S2), assuming that areas would be inundated when the storm tide exceeds the ground elevation.  
+
+In Louisiana, the actual seawall heights vary significantly along the coastline, ranging from 2 to 5 meters and often changing over short distances. Due to the difficulty in acquiring the precise data,
+
+<--- Page Split --->
+
+our simulations do not incorporate partially available measurements (43 to 45). We assigned the current seawall level based on estimated 100- year flood level (estimated in ref. 14) for each coastal county in Louisiana, according to the typical design guidance. For example, the 100- year flood level for New Orleans is approximately 3.4 meters above the North American Vertical Datum of 1988 (NAVD 88) (13). This approximation introduces a degree of inaccuracy into our flood modeling. Acknowledging this limitation, we subsequently focused on binary flood data—whether a flood occurs or not— when developing our power system damage and recovery models. We observe that the occurrence of flooding is a critical factor that significantly hinders the restoration efforts of the power system in coastal counties. However, the inundation depth of the flooding appears to have a less substantial impact, as indicated in the sensitivity analysis in Fig. S1. When compared to the areas affected by the TC’s wind and rainfall, the flooded regions are generally smaller. Therefore, the majority of the structural damage to the power system may not be caused by flooding.  
+
+The future coastal levee plan is uncertain. In the future climate simulations, we assume the coastal levee will be elevated to the historical 100- year return level plus one percent quantile SLR ( \(\sim 2\) meters averaging over SSP2 4.5 and SSP5 8.5 for New Orleans calculated from ref. 14). This design strategy is commonly used by governmental agencies to plan the seawall height, and it is within the framework proposed by the U.S. Army Corps of Engineers for the New Orleans Region, Lafayette, and Lake Charles (Error! Reference source not found.). A sensitivity test was performed on future compound hazard risks given different elevations of the coastal levee from 0- 3 meters above the current level. If the levee was not elevated, the surge impact on the compound hazard risk would be significantly higher. On the other hand, when the levee is elevated by higher than 2 meters, the estimated compound risk is not sensitive to the variation of the assumed levee height (Supplementary, Fig. S1). The generated wind, rainfall, and coastal flood conditions from each sampled storm drive the power grid outage and recovery analysis.  
+
+## Heatwave Projections  
+
+Following ref. 5, the daily HI is determined as a function of daily maximum near- surface (2 m) air temperature, daily mean specific humidity, and daily mean surface pressure. To maintain consistency with the TC simulation, we obtain these data for Louisiana from the NCEP reanalysis and the six GCMs stated above during and after landfall for each sampled synthetic storm (each synthetic storm is associated with a climatological time of occurrence and development). The future HI projected by the GCM is bias- corrected (11) by adding the difference between the NCEP reanalysis and the GCM- estimated historical HI. According to the historical analysis in ref. 4, the HI will drop upon TC landfall and will recover to the ambient average within around ten days. To account for this dependence between TCs and heatwaves, we add the composite of the impact of TC passage to the meteorological variables used to calculate the HI, where the composite impact is estimated based on historical data (Fig. 3a in ref. 4).  
+
+## Sea Level Rise Projections  
+
+We employ sea- level projections produced by the Intergovernmental Panel on Climate Change Sixth Assessment Report (AR6; 15,17) using the Framework for Assessing Changes To Sea- level (FACTS; 16). Localized probabilistic SLR projections under the SSP5 8.5 and SSP2 4.5 emission scenarios with ‘medium confidence’ are incorporated in this analysis (there are two confidence levels in the datasets, which are low and medium levels). The local sea level projection takes into account ground uplift or subsidence, oceanographic effects, and spatially variable responses of the geoid and the lithosphere to shrinking land ice. The projection of SLR was developed for tide
+
+<--- Page Split --->
+
+gauge stations. For each TC sequence realization, we first sample a SLR time series (a realization) from the projection. Then, we linearly interpolate the SLR projection to the locations of coastal counties, and we estimate the storm surge relative to the current sea level by adding the SLR to the storm tide level at each time point for each county.  
+
+## Power Outage and Recovery Model  
+
+We apply a physics- based power system model, which explicitly simulates component level damage to predict the total power outage, accounting for the effects of future evolving factors, e.g., climate change, infrastructure upgrade, and utility maintenance. The physics- driven modeling of the power system allows us to better understand the impact of climate change and effectiveness of risk mitigation measures compared to if we used purely data- driven models (47,48).  
+
+Specifically, we extend the power grid outage and recovery model developed by refs. 11 and 18 to simulate TC impact on the electric power system in Louisiana (power topology shown in Fig. S3- 5). The power grid failure model first applies probabilistic fragility functions to estimate the damage states of five main vulnerable component types of the power network: transmission substations, transmission lines, distribution nodes, distribution lines, and local distribution circuits. Component failures alter the power grid topology and may separate the power grid into disconnected sub- grids. A DC flow simulation is then performed to capture the power availability in each sub- grid (similar to approaches in 49,50,51,52). The power system is open and connects with systems outside the study area via transmission lines; the performance of the power grid outside the study area is assumed to be under normal operation.  
+
+The fragility curves in refs. Error! Reference source not found. and 18 only considered the wind damage. Here we extend the fragility functions to consider the effects of flood and rainfall. For example, the probability of failure of a substation given specific wind, rainfall, and surge flood levels is estimated based on a log- normal fragility function as:  
+
+\[P(D\geq d_t|H = h = w + \alpha \cdot r + \beta \cdot f) = \int_0^h\frac{1}{\sqrt{2\pi}\sigma_t x}\exp \left(-\frac{(\ln x - \mu_i)^2}{2\sigma_t^2}\right)dx \quad (1)\]  
+
+where hazard (H/h) is considered as a linear combination of wind speed (w), rainfall amount (r) and flood condition (f, flooded or not; Boolean variable) with two parameters \(\alpha\) and \(\beta\) . With the shape \((\sigma_i)\) and location \((\mu_i)\) parameters, the log- normal distribution describes the probability of potential damage (D) in each of four states (di), i.e., i = {low, moderate, severe, complete} damage. Fragility function refers to the latent distribution of a component's ability to withstand outer forces (hazard). Some components may not withstand any force at all, while others can withstand very large outer force. Given a certain outer force, the probability of damage to the component is equal to the integral of fragility from 0 to that force level, i.e., the probability that the strength of the component is lower than outer force. The fragility functions for other components (support structures, distribution nodes, poles, conductors, and circuits) are similarly modeled with exponential, logistic, or uniform distributions. These fragility functions are similar to those in refs. Error! Reference source not found. and 18 except that the effects of rainfall and flood are incorporated. The parameters are estimated by the Markov chain Monte Carlo (MCMC) method to minimize the mean squared error between simulated and observed county- level power outages.
+
+<--- Page Split --->
+
+The recovery model, developed based on emergency response plans and operational data, applies estimated recovery resources based on a priority- oriented strategy to repair damaged transmission substations, transmission lines, and critical facilities vital to public safety, health, and welfare before local distribution networks (11,18). Debris should be removed before utilities become able to reinstate the power system. This debris- cleaning time is sampled from a uniform distribution between 48 to 72 hours (estimated from utility reports, 2). We also account for that, within the debris- cleaning period and without structural failure of the distribution system, residents may turn on the main power switch themselves 24 to 48 hours after being flooded (53).  
+
+There is a chance that TCs will make landfill in sequence and the second TC come before the damage caused by the first TC is fully recovered (54). We account for this temporal compounding effect in our power system outage and recovery analysis. For each sampled sequential hazard time series, the initial state of the power system when a TC arrives is set based on the condition of the restoration state from the previous TC. If the power system is indeed not fully recovered from the previous TC impact, the emergency response plans following the second TC are also adjusted considering the recovery process for the first TC. Specifically, the response plans will re- evaluate and prioritize the restoration tasks and redirect the repair efforts based on this updated priority list.  
+
+The power grid outage and recovery models were calibrated (to determine the model parameters) for the study area using observed power outage data for Hurricane Ida and Laura using simulated wind and observed rainfall (55) and flood (56). The same wind field modeling method applied to the synthetic storms is used for these two historical storms with storm characteristics (i.e., track, intensity, and size) taken from the extended best track data (57).  
+
+## References  
+
+1. Hurricane Response Record, Office of Cybersecurity, Energy Security, and Emergency Response, U.S. department of energy (2021).  
+2. Hurricane Ida Power Restoration, Utility Report, Entergy (2021)  
+3. Nicholas Bogel-Burroughs and Katy Reckdahl. The Greatest Killer in New Orleans Wasn’t the Hurricane. It Was the Heat.. New York Times (2021).  
+4. National Weather Service, New Orleans, New Orleans/Baton Rouge (2021)  
+5. K.L. Ebi, A. Capon, P. Berry, C. Broderick, R. de Dear, G. Havenith, Y. Honda, R.S. Kovats, W. Ma, A. Malik, and N.B. Morris. "Hot Weather and Heat Extremes: Health Risks." The Lancet, 398(10301), pp. 698-708 (2021).  
+6. D.H.K. Lee. "Seventy-five Years of Searching for a Heat Index." Environ. Res. 22(2), 331-356 (1980).  
+7. J. Zscheischler, S. Westra, B.J. Van Den Hurk, S.I. Seneviratne, P.J. Ward, A. Pitman, A. AghaKouchak, D.N. Bresch, M. Leonard, T. Wahl and X. Zhang "Future climate risk from compound events." Nat. Clim. Chang. 8, no. 6: 469-477(2018).  
+8. Gori, A., Lin, N., Xi, D. et al. Tropical cyclone climatology change greatly exacerbates US extreme rainfall–surge hazard. Nat. Clim. Chang. (2022).  
+9. T. Houser and P. Marsters. The world’s second largest blackout. Rhodium Group, 385(2018).
+
+<--- Page Split --->
+
+10. T. Matthews, R.L.Wilby and C. Murphy "An emerging tropical cyclone–deadly heat compound hazard". Nat. Clim. Chang., 9(8), 602-606(2019).  
+
+11. K. Feng, M. Ouyang, and N. Lin."Tropical cyclone-blackout-heatwave compound hazard resilience in a changing climate." Nature communications 13, no. 1: 4421(2022).  
+
+12. Louisiana Department of Health verifies one additional hurricane-related death, bringing toll to 27. Louisiana Dep. Heal. (2020). Published online Sept 9. Available at: https://ldh.la.gov/index.cfm/newsroom/detail/5761 (Accessed Oct 21, 2020).  
+
+13. B.C. O'Neill, C. Tebaldi, D.P. Van Vuuren, V. Eyring, P. Friedlingstein, G. Hurtt, R. Knutti, E. Kriegler, J.F. Lamarque, J. Lowe, and G.A. Meehl."The scenario model intercomparison project (ScenarioMIP) for CMIP6." Geosci. Model Dev. 9, 3461-3482 (2016).  
+
+14. D. Xi, N. Lin, and A. Gori. "Increasing Sequential Tropical Cyclone Hazards along the US East and Gulf Coasts." Nat. Clim. Chang. 13(3), 258-265 (2023).  
+
+15. Fox-Kemper, B., et al. Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA, pp. 1211-1362, doi:10.1017/9781009157896.011.  
+
+16. Kopp, R. E., Garner, G. G., Hermans, T. H. J., Jha, S., Kumar, P., Reedy, A., Slangen, A. B. A., Turilli, M., Edwards, T. L., Gregory, J. M., Koubbe, G., Levermann, A., Merzky, A., Nowicki, S., Palmer, M. D., & Smith, C. (2023). The Framework for Assessing Changes To Sea-Level (FACTS) v1.0: A platform for characterizing parametric and structural uncertainty in future global, relative, and extreme sea-level change. Geoscientific Model Development, 16, 7461-7489.  
+
+17. Garner, G. G., T. Hermans, R. E. Kopp, A. B. A. Slangen, T. L. Edwards, A. Levermann, S. Nowicki, M. D. Palmer, C. Smith, B. Fox-Kemper, H. T. Hewitt, C. Xiao, G. Aðalgeirsdóttir, S. S. Drijfhout, T. L. Edwards, N. R. Golledge, M. Hemer, G. Krinner, A. Mix, D. Notz, S. Nowicki, I. S. Nurhati, L. Ruiz, J-B. Sallée, Y. Yu, L. Hua, T. Palmer, B. Pearson, 2021. IPCC AR6 Sea Level Projections. Version 20210809.  
+
+18. M. Ouyang, L. Duenas-Osorio, Multi-dimensional hurricane resilience assessment of electric power systems. Struct. Saf. 48, 15-24 (2014).  
+
+19. Entergy, Restoration Times for Louisiana, including Hardest Hit Areas (2021).  
+
+20. U.S. Energy Information Administration, Annual Electric Power Industry Report, Form EIA-861 detailed data files(2021)  
+
+21. J.L. Gastwirth (1972). "The estimation of the Lorenz curve and Gini index." The Review of Economics and Statistics, 306-316.  
+
+22. A.H. Sobel, A.A. Wing, S.J. Camargo, C.M. Patricola, G.A. Vecchi, C.Y. Lee, and M.K. Tippett. "Tropical cyclone frequency." Earth's Future 9, no. 12: e2021EF002275(2021).  
+
+23. T. Knutson, S.J. Camargo, J.C. Chan, K. Emanuel, C.H. Ho, J. Kossin, M. Mohapatra, M. Satoh, M. Sugi, K. Walsh, and L. Wu. "Tropical cyclones and climate change assessment: Part II: Projected response to anthropogenic warming." Bull. Amer. Meteor. Soc. 101, no. 3: E303-E322(2020).  
+
+24. J.A. Bennett, C.N. Trevisan, J.F. DeCarolis, C. Ortiz-García, M. Pérez-Lugo, B.T. Etienne, and A.F. Clarens. "Extending energy system modelling to include extreme
+
+<--- Page Split --->
+
+weather risks and application to hurricane events in Puerto Rico." Nat. Energy 6, no. 3: 240- 249 (2021).  
+
+25. M.E. Hauer. "Migration induced by sea-level rise could reshape the US population landscape." Nat. Clim. Chang. 7, 321 (2017).  
+
+26. K. Feng, N. Lin, S. Xian, and M.V. Chester. "Can we evacuate from hurricanes with electric vehicles?". Transp. Res. Part D: Transp. Environ. 86, p.102458 (2020).  
+
+27. S. E. Gill, J. F. Handley, A. R. Ennos, and S. Pauleit. "Adapting Cities for Climate Change: The Role of the Green Infrastructure." Built Environ. 33(1), 115-133 (2007).  
+
+28. C. Wamsler, E. Brink, and O. Rantala. "Climate Change, Adaptation, and Formal Education: The Role of Schooling for Increasing Societies' Adaptive Capacities in El Salvador and Brazil." Ecol. Soc. 17(2), 2 (2012).  
+
+29. J. Birkmann and K. von Teichman. "Integrating Disaster Risk Reduction and Climate Change Adaptation: Key Challenges—Scales, Knowledge, and Norms." Sustain. Sci. 5(2), 171-184 (2010).  
+
+30. K.C. Seto, B. Güneralp, and L.R. Hutyra. "Global Forecasts of Urban Expansion to 2030 and Direct Impacts on Biodiversity and Carbon Pools." Proc. Natl. Acad. Sci. U.S.A. 109(40), 16083-16088 (2012).  
+
+31. DT Cohen, 60 million live in the path of hurricanes. U.S. census bureau. (2018)  
+
+32. Xu, Luo, Kairui Feng, Ning Lin, A. T. D. Perera, H. Vincent Poor, Le Xie, Chuanyi Ji, X. Andy Sun, Qinglai Guo, and Mark O'Malley. "Resilience of renewable power systems under climate risks." Nature Reviews Electrical Engineering 1, no. 1 (2024): 53-66.  
+
+33. K. Emanuel, S. Ravela, E. Vivant, and C. Risi. "A Statistical Deterministic Approach to Hurricane Risk Assessment." Bull. Am. Meteorol. Soc. 87, S1-S5 (2006).  
+
+34. R. Marsooli, N. Lin, K. Emanuel, and K. Feng. "Climate change exacerbates hurricane flood hazards along US Atlantic and Gulf Coasts in spatially varying patterns." Nat. Commun. 10, 1-9 (2019).  
+
+35. N. Lin, K. Emanuel, M. Oppenheimer, and E. Vanmarcke. "Physically based assessment of hurricane surge threat under climate change." Nat. Clim. Chang. 2, 462-467 (2012).  
+
+36. R.A. Luettich, J.J. Westerink, and N.W. Scheffner. "ADCIRC: an advanced three-dimensional circulation model for shelves, coasts, and estuaries. Report 1, Theory and methodology of ADCIRC-2DD1 and ADCIRC-3DL." (1992).  
+
+37. J.J. Westerink, R.A. Luettich, C.A. Blain, and N.W. Scheffner. "ADCIRC: An advanced three-dimensional circulation model for shelves, coasts, and estuaries. Report 2: User's Manual for ADCIRC-2DD1." (1992).  
+
+38. D. Xi, N. Lin, and J. Smith. "Evaluation of a physics-based tropical cyclone rainfall model for risk assessment." J. Hydrometeorol. 21, no. 9: 2197-2218 (2020).  
+
+39. K. Emanuel. "Assessing the present and future probability of Hurricane Harvey's rainfall." Proc. Natl. Acad. Sci. 114(48): 12681-12684 (2017).  
+
+40. G.J. Holland. "An analytic model of the wind and pressure profiles in hurricanes." Mon. Weather Rev. 108(8): 1212-1218 (1980).
+
+<--- Page Split --->
+
+41. N. Lin and D. Chavas. "On hurricane parametric wind and applications in storm surge modeling." J. Geophys. Res. Atmos. 117(D9) (2012).  
+
+42. USGS (U.S. Geological Survey). Digital Elevation Model(2022).  
+
+43. C. Fagnant, A. Gori, K.B. Ensor, A. Sebastian, and P.B. Bedient. "Characterizing Spatiotemporal Trends in Extreme Precipitation Across the Southern Texas Coast." Nat. Hazards (2020).  
+
+44. M. Palaiseanu-Lovejoy, C.A. Thatcher, and J.A. Barras. "Levee crest elevation profiles derived from airborne lidar-based high resolution digital elevation models in south Louisiana." ISPRS J. Photogramm. Remote Sens. 91: 114-126 (2014).  
+
+45. D.R. Johnson, J.R. Fischbach, and D.S. Ortiz. "Estimating surge-based flood risk with the coastal Louisiana risk assessment model." J. Coast. Res. (67 (10067)): 109-126 (2013).  
+
+46. US Army Corps of Engineers, Hurricane & Storm Damage Risk Reduction System, New Orleans District(2021).  
+
+47. R. Nateghi, S.D. Guikema, and S.M. Quiring. "Comparison and validation of statistical methods for predicting power outage durations in the event of hurricanes." Risk Anal. 31(12): 1897-1906 (2011).  
+
+48. C. Ji, Y. Wei, H. Mei, J. Calzada, M. Carey, S. Church, T. Hayes, B. Nugent, G. Stella, M. Wallace, and J. White. "Large-scale Data Analysis of Power Grid Resilience Across Multiple US Service Regions." Nat. Energy, 1(5), pp.1-8 (2016).  
+
+49. Y. Yang, T. Nishikawa, and A.E. Motter. "Vulnerability and Cosusceptibility Determine the Size of Network Cascades." Phys. Rev. Lett., 118(4), 048301 (2017).  
+
+50. Y. Yang, T. Nishikawa, and A.E. Motter. "Small Vulnerable Sets Determine Large Network Cascades in Power Grids." Science, 358(2017).  
+
+51. C. Zhai, T. Y. J. Chen, A. G. White, and S. D. Guikema. "Power Outage Prediction for Natural Hazards Using Synthetic Power Distribution Systems." Reliab. Eng. Syst. Saf., 208, 107348 (2021).  
+
+52. P. Hines, E. Cotilla-Sanchez, and S. Blumsack. "Do Topological Models Provide Good Information about Electricity Infrastructure Vulnerability?" Chaos 20, 033122 (2010).  
+
+53. Seven Steps to Restore Power After a Flood. New Equipment Digest (2017).  
+
+54. D. Xi and N. Lin. "Sequential Landfall of Tropical Cyclones in the United States: From Historical Records to Climate Projections." Geophys. Res. Lett. (2021).  
+
+55. J. Du. "NCEP/EMC 4KM Gridded Data (GRIB) Stage IV Data. Version 1.0." UCAR/NCAR - Earth Observing Laboratory, 2011. https://doi.org/10.5065/D6PG1QDD.  
+
+56. Dartmouth Flood Observatory.(2021).  
+
+57. J. Demuth, M. DeMaria, and J.A. Knaff. "Improvement of Advanced Microwave Sounder Unit Tropical Cyclone Intensity and Size Estimation Algorithms." J. Appl. Meteor. 45, 1573-1581 (2006).
+
+<--- Page Split --->
+
+## Acknowledgments:  
+
+We thank the projection authors for developing and making the sea- level rise projections available, multiple funding agencies for supporting the development of the projections, and the NASA Sea Level Change Team for developing and hosting the IPCC AR6 Sea Level Projection Tool. K.F., N.L., A.G. and D.X. are supported by the U.S. National Science Foundation (1652448 and 2103754 as part of the Megalopolitan Coastal Transformation Hub) and C3. ai Digital Transformation Institute (C3. ai DTI Research Award). K.F. is supported by the HMEI- STEP Graduate Fellowship. M. Openheimer is supported by U.S. NSF grant 2103754. M. Ouyang is supported by the National Natural Science Foundation of China, 72074089, 51938004, 71821001.  
+
+## Author contributions:  
+
+Conceptualization: KF, NL, MOY, MOP  Methodology: KF, NL, AG, DX, MOY  Writing—original draft: KF, NL, MOP  Writing—review & editing: AG, DX, MOY  
+
+## Competing interests:  
+
+The authors declare no competing interests.  
+
+## Data and materials availability:  
+
+The data generated from this study will be deposited to the NSF DesignSafe- CI online; a link to the data set will be provided upon publication.  
+
+## Figures and Tables
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1. Simulated and observed the total power outage and recovery process in Louisiana for Hurricane. a) Ida and b) Laura. The red curve shows median values, with 5% to 95% quantile range shown by shade and the blue curve shows the observation. The yellow curve in (a) shows the percent of customers impacted by heatwaves (value reads the right axis). c) Comparison of observed and simulated spatial distribution of power outage for Hurricane Ida. </center>
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure. 2. The return period of hurricane events by various metrics for Louisiana. a) interruption hours under power outage, b) interruption hours under blackout-heatwave compound hazard. c) summary statistics for the return period and impact for Ida-level events under different climates. The red curve shows median values for SSP5 8.5 for the future climate, with the 5% to 95% quantile range shown by shade, the yellow curve represents SSP2 4.5 for the future, and the blue for the historical climate. The dashed lines highlight Hurricane Ida’s power outage and compound hazard return levels. </center>
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Figure. 3. Estimated average interruption days of compound hazard. a) historical average interruption, b) future SSP 5 8.5 average interruption, and c) future SSP2 4.5 average interruption for each county in Louisiana for a compound hazard event with a 278-year return period. d) Distribution of Orleans Parish customers' compound hazard duration under a 278-year return period event. The solid lines show the percentage of residents affected by the compound hazard up to a certain length in the historical and future climates. </center>
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Figure. 4. Relative impacts of climate change factors on Ida’s compound hazard return period. a). Relative impact of each climate change factor assuming a consistent TC frequency. b). Sensitivity to TC frequency change. Note that the combined impact of all climate factors on Ida’s compound hazard return period is highly non-linear and thus the sum of the relative impact of individual factors does not equal the total impact. The solid line below the points shows \(\pm 1\sigma\) of uncertainty level. </center>
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+- smncidacompoundMarch13.pdf
+
+<--- Page Split --->

preprint/preprint__2b8c98fb229297a1262aa712f9c884323c7fbe1c918252cec11c83d0130a8dc7/preprint__2b8c98fb229297a1262aa712f9c884323c7fbe1c918252cec11c83d0130a8dc7_det.mmd

@@ -0,0 +1,492 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 839, 175]]<|/det|>
+# Hurricane Ida's blackout-heatwave compound hazard risk in a changing climate  
+
+<|ref|>text<|/ref|><|det|>[[44, 196, 110, 214]]<|/det|>
+Ning Lin  
+
+<|ref|>text<|/ref|><|det|>[[55, 223, 256, 240]]<|/det|>
+nlin@princeton.edu  
+
+<|ref|>text<|/ref|><|det|>[[44, 269, 592, 288]]<|/det|>
+Princeton University https://orcid.org/0000- 0002- 5571- 1606  
+
+<|ref|>text<|/ref|><|det|>[[44, 294, 144, 312]]<|/det|>
+Kairui Feng  
+
+<|ref|>text<|/ref|><|det|>[[53, 316, 592, 334]]<|/det|>
+Princeton University https://orcid.org/0000- 0001- 8978- 2480  
+
+<|ref|>text<|/ref|><|det|>[[44, 340, 163, 357]]<|/det|>
+Avantika Gori  
+
+<|ref|>text<|/ref|><|det|>[[53, 363, 232, 380]]<|/det|>
+Princeton University  
+
+<|ref|>text<|/ref|><|det|>[[44, 387, 118, 404]]<|/det|>
+Dazhi Xi  
+
+<|ref|>text<|/ref|><|det|>[[53, 408, 590, 427]]<|/det|>
+Princeton University https://orcid.org/0000- 0002- 4096- 8441  
+
+<|ref|>text<|/ref|><|det|>[[44, 432, 150, 450]]<|/det|>
+Min Ouyang  
+
+<|ref|>text<|/ref|><|det|>[[53, 454, 835, 473]]<|/det|>
+Huazhong University of Science and Technology https://orcid.org/0000- 0002- 3190- 4390  
+
+<|ref|>text<|/ref|><|det|>[[44, 479, 238, 496]]<|/det|>
+Michael Oppenheimer  
+
+<|ref|>text<|/ref|><|det|>[[53, 500, 592, 519]]<|/det|>
+Princeton University https://orcid.org/0000- 0002- 9708- 5914  
+
+<|ref|>text<|/ref|><|det|>[[44, 560, 104, 578]]<|/det|>
+Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 599, 137, 617]]<|/det|>
+Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[44, 636, 315, 655]]<|/det|>
+Posted Date: March 29th, 2024  
+
+<|ref|>text<|/ref|><|det|>[[44, 675, 475, 694]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 4096843/v1  
+
+<|ref|>text<|/ref|><|det|>[[42, 712, 914, 754]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 772, 535, 792]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 828, 912, 871]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on May 15th, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 59737- 8.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[63, 112, 106, 128]]<|/det|>
+## Title  
+
+<|ref|>text<|/ref|><|det|>[[121, 146, 752, 165]]<|/det|>
+Hurricane Ida's blackout- heatwave compound hazard risk in a changing climate  
+
+<|ref|>sub_title<|/ref|><|det|>[[63, 184, 135, 200]]<|/det|>
+## Authors  
+
+<|ref|>text<|/ref|><|det|>[[121, 216, 750, 253]]<|/det|>
+Kairui Feng \(^{1,2}\) , Ning Lin \(^{1,*}\) , Avantika Gori \(^{1}\) , Dazhi Xi \(^{1,3}\) , Min Ouyang \(^{4}\) , Michael Oppenheimer \(^{2,3,5}\)  
+
+<|ref|>sub_title<|/ref|><|det|>[[63, 289, 159, 305]]<|/det|>
+## Affiliations  
+
+<|ref|>text<|/ref|><|det|>[[120, 322, 828, 358]]<|/det|>
+\(^{1}\) Department of Civil and Environmental Engineering, Princeton University, New Jersey, U.S.  
+
+<|ref|>text<|/ref|><|det|>[[120, 373, 762, 393]]<|/det|>
+\(^{2}\) High Meadows Environmental Institute, Princeton University, New Jersey, U.S.  
+
+<|ref|>text<|/ref|><|det|>[[120, 409, 830, 429]]<|/det|>
+\(^{3}\) Program in Atmospheric and Oceanic Sciences, Princeton University, New Jersey, U.S.  
+
+<|ref|>text<|/ref|><|det|>[[120, 445, 816, 483]]<|/det|>
+\(^{4}\) School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, China.  
+
+<|ref|>text<|/ref|><|det|>[[120, 498, 668, 518]]<|/det|>
+\(^{5}\) Department of Geosciences, Princeton University, New Jersey, U.S.  
+
+<|ref|>text<|/ref|><|det|>[[121, 534, 288, 552]]<|/det|>
+\(^{*}\) nlin@princeton.edu  
+
+<|ref|>sub_title<|/ref|><|det|>[[63, 588, 139, 604]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[59, 621, 825, 848]]<|/det|>
+The emerging tropical cyclone (TC)- blackout- heatwave compound hazard under climate change are not well understood. In this study, we employ future projections of TCs, sea levels, and heatwaves, in conjunction with power system resilience modeling, to evaluate historical and future TC- blackout- heatwave compound hazard risks in Louisiana, US. We find that the return period for a compound hazard event comparable to Hurricane Ida (2021), with approximately 35 million customer hours of simultaneous power outage and heatwave exposure in Louisiana, is around 278 years in the historical climate (1980- 2005). Under the emissions scenario SSP5 8.5 (SSP2 4.5), this return period may decrease by a factor of \(\sim 17 \times (10x)\) to 16.2 (28.4) years in the future climate (2070- 2100). The significant increase in risk can be primarily attributed to projected escalations in heatwaves, which result in an approximate 5(2)- fold decrease in compound hazard return period, and in TC activity, which cause an estimated 2(1)- fold decrease in the return period. The findings contribute to our knowledge of and adaptation to compound climate hazards.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[63, 61, 173, 78]]<|/det|>
+## Introduction  
+
+<|ref|>text<|/ref|><|det|>[[62, 85, 836, 245]]<|/det|>
+In August 2021, Hurricane Ida, a Category 4 storm, struck Louisiana with intense winds, heavy rainfall, and significant storm surges, resulting in widespread flooding and damage to the state's infrastructure systems. Subsequent to the hurricane's landfall, the state experienced a loss of approximately 200 million customer hours of electricity, affecting roughly 2.15 million customers for an average power outage duration of 96 hours. Data from the U.S. Department of Energy reveals that Hurricane Ida caused the most extensive power outage in Louisiana's history, largely surpassing Hurricane Katrina (Category 5; 2005) and Hurricane Laura (Category 4; 2020), which led to losses of approximately 140 million and 100 million customer hours of electricity, respectively (1,2,3).  
+
+<|ref|>text<|/ref|><|det|>[[62, 275, 820, 433]]<|/det|>
+Furthermore, a prolonged heatwave occurred in the aftermath of Hurricane Ida, particularly affecting households that lost power and thus had no air conditioning (3). Consequently, Louisiana residents experienced a total of 35 million hours of compound blackout- heatwave hazard risk (with a heat index surpassing \(37.8^{\circ}\mathrm{C} / 100^{\circ}\mathrm{F}\) , in accordance with Louisiana's heat advisory criteria, 4). Customers exposed to the compound hazard endured an average of approximately 98 hours of heatwave conditions (4). Prolonged heat exposure can cause hospitalization and mortality risks (5), especially among vulnerable populations, for example, elders (6). Understanding how often Ida- like compound blackout- heatwave events may occur is critical for the development of coastal risk mitigation strategies.  
+
+<|ref|>text<|/ref|><|det|>[[62, 464, 833, 691]]<|/det|>
+A "compound climate event" can result in significant impacts due to the combination of climate drivers and hazards such as floods, wildfires, heatwaves, and droughts (7). Traditional risk assessment methods typically consider one hazard at a time, potentially leading to an underestimation of risk, since the physical drivers causing extreme events may exhibit spatial and/or temporal dependencies and interact to exacerbate the overall impact. Hurricanes or intense tropical cyclones (TCs), as drivers of extreme wind, rainfall, and storm surge, inherently lead to compound impacts on coastal regions (8) and are responsible for nine of the ten largest power outages in the United States over the past two decades (9). While extreme winds are the primary source of damage to power systems, the presence of storm surges and heavy rainfall resulted in extensive flood inundation during Hurricane Ida, which caused additional physical damage and hindered power system resilience as repair crews were unable to access affected areas (2). In addition, sea- level rise (SLR) may intensify coastal flood inundation by extending and prolonging the flood coverage, further exacerbating power system damage and delaying recovery operations.  
+
+<|ref|>text<|/ref|><|det|>[[62, 722, 836, 934]]<|/det|>
+Due to the seasonal peak of intense heat being ahead of that of major TCs, TC- heatwave compound events have so far been rare worldwide (10). However, a previous study (11) found that for Harris County, TX (a portion of Houston situated on higher ground and affected by hurricane winds), the TC- blackout- heatwave compound hazard risk would increase by a factor as large as 23 over the course of \(21^{\mathrm{st}}\) century under the high emissions scenario RCP8.5. In recent years, TC- heatwave compound hazards have happened in the Gulf Coast region. Hurricane Ida may represent the first hurricane landfall on the mainland United States associated with a long- lasting (i.e., multi- day), large- scale blackout- heatwave compound hazard (during Hurricane Laura the state- average heat index was also high but did not reach the threshold of \(100^{\circ}\mathrm{F}\) ). During Hurricane Ida and Laura, at least eleven and eight Louisianans, respectively, died of heat- related illnesses (1,12). Investigating this emerging compound threat, possibly enhanced by climate change, will contribute to our knowledge of and adaptation to compound climate hazards.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[62, 87, 833, 210]]<|/det|>
+In this study, we integrate hazard projection and power system analysis to examine TC- blackout- heatwave compound hazard risks for Louisiana over the 21st century under the combined influence of SLR and changes in heatwave and storm climatology. We highlight the change in the return period/recurrence interval of Ida- like compound events from the historical to future climates. We further quantify the relative importance of the change in various climatological variables (i.e., heat stress, sea level, storm frequency, storm intensity) in driving the changes in the compound hazard risk.  
+
+<|ref|>text<|/ref|><|det|>[[61, 227, 835, 593]]<|/det|>
+Our framework is an extension of the previous study (11) to incorporate multiple hazards, including storm surge, rainfall and SLR, in addition to wind and heatwave, to more comprehensively model TC- blackout- heatwave compound hazard risk, for the entire State of Louisiana under various climate conditions. Specifically, we combine projections of heatwaves (13), TC hazards (including wind, storm surge, and rainfall, 14), and SLR (15,16,17), driven by CMIP6 GCMs (13) under both the high (Shared Socioeconomic Pathway 5 8.5; SSP5 8.5) and moderate (SSP2 4.5) emission scenarios. We generate a large number of compound hazard events based on the combined projections for historical (1980- 2005) and future (2081- 2100) climates to estimate how hazard probabilities may change over the 21st century (Methods). Then we utilize a physics- based power outage and restoration model for Louisiana to simulate wind/surge/rainfall- induced power system failure and recovery for each hazard event, to estimate the probabilities of TC- blackout- heatwave hazards. We extends the existing wind- impact- only simulation method (county level; 11,18) to a wind- rainfall- surge coupled framework for power damage and recovery process modeling to consider a larger study area including coastal regions (state level; Methods). Considering the uncertainty surrounding the impact of climate change on the frequency of TCs making landfall along the Gulf Coast, we assume a constant TC frequency but also assess the sensitivity of the compound hazard risk to TC frequency projection. To focus on the impact of climate change, we assume that the power system, population distribution, and recovery plans in the study region will remain unchanged. However, we assume that the coastal levees will be elevated following a design based on the return period of storm tides, as this enhancement may be considered necessary to prevent the region from frequent inundation due to SLR (Methods).  
+
+<|ref|>sub_title<|/ref|><|det|>[[64, 610, 126, 626]]<|/det|>
+## Results  
+
+<|ref|>sub_title<|/ref|><|det|>[[64, 645, 201, 662]]<|/det|>
+## Historical Cases  
+
+<|ref|>text<|/ref|><|det|>[[61, 679, 833, 942]]<|/det|>
+We first examine power outage simulations of the historical cases of Hurricanes Ida and Laura, which are the two major events over the last decade that caused widespread power disruptions in Louisiana (1). Ida devastated the eastern half of Louisiana, which is more densely populated (including the city of New Orleans), whereas Laura grazed the western side. Ida destroyed 31,000 poles (reported by local utility company Entergy, 19) that carry lower- voltage distribution lines in the neighborhoods, twice as many as those in Hurricane Laura (14,000 poles) and Katrina (2005; 17,000 poles). As shown in Fig. 1a and b, the model's estimates for the overall impact of Hurricanes Ida and Laura on Louisiana compare relatively well with the observation. Hurricane Ida led to 47% (48% in simulation) of customers being out of power within the first 24 hours and it took \(\sim 10\) days (11 days in simulation) for 90% of customers to restore power. Meanwhile, up to 60% of Louisiana residents were under heatwave conditions within 6 days after Hurricane Ida's landfall. On average, 42% of customers experienced compound power outage- heatwave hazards for at least a day after the hurricane's landfall, based on the overlap of county- level power outage and heat index data. Hurricane Laura led to 27% (29% in simulation) of customers being out of electricity initially and it took \(\sim 6\) days (8 days in simulation) for 90% of customers to restore
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[62, 60, 830, 255]]<|/det|>
+power. To measure the overall severity of the blackout associated with each TC, we compute the cumulative interruption hours of customers throughout Louisiana (the total of all customers' power outage duration), a commonly used metric in evaluating power system reliability (20). The model estimates in total 189 (156- 242; \(\pm 3\sigma\) ) million power interruption hours for Ida, which is consistent with the observed 206 million hours, and 110 (77- 153) million power interruption hours for Laura, which compares relatively well with the observed 99 million interruption hours. As a comparison, Hurricane Katrina led to \(\sim 140\) million power interruption hours. The model estimation for the spatial and temporal distribution of power outage also correlates well with observations (the average relative error is \(< 10\%\) between the modeled and observed county- level power outage), as illustrated in Fig. 1c for the peak power outages within 24 hours and 5 days after landfall power outages at the county level for Hurricane Ida.  
+
+<|ref|>sub_title<|/ref|><|det|>[[65, 272, 352, 290]]<|/det|>
+## Blackout and Compound Hazards  
+
+<|ref|>text<|/ref|><|det|>[[62, 305, 839, 708]]<|/det|>
+Integrating power outage and recovery modeling with projections of future TC, SLR, and heatwaves, we examine the risk of TC- induced blackout- heatwave compound hazards in Louisiana. We generate 10,000 simulations of synthetic hazard events for each of the historical (1980- 2005) and future (2081- 2100) SSP5 8.5 and SSP2 4.5 scenarios. Each stochastic simulation includes a continuous 20- year sequence of TC occurrences, along with the physical simulation of TC tracks, wind speeds, rainfall amounts, storm surge levels, and heatwaves. We track each customer's exposure (i.e., duration) to blackout or compound blackout- heatwave hazard in the power outage and recovery modeling process for each synthetic hazard event. Then, we integrate the customer- level results to obtain state- level statistics and estimate the return periods (i.e., reciprocal of annual exceedance probability) of event total interruption hours for the historical and future climates. As demonstrated by the substantial shift of the return period curves (Fig. 2), the power outage risk will increase dramatically from the historical to the future climate. Specifically, the historical return period of a power outage of 206 million customer hours, as in Hurricane Ida, is 64 years. Under the SSP5 8.5 (SSP2 4.5) scenario, the return period of Ida's total power outage is estimated to be 35.8 (38.2) years in the future. The power outage with Ida's return period of 64 years is estimated to be about 413 million (265 million) customer hours in the future. The return period of a TC- blackout- heatwave compound hazard of 35 million customer hours, as in Hurricane Ida, is 278 years in the historical climate. The return period of Ida's compound hazard of 35 million customer hours is estimated to be 16.2 (28.4) years in the future climate ( \(\sim 17x\) (10x) decrease in return period). The compound hazard with Ida's return period of 278 years is estimated to be about 435 million (138 million) customer hours in the future climate, which corresponds to an average 8.8 (2.8)- day blackout- heatwave compound hazard experience for every of Louisiana's 2.13 million customers.  
+
+<|ref|>text<|/ref|><|det|>[[62, 723, 836, 934]]<|/det|>
+Compared to that in SSP2 4.5, the power outage level is similar to that in SSP5 8.5 at Ida's return period or lower, although the power outage risk becomes significantly higher in SSP5 8.5 at higher return periods (Fig. 2a), due to higher extreme TC hazards and SLR in SSP5 8.5. The difference between the two emissions scenarios is larger for the compound hazard risks. Specifically, compared to that in SSP2 4.5, the compound hazard risk is slightly higher in SSP5 8.5 at Ida's return period or lower, and the compound hazard risk becomes dramatically higher in SSP5 8.5 at higher return periods (Fig. 2a), due to combined effects of larger increases in extreme heatwaves, TC hazards, SLR in SSP5 8.5. Specifically, the frequency of extremely severe compound events, such as those with triple the impact of Hurricane Ida (i.e., 100 million customer- hours of compound hazards), is expected to be 2.5 times higher under the high emission scenario (SSP5 8.5) compared to the moderate emission scenario (SSP2 4.5). However, for less severe events, such as those with a third of Ida's impact (i.e., 10 million customer- hours of
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[63, 60, 836, 150]]<|/det|>
+compound hazards), we do not find a statistically significant difference in event frequency between the two emission scenarios. These findings suggest that the combined effects of global warming and increasing hurricane intensity significantly amplify the risk of the most extreme compound events. Nonetheless, the moderate emission scenario may still lead to a similar level of compound hazard risk for events of Ida's magnitude as the high emission scenario.  
+
+<|ref|>sub_title<|/ref|><|det|>[[63, 202, 423, 220]]<|/det|>
+## Spatial Pattern of Compound Hazard Risk  
+
+<|ref|>text<|/ref|><|det|>[[62, 237, 835, 448]]<|/det|>
+To investigate the spatial distribution of the compound hazard risk, we estimate the county- average compound hazard interruption days for each synthetic hazard event. Fig. 3 shows the compound hazard interruption days with Ida's return period of 278 years for each county in Louisiana in the historical and future climates. The coastal counties face a greater compound hazard risk than inland counties for both historical and future climates (Figs. 3a- 3c). For example, the counties with an average compound hazard impact larger than 20 days in the future climates are mostly coastal counties. Coastal counties often face a greater compound hazard risk since hurricane winds reach peak strength before the storm makes landfall, and storms tend to weaken significantly as they move inland, causing less damage to the inland power infrastructure. Moreover, the floods induced by storm surge and/or heavy rainfall can severely damage coastal power sectors. The flooding also hampers the recovery efforts of local contractors by submerging electrical components in water and obstructing local traffic and logistics with debris.  
+
+<|ref|>text<|/ref|><|det|>[[61, 463, 836, 725]]<|/det|>
+The general spatial disparities in compound hazard risks are also significant and will increase with climate change. For example, in the historical climate, the most impacted county may face on average a 1.8- day compound hazard with Ida's return period (278 years), and the least impacted county does not face any compound hazard risk (Fig. 3a). In the future under the SSP5 8.5 (SSP2 4.5) scenario, the county with the greatest impact may face the compound hazard of on average 12.7 (3.1) days with Ida's return period (278 years), and the least impacted area will face the compound hazard of on average 1.1 (0.1) days (Figs. 3b- 3c). To quantify the spatial disparity in compound hazard risks, we employ the Gini coefficient, which is a measure of statistical dispersion often used to represent income inequality, wealth inequality, or consumption inequality within a nation or a social group (21). It ranges from 0 to 1, where 0 represents perfect equality (i.e., every county has the same average compound hazard duration) and 1 represents full inequality (i.e., one county faces the compound hazard while others do not). In the historical climate, the Gini coefficient is around 0.312; however, it becomes 0.632 (0.411) in the future climate under SSP5 8.5 (SSP2 4.5). The results indicate that warming is likely to exacerbate the existing disparities and inequalities of TC- blackout- heatwave compound hazard risk in Louisiana.  
+
+<|ref|>text<|/ref|><|det|>[[61, 741, 836, 918]]<|/det|>
+We also investigate the distribution of compound hazard duration for residents within densely populated counties. In Fig. 3d, we show the distribution of compound hazard interruption days for affected customers in Orleans Parish for Ida's return period of 278 years. In the historical climate, only 5% of affected customers may face a \(>120\) - hour (5- day) compound hazard. However, in the future climate under the SSP5 8.5 (SSP2 4.5) scenario, over 70% (50%) of affected customers will confront a \(>120\) - hour (5- day) compound hazard, over 25% (3%) of affected customers will confront a \(>240\) - hour (10- day) compound hazard, and 9% (0%) of affected customers will encounter the compound hazard over 360 hours (15 days). Hence, climate change not only increases the average compound hazard impact but also intensifies the tail risk that vulnerable residents may encounter, especially under the high emission scenario.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[63, 63, 458, 81]]<|/det|>
+## Drivers of Changes in Compound Hazard Risk  
+
+<|ref|>text<|/ref|><|det|>[[62, 96, 836, 515]]<|/det|>
+The change in the compound hazard risk is driven by the change in three climate factors: 1) heatwaves (heat index) 2) TC climatology and 3) the sea level. As we assume that TC frequency remains unchanged in the future, the changes in TC climatology include changes in TC characteristics, particularly intensity (which drives changes in wind, storm tide, and rainfall). To determine the relative effect of the changes in these factors, we estimate the changes in the compound hazard risk due to changes in temperature, SLR, and TC intensity, respectively, by adjusting each variable to its future value or distribution and calculating the resulting return period of Ida's compound hazard (i.e., 35 million customer hours of simultaneous power outage and heatwave impact), as shown in Fig. 4a. As discussed above, if all climate change factors are considered, Ida's return period would drop from 278 years to 16.2 (28.4) years from historical climate to future climate under SSP5 8.5 (SSP2 4.5). The change in heatwaves is the largest contributor to the change in Ida's return period from the historical climate to the future climate; due to solely the heatwave change, Ida's return period would drop from 278 years in the historical climate to 47.6 (84.3) years in the future climate, which is a \(\sim 5x\) (3x) return period decrease. This large impact is induced by the dramatic change in temperature and humidity: the annual number of heatwave days \((>37.8^{\circ}C)\) will increase by 8 times from the historical climate to the future climate (ensemble average of the six GCMs) for the study region. The contribution TC intensity change reduces Ida's return period to 167.8 (185.2) years in the future climate. The impact of SLR on Ida's return period is relatively small, reducing Ida's return period to 251.2 (263.3) years in the future climate. SLR appears to have a relatively low impact because we assume the levees along the coast will be elevated. Also, the impact of SLR is limited to coastal regions and it is averaged out when the compound hazard impact is calculated for the entire state. The contribution of the various climatological drivers to future compound hazard risk is consistent across the two different emission scenarios.  
+
+<|ref|>text<|/ref|><|det|>[[62, 531, 833, 741]]<|/det|>
+Considering that the projection of TC frequency is subject to significant uncertainty (22,23), we assumed a constant TC frequency in above analyses. Here we investigate the sensitivity of the estimated compound hazard risk to the projection of TC frequency change. In one scenario, we apply the TC frequency in ref. 14, which projects relatively high increases in TC frequency in the future climates under SSP5 8.5 (SSP2 4.5), and Ida's return period would drop to 7.9 (15.2) years, compared to 16.2 (28.4) years when accounting for all climate change factors except TC frequency change. In another scenario, we consider a \(30\%\) decrease in TC frequency, the lower bound of TC frequency projections ensembled in ref. 23. Ida's return period would become 23.1 (40.6) years, which is a similarly dramatic decrease from 278 years in the historical climate, compared to the case when TC frequency was hold constant. This sensitivity test indicates the relatively small impact of TC frequency change compared to the combined effects of other climate change factors.  
+
+<|ref|>sub_title<|/ref|><|det|>[[64, 795, 153, 812]]<|/det|>
+## Discussion  
+
+<|ref|>text<|/ref|><|det|>[[63, 829, 839, 935]]<|/det|>
+This analysis highlights the substantial increase in the frequency of Ida- level extreme power outage- heatwave compound hazards over time, resulting from the combined effect of temperature increase, sea- level rise, and storm climatology changes under the SSP5 8.5 (SSP 2 4.5) climate change scenario. Linear interpolation reveals that the return period of Hurricane Ida has decreased from 278 years around 2000 to 225.6 (228.0) years in the 2020s, indicating a \(19\%\) reduction in the return period over the past 20 years. This real- life observation of an emerging climate compound
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[63, 60, 771, 97]]<|/det|>
+hazard motivates further research on projecting future compound climate hazard risks and developing strategies to mitigate climate risks for various regions around the world.  
+
+<|ref|>text<|/ref|><|det|>[[63, 113, 835, 272]]<|/det|>
+When examining the impact of various climate scenarios, such as the high emission scenario SSP5 8.5 and the moderate emission scenario SSP2 4.5, it appears that the risk associated with Ida- or smaller- scale compound hazard events may not exhibit substantial difference. This result indicates that utility companies urgently need to prepare for the compound events to prevent major impacts. On the other hand, for less frequent events, the impact of these compound hazards is expected to be significantly lower under the moderate emissions scenario. Moreover, the duration of interruptions caused by compound hazards will also be reduced with moderate emissions. This result highlights the importance of strengthening climate change mitigation policy to reduce the impact of extreme climate hazards.  
+
+<|ref|>text<|/ref|><|det|>[[62, 289, 835, 464]]<|/det|>
+We do not consider the potential change in the power grid or its operation in the future. In the future, localized solutions, including backup generators and solar panels, can provide temporary support to residents who lose power from the main grid, thus mitigating the impacts of compound hazards (24). These solutions can help reduce the exposure of vulnerable populations to the effects of power outages and extreme heat, thereby lessening the overall impact of compound hazard events. However, backup generators and solar panels may be cost- prohibitive for many middle and low- income communities, limiting their effectiveness in reducing heat stress. From the main power grid design perspective, adopting effective strategies like burying distribution networks (5) and developing distributed power systems (24) can bolster the resilience of power infrastructure against extreme weather events.  
+
+<|ref|>text<|/ref|><|det|>[[62, 481, 836, 673]]<|/det|>
+We also do not consider the potential changes in demographics and human habitat. As extreme climate events become more frequent, coastal megacities are also expected to develop rapidly (25,26). Encouraging climate- resilient urban design principles that prioritize green spaces, water management systems, and heat- resistant building materials can enhance cities' resilience against compound hazard risks (27). Furthermore, the implementation of advanced early warning systems and preparedness measures, combined with public awareness campaigns, can help minimize potential impacts on vulnerable communities (28,29). Moreover, changes in population patterns, such as urbanization in low- elevation coastal zones and the concentration of populations in areas vulnerable to climate hazards, can also influence the severity and duration of compound hazards, emphasizing the need to account for these demographic shifts when devising adaptation strategies (25,30).  
+
+<|ref|>text<|/ref|><|det|>[[62, 689, 828, 935]]<|/det|>
+Quantifying the reliability and resilience of infrastructure systems under the impact of future compound hazards is essential for climate change adaptation. Developing an integrated risk assessment framework that combines climatology, civil and electric engineering, urban planning, and social sciences is crucial for comprehensively understanding the interconnected nature of compound hazard risks and their societal impacts, and the formulation of effective mitigation strategies (31,32). For example, conventional statistical methods may fail to detect significant changes in compound hazard risk, especially for the most extremes. Our analysis shows that the intensity of relatively frequent hazards may not change significantly in the future, especially under moderate emissions scenarios. However, if such a conclusion for frequent, observable events is statistically extrapolated to that for extreme events, we may transform events like "Ida" or extreme, which could have been foreseen and prepared for, into "black swan" events—unpredictable extreme disasters with unimaginable losses. Only physics- based modeling integrating climate and hazards projection and infrastructure/social system analysis may provide reliable estimates of future risks. This multidisciplinary perspective is essential for capturing the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[63, 60, 833, 202]]<|/det|>
+complex interactions between different hazards and their cascading effects on infrastructure systems and society as a whole, ultimately enabling the development of robust and resilient strategies to mitigate the impacts of compound hazards. Also, given various uncertainties in climate projection to social development, there is a need for continuous refinement and updating of risk analysis techniques as improved modeling approaches and new data become available. By adopting a comprehensive approach that integrates various disciplines and continuously enhances our understanding of compound hazard risks, we can work towards developing effective adaptation strategies for a sustainable future in the face of a changing climate.  
+
+<|ref|>sub_title<|/ref|><|det|>[[65, 238, 261, 255]]<|/det|>
+## Materials and Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[64, 271, 541, 290]]<|/det|>
+## Synthetic TCs, Storm Surge, Tide and Rainfall Modeling  
+
+<|ref|>text<|/ref|><|det|>[[62, 305, 836, 586]]<|/det|>
+We use the synthetic TC hazard dataset generated in ref. 14 for the North Atlantic basin and select the TC tracks passing within \(200\mathrm{km}\) of Louisiana. The dataset contains synthetic TC tracks generated with the statistical- deterministic TC model (33), which has been applied to TC hazard assessment (8,34,35). The synthetic TC tracks for the historical period (between 1980 and 2005) were generated based on the National Centers for Environmental Prediction (NCEP) reanalysis. The dataset also contains bias- corrected and weighted- average climate projections of TCs for the future period (2070 to 2100) under combined Shared Socioeconomic Pathway (SSP) emissions scenarios, SSP5 8.5 and SSP2 4.5, based on six CMIP6 climate models: CanESM5, CNRM- CM6- 1, UKESM1- 0- LL, EC- Earth3, IP- SL- CM6A- LR and MIROC6. The TC storm tides were modeled in ref. 13 using the Advanced Circulation (ADCIRC) hydrodynamic model (36,37). We extract peak storm tides at nodes ( \(\sim 1\mathrm{km}\) resolution) along the coastline of Louisiana for each TC and match these to the county level. The rain fields were simulated in ref. 14 for each synthetic TC using the physics- based Tropical Cyclone Rainfall (TCR) model (38). We apply area- averaged TCR estimates at the county level, and we employ the maximum 24- hour rainfall accumulation from each storm event, since the 24- hour storm duration is often utilized for rainfall risk assessment (38).  
+
+<|ref|>text<|/ref|><|det|>[[62, 602, 816, 743]]<|/det|>
+To perform sequential risk analysis, we generate 10,000 stochastic samples of storm sequences for both historical and future climate periods. These samples were derived by selecting storms (according to a Poisson process with a rate as the TC annual frequency) and their associated hazards from the TC hazard dataset (14) described above. Each stochastic sample consists of 20 consecutive years of TC activity. For the primary analysis in this study, we maintain a constant TC frequency in the future climate. For the sensitivity analysis, we consider the increased TC frequency projected by the statistical- deterministic TC model in ref. 14 and the decreased TC frequency by up to \(30\%\) projected by a range climate models ensembled in ref. 23.  
+
+<|ref|>text<|/ref|><|det|>[[62, 758, 828, 881]]<|/det|>
+For each sampled storm, we generate the spatial- temporal wind field, employing the classical Holland wind profile (39) and accounting for the effects of surface friction and large- scale background wind following ref. 40, and converting one- minute mean winds to 3- second wind gusts using gust factors (41). We estimate the coastal flood area by comparing the height of peak storm tide (if over levee height) to the ground surface elevation specified by the USGS thirty- meter DEM (42; Fig. S2), assuming that areas would be inundated when the storm tide exceeds the ground elevation.  
+
+<|ref|>text<|/ref|><|det|>[[62, 900, 833, 935]]<|/det|>
+In Louisiana, the actual seawall heights vary significantly along the coastline, ranging from 2 to 5 meters and often changing over short distances. Due to the difficulty in acquiring the precise data,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[62, 60, 825, 290]]<|/det|>
+our simulations do not incorporate partially available measurements (43 to 45). We assigned the current seawall level based on estimated 100- year flood level (estimated in ref. 14) for each coastal county in Louisiana, according to the typical design guidance. For example, the 100- year flood level for New Orleans is approximately 3.4 meters above the North American Vertical Datum of 1988 (NAVD 88) (13). This approximation introduces a degree of inaccuracy into our flood modeling. Acknowledging this limitation, we subsequently focused on binary flood data—whether a flood occurs or not— when developing our power system damage and recovery models. We observe that the occurrence of flooding is a critical factor that significantly hinders the restoration efforts of the power system in coastal counties. However, the inundation depth of the flooding appears to have a less substantial impact, as indicated in the sensitivity analysis in Fig. S1. When compared to the areas affected by the TC’s wind and rainfall, the flooded regions are generally smaller. Therefore, the majority of the structural damage to the power system may not be caused by flooding.  
+
+<|ref|>text<|/ref|><|det|>[[62, 306, 833, 516]]<|/det|>
+The future coastal levee plan is uncertain. In the future climate simulations, we assume the coastal levee will be elevated to the historical 100- year return level plus one percent quantile SLR ( \(\sim 2\) meters averaging over SSP2 4.5 and SSP5 8.5 for New Orleans calculated from ref. 14). This design strategy is commonly used by governmental agencies to plan the seawall height, and it is within the framework proposed by the U.S. Army Corps of Engineers for the New Orleans Region, Lafayette, and Lake Charles (Error! Reference source not found.). A sensitivity test was performed on future compound hazard risks given different elevations of the coastal levee from 0- 3 meters above the current level. If the levee was not elevated, the surge impact on the compound hazard risk would be significantly higher. On the other hand, when the levee is elevated by higher than 2 meters, the estimated compound risk is not sensitive to the variation of the assumed levee height (Supplementary, Fig. S1). The generated wind, rainfall, and coastal flood conditions from each sampled storm drive the power grid outage and recovery analysis.  
+
+<|ref|>sub_title<|/ref|><|det|>[[64, 533, 248, 551]]<|/det|>
+## Heatwave Projections  
+
+<|ref|>text<|/ref|><|det|>[[62, 567, 837, 760]]<|/det|>
+Following ref. 5, the daily HI is determined as a function of daily maximum near- surface (2 m) air temperature, daily mean specific humidity, and daily mean surface pressure. To maintain consistency with the TC simulation, we obtain these data for Louisiana from the NCEP reanalysis and the six GCMs stated above during and after landfall for each sampled synthetic storm (each synthetic storm is associated with a climatological time of occurrence and development). The future HI projected by the GCM is bias- corrected (11) by adding the difference between the NCEP reanalysis and the GCM- estimated historical HI. According to the historical analysis in ref. 4, the HI will drop upon TC landfall and will recover to the ambient average within around ten days. To account for this dependence between TCs and heatwaves, we add the composite of the impact of TC passage to the meteorological variables used to calculate the HI, where the composite impact is estimated based on historical data (Fig. 3a in ref. 4).  
+
+<|ref|>sub_title<|/ref|><|det|>[[64, 777, 285, 795]]<|/det|>
+## Sea Level Rise Projections  
+
+<|ref|>text<|/ref|><|det|>[[62, 811, 837, 935]]<|/det|>
+We employ sea- level projections produced by the Intergovernmental Panel on Climate Change Sixth Assessment Report (AR6; 15,17) using the Framework for Assessing Changes To Sea- level (FACTS; 16). Localized probabilistic SLR projections under the SSP5 8.5 and SSP2 4.5 emission scenarios with ‘medium confidence’ are incorporated in this analysis (there are two confidence levels in the datasets, which are low and medium levels). The local sea level projection takes into account ground uplift or subsidence, oceanographic effects, and spatially variable responses of the geoid and the lithosphere to shrinking land ice. The projection of SLR was developed for tide
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[62, 60, 837, 133]]<|/det|>
+gauge stations. For each TC sequence realization, we first sample a SLR time series (a realization) from the projection. Then, we linearly interpolate the SLR projection to the locations of coastal counties, and we estimate the storm surge relative to the current sea level by adding the SLR to the storm tide level at each time point for each county.  
+
+<|ref|>sub_title<|/ref|><|det|>[[65, 149, 364, 168]]<|/det|>
+## Power Outage and Recovery Model  
+
+<|ref|>text<|/ref|><|det|>[[62, 184, 808, 290]]<|/det|>
+We apply a physics- based power system model, which explicitly simulates component level damage to predict the total power outage, accounting for the effects of future evolving factors, e.g., climate change, infrastructure upgrade, and utility maintenance. The physics- driven modeling of the power system allows us to better understand the impact of climate change and effectiveness of risk mitigation measures compared to if we used purely data- driven models (47,48).  
+
+<|ref|>text<|/ref|><|det|>[[62, 306, 833, 481]]<|/det|>
+Specifically, we extend the power grid outage and recovery model developed by refs. 11 and 18 to simulate TC impact on the electric power system in Louisiana (power topology shown in Fig. S3- 5). The power grid failure model first applies probabilistic fragility functions to estimate the damage states of five main vulnerable component types of the power network: transmission substations, transmission lines, distribution nodes, distribution lines, and local distribution circuits. Component failures alter the power grid topology and may separate the power grid into disconnected sub- grids. A DC flow simulation is then performed to capture the power availability in each sub- grid (similar to approaches in 49,50,51,52). The power system is open and connects with systems outside the study area via transmission lines; the performance of the power grid outside the study area is assumed to be under normal operation.  
+
+<|ref|>text<|/ref|><|det|>[[62, 497, 825, 569]]<|/det|>
+The fragility curves in refs. Error! Reference source not found. and 18 only considered the wind damage. Here we extend the fragility functions to consider the effects of flood and rainfall. For example, the probability of failure of a substation given specific wind, rainfall, and surge flood levels is estimated based on a log- normal fragility function as:  
+
+<|ref|>equation<|/ref|><|det|>[[123, 583, 833, 617]]<|/det|>
+\[P(D\geq d_t|H = h = w + \alpha \cdot r + \beta \cdot f) = \int_0^h\frac{1}{\sqrt{2\pi}\sigma_t x}\exp \left(-\frac{(\ln x - \mu_i)^2}{2\sigma_t^2}\right)dx \quad (1)\]  
+
+<|ref|>text<|/ref|><|det|>[[62, 633, 833, 899]]<|/det|>
+where hazard (H/h) is considered as a linear combination of wind speed (w), rainfall amount (r) and flood condition (f, flooded or not; Boolean variable) with two parameters \(\alpha\) and \(\beta\) . With the shape \((\sigma_i)\) and location \((\mu_i)\) parameters, the log- normal distribution describes the probability of potential damage (D) in each of four states (di), i.e., i = {low, moderate, severe, complete} damage. Fragility function refers to the latent distribution of a component's ability to withstand outer forces (hazard). Some components may not withstand any force at all, while others can withstand very large outer force. Given a certain outer force, the probability of damage to the component is equal to the integral of fragility from 0 to that force level, i.e., the probability that the strength of the component is lower than outer force. The fragility functions for other components (support structures, distribution nodes, poles, conductors, and circuits) are similarly modeled with exponential, logistic, or uniform distributions. These fragility functions are similar to those in refs. Error! Reference source not found. and 18 except that the effects of rainfall and flood are incorporated. The parameters are estimated by the Markov chain Monte Carlo (MCMC) method to minimize the mean squared error between simulated and observed county- level power outages.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[62, 60, 835, 201]]<|/det|>
+The recovery model, developed based on emergency response plans and operational data, applies estimated recovery resources based on a priority- oriented strategy to repair damaged transmission substations, transmission lines, and critical facilities vital to public safety, health, and welfare before local distribution networks (11,18). Debris should be removed before utilities become able to reinstate the power system. This debris- cleaning time is sampled from a uniform distribution between 48 to 72 hours (estimated from utility reports, 2). We also account for that, within the debris- cleaning period and without structural failure of the distribution system, residents may turn on the main power switch themselves 24 to 48 hours after being flooded (53).  
+
+<|ref|>text<|/ref|><|det|>[[62, 217, 836, 375]]<|/det|>
+There is a chance that TCs will make landfill in sequence and the second TC come before the damage caused by the first TC is fully recovered (54). We account for this temporal compounding effect in our power system outage and recovery analysis. For each sampled sequential hazard time series, the initial state of the power system when a TC arrives is set based on the condition of the restoration state from the previous TC. If the power system is indeed not fully recovered from the previous TC impact, the emergency response plans following the second TC are also adjusted considering the recovery process for the first TC. Specifically, the response plans will re- evaluate and prioritize the restoration tasks and redirect the repair efforts based on this updated priority list.  
+
+<|ref|>text<|/ref|><|det|>[[62, 392, 830, 480]]<|/det|>
+The power grid outage and recovery models were calibrated (to determine the model parameters) for the study area using observed power outage data for Hurricane Ida and Laura using simulated wind and observed rainfall (55) and flood (56). The same wind field modeling method applied to the synthetic storms is used for these two historical storms with storm characteristics (i.e., track, intensity, and size) taken from the extended best track data (57).  
+
+<|ref|>sub_title<|/ref|><|det|>[[62, 515, 157, 531]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[110, 540, 832, 914]]<|/det|>
+1. Hurricane Response Record, Office of Cybersecurity, Energy Security, and Emergency Response, U.S. department of energy (2021).  
+2. Hurricane Ida Power Restoration, Utility Report, Entergy (2021)  
+3. Nicholas Bogel-Burroughs and Katy Reckdahl. The Greatest Killer in New Orleans Wasn’t the Hurricane. It Was the Heat.. New York Times (2021).  
+4. National Weather Service, New Orleans, New Orleans/Baton Rouge (2021)  
+5. K.L. Ebi, A. Capon, P. Berry, C. Broderick, R. de Dear, G. Havenith, Y. Honda, R.S. Kovats, W. Ma, A. Malik, and N.B. Morris. "Hot Weather and Heat Extremes: Health Risks." The Lancet, 398(10301), pp. 698-708 (2021).  
+6. D.H.K. Lee. "Seventy-five Years of Searching for a Heat Index." Environ. Res. 22(2), 331-356 (1980).  
+7. J. Zscheischler, S. Westra, B.J. Van Den Hurk, S.I. Seneviratne, P.J. Ward, A. Pitman, A. AghaKouchak, D.N. Bresch, M. Leonard, T. Wahl and X. Zhang "Future climate risk from compound events." Nat. Clim. Chang. 8, no. 6: 469-477(2018).  
+8. Gori, A., Lin, N., Xi, D. et al. Tropical cyclone climatology change greatly exacerbates US extreme rainfall–surge hazard. Nat. Clim. Chang. (2022).  
+9. T. Houser and P. Marsters. The world’s second largest blackout. Rhodium Group, 385(2018).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 60, 816, 97]]<|/det|>
+10. T. Matthews, R.L.Wilby and C. Murphy "An emerging tropical cyclone–deadly heat compound hazard". Nat. Clim. Chang., 9(8), 602-606(2019).  
+
+<|ref|>text<|/ref|><|det|>[[115, 104, 792, 157]]<|/det|>
+11. K. Feng, M. Ouyang, and N. Lin."Tropical cyclone-blackout-heatwave compound hazard resilience in a changing climate." Nature communications 13, no. 1: 4421(2022).  
+
+<|ref|>text<|/ref|><|det|>[[115, 164, 830, 217]]<|/det|>
+12. Louisiana Department of Health verifies one additional hurricane-related death, bringing toll to 27. Louisiana Dep. Heal. (2020). Published online Sept 9. Available at: https://ldh.la.gov/index.cfm/newsroom/detail/5761 (Accessed Oct 21, 2020).  
+
+<|ref|>text<|/ref|><|det|>[[115, 223, 835, 293]]<|/det|>
+13. B.C. O'Neill, C. Tebaldi, D.P. Van Vuuren, V. Eyring, P. Friedlingstein, G. Hurtt, R. Knutti, E. Kriegler, J.F. Lamarque, J. Lowe, and G.A. Meehl."The scenario model intercomparison project (ScenarioMIP) for CMIP6." Geosci. Model Dev. 9, 3461-3482 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[115, 300, 830, 336]]<|/det|>
+14. D. Xi, N. Lin, and A. Gori. "Increasing Sequential Tropical Cyclone Hazards along the US East and Gulf Coasts." Nat. Clim. Chang. 13(3), 258-265 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 336, 825, 371]]<|/det|>
+15. Fox-Kemper, B., et al. Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA, pp. 1211-1362, doi:10.1017/9781009157896.011.  
+
+<|ref|>text<|/ref|><|det|>[[115, 377, 825, 483]]<|/det|>
+16. Kopp, R. E., Garner, G. G., Hermans, T. H. J., Jha, S., Kumar, P., Reedy, A., Slangen, A. B. A., Turilli, M., Edwards, T. L., Gregory, J. M., Koubbe, G., Levermann, A., Merzky, A., Nowicki, S., Palmer, M. D., & Smith, C. (2023). The Framework for Assessing Changes To Sea-Level (FACTS) v1.0: A platform for characterizing parametric and structural uncertainty in future global, relative, and extreme sea-level change. Geoscientific Model Development, 16, 7461-7489.  
+
+<|ref|>text<|/ref|><|det|>[[115, 489, 830, 593]]<|/det|>
+17. Garner, G. G., T. Hermans, R. E. Kopp, A. B. A. Slangen, T. L. Edwards, A. Levermann, S. Nowicki, M. D. Palmer, C. Smith, B. Fox-Kemper, H. T. Hewitt, C. Xiao, G. Aðalgeirsdóttir, S. S. Drijfhout, T. L. Edwards, N. R. Golledge, M. Hemer, G. Krinner, A. Mix, D. Notz, S. Nowicki, I. S. Nurhati, L. Ruiz, J-B. Sallée, Y. Yu, L. Hua, T. Palmer, B. Pearson, 2021. IPCC AR6 Sea Level Projections. Version 20210809.  
+
+<|ref|>text<|/ref|><|det|>[[115, 600, 819, 636]]<|/det|>
+18. M. Ouyang, L. Duenas-Osorio, Multi-dimensional hurricane resilience assessment of electric power systems. Struct. Saf. 48, 15-24 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[115, 643, 755, 662]]<|/det|>
+19. Entergy, Restoration Times for Louisiana, including Hardest Hit Areas (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 668, 835, 704]]<|/det|>
+20. U.S. Energy Information Administration, Annual Electric Power Industry Report, Form EIA-861 detailed data files(2021)  
+
+<|ref|>text<|/ref|><|det|>[[115, 711, 781, 746]]<|/det|>
+21. J.L. Gastwirth (1972). "The estimation of the Lorenz curve and Gini index." The Review of Economics and Statistics, 306-316.  
+
+<|ref|>text<|/ref|><|det|>[[115, 753, 800, 805]]<|/det|>
+22. A.H. Sobel, A.A. Wing, S.J. Camargo, C.M. Patricola, G.A. Vecchi, C.Y. Lee, and M.K. Tippett. "Tropical cyclone frequency." Earth's Future 9, no. 12: e2021EF002275(2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 813, 825, 882]]<|/det|>
+23. T. Knutson, S.J. Camargo, J.C. Chan, K. Emanuel, C.H. Ho, J. Kossin, M. Mohapatra, M. Satoh, M. Sugi, K. Walsh, and L. Wu. "Tropical cyclones and climate change assessment: Part II: Projected response to anthropogenic warming." Bull. Amer. Meteor. Soc. 101, no. 3: E303-E322(2020).  
+
+<|ref|>text<|/ref|><|det|>[[115, 890, 800, 925]]<|/det|>
+24. J.A. Bennett, C.N. Trevisan, J.F. DeCarolis, C. Ortiz-García, M. Pérez-Lugo, B.T. Etienne, and A.F. Clarens. "Extending energy system modelling to include extreme
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[141, 61, 830, 96]]<|/det|>
+weather risks and application to hurricane events in Puerto Rico." Nat. Energy 6, no. 3: 240- 249 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[112, 103, 800, 140]]<|/det|>
+25. M.E. Hauer. "Migration induced by sea-level rise could reshape the US population landscape." Nat. Clim. Chang. 7, 321 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[112, 146, 812, 183]]<|/det|>
+26. K. Feng, N. Lin, S. Xian, and M.V. Chester. "Can we evacuate from hurricanes with electric vehicles?". Transp. Res. Part D: Transp. Environ. 86, p.102458 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[112, 189, 820, 225]]<|/det|>
+27. S. E. Gill, J. F. Handley, A. R. Ennos, and S. Pauleit. "Adapting Cities for Climate Change: The Role of the Green Infrastructure." Built Environ. 33(1), 115-133 (2007).  
+
+<|ref|>text<|/ref|><|det|>[[112, 231, 820, 284]]<|/det|>
+28. C. Wamsler, E. Brink, and O. Rantala. "Climate Change, Adaptation, and Formal Education: The Role of Schooling for Increasing Societies' Adaptive Capacities in El Salvador and Brazil." Ecol. Soc. 17(2), 2 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[112, 291, 820, 344]]<|/det|>
+29. J. Birkmann and K. von Teichman. "Integrating Disaster Risk Reduction and Climate Change Adaptation: Key Challenges—Scales, Knowledge, and Norms." Sustain. Sci. 5(2), 171-184 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[112, 351, 800, 404]]<|/det|>
+30. K.C. Seto, B. Güneralp, and L.R. Hutyra. "Global Forecasts of Urban Expansion to 2030 and Direct Impacts on Biodiversity and Carbon Pools." Proc. Natl. Acad. Sci. U.S.A. 109(40), 16083-16088 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[112, 410, 773, 429]]<|/det|>
+31. DT Cohen, 60 million live in the path of hurricanes. U.S. census bureau. (2018)  
+
+<|ref|>text<|/ref|><|det|>[[112, 435, 830, 504]]<|/det|>
+32. Xu, Luo, Kairui Feng, Ning Lin, A. T. D. Perera, H. Vincent Poor, Le Xie, Chuanyi Ji, X. Andy Sun, Qinglai Guo, and Mark O'Malley. "Resilience of renewable power systems under climate risks." Nature Reviews Electrical Engineering 1, no. 1 (2024): 53-66.  
+
+<|ref|>text<|/ref|><|det|>[[112, 512, 836, 548]]<|/det|>
+33. K. Emanuel, S. Ravela, E. Vivant, and C. Risi. "A Statistical Deterministic Approach to Hurricane Risk Assessment." Bull. Am. Meteorol. Soc. 87, S1-S5 (2006).  
+
+<|ref|>text<|/ref|><|det|>[[112, 555, 830, 608]]<|/det|>
+34. R. Marsooli, N. Lin, K. Emanuel, and K. Feng. "Climate change exacerbates hurricane flood hazards along US Atlantic and Gulf Coasts in spatially varying patterns." Nat. Commun. 10, 1-9 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[112, 615, 830, 667]]<|/det|>
+35. N. Lin, K. Emanuel, M. Oppenheimer, and E. Vanmarcke. "Physically based assessment of hurricane surge threat under climate change." Nat. Clim. Chang. 2, 462-467 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[112, 674, 825, 728]]<|/det|>
+36. R.A. Luettich, J.J. Westerink, and N.W. Scheffner. "ADCIRC: an advanced three-dimensional circulation model for shelves, coasts, and estuaries. Report 1, Theory and methodology of ADCIRC-2DD1 and ADCIRC-3DL." (1992).  
+
+<|ref|>text<|/ref|><|det|>[[112, 734, 833, 787]]<|/det|>
+37. J.J. Westerink, R.A. Luettich, C.A. Blain, and N.W. Scheffner. "ADCIRC: An advanced three-dimensional circulation model for shelves, coasts, and estuaries. Report 2: User's Manual for ADCIRC-2DD1." (1992).  
+
+<|ref|>text<|/ref|><|det|>[[112, 794, 805, 829]]<|/det|>
+38. D. Xi, N. Lin, and J. Smith. "Evaluation of a physics-based tropical cyclone rainfall model for risk assessment." J. Hydrometeorol. 21, no. 9: 2197-2218 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[112, 836, 785, 872]]<|/det|>
+39. K. Emanuel. "Assessing the present and future probability of Hurricane Harvey's rainfall." Proc. Natl. Acad. Sci. 114(48): 12681-12684 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[112, 878, 790, 915]]<|/det|>
+40. G.J. Holland. "An analytic model of the wind and pressure profiles in hurricanes." Mon. Weather Rev. 108(8): 1212-1218 (1980).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 60, 828, 97]]<|/det|>
+41. N. Lin and D. Chavas. "On hurricane parametric wind and applications in storm surge modeling." J. Geophys. Res. Atmos. 117(D9) (2012).  
+
+<|ref|>text<|/ref|><|det|>[[112, 103, 658, 122]]<|/det|>
+42. USGS (U.S. Geological Survey). Digital Elevation Model(2022).  
+
+<|ref|>text<|/ref|><|det|>[[112, 128, 802, 182]]<|/det|>
+43. C. Fagnant, A. Gori, K.B. Ensor, A. Sebastian, and P.B. Bedient. "Characterizing Spatiotemporal Trends in Extreme Precipitation Across the Southern Texas Coast." Nat. Hazards (2020).  
+
+<|ref|>text<|/ref|><|det|>[[112, 188, 826, 241]]<|/det|>
+44. M. Palaiseanu-Lovejoy, C.A. Thatcher, and J.A. Barras. "Levee crest elevation profiles derived from airborne lidar-based high resolution digital elevation models in south Louisiana." ISPRS J. Photogramm. Remote Sens. 91: 114-126 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[112, 248, 825, 300]]<|/det|>
+45. D.R. Johnson, J.R. Fischbach, and D.S. Ortiz. "Estimating surge-based flood risk with the coastal Louisiana risk assessment model." J. Coast. Res. (67 (10067)): 109-126 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[112, 308, 805, 343]]<|/det|>
+46. US Army Corps of Engineers, Hurricane & Storm Damage Risk Reduction System, New Orleans District(2021).  
+
+<|ref|>text<|/ref|><|det|>[[112, 350, 832, 403]]<|/det|>
+47. R. Nateghi, S.D. Guikema, and S.M. Quiring. "Comparison and validation of statistical methods for predicting power outage durations in the event of hurricanes." Risk Anal. 31(12): 1897-1906 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[112, 410, 835, 463]]<|/det|>
+48. C. Ji, Y. Wei, H. Mei, J. Calzada, M. Carey, S. Church, T. Hayes, B. Nugent, G. Stella, M. Wallace, and J. White. "Large-scale Data Analysis of Power Grid Resilience Across Multiple US Service Regions." Nat. Energy, 1(5), pp.1-8 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[112, 470, 805, 504]]<|/det|>
+49. Y. Yang, T. Nishikawa, and A.E. Motter. "Vulnerability and Cosusceptibility Determine the Size of Network Cascades." Phys. Rev. Lett., 118(4), 048301 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[112, 511, 800, 547]]<|/det|>
+50. Y. Yang, T. Nishikawa, and A.E. Motter. "Small Vulnerable Sets Determine Large Network Cascades in Power Grids." Science, 358(2017).  
+
+<|ref|>text<|/ref|><|det|>[[112, 554, 830, 606]]<|/det|>
+51. C. Zhai, T. Y. J. Chen, A. G. White, and S. D. Guikema. "Power Outage Prediction for Natural Hazards Using Synthetic Power Distribution Systems." Reliab. Eng. Syst. Saf., 208, 107348 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[112, 613, 830, 648]]<|/det|>
+52. P. Hines, E. Cotilla-Sanchez, and S. Blumsack. "Do Topological Models Provide Good Information about Electricity Infrastructure Vulnerability?" Chaos 20, 033122 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[112, 656, 748, 675]]<|/det|>
+53. Seven Steps to Restore Power After a Flood. New Equipment Digest (2017).  
+
+<|ref|>text<|/ref|><|det|>[[112, 681, 833, 716]]<|/det|>
+54. D. Xi and N. Lin. "Sequential Landfall of Tropical Cyclones in the United States: From Historical Records to Climate Projections." Geophys. Res. Lett. (2021).  
+
+<|ref|>text<|/ref|><|det|>[[112, 723, 756, 776]]<|/det|>
+55. J. Du. "NCEP/EMC 4KM Gridded Data (GRIB) Stage IV Data. Version 1.0." UCAR/NCAR - Earth Observing Laboratory, 2011. https://doi.org/10.5065/D6PG1QDD.  
+
+<|ref|>text<|/ref|><|det|>[[112, 784, 445, 800]]<|/det|>
+56. Dartmouth Flood Observatory.(2021).  
+
+<|ref|>text<|/ref|><|det|>[[112, 802, 802, 853]]<|/det|>
+57. J. Demuth, M. DeMaria, and J.A. Knaff. "Improvement of Advanced Microwave Sounder Unit Tropical Cyclone Intensity and Size Estimation Algorithms." J. Appl. Meteor. 45, 1573-1581 (2006).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[65, 133, 225, 150]]<|/det|>
+## Acknowledgments:  
+
+<|ref|>text<|/ref|><|det|>[[62, 165, 836, 308]]<|/det|>
+We thank the projection authors for developing and making the sea- level rise projections available, multiple funding agencies for supporting the development of the projections, and the NASA Sea Level Change Team for developing and hosting the IPCC AR6 Sea Level Projection Tool. K.F., N.L., A.G. and D.X. are supported by the U.S. National Science Foundation (1652448 and 2103754 as part of the Megalopolitan Coastal Transformation Hub) and C3. ai Digital Transformation Institute (C3. ai DTI Research Award). K.F. is supported by the HMEI- STEP Graduate Fellowship. M. Openheimer is supported by U.S. NSF grant 2103754. M. Ouyang is supported by the National Natural Science Foundation of China, 72074089, 51938004, 71821001.  
+
+<|ref|>sub_title<|/ref|><|det|>[[65, 324, 250, 340]]<|/det|>
+## Author contributions:  
+
+<|ref|>text<|/ref|><|det|>[[65, 341, 412, 410]]<|/det|>
+Conceptualization: KF, NL, MOY, MOP  Methodology: KF, NL, AG, DX, MOY  Writing—original draft: KF, NL, MOP  Writing—review & editing: AG, DX, MOY  
+
+<|ref|>sub_title<|/ref|><|det|>[[65, 428, 240, 445]]<|/det|>
+## Competing interests:  
+
+<|ref|>text<|/ref|><|det|>[[65, 446, 407, 463]]<|/det|>
+The authors declare no competing interests.  
+
+<|ref|>sub_title<|/ref|><|det|>[[65, 480, 330, 496]]<|/det|>
+## Data and materials availability:  
+
+<|ref|>text<|/ref|><|det|>[[64, 496, 820, 532]]<|/det|>
+The data generated from this study will be deposited to the NSF DesignSafe- CI online; a link to the data set will be provided upon publication.  
+
+<|ref|>sub_title<|/ref|><|det|>[[65, 567, 226, 584]]<|/det|>
+## Figures and Tables
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[128, 66, 770, 420]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[121, 444, 828, 551]]<|/det|>
+<center>Figure 1. Simulated and observed the total power outage and recovery process in Louisiana for Hurricane. a) Ida and b) Laura. The red curve shows median values, with 5% to 95% quantile range shown by shade and the blue curve shows the observation. The yellow curve in (a) shows the percent of customers impacted by heatwaves (value reads the right axis). c) Comparison of observed and simulated spatial distribution of power outage for Hurricane Ida. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[155, 72, 720, 747]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[120, 764, 835, 886]]<|/det|>
+<center>Figure. 2. The return period of hurricane events by various metrics for Louisiana. a) interruption hours under power outage, b) interruption hours under blackout-heatwave compound hazard. c) summary statistics for the return period and impact for Ida-level events under different climates. The red curve shows median values for SSP5 8.5 for the future climate, with the 5% to 95% quantile range shown by shade, the yellow curve represents SSP2 4.5 for the future, and the blue for the historical climate. The dashed lines highlight Hurricane Ida’s power outage and compound hazard return levels. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[140, 70, 777, 560]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[121, 579, 805, 701]]<|/det|>
+<center>Figure. 3. Estimated average interruption days of compound hazard. a) historical average interruption, b) future SSP 5 8.5 average interruption, and c) future SSP2 4.5 average interruption for each county in Louisiana for a compound hazard event with a 278-year return period. d) Distribution of Orleans Parish customers' compound hazard duration under a 278-year return period event. The solid lines show the percentage of residents affected by the compound hazard up to a certain length in the historical and future climates. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[135, 60, 480, 321]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[120, 342, 821, 449]]<|/det|>
+<center>Figure. 4. Relative impacts of climate change factors on Ida’s compound hazard return period. a). Relative impact of each climate change factor assuming a consistent TC frequency. b). Sensitivity to TC frequency change. Note that the combined impact of all climate factors on Ida’s compound hazard return period is highly non-linear and thus the sum of the relative impact of individual factors does not equal the total impact. The solid line below the points shows \(\pm 1\sigma\) of uncertainty level. </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[43, 93, 768, 112]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 366, 149]]<|/det|>
+- smncidacompoundMarch13.pdf
+
+<--- Page Split --->

preprint/preprint__2bad4797b5cc24f4647b5de99a0eb4e23c0e77ca443e03e3c2737e1043560bec/images_list.json

@@ -0,0 +1,62 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1 | Atomic structures of Zr-ZrO2 nanoparticles in 3D. a, Experimental 3D atomic model of the Zr1 NP with a-ZrO2 in blue, c-ZrO2 in green and metal core in red. b, Normalized BOO parameters of all atoms. The red dashed curve is a criterion to distinguish the disordered atoms (32% in total, atoms below the curve) and ordered atoms (68% in total, atoms above the curve). The standard BCC, HCP and FCC parameters are marked as black dots for reference. c, Zr-Zr PDFs of the c-ZrO2 (top panel) and a-ZrO2 (bottom panel), with Zr1 in red, Zr2 in blue and Zr3 in black. The gray peaks show the peak positions of the standard PDF of cubic phase ZrO2 (top panel) and monoclinic phase ZrO2 (bottom panel) for comparison. d, The NP consists of a-ZrO2, c-ZrO2 and a metal core grain. e, Magnified atomic structure of the pure Zr metal core viewing from (110) direction.",
+    "footnote": [],
+    "bbox": [
+      [
+        122,
+        386,
+        870,
+        646
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2 | Atomic concentration and the degree of oxidation of Zr-ZrO \\(_2\\) NP. a,b, Atomic concentration \\(\\rho_{\\mathrm{N}}\\) distribution (a) and the degree of oxidation (b) of all the atoms in Zr1. Each slice has a thickness of \\(5.3\\mathrm{\\AA}\\) . To increase the signal-to-noise ratio, we convolved the degree of oxidation with a 2-Å-wide 3D Gaussian kernel, but this also reduces the 3D spatial resolution of oxidation map to \\(\\sim 4\\mathrm{\\AA}\\) . c, Distribution of \\(\\rho_{\\mathrm{N}}\\) in c-ZrO \\(_2\\) (blue), a-ZrO \\(_2\\) (red) and metal core (yellow) phase. c-ZrO \\(_2\\) has a slightly larger \\(\\rho_{\\mathrm{N}}\\) distribution by \\(1\\%\\) (dashed lines) than a-ZrO \\(_2\\) . The inset figure shows the magnified histogram of metal core. d, The \\(\\rho_{\\mathrm{N}}\\) distribution of metal/c-ZrO \\(_2\\) as a function of the distance from the surface of metal core. The dashed lines show the standard \\(\\rho_{\\mathrm{N}}\\) in Zr metal (red) and cubic phase ZrO \\(_2\\) (green). e, A slice through the Zr1 NP as the red rectangle marked in (b), showing the degree of oxidation at different regions. f, 3D surface rendering of local degree of oxidation and corresponding atomic concentration \\(\\rho_{\\mathrm{N}}\\) , showing the strong correlation. The cutout is \\(25\\times 25\\times 25\\mathrm{\\AA}^3\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        118,
+        72,
+        881,
+        490
+      ]
+    ],
+    "page_idx": 7
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3 | 3D atomic metal-oxide interfaces. a-d, 3D surface renderings of three major phase, showing the contour (a) of metal core (red), c-ZrO2 (green) and a-ZrO2 (blue) of Zr1. Three planes are going through the Zr1 in different directions. The sliced atomic models (4-atom-layers in thickness) highlights three different types of interfaces, i.e., semi-coherent interface between metal core and c-ZrO2 (b; in light blue frame), incoherent interface between metal core and c-ZrO2 (c; in green frame) and incoherent interface between metal core and a-ZrO2 (d; in orange frame). e-g, Experimental semi-coherent interface structures specified by the rectangle region in (b) and (h-i) ideal model built with ideal FCC Zr metal and cubic ZrO2. e, The semi-coherent interface viewing from metal [110] direction. There is a bending of \\(\\sim 11^{\\circ}\\) between metal and interfacial layers in metal [112] direction (angle between red line and blue line), and a bending of \\(\\sim 8^{\\circ}\\) between interfacial layers and c-ZrO2 in oxide [110] direction (angle in blue line and ivory line). The coordination tripods in red and ivory boxes shows the spatial crystal orientation of metal and oxide, respectively. f, The semi-coherent interface viewing from metal [101] direction (by rotating the cutout in (e) \\(120^{\\circ}\\) counter clockwise), showing a twisting of \\(\\sim 4^{\\circ}\\) in metal [110] direction (angle between red line and ivory line). g, One atomic plane extracted from the semi-coherent interface (the highlighted area in red in e), viewing from metal [111] direction. In this direction, the oxide shows the (002) plane. The color of the atomic bonding shows the Zr-Zr bond length. The Zr-Zr bond lengths in metal and oxide are close to \\(3.3 \\mathring{\\mathrm{A}}\\) and \\(3.6 \\mathring{\\mathrm{A}}\\) , respectively. The Zr-Zr bond lengths in the interfacial layers are longer. h, The ideal model of an interface structure between ideal FCC Zr metal and ideal cubic ZrO2, showing a \\(15^{\\circ}\\) of wedge if no bending exists, some of the atoms cannot be bonded this way. To minimize the energy and maintain the coherency, a bending of \\(15^{\\circ}\\) (i) is needed to release the stress. The structure changed from metal to oxide shows in (j). The oxygen atoms are",
+    "footnote": [],
+    "bbox": [
+      [
+        123,
+        68,
+        877,
+        545
+      ]
+    ],
+    "page_idx": 9
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4 | Porous structures during oxidation. a, A 2.4-Å-thick slice from the reconstructed volume of Zr1, with vacancy (triangle), nano-pores (rectangle) and the largest pore (circle) highlight. b, Volume distribution of all the voids. We define the voids with volume no larger than filling two Zr atoms (125 Å3, Methods) as vacancies, the voids with volume between 125 and 4500 Å3 as nano-pores. We consider the largest pore with volume of 34000 Å3 independently as it touches and separates all three phases. Dashed lines show the boundaries between three types of voids we define. c, The surface renderings of all vacancies in c-ZrO2 (in green) and a-ZrO2 (in blue). The outline of whole NP is plotted with gray contour. d-e, Statistics of vacancies. (d) The fractions of vacancies in c-ZrO2 and a-ZrO2. (e) The radially normalized density distribution of vacancies as a function of distance from the surface core to the surface. f, The surface renderings of all nano-pores in c-ZrO2 (in green), in a-ZrO2 (in blue) and in between c-ZrO2 and a-ZrO2 (in orange). g, The surface rendering of the largest pore. The boundary atoms composed of amorphous and crystalline atoms are colored by blue and green, respectively. h, One interface between two c-ZrO2 regions with distorted interfacial Zr-Zr bonds, amorphous region and nano-pores. The crystal and amorphous atoms distinguished by BOO analysis are colored as green and blue, respectively. The contour of the nano-pore is colored as orange. i, Three representative slices show a 7.8-Å-thick (approximately five atomic layers) cross section of the nano-pores and surrounding atoms.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        90,
+        875,
+        540
+      ]
+    ],
+    "page_idx": 11
+  }
+]

preprint/preprint__2bad4797b5cc24f4647b5de99a0eb4e23c0e77ca443e03e3c2737e1043560bec/preprint__2bad4797b5cc24f4647b5de99a0eb4e23c0e77ca443e03e3c2737e1043560bec.mmd

@@ -0,0 +1,318 @@
+
+# Three-dimensional atomic interface between metal and oxide in Zr-ZrO2 nanoparticles  
+
+Jihan Zhou jhzhou@pku.edu.cn  
+
+Peking University https://orcid.org/0009- 0006- 8069- 0356  
+
+Yao Zhang Peking University https://orcid.org/0000- 0003- 3409- 6197  
+
+Zezhou Li Peking University https://orcid.org/0009- 0002- 4059- 4889  
+
+Xing Tong Songshan Lake Materials Laboratory  
+
+Zhiheng Xie Peking University  
+
+Siwei Huang Peking University  
+
+Yue- E Zhang Songshan Lake Materials Laboratory  
+
+Hai- Bo Ke Songshan Lake Materials Laboratory  
+
+Wei- Hua Wang  
+
+Chinese academy of science, Beijing https://orcid.org/0000- 0002- 9155- 2462  
+
+## Article  
+
+Keywords:  
+
+Posted Date: March 13th, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 3972857/v1  
+
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.
+
+<--- Page Split --->
+
+Version of Record: A version of this preprint was published at Nature Communications on September 2nd, 2024. See the published version at https://doi.org/10.1038/s41467-024-52026-w.
+
+<--- Page Split --->
+
+# Three-dimensional atomic interface between metal and oxide in Zr-ZrO₂  
+
+## nanoparticles  
+
+Yao Zhang \(^{1,5}\) , Zezhou Li \(^{1,5}\) , Xing Tong \(^{2,5}\) , Zhiheng Xie \(^{1}\) , Siwei Huang \(^{1}\) , Yue- E Zhang \(^{2,3}\) , Hai- Bo Ke \(^{2*}\) , Wei- Hua Wang \(^{2,4}\) , Jihan Zhou \(^{1*}\)  
+
+\(^{1}\) Beijing National Laboratory for Molecular Sciences, Center for Integrated Spectroscopy, College of Chemistry and Molecular Engineering, Peking University; Beijing, 100871, China. \(^{2}\) Songshan Lake Materials Laboratory, Dongguan 523808, China. \(^{3}\) College of Physics, Liaoning University, Shenyang 110036, China. \(^{4}\) Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China. \(^{5}\) These authors contributed equally to this work. \*Correspondence and requests for materials should be addressed to H.- B. K. (email: kehaibo@sslab.org.cn) and J. Z. (email: jhzhou@pku.edu.cn)  
+
+## Abstract  
+
+Metal- oxide interfaces with poor coherency have unique properties comparing to the bulk materials and offer broad applications in the fields of heterogeneous catalysis, battery, and electronics. However, current understanding of the three- dimensional (3D) atomic metal- oxide interfaces remains limited because of their inherent structural complexity and limitations of conventional two- dimensional imaging techniques. Here, we determine the 3D atomic structure of metal- oxide interfaces in zirconium- zirconia nanoparticles using atomic- resolution electron tomography. We quantitatively analyze the atomic concentration and the degree of oxidation, and find the coherency and translational symmetry of the interfaces are broken. Moreover, we observe porous structures such as Zr vacancies and nano- pores and investigate their distribution. Our findings provide a clear 3D atomic picture of metal- oxide interface with direct experimental evidence. We anticipate this work could encourage future studies on fundamental problems of oxides such as interfacial structures in semiconductor and atomic motion during oxidation process.
+
+<--- Page Split --->
+
+## Introduction  
+
+The oxidation is ubiquitous and involved in many processes on a daily basis. Most metals spontaneously form an oxidation layer on their surfaces. The metal- oxide interface plays a critical role in broad applications ranging from heterocatalysis \(^{1,2}\) , batteries \(^{3,4}\) , and electronics \(^{5,6}\) . The thermodynamics and kinetics of oxidation process have been extensively studied over the years \(^{7 - 9}\) . Plenty of research has been focused on the metal's work function \(^{10}\) , the transport of metal or oxygen species \(^{11 - 13}\) , and the rate of oxidation \(^{14}\) . A number of theories have been proposed upon these studies to understand the oxidation behavior. For example, Kirkendall effect is used to explain the formation of oxidized pores \(^{15 - 17}\) ; Wagner proposed the law of oxidation kinetics in which the oxidation rate is controlled by the transport of ion under electrochemical potential gradient based on several assumptions \(^{12,14,18}\) . However, owing to the lack of direct observation of the metal- oxide interface at nanoscale or atomic scale, most of the classical oxidation theories have been limited to the macroscopic scale. Properties of the interface including catalytic activity, phonon dispersion and electron transportation are strongly related to the local atomic arrangements of the metal- oxide interface such as coordination numbers and atomic bond lengths \(^{19 - 24}\) . It is therefore essential to determine the three- dimensional (3D) atomic arrangements and understand the detailed oxidation structure of metal.  
+
+With the recent development in aberration corrected transmission electron microscopy (TEM), local structures of metal- oxide interfaces such as \(\mathrm{Cu - Cu_2O^{25,26}}\) , \(\mathrm{Ag - Ag_2O^{27}}\) , and \(\mathrm{Ni - NiO^{28}}\) have been studied at nanoscale or atomic scale; several of them were probed in situ by atomic resolution imaging and theoretical simulation \(^{25 - 29}\) . Semi- coherent and incoherent interfaces between metal and oxide have been observed at sub- angstrom resolution from two dimensional (2D) projections \(^{26}\) . Luo et al. discovered the periodic dislocation in \(\mathrm{Cu - Cu_2O}\) semi- coherent interface, suggesting the mechanism of strain release by defects between metal and oxide \(^{25}\) . Zhu et al. tracked the formation of voids in \(\mathrm{Ni - NiO}\) nanoparticles at nanoscale, identifying a two- stage oxidation mechanism including early- stage nucleation and then the Wagner oxidation \(^{28}\) . However, since the oxidized interfaces are usually non- epitaxial and inherently disordered due to the lattice mismatch, the atomic arrangements between some metal and its oxidation layer cannot be clearly elucidated using high resolution TEM or crystallography. Conventional 3D characterization methods such as atom probe tomography \(^{30,31}\) , electron tomography \(^{28,32 - 35}\) , and depth sectioning \(^{36,37}\) have been used to study the 3D morphological structures of the oxidized interfaces, and these techniques could overcome the limitation of single images which only provide the projected information of the 3D structures in 2D. However, the resolution of these techniques has limited to nanometer scale. Thus, determining the 3D atomic arrangements of the metal- oxide interface remains a major challenge. Although it remains notoriously difficult to imaging and identify each of the oxygen atoms of oxides in 3D, especially in high- angle annular dark- field scanning transmission electron microscopy (HAADF- STEM) mode, atomic resolution electron tomography (AET), which is an effective tool for determining the 3D atomic structure of nanomaterials \(^{32 - 35}\) , can in principle resolve the positions of heavy metal atoms in oxides and therefore give important structural information on this long- standing problem.  
+
+Here using \(\mathrm{Zr - ZrO_2}\) as a model system, we determine the 3D atomic structure of the metal- oxide interface using AET. We choose \(\mathrm{Zr - ZrO_2}\) for two reasons, first, \(\mathrm{Zr}\) can form oxide spontaneously in air and the oxidation process is moderate \(^{10}\) ; second, the \(\mathrm{Zr - O}\) bonding is extremely strong among all the common metal oxides and \(\mathrm{ZrO_2}\) has extreme chemical stability,
+
+<--- Page Split --->
+
+the \(\mathrm{Zr - ZrO_2}\) interface can maintain its atomic structures after electron illumination at a dose rate of \(6 \times 10^{5} \mathrm{e} \cdot \mathrm{\AA}^{- 2}\) , which is essential for electron tomography experiment. By determining all the Zr atomic positions in \(\mathrm{Zr - ZrO_2}\) nanoparticles (NPs), we obtained the 3D atomic structure of a partially oxidized Zr NP; it has an uncommon face-centered cubic (FCC) Zr metal crystal nucleus as the core and amorphous/crystalline \(\mathrm{ZrO_2}\) as the shell. We observed the atomic packing heterogeneity and numbers of Zr vacancies and small nano- cracks in the oxidation shell. Instead of forming a coherent interface, most of the atoms at the \(\mathrm{Zr - ZrO_2}\) interfaces connect with each other semi- coherently or incoherently. The degree of oxidation is decreasing while Zr packing density is increasing from the oxide surface to the metal core. We discovered a bidirectional distortion including bending and twisting at the semi- coherent metal- oxide interface. Moreover, we identify numbers of voids in the oxides including Zr vacancies, nanopores and large pores; the oxidation process is related to the distribution of the voids. These findings expand our understanding of the atomic structures of metal- oxide interfaces with poor coherency, encourage future studies on oxidation process at 3D atomic resolution, and further inspire the designing and modeling of atomic metal- oxide interface in surface engineering, heterogeneous catalysis and semiconductor.  
+
+## Results  
+
+## Atomic structures of \(\mathrm{Zr - ZrO_2}\) nanoparticles in 3D  
+
+NPs made of different monatomic metals with both disordered and crystalline structures can be achieved using fast- cooling vitrification process \(^{38,39}\) . Zr NPs were synthesized using pulse laser ablation of pure Zr target (purity \(>99.95\%\) ) in ethanol (Methods). By naturally oxidizing the freshly prepared Zr NPs in air, we obtained \(\mathrm{Zr - ZrO_2}\) NPs at different stages of the oxidation process. Some of the \(\mathrm{Zr - ZrO_2}\) NPs have an oxidized shell and a metal core (Supplementary Fig. 1). To confirm the oxidation, we used high resolution HAADF- STEM, energy dispersive spectroscopy (EDS) and electron energy loss spectroscopy (EELS) to characterize the \(\mathrm{Zr - ZrO_2}\) NPs (Supplementary Fig. 2 and Supplementary Fig. 3, respectively); it is notable that the edges of the NPs have a high degree of oxidation since the oxygen signal of EELS at the edge are stronger while HAADF- STEM intensities are weaker.  
+
+We resolved the 3D atomic structures of all Zr atoms in several \(\mathrm{Zr - ZrO_2}\) NPs using AET. In short, tomography tilt series (Supplementary Figs. 4- 6) were acquired from three \(\mathrm{Zr - ZrO_2}\) NPs at different stages of the oxidation process using an aberration- corrected TEM in HAADF- STEM mode (Supplementary Table 1). After imaging processing including denoising, background subtracting and alignment (Methods), the tilt series were reconstructed using algorithm described elsewhere \(^{33 - 35}\) . The 3D atomic coordinates of all Zr were traced from the computed reconstructions (Methods). We chose a partially oxidized \(\mathrm{Zr - ZrO_2}\) (named Zr1) as our main interest to elucidate the metal- oxide interfaces. The other two particles are fully oxidized without obvious metal core (named as Zr2 and Zr3; Supplementary Fig. 7). Fig. 1a and Supplementary Movie 1 show the experimental 3D atomic model of Zr1, showing there are ordered crystalline grains and disordered structures in the particle. We calculated the normalized bond orientational order (BOO) parameters for all the atoms to quantify the disorder (Fig. 1b and Methods); about \(32\%\) of all the Zr atoms are disordered. The particle has complicated phases, composing of a central metal grain, crystalline oxide grains (c- \(\mathrm{ZrO_2}\) ) and an amorphous oxide phase (a- \(\mathrm{ZrO_2}\) ) (Fig. 1d and Supplementary Movie 2). We calculated the Zr- Zr partial pair
+
+<--- Page Split --->
+
+distribution functions (PDFs) of the atoms in c- ZrO2 and a- ZrO2 separately (Methods). The c- ZrO2 show a well- matched cubic phase ZrO2 structure instead of monoclinic phase ZrO2 in both Zr1 and Zr2 (Fig. 1c). The PDFs of the a- ZrO2 atoms in all three NPs exhibit similar shape; and the first peak position is located at 3.45 Å, which is close to the first peak position of monoclinic phase ZrO2 (Fig. 1c). All the positions of main peaks and valleys in our Zr- Zr PDFs obtained from the atomic coordinates of our a- ZrO2 structures agree with those obtained from synchrotron X- ray diffraction<sup>40</sup>. The most populated Zr- Zr bond lengths in c- ZrO2 and a- ZrO2 are 3.6 Å and 3.45 Å, respectively (Supplementary Fig. 8). Interestingly, there is a small Zr metal core inside Zr1, confirmed by polyhedron template matching<sup>41</sup> and atomic concentration analysis (Methods). Fig. 1e shows the atomic structure of the pure Zr metal core viewing from (110) direction. The metal core has a distorted FCC structure with averaged Zr- Zr bond length being 3.3 Å, slightly longer than the standard value in Zr metal (3.2 Å). These observations are different from the bulk behavior, where Zr typically forms hexagonal close- packed (HCP) or body- centered cubic (BCC) structures; and usually forms monoclinic phase ZrO2 after natural oxidation<sup>42</sup>. This discrepancy highlights the distinctive behavior of materials at the nanoscale.  
+
+![](images/Figure_1.jpg)
+
+<center>Fig. 1 | Atomic structures of Zr-ZrO2 nanoparticles in 3D. a, Experimental 3D atomic model of the Zr1 NP with a-ZrO2 in blue, c-ZrO2 in green and metal core in red. b, Normalized BOO parameters of all atoms. The red dashed curve is a criterion to distinguish the disordered atoms (32% in total, atoms below the curve) and ordered atoms (68% in total, atoms above the curve). The standard BCC, HCP and FCC parameters are marked as black dots for reference. c, Zr-Zr PDFs of the c-ZrO2 (top panel) and a-ZrO2 (bottom panel), with Zr1 in red, Zr2 in blue and Zr3 in black. The gray peaks show the peak positions of the standard PDF of cubic phase ZrO2 (top panel) and monoclinic phase ZrO2 (bottom panel) for comparison. d, The NP consists of a-ZrO2, c-ZrO2 and a metal core grain. e, Magnified atomic structure of the pure Zr metal core viewing from (110) direction. </center>  
+
+## Atomic concentration and the degree of oxidation  
+
+To compare the local atomic packing density of Zr in all phases, we obtained the compactness of the NP by determining the Zr atomic concentration ( \(\rho_{\mathrm{N}}\) ) of all the regions present in Zr1 NP
+
+<--- Page Split --->
+
+(Methods). Fig. 2a shows the 3D \(\rho_{\mathrm{N}}\) distribution of Zr1. The low packing density regions are not related to any voids in the NP as we exclude all the voids from consideration when performing calculation (Methods). The averaged \(\rho_{\mathrm{N}}\) of metal core is \(3.85 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3}\) (Fig. 2c), close to the \(\rho_{\mathrm{N}}\) of ideal close-packed metallic Zr \((3.9 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3})\) . The \(\rho_{\mathrm{N}}\) of oxides (both c- ZrO2 and a- ZrO2 phases) are significantly lower than that of pure metal, being \(2.92 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3}\) and \(2.89 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3}\) , respectively. They are comparable to the \(\rho_{\mathrm{N}}\) of ideal c- phase ZrO2 \((3.0 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3})\) . We also observed 3D local \(\rho_{\mathrm{N}}\) heterogeneity in the oxides particularly distributed around the metal- oxide interfaces. Fig. 2d shows the \(\rho_{\mathrm{N}}\) distribution as a function of the distance from the surface of metal core (metal to c- ZrO2). The gradually decrease in \(\rho_{\mathrm{N}}\) suggests the metal- oxides interfaces are atomically smooth interface. The packing density gradient is attributed to the gradual change of the degree of oxidation of the Zr metal. Our PDFs and Zr- Zr bond length analysis suggest that c- ZrO2 is c- phase, and a- ZrO2 mainly forms the tetrahedral structure locally; \(^{40}\) oxygen should locate in tetrahedral sites in both phases (Supplementary Fig. 9). Next, we quantified the degree of oxidation by geometrically filling oxygen into the tetrahedral sites (Methods). Since EDS and EELS measurements in other similar Zr- ZrO2 NPs suggest the oxide grain are almost fully oxidized which is confirmed by our atomic concentration analysis (Fig. 2c), to satisfy the stoichiometric ratio of ZrO2, oxygen can be filled in eight tetrahedral sites (5.5 \(\mathring{\mathrm{A}}^{3}\) ) of the oxide (Supplementary Fig. 9a); but those tetrahedral sites (4.2 \(\mathring{\mathrm{A}}^{3}\) ) in Zr metal are too small (Supplementary Fig. 9b). Fig. 2b and Supplementary Movie 3 shows the 3D oxidation maps of Zr1. The degree of oxidation distribution in all the phases are shown in Fig. 2e and Supplementary Fig. 10, where the degree of oxidation increases along with the decrease of Zr \(\rho_{\mathrm{N}}\) . The c- ZrO2 and a- ZrO2 grains are fully oxidized in their surfaces; and they become less oxidized as closer to their interfaces with the central Zr core (Fig. 2e). The experimentally measured tetrahedral sites in central Zr core are too small to be filled with oxygen, confirming the core is barely oxidized. A cubic cutout of the 3D oxidation maps reveals that the degree of oxidation is strongly correlated to the atomic packing density of Zr; a highly oxidized region always has a lower Zr \(\rho_{\mathrm{N}}\) (Fig. 2f). It’s notable that some other Zr NPs are completely oxidized to cubic phase ZrO2 and/or amorphous ZrO2 (Supplementary Fig. 11) even from the same batch of oxidation. These results indicate the oxidation process is kinetics controlled, in which we observed several intermediate states of oxidized Zr- ZrO2.
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Fig. 2 | Atomic concentration and the degree of oxidation of Zr-ZrO \(_2\) NP. a,b, Atomic concentration \(\rho_{\mathrm{N}}\) distribution (a) and the degree of oxidation (b) of all the atoms in Zr1. Each slice has a thickness of \(5.3\mathrm{\AA}\) . To increase the signal-to-noise ratio, we convolved the degree of oxidation with a 2-Å-wide 3D Gaussian kernel, but this also reduces the 3D spatial resolution of oxidation map to \(\sim 4\mathrm{\AA}\) . c, Distribution of \(\rho_{\mathrm{N}}\) in c-ZrO \(_2\) (blue), a-ZrO \(_2\) (red) and metal core (yellow) phase. c-ZrO \(_2\) has a slightly larger \(\rho_{\mathrm{N}}\) distribution by \(1\%\) (dashed lines) than a-ZrO \(_2\) . The inset figure shows the magnified histogram of metal core. d, The \(\rho_{\mathrm{N}}\) distribution of metal/c-ZrO \(_2\) as a function of the distance from the surface of metal core. The dashed lines show the standard \(\rho_{\mathrm{N}}\) in Zr metal (red) and cubic phase ZrO \(_2\) (green). e, A slice through the Zr1 NP as the red rectangle marked in (b), showing the degree of oxidation at different regions. f, 3D surface rendering of local degree of oxidation and corresponding atomic concentration \(\rho_{\mathrm{N}}\) , showing the strong correlation. The cutout is \(25\times 25\times 25\mathrm{\AA}^3\) . </center>  
+
+## 3D atomic metal-oxide interfaces  
+
+Coherency of the metal- oxide interface affects many properties including strain, diffusion and band structure \(^{26,43,44}\) . It is extremely difficult to identify the atomic arrangement of semiconductor or incoherent metal- oxide interfaces from 2D projected images. To probe the 3D structure of metal- oxide interface at atomic level, we focus on the atomic Zr- Zr bonding of the interfaces with a range of \(\sim 10\mathrm{\AA}\) based on the packing density between metal core and oxide phases (Fig. 2d). Fig. 3a presents the 3D surface renderings of three major phase, showing the contour of metal core, c- ZrO \(_2\) and a- ZrO \(_2\) phase. Three slices with four atomic layers in thickness through the metal core show the Zr- Zr bonding of metal- oxide interfaces (Figs. 3b-
+
+<--- Page Split --->
+
+3d). We found several types of interfaces including semi-coherent and incoherent interfaces between metal and c- ZrO₂, and incoherent interface between metal and a- ZrO₂. The white rectangles in Figs. 3b- 3d highlight three cutouts from the atomic structures of a semi- coherent (Fig. 3e) and an incoherent interface (Fig. 3k) between metal and c- ZrO₂, and an incoherent interface (Fig. 3l) between metal and a- ZrO₂, respectively.  
+
+In the semi- coherent interface, four layers of metal Zr atoms (marked in deep red) from the metal \([1\bar{1} 0]\) direction correspond to four layers of Zr atoms (marked as ivory) in the oxide (Fig. 3e). To see the atomic connections in a single corresponding layer, one plane in the cutout is extracted and viewed from \([\bar{1}\bar{1} 1]\) direction of metal (Fig. 3g, Supplementary Fig. 12). Metal \((\bar{1}\bar{1} 1)\) plane is almost coplanar with oxide (002) plane; and the interface is about two atomic layers in thickness (blue atoms in Fig. 3g) and primarily connects metal (111) face with oxide \((11\bar{1})\) face. The Zr- Zr bond lengths increase from metal side \((\sim 3.3 \text{Å})\) to oxide side \((\sim 3.6 \text{Å})\) . The interface has a long Zr- Zr distance which is due to partially oxidation. Moreover, there is an angular mismatch of \(\sim 11^{\circ}\) between metal planes and the interfacial planes in metal [112] direction (oxide [110] direction), making the interface bending towards the oxide (Fig. 3e). To better illustrate the origin of the angular mismatch, we build an ideal model of Zr crystal grain and connect it to a cubic phase ZrO₂ from the same crystal orientation (Fig. 3h). Since the metal and oxide grains have different crystal orientations, there is a \(15^{\circ}\) of wedge through direct connection (angular mismatch in Fig. 3h). To minimize the interfacial energy while maintain the coherency, the oxide has to adopt a bending of \(15^{\circ}\) to fill the wedge (Fig. 3i). At the interface, the maximum numbers of filling oxygen are four instead of eight (Fig. 3j), which means the interface is partially oxidized and the maximum stoichiometric ratio is ZrO. Besides, it is notable that there is a gap angle of \(\sim 8^{\circ}\) between the Zr (100) planes in the oxide and those in the interface (Fig. 3e), alleviating the overall strain in the whole NPs. By rotating this cutout \(120^{\circ}\) counter- clockwise, we observed another angular mismatch of \(\sim 4^{\circ}\) between the metal \((\bar{1}\bar{1} 1)\) planes and oxide (002) planes in metal \([1\bar{1} 0]\) direction (oxide \([1\bar{1} 0]\) direction; Fig. 3f), which is perpendicular to metal [112] direction. It is considerable to have angular mismatch when two adjacent crystal grains having different crystal plane spacing. The (111) spacing of Zr metal is \(2.694 \text{Å}\) while (200) spacing of Zr oxide is \(2.546 \text{Å}\) . To compensate the spacing mismatch and to maintain the coherency, a certain degree (approximately \(4^{\circ}\) ) of twisting between metal and oxide is preferred (Supplementary Fig. 13)⁴⁵.  
+
+Most of the metal- oxide interfaces are incoherent in the whole particle. Figs. 3k and 3l show the incoherent interfaces of metal/c- ZrO₂ and metal/a- ZrO₂, respectively. Although the atomic bonding become more distorted and disordered in the incoherent interfaces between metal and c- ZrO₂, most of the metal core {111} faces still correspond to oxide {111} faces (Fig. 3k and Supplementary Fig. 14). Zr atoms form an incoherent boundary with lower coordination number and longer bond length than crystalline region (Supplementary Fig. 15 and Supplementary Fig. 16). Those metal- oxide incoherent interfaces introduce a number of defects which are distributed around the metal core. Many Zr defects are found in those incoherent interfaces. These observations indicate that when semi- coherent interface forms, a significant amount of strain could occur due to lattice and/or angular mismatch during oxidation. Once the strain caused by bending or twisting is too large, some of the semi- coherent interfaces could possibly turn to disordered structures through amorphization⁴⁶, where the coherency of interface is completely broken.
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Fig. 3 | 3D atomic metal-oxide interfaces. a-d, 3D surface renderings of three major phase, showing the contour (a) of metal core (red), c-ZrO2 (green) and a-ZrO2 (blue) of Zr1. Three planes are going through the Zr1 in different directions. The sliced atomic models (4-atom-layers in thickness) highlights three different types of interfaces, i.e., semi-coherent interface between metal core and c-ZrO2 (b; in light blue frame), incoherent interface between metal core and c-ZrO2 (c; in green frame) and incoherent interface between metal core and a-ZrO2 (d; in orange frame). e-g, Experimental semi-coherent interface structures specified by the rectangle region in (b) and (h-i) ideal model built with ideal FCC Zr metal and cubic ZrO2. e, The semi-coherent interface viewing from metal [110] direction. There is a bending of \(\sim 11^{\circ}\) between metal and interfacial layers in metal [112] direction (angle between red line and blue line), and a bending of \(\sim 8^{\circ}\) between interfacial layers and c-ZrO2 in oxide [110] direction (angle in blue line and ivory line). The coordination tripods in red and ivory boxes shows the spatial crystal orientation of metal and oxide, respectively. f, The semi-coherent interface viewing from metal [101] direction (by rotating the cutout in (e) \(120^{\circ}\) counter clockwise), showing a twisting of \(\sim 4^{\circ}\) in metal [110] direction (angle between red line and ivory line). g, One atomic plane extracted from the semi-coherent interface (the highlighted area in red in e), viewing from metal [111] direction. In this direction, the oxide shows the (002) plane. The color of the atomic bonding shows the Zr-Zr bond length. The Zr-Zr bond lengths in metal and oxide are close to \(3.3 \mathring{\mathrm{A}}\) and \(3.6 \mathring{\mathrm{A}}\) , respectively. The Zr-Zr bond lengths in the interfacial layers are longer. h, The ideal model of an interface structure between ideal FCC Zr metal and ideal cubic ZrO2, showing a \(15^{\circ}\) of wedge if no bending exists, some of the atoms cannot be bonded this way. To minimize the energy and maintain the coherency, a bending of \(15^{\circ}\) (i) is needed to release the stress. The structure changed from metal to oxide shows in (j). The oxygen atoms are </center>
+
+<--- Page Split --->
+
+colored in red. k- l, Incoherent interface structures specified by the rectangle region in (c) and (d), showing the metal/c- ZrO₂ interface (k) and metal/a- ZrO₂ interface (l). In panel e- l, the metal atoms, interfacial atoms and oxide atoms are colored in deep red, blue and ivory, respectively. The Zr atoms are bonded with their first- nearest Zr neighbors and linked with lines (Methods).  
+
+## Porous structures during oxidation  
+
+Porous structures of the oxide film formed on the surface of metal are usually associated with metal corrosion<sup>11,17,27–29</sup>. We observed numbers of porous structures in the Zr- ZrO₂ particles. Fig. 4a shows a 2.4- Å- thick slice from the reconstruction volume of Zr1; in which significant number of voids, such as Zr vacancies (triangle), nano- pores (rectangle) and the largest pore (circle) are observed in the particle. From the 3D intensity and surface renderings of three consecutive atomic layers, a single Zr vacancy defect can be clearly located (Supplementary Fig. 17). To determine all the voids and evaluate their occupied volume, we employed Voronoi analysis by measuring the distance of Voronoi vertices to atoms (Methods). Fig. 4b shows the histogram of volume distribution of all the voids, which we define as Zr vacancies, nano- pores and an extremely large nanoscale pore throughout the particle. The porosity is 17% and 14% in Zr1 and Zr2, respectively. Fig. 4c and Supplementary Movie 4 show the distribution of all Zr vacancies in Zr1. No vacancy is found in the metal core. More than 110 vacancies are distributed in the particle and all the Zr vacancies contribute 8.4% of the total porosity. Slightly more vacancies are found in a- ZrO₂ than in c- ZrO₂ (Fig. 4d). We plot the density of Zr vacancies from the boundary of the metal core to the surface of the particle (Fig. 4e). Most of the vacancies are distributed in the range of 15 Å between metal core and oxide, which corresponds to the region where Zr packing density exponentially decreases (Fig. 2d). It's notable that we exclude the vacancies from calculating the Zr packing density, the lower \(\rho_{\mathrm{N}}\) of interface is independent with the rich vacancies surrounding the metal core. We found 41 nano- pores in the volume range between 125 and 4500 ų. They are mostly irregular and with a relatively large length- to- radius ratio. Fig. 4f and Supplementary Movie 4 show the distribution of all nano- pores; they mostly sit at the boundaries between c- ZrO₂ and a- ZrO₂ regions. The largest pore is more than 34000 ų and penetrates throughout the whole particle, providing possible pathway for further oxidation (Fig. 4g and Supplementary Movie 4). This pore predominantly sits in the a- ZrO₂ regions, connecting and separating all three phases; it terminates at the c- ZrO₂ region, releasing a large amount of strain. It's interesting that the Zr- Zr bonds are extremely distorted at the boundary between two c- ZrO₂ domains (Figs. 4h and 4i), where some of the Zr atoms turns to be completely amorphous to release strain. Several nano- pores are coincidentally observed at this region near the small amorphous ZrO₂ domain. These findings indicate that when the strain reaches to a certain point, possibly higher than the fracture point, the Zr- Zr crystal bonding could turn distorted and amorphous first, and then rupture to form defects to release the strain. It is generally believed that a compact layer of amorphous oxide at the micrometers scale can protect the interior of metal from further oxidation in aluminum<sup>47,48</sup>. While our results reveals that in zirconium oxide, the amorphous oxide regions are substantially more porous than those in the crystalline regions, the voids would further advance the oxidation of Zr metal. We observed variety of voids in Zr NPs at atomic scale to nanometer scale, which are highly related to the structure of incoherent interface. The rearrangements of all the atom positions including distortion, amorphization and the rupture of bonding are possibly due to the massive mass transportation during oxidation at the metal- oxide interfaces facilitated by these voids.
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Fig. 4 | Porous structures during oxidation. a, A 2.4-Å-thick slice from the reconstructed volume of Zr1, with vacancy (triangle), nano-pores (rectangle) and the largest pore (circle) highlight. b, Volume distribution of all the voids. We define the voids with volume no larger than filling two Zr atoms (125 Å3, Methods) as vacancies, the voids with volume between 125 and 4500 Å3 as nano-pores. We consider the largest pore with volume of 34000 Å3 independently as it touches and separates all three phases. Dashed lines show the boundaries between three types of voids we define. c, The surface renderings of all vacancies in c-ZrO2 (in green) and a-ZrO2 (in blue). The outline of whole NP is plotted with gray contour. d-e, Statistics of vacancies. (d) The fractions of vacancies in c-ZrO2 and a-ZrO2. (e) The radially normalized density distribution of vacancies as a function of distance from the surface core to the surface. f, The surface renderings of all nano-pores in c-ZrO2 (in green), in a-ZrO2 (in blue) and in between c-ZrO2 and a-ZrO2 (in orange). g, The surface rendering of the largest pore. The boundary atoms composed of amorphous and crystalline atoms are colored by blue and green, respectively. h, One interface between two c-ZrO2 regions with distorted interfacial Zr-Zr bonds, amorphous region and nano-pores. The crystal and amorphous atoms distinguished by BOO analysis are colored as green and blue, respectively. The contour of the nano-pore is colored as orange. i, Three representative slices show a 7.8-Å-thick (approximately five atomic layers) cross section of the nano-pores and surrounding atoms. </center>  
+
+## Conclusion
+
+<--- Page Split --->
+
+In conclusion, we determined the 3D atomic structure of metal- oxide interfaces in Zr- ZrO₂ NP for the first time using atomic resolution electron tomography. We quantitatively measured the atomic packing density and the degree of oxidation from our experimental model of metal- oxide interface. The degree of oxidation from metal to oxide increases gradually, resulting a diffuse interface between FCC Zr core and ZrO₂. The Zr metal connects with its oxide via {111} planes; and the semi- coherent interface has severe distortion including bending and twisting. The significant stress in the interface is relieved through low coordination and defects. Numbers of defects including vacancies and nano- pores together leverage the mass transportation during oxidation. We anticipate that our new findings will fulfill the dearth of 3D atomic structure of metal- oxide interface and advance the study of fundamental problems of metal- oxide interfaces such as oxidation kinetics, diffusion and defect evolution in variety of materials.  
+
+## Methods  
+
+## Synthesis of zirconium nanoparticles  
+
+The Zr NPs were synthesized in liquid using laser ablation methods. An all- solid- state ultraviolet laser with a wavelength of 355 nm was employed for laser ablation in ethanol, with a pulse width of 7 ps, a max pulse energy of \(18 \mu \mathrm{J}\) , a repetition frequency of \(800 \mathrm{kHz}\) , and a beam spot diameter of \(20 \mu \mathrm{m}\) . Before being placed in a clean beaker, the bulk Zr target (purity \(>99.95\%\) ) was washed by acetone (99.5%) and ethanol (99.9%). The dissolved oxygen in liquid is eliminated to the minimization by nitrogen flow for 60 min with a flow rate of 4 L/min. Then, the Zr target was submerged into ethanol and the laser beam was accurately focused vertically on the surface of the bulk Zr through the ethanol in a closed chamber. The laser ablation was continued for 1 min, and the produced Zr NPs were dispersed using ultrasonic agitation and subsequently isolated via centrifugation to be collected in the ethanol solution. The freshly- prepared Zr NPs were then placed in air for one month to obtain a naturally oxidized layer several nanometers thick. The detailed methods of synthesis are described elsewhere. The final Zr- ZrO₂ NPs were drop cast onto 7- nm- thick \(\mathrm{Si}_3\mathrm{N}_4\) membranes using atomizer for TEM experiment.  
+
+## Atomic-resolution electron tomography  
+
+EELS maps were collected on an aberration- corrected Thermo Fisher Scientific Spectra 300 microscope operated at \(300 \mathrm{kV}\) using a Gatan Continuum GIF with a K3 direct electron detector in the Bay Area Centre for Electron Microscopy at Songshan Lake Materials Laboratory. EDS maps were collected on an aberration- corrected Thermo Fisher Scientific Themis Z microscope at \(300 \mathrm{kV}\) with a \(50 \mathrm{pA}\) beam current and a total acquisition time of approximately 5 min in Analytical Instrumentation Center at Peking University.  
+
+Tomographic tilt series are acquired by Thermo Fisher Scientific Titan microscope with spherical aberration correction at Electron Microscopy Laboratory of Peking University. The acceleration voltage was \(300 \mathrm{kV}\) and the imaging mode was HAADF mode. The tomographic tilt series were acquired at very low dose rate \((< 5 \times 10^5 \mathrm{e} / \mathrm{\AA}^2)\) to protect the structure of oxide. For each tilt angle, three sequential images with a dwell time of \(2 \mathrm{to} 4 \mu \mathrm{s}\) were acquired and registered using normalized cross- correlation, and then the averaged to enhance the signal- to- noise ratio.
+
+<--- Page Split --->
+
+Acquired images were drift corrected, denoised and aligned before reconstruction. Linear drift from the sample or stage was corrected during the image registration. Block- matching and 3D filtering (BM3D) is employed to denoise the images after drift correction<sup>49</sup>. And then, the background was estimated using the discrete Laplacian function of MATLAB and subtracted. In the direction perpendicular to the tilt axis, the images were aligned by maximizing the cross- correlation between the common lines. Along the tilt axis, the images were aligned using the center- of- mass method.  
+
+After image processing, the 3D reconstruction was computed from experimental tilt series using Real Space Iterative Reconstruction (RESIRE) algorithm<sup>33</sup>.  
+
+After reconstruction, atom tracing was performed to determine the 3D atomic coordinates. First, we interpolated reconstructed volume with spline method. All the local maxima in the reconstruction were identified as the rough atomic coordinates. Then, the coordinates were optimized according to the local volume of 1.7 Å \(\times\) 1.7 Å \(\times\) 1.7 Å with a polynomial fitting method. To separate the non- atoms from the potential atoms, K- means clustering method was employed based on the integrated intensity of the local volume (1.7 Å \(\times\) 1.7 Å \(\times\) 1.7 Å). For every potential atom, a minimum distance of 2 Å to its nearest atom should be satisfied. By carefully comparing the individual atom in the potential atomic models with the reconstructed volume, we manually corrected the atomic coordinates of unidentified or misidentified atoms (typically \(< 1\%\) ). The more detailed atom tracing procedure is described elsewhere.<sup>33</sup>  
+
+## Calculation of PDF  
+
+We calculated the PDF curve from experimental 3D atomic model by  
+
+\[g(r) = \frac{1}{N^2}\sum_{i = 1}^{N}\sum_{j = 1}^{N}\langle \delta \big(|\mathbf{r}_{ij}| - r\big)\rangle\]  
+
+where \(N\) is the total number of atoms; \(\delta\) is the Dirac delta function; \(\langle \cdot \rangle\) is the notation for expectation; \(|\mathbf{r}_{ij}|\) is the distance between the \(i\) - th atom and the \(j\) - th atom. To get a smoother PDF curve, a Gaussian kernel function with a \(\sigma\) of 1.5 Å was applied to convolute with original \(g(r)\). Finally, the PDF was scaled to approach one at the large pair distances. Using this procedure, we calculate the c- ZrO<sub>2</sub> and a- ZrO<sub>2</sub> separately in three NPs. From the PDF, we determined the first valley position as 4.5 Å, corresponding to the first- nearest- neighbor shell distance. The distance was used as a cutoff for BOO and alpha shape calculation (see the sections below).  
+
+## Local BOO parameters  
+
+We calculated the normalized local BOO parameters to distinguish the order and disorder of all the atoms. The normalized BOO parameter is defined as \(\sqrt{\bar{Q}_4^2 + \bar{Q}_6^2} /\sqrt{\bar{Q}_4^2}_{\mathrm{FCC}} + \bar{Q}_6^2_{\mathrm{FCC}}\) , where the \(\bar{Q}_4\) and \(\bar{Q}_6\) values were computed based on the procedure described elsewhere, using 4.5 Å (the first- nearest- neighbor shell distance) as a constraint<sup>50</sup>. The \(\bar{Q}_4_{\mathrm{FCC}}\) and \(\bar{Q}_6_{\mathrm{FCC}}\) are the reference values of the standard FCC structure. We separated the amorphous part from
+
+<--- Page Split --->
+
+crystalline part according to the criterion of the normalized BOO less than \(0.5^{34}\) .  
+
+## Determination of voids  
+
+Delaunay triangulation and Voronoi tessellation were performed to determine the voids. Delaunay triangulation, Voronoi tessellation and alpha shape were performed with the built- in functions of MATLAB (namely, 'delaunayTriangulation', 'voronoin', and 'alphaShape'). The initial spatial region of NP was calculated by alpha shape with \(\alpha = 4.5 \mathrm{\AA}\) (the first- nearest- neighbor shell distance). Then, we calculated the space that accommodates at least one Zr atom in the initial particle region with the following steps:  
+
+(1) The initial particle region was divided into tetrahedra by Delaunay triangulation.  
+
+(2) We determined whether a tetrahedron is void. We calculated the radius of circumscribed sphere for each tetrahedron. The radius represents the maximized sphere that can fit within the NP without intersecting with the center of any atom. Tetrahedra with a radius larger than \(3.19 \mathrm{\AA}\) were classified as voids. This criterion of radius was obtained based on the standard cubic phase of \(\mathrm{ZrO_2}\) with one vacancy.  
+
+(3) We grouped neighboring voids together to form larger voids. Two voids that share a common face are considered neighboring and thus combined into a single, larger void. The volume of these larger voids was calculated by summing the volumes of each void of component.  
+
+(4) We classified the voids into vacancies, nano-pores and the largest pore based on volume. We define the voids with volume filling one or two Zr atoms (45-125 \(\mathrm{\AA}^3\) ) as vacancies, the voids with volume between 125 and 4500 \(\mathrm{\AA}^3\) as nano-pores. The largest pore with volume of 34000 \(\mathrm{\AA}^3\) was considered independently.  
+
+Finally, the contours of voids were displayed with Laplacian or HC smoothing conducted by GIBBON \(^{51}\) .  
+
+## Atomic concentration  
+
+Topological bonds were determined based on the Voronoi tessellation. Two atoms are considered topologically bonded if their corresponding Voronoi polygons share a common face. In constructing the Voronoi polygons, we removed those surfaces with area less than \(1\%\) of the total area of the polygon surfaces \(^{52}\) . Additionally, this bond must also be shorter than \(4.5 \mathrm{\AA}\) , corresponding to the first- nearest- neighbor shell distance. The atomic concentration was calculated by \(\rho_{\mathrm{N}} = 1 / V\) , where \(V\) is the volume of Voronoi cell of an atom.  
+
+## The degree of oxidation  
+
+The oxidation state was determined using Delaunay triangulation. First, the distortion of Delaunay tetrahedra was considered. The distortion parameter was calculated by \(\delta = e_{\mathrm{max}} / e_{\mathrm{avg}} - 1\) , where \(e_{\mathrm{max}}\) and \(e_{\mathrm{avg}}\) are the maximum and average edge lengths of tetrahedron \(^{33}\) . We removed the tetrahedron with a distortion parameter larger than 0.255. Then, the volume of
+
+<--- Page Split --->
+
+remaining Delaunay tetrahedra was calculated. If the volume of a tetrahedron is larger than 4.68 \(\mathring{\mathrm{A}}^3\) (the averaged tetrahedron volume of FCC Zr lattice and c- ZrO2 lattice), an oxygen atom was placed inside. Finally, the degree of oxidation for each Zr atom was calculated by the fraction of its surrounding tetrahedra that accommodate one oxygen atom.  
+
+Acknowledgments: We thank the support of High- performance Computing Platform of Peking University. We thank the Electron Microscopy Laboratory at Peking University, Bay Area Centre for Electron Microscopy at Songshan Lake Materials Laboratory and Analytical Instrumentation Center at Peking University for the use of the aberration- corrected electron microscope. This work was supported by the National Natural Science Foundation of China (Grant No. 22172003, 52071222) and Guangdong Major Project of Basic and Applied Basic Research, China (Grant No. 2019B030302010).  
+
+## Author contributions:  
+
+J.Z. conceived the idea and directed the study. Z.L., Z.X. and Y.Z. performed TEM experiment and acquired the data. Y.Z. and S.H. performed the imaging processing, reconstructions, and atom tracing. Y.Z., Z.L., S.H. conducted/discussed data analysis under the direction of J.Z., T.X. and Y.- E. Z. synthesized Zr NPs under the direction of H.- B. K. and W.- H. W. Y.Z., Z.L. and J.Z. wrote the manuscript. All authors commented on the manuscript.  
+
+Competing interests: The authors declare no competing interests.  
+
+Data availability: All data are available upon reasonable request.  
+
+## Reference  
+
+1. Lee, S. W. et al. Controlling hot electron flux and catalytic selectivity with nanoscale metal-oxide interfaces. Nat. Commun. 12, 40 (2021).  
+2. Han, Z., Yang, F. & Li, Y. Dynamics of metal-support interface revealed by environmental transmission electron microscopy. Matter 5, 2531–2533 (2022).  
+3. Lowe, J. S. & Siegel, D. J. Modeling the interface between lithium metal and its native oxide. ACS Appl. Mater. Interfaces 12, 46015–46026 (2020).  
+4. Li, Q. et al. Extra storage capacity in transition metal oxide lithium-ion batteries revealed by in situ magnetometry. Nat. Mater. 20, 76–83 (2021).  
+5. Foerster, B. et al. Interfacial states cause equal decay of plasmons and hot electrons at gold–metal oxide interfaces. Nano Lett. 20, 3338–3343 (2020).  
+6. Tang, W. et al. The roles of metal oxidation states in perovskite semiconductors. Matter 6, 3782–3802 (2023).  
+7. Smialek, J. L. Invited review paper in commemoration of over 50 years of oxidation of metals: alumina scale adhesion mechanisms: a retrospective assessment. Oxid. Met. 97, 1–50 (2022).  
+8. Hu, J. et al. Microstructure and oxidation behavior of the Y/Ta/Hf co-doped AlCoCrFeNi high-entropy alloys in air at 1100 °C. Corros. Sci. 212, 110930 (2023).  
+9. Backman, L., Gild, J., Luo, J. & Opila, E. J. Part I: Theoretical predictions of preferential oxidation in refractory high entropy materials. Acta Mater. 197, 20–27 (2020).  
+10. Jeurgens, L. P. H., Lyapin, A. & Mittemeijer, E. J. The mechanism of low-temperature oxidation of zirconium. Acta Mater. 53, 4871–4879 (2005).
+
+<--- Page Split --->
+
+11. Birks, N., Meier, G. H. & Pettit, F. S. Introduction to the High-Temperature Oxidation of Metals, 2nd Edition. (Cambridge University Press, 2006).  
+
+12. Sun, X. et al. Dislocation-induced stop-and-go kinetics of interfacial transformations. Nature 607, 708-713 (2022).  
+
+13. Zhu, L. et al. Visualizing anisotropic oxygen diffusion in ceria under activated conditions. Phys. Rev. Lett. 124, 056002 (2020).  
+
+14. Morin, F., Beranger, G. & Lacombe, P. Limits of application for Wagner's oxidation theory. Oxid. Met. 4, 51-62 (1972).  
+
+15. Railsback, J. G., Johnston-Peck, A. C., Wang, J. & Tracy, J. B. Size-dependent nanoscale Kirkendall effect during the oxidation of nickel nanoparticles. ACS Nano 4, 1913-1920 (2010).  
+
+16. You, R. et al. Revealing temperature-dependent oxidation dynamics of Ni nanoparticles via ambient pressure transmission electron microscopy. Nano Lett. 23, 7260-7266 (2023).  
+
+17. Vijayakumar, J. et al. Absence of a pressure gap and atomistic mechanism of the oxidation of pure Co nanoparticles. Nat. Commun. 14, 174 (2023).  
+
+18. Gesmundo, F. & Viani, F. Application of Wagner's theory to the parabolic growth of oxides containing different kinds of defects: I. pure oxides. J. Electrochem. Soc. 128, 460-469 (1981).  
+
+19. Neyman, K. M. & Kozlov, S. M. Quantifying interactions on interfaces between metal particles and oxide supports in catalytic nanomaterials. NPG Asia Mater. 14, 59 (2022).  
+
+20. Tea, E., Huang, J., Li, G. & Hin, C. Atomic bonding and electrical potential at metal/oxide interfaces, a first principle study. J. Chem. Phys. 146, 124706 (2017).  
+
+21. Zou, S. et al. Grafting nanometer metal/oxide interface towards enhanced low-temperature acetylene semi-hydrogenation. Nat. Commun. 12, 5770 (2021).  
+
+22. Pacchioni, G. Electronic interactions and charge transfers of metal atoms and clusters on oxide surfaces. Phys. Chem. Chem. Phys. 15, 1737 (2013).  
+
+23. Wilson, R. B., Apgar, B. A., Hsieh, W.-P., Martin, L. W. & Cahill, D. G. Thermal conductance of strongly bonded metal-oxide interfaces. Phys. Rev. B 91, 115414 (2015).  
+
+24. Greiner, M. T., Chai, L., Helander, M. G., Tang, W. & Lu, Z. Metal/metal-oxide interfaces: How metal contacts affect the work function and band structure of MoO3. Adv. Funct. Mater. 23, 215-226 (2013).  
+
+25. Nian, Y. et al. Atomic-scale dynamic interaction of H2O molecules with Cu surface. Phys. Rev. Lett. 125, 156101 (2020).  
+
+26. Dong, Z., Ren, Q., Wang, S., Zhang, L. & Luo, L. Atomic-scale interfacial phase transformation governed Cu oxidation in water vapor. J. Phys. Chem. Lett. 12, 6996-7001 (2021).  
+
+27. Zhu, Q. et al. Defect-driven selective metal oxidation at atomic scale. Nat. Commun. 12, 558 (2021).  
+
+28. Sainju, R. et al. In situ studies of single-nanoparticle-level nickel thermal oxidation: from early oxide nucleation to diffusion-balanced oxide thickening. ACS Nano 16, 6468-6479 (2022).  
+
+29. Xia, W. et al. Bimetallic nanoparticle oxidation in three dimensions by chemically sensitive electron tomography and in situ transmission electron microscopy. ACS Nano 12, 7866-7874 (2018).  
+
+30. Ni, N. et al. How the crystallography and nanoscale chemistry of the metal/oxide interface develops during the aqueous oxidation of zirconium cladding alloys. Acta Mater. 60, 7132-7149 (2012).  
+
+31. Dong, Y., Motta, A. T. & Marquis, E. A. Atom probe tomography study of alloying element distributions in Zr alloys and their oxides. J. Nucl. Mater. 442, 270-281 (2013).  
+
+32. Zhou, J. et al. Observing crystal nucleation in four dimensions using atomic electron tomography. Nature 570, 500-503 (2019).  
+
+33. Yuan, Y. et al. Three-dimensional atomic packing in amorphous solids with liquid-like structure. Nat. Mater. 21, 95-102 (2022).  
+
+34. Yang, Y. et al. Determining the three-dimensional atomic structure of an amorphous solid. Nature 592,
+
+<--- Page Split --->
+
+60- 64 (2021).  
+
+35. Li, Z. et al. Probing the atomically diffuse interfaces in Pd@Pt core-shell nanoparticles in three dimensions. Nat. Commun. 14, 2934 (2023).  
+
+36. Ishikawa, R., Shibata, N., Taniguchi, T. & Ikuhara, Y. Three-dimensional imaging of a single dopant in a crystal. Phys. Rev. Applied 13, 034064 (2020).  
+
+37. Ishikawa, R. et al. High spatiotemporal-resolution imaging in the scanning transmission electron microscope. Microscopy 69, 240-247 (2020).  
+
+38. Liang, S.-X., Zhang, L.-C., Reichenberger, S. & Barcikowski, S. Design and perspective of amorphous metal nanoparticles from laser synthesis and processing. Phys. Chem. Chem. Phys. 23, 11121-11154 (2021).  
+
+39. Liu, Y. et al. Fabrication of amorphous PdNiCuP nanoparticles as efficient bifunctional and highly durable electrocatalyst for methanol and formic acid oxidation. J. Mater. Sci. Technol. 122, 148-155 (2022).  
+
+40. Zhang, F. et al. In situ study of the crystallization from amorphous to cubic zirconium oxide: riveted and reverse Monte Carlo analyses. Chem. Mater. 19, 3118-3126 (2007).  
+
+41. Larsen, P. M., Schmidt, S. & Schiotz, J. Robust structural identification via polyhedral template matching. Modelling Simul. Mater. Sci. Eng. 24, 055007 (2016).  
+
+42. Nicholls, R. J. et al. Crystal structure of the ZrO phase at zirconium/zirconium oxide interfaces. Adv. Eng. Mater. 17, 211-215 (2015).  
+
+43. Dholabhai, P. P., Martinez, E., Brown, N. T. & Uberuaga, B. P. On the mobility of carriers at semiconductor oxide heterointerfaces. Phys. Chem. Chem. Phys. 19, 23122-23130 (2017).  
+
+44. Pilania, G., Dholabhai, P. P. & Uberuaga, B. P. Role of symmetry, geometry, and termination chemistry on misfit dislocation patterns at semicoherent heterointerfaces. Matter 2, 1324-1337 (2020).  
+
+45. Shao, S., Wang, J. & Misra, A. Energy minimization mechanisms of semi-coherent interfaces. J. Appl. Phys. 116, 023508 (2014).  
+
+46. Sun, Z. et al. Janus Icosahedral Particles: Amorphization Driven by Three-Dimensional Atomic Misfit and Edge Dislocation Compensation. http://arxiv.org/abs/2310.16474 (2023) doi:10.21203/rs.3.rs-3439840/v1.  
+
+47. McGowan, J. C. Estimation of the approximate ionization constants of inorganic bases. Nature 168, 601-602 (1951).  
+
+48. Nguyen, L. et al. Atomic-scale insights into the oxidation of aluminum. ACS Appl. Mater. Interfaces 10, 2230-2235 (2018).  
+
+49. Makinen, Y., Azzari, L. & Foi, A. Collaborative filtering of correlated noise: Exact transform-domain variance for improved shrinkage and patch matching. IEEE Trans. on Image Process. 29, 8339-8354 (2020).  
+
+50. Lechner, W. & Dellago, C. Accurate determination of crystal structures based on averaged local bond order parameters. J. Chem. Phys. 129, 114707 (2008).  
+
+51. M Moerman, K. GIBBON: The geometry and image-based bioengineering add-on. JOSS 3, 506 (2018).  
+
+52. Sheng, H. W., Luo, W. K., Alamgir, F. M., Bai, J. M. & Ma, E. Atomic packing and short-to-medium-range order in metallic glasses. Nature 439, 419-425 (2006).
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+- ZrZrO2interfaceNatureSupplementaryInformation20240219final.pdf- MovieS1.mp4- MovieS2.mp4- MovieS3.mp4- MovieS4.mp4
+
+<--- Page Split --->

preprint/preprint__2bad4797b5cc24f4647b5de99a0eb4e23c0e77ca443e03e3c2737e1043560bec/preprint__2bad4797b5cc24f4647b5de99a0eb4e23c0e77ca443e03e3c2737e1043560bec_det.mmd

@@ -0,0 +1,431 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 930, 175]]<|/det|>
+# Three-dimensional atomic interface between metal and oxide in Zr-ZrO2 nanoparticles  
+
+<|ref|>text<|/ref|><|det|>[[44, 196, 245, 242]]<|/det|>
+Jihan Zhou jhzhou@pku.edu.cn  
+
+<|ref|>text<|/ref|><|det|>[[44, 268, 568, 288]]<|/det|>
+Peking University https://orcid.org/0009- 0006- 8069- 0356  
+
+<|ref|>text<|/ref|><|det|>[[44, 293, 568, 335]]<|/det|>
+Yao Zhang Peking University https://orcid.org/0000- 0003- 3409- 6197  
+
+<|ref|>text<|/ref|><|det|>[[44, 340, 568, 382]]<|/det|>
+Zezhou Li Peking University https://orcid.org/0009- 0002- 4059- 4889  
+
+<|ref|>text<|/ref|><|det|>[[44, 387, 377, 428]]<|/det|>
+Xing Tong Songshan Lake Materials Laboratory  
+
+<|ref|>text<|/ref|><|det|>[[44, 433, 208, 473]]<|/det|>
+Zhiheng Xie Peking University  
+
+<|ref|>text<|/ref|><|det|>[[44, 479, 208, 519]]<|/det|>
+Siwei Huang Peking University  
+
+<|ref|>text<|/ref|><|det|>[[44, 525, 377, 565]]<|/det|>
+Yue- E Zhang Songshan Lake Materials Laboratory  
+
+<|ref|>text<|/ref|><|det|>[[44, 571, 377, 611]]<|/det|>
+Hai- Bo Ke Songshan Lake Materials Laboratory  
+
+<|ref|>text<|/ref|><|det|>[[44, 617, 175, 636]]<|/det|>
+Wei- Hua Wang  
+
+<|ref|>text<|/ref|><|det|>[[50, 640, 733, 659]]<|/det|>
+Chinese academy of science, Beijing https://orcid.org/0000- 0002- 9155- 2462  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 700, 103, 717]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 737, 137, 755]]<|/det|>
+Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[44, 775, 315, 794]]<|/det|>
+Posted Date: March 13th, 2024  
+
+<|ref|>text<|/ref|><|det|>[[44, 814, 474, 833]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 3972857/v1  
+
+<|ref|>text<|/ref|><|det|>[[42, 851, 914, 893]]<|/det|>
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 911, 535, 931]]<|/det|>
+Additional Declarations: There is NO Competing Interest.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 45, 920, 88]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on September 2nd, 2024. See the published version at https://doi.org/10.1038/s41467-024-52026-w.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[133, 82, 864, 103]]<|/det|>
+# Three-dimensional atomic interface between metal and oxide in Zr-ZrO₂  
+
+<|ref|>sub_title<|/ref|><|det|>[[429, 120, 567, 139]]<|/det|>
+## nanoparticles  
+
+<|ref|>text<|/ref|><|det|>[[123, 153, 872, 191]]<|/det|>
+Yao Zhang \(^{1,5}\) , Zezhou Li \(^{1,5}\) , Xing Tong \(^{2,5}\) , Zhiheng Xie \(^{1}\) , Siwei Huang \(^{1}\) , Yue- E Zhang \(^{2,3}\) , Hai- Bo Ke \(^{2*}\) , Wei- Hua Wang \(^{2,4}\) , Jihan Zhou \(^{1*}\)  
+
+<|ref|>text<|/ref|><|det|>[[115, 217, 881, 365]]<|/det|>
+\(^{1}\) Beijing National Laboratory for Molecular Sciences, Center for Integrated Spectroscopy, College of Chemistry and Molecular Engineering, Peking University; Beijing, 100871, China. \(^{2}\) Songshan Lake Materials Laboratory, Dongguan 523808, China. \(^{3}\) College of Physics, Liaoning University, Shenyang 110036, China. \(^{4}\) Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China. \(^{5}\) These authors contributed equally to this work. \*Correspondence and requests for materials should be addressed to H.- B. K. (email: kehaibo@sslab.org.cn) and J. Z. (email: jhzhou@pku.edu.cn)  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 392, 194, 408]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[115, 416, 883, 656]]<|/det|>
+Metal- oxide interfaces with poor coherency have unique properties comparing to the bulk materials and offer broad applications in the fields of heterogeneous catalysis, battery, and electronics. However, current understanding of the three- dimensional (3D) atomic metal- oxide interfaces remains limited because of their inherent structural complexity and limitations of conventional two- dimensional imaging techniques. Here, we determine the 3D atomic structure of metal- oxide interfaces in zirconium- zirconia nanoparticles using atomic- resolution electron tomography. We quantitatively analyze the atomic concentration and the degree of oxidation, and find the coherency and translational symmetry of the interfaces are broken. Moreover, we observe porous structures such as Zr vacancies and nano- pores and investigate their distribution. Our findings provide a clear 3D atomic picture of metal- oxide interface with direct experimental evidence. We anticipate this work could encourage future studies on fundamental problems of oxides such as interfacial structures in semiconductor and atomic motion during oxidation process.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[117, 68, 228, 83]]<|/det|>
+## Introduction  
+
+<|ref|>text<|/ref|><|det|>[[116, 92, 882, 390]]<|/det|>
+The oxidation is ubiquitous and involved in many processes on a daily basis. Most metals spontaneously form an oxidation layer on their surfaces. The metal- oxide interface plays a critical role in broad applications ranging from heterocatalysis \(^{1,2}\) , batteries \(^{3,4}\) , and electronics \(^{5,6}\) . The thermodynamics and kinetics of oxidation process have been extensively studied over the years \(^{7 - 9}\) . Plenty of research has been focused on the metal's work function \(^{10}\) , the transport of metal or oxygen species \(^{11 - 13}\) , and the rate of oxidation \(^{14}\) . A number of theories have been proposed upon these studies to understand the oxidation behavior. For example, Kirkendall effect is used to explain the formation of oxidized pores \(^{15 - 17}\) ; Wagner proposed the law of oxidation kinetics in which the oxidation rate is controlled by the transport of ion under electrochemical potential gradient based on several assumptions \(^{12,14,18}\) . However, owing to the lack of direct observation of the metal- oxide interface at nanoscale or atomic scale, most of the classical oxidation theories have been limited to the macroscopic scale. Properties of the interface including catalytic activity, phonon dispersion and electron transportation are strongly related to the local atomic arrangements of the metal- oxide interface such as coordination numbers and atomic bond lengths \(^{19 - 24}\) . It is therefore essential to determine the three- dimensional (3D) atomic arrangements and understand the detailed oxidation structure of metal.  
+
+<|ref|>text<|/ref|><|det|>[[115, 396, 882, 840]]<|/det|>
+With the recent development in aberration corrected transmission electron microscopy (TEM), local structures of metal- oxide interfaces such as \(\mathrm{Cu - Cu_2O^{25,26}}\) , \(\mathrm{Ag - Ag_2O^{27}}\) , and \(\mathrm{Ni - NiO^{28}}\) have been studied at nanoscale or atomic scale; several of them were probed in situ by atomic resolution imaging and theoretical simulation \(^{25 - 29}\) . Semi- coherent and incoherent interfaces between metal and oxide have been observed at sub- angstrom resolution from two dimensional (2D) projections \(^{26}\) . Luo et al. discovered the periodic dislocation in \(\mathrm{Cu - Cu_2O}\) semi- coherent interface, suggesting the mechanism of strain release by defects between metal and oxide \(^{25}\) . Zhu et al. tracked the formation of voids in \(\mathrm{Ni - NiO}\) nanoparticles at nanoscale, identifying a two- stage oxidation mechanism including early- stage nucleation and then the Wagner oxidation \(^{28}\) . However, since the oxidized interfaces are usually non- epitaxial and inherently disordered due to the lattice mismatch, the atomic arrangements between some metal and its oxidation layer cannot be clearly elucidated using high resolution TEM or crystallography. Conventional 3D characterization methods such as atom probe tomography \(^{30,31}\) , electron tomography \(^{28,32 - 35}\) , and depth sectioning \(^{36,37}\) have been used to study the 3D morphological structures of the oxidized interfaces, and these techniques could overcome the limitation of single images which only provide the projected information of the 3D structures in 2D. However, the resolution of these techniques has limited to nanometer scale. Thus, determining the 3D atomic arrangements of the metal- oxide interface remains a major challenge. Although it remains notoriously difficult to imaging and identify each of the oxygen atoms of oxides in 3D, especially in high- angle annular dark- field scanning transmission electron microscopy (HAADF- STEM) mode, atomic resolution electron tomography (AET), which is an effective tool for determining the 3D atomic structure of nanomaterials \(^{32 - 35}\) , can in principle resolve the positions of heavy metal atoms in oxides and therefore give important structural information on this long- standing problem.  
+
+<|ref|>text<|/ref|><|det|>[[117, 848, 881, 921]]<|/det|>
+Here using \(\mathrm{Zr - ZrO_2}\) as a model system, we determine the 3D atomic structure of the metal- oxide interface using AET. We choose \(\mathrm{Zr - ZrO_2}\) for two reasons, first, \(\mathrm{Zr}\) can form oxide spontaneously in air and the oxidation process is moderate \(^{10}\) ; second, the \(\mathrm{Zr - O}\) bonding is extremely strong among all the common metal oxides and \(\mathrm{ZrO_2}\) has extreme chemical stability,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 66, 883, 362]]<|/det|>
+the \(\mathrm{Zr - ZrO_2}\) interface can maintain its atomic structures after electron illumination at a dose rate of \(6 \times 10^{5} \mathrm{e} \cdot \mathrm{\AA}^{- 2}\) , which is essential for electron tomography experiment. By determining all the Zr atomic positions in \(\mathrm{Zr - ZrO_2}\) nanoparticles (NPs), we obtained the 3D atomic structure of a partially oxidized Zr NP; it has an uncommon face-centered cubic (FCC) Zr metal crystal nucleus as the core and amorphous/crystalline \(\mathrm{ZrO_2}\) as the shell. We observed the atomic packing heterogeneity and numbers of Zr vacancies and small nano- cracks in the oxidation shell. Instead of forming a coherent interface, most of the atoms at the \(\mathrm{Zr - ZrO_2}\) interfaces connect with each other semi- coherently or incoherently. The degree of oxidation is decreasing while Zr packing density is increasing from the oxide surface to the metal core. We discovered a bidirectional distortion including bending and twisting at the semi- coherent metal- oxide interface. Moreover, we identify numbers of voids in the oxides including Zr vacancies, nanopores and large pores; the oxidation process is related to the distribution of the voids. These findings expand our understanding of the atomic structures of metal- oxide interfaces with poor coherency, encourage future studies on oxidation process at 3D atomic resolution, and further inspire the designing and modeling of atomic metal- oxide interface in surface engineering, heterogeneous catalysis and semiconductor.  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 396, 181, 412]]<|/det|>
+## Results  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 421, 546, 439]]<|/det|>
+## Atomic structures of \(\mathrm{Zr - ZrO_2}\) nanoparticles in 3D  
+
+<|ref|>text<|/ref|><|det|>[[115, 447, 882, 630]]<|/det|>
+NPs made of different monatomic metals with both disordered and crystalline structures can be achieved using fast- cooling vitrification process \(^{38,39}\) . Zr NPs were synthesized using pulse laser ablation of pure Zr target (purity \(>99.95\%\) ) in ethanol (Methods). By naturally oxidizing the freshly prepared Zr NPs in air, we obtained \(\mathrm{Zr - ZrO_2}\) NPs at different stages of the oxidation process. Some of the \(\mathrm{Zr - ZrO_2}\) NPs have an oxidized shell and a metal core (Supplementary Fig. 1). To confirm the oxidation, we used high resolution HAADF- STEM, energy dispersive spectroscopy (EDS) and electron energy loss spectroscopy (EELS) to characterize the \(\mathrm{Zr - ZrO_2}\) NPs (Supplementary Fig. 2 and Supplementary Fig. 3, respectively); it is notable that the edges of the NPs have a high degree of oxidation since the oxygen signal of EELS at the edge are stronger while HAADF- STEM intensities are weaker.  
+
+<|ref|>text<|/ref|><|det|>[[115, 639, 882, 917]]<|/det|>
+We resolved the 3D atomic structures of all Zr atoms in several \(\mathrm{Zr - ZrO_2}\) NPs using AET. In short, tomography tilt series (Supplementary Figs. 4- 6) were acquired from three \(\mathrm{Zr - ZrO_2}\) NPs at different stages of the oxidation process using an aberration- corrected TEM in HAADF- STEM mode (Supplementary Table 1). After imaging processing including denoising, background subtracting and alignment (Methods), the tilt series were reconstructed using algorithm described elsewhere \(^{33 - 35}\) . The 3D atomic coordinates of all Zr were traced from the computed reconstructions (Methods). We chose a partially oxidized \(\mathrm{Zr - ZrO_2}\) (named Zr1) as our main interest to elucidate the metal- oxide interfaces. The other two particles are fully oxidized without obvious metal core (named as Zr2 and Zr3; Supplementary Fig. 7). Fig. 1a and Supplementary Movie 1 show the experimental 3D atomic model of Zr1, showing there are ordered crystalline grains and disordered structures in the particle. We calculated the normalized bond orientational order (BOO) parameters for all the atoms to quantify the disorder (Fig. 1b and Methods); about \(32\%\) of all the Zr atoms are disordered. The particle has complicated phases, composing of a central metal grain, crystalline oxide grains (c- \(\mathrm{ZrO_2}\) ) and an amorphous oxide phase (a- \(\mathrm{ZrO_2}\) ) (Fig. 1d and Supplementary Movie 2). We calculated the Zr- Zr partial pair
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 66, 883, 345]]<|/det|>
+distribution functions (PDFs) of the atoms in c- ZrO2 and a- ZrO2 separately (Methods). The c- ZrO2 show a well- matched cubic phase ZrO2 structure instead of monoclinic phase ZrO2 in both Zr1 and Zr2 (Fig. 1c). The PDFs of the a- ZrO2 atoms in all three NPs exhibit similar shape; and the first peak position is located at 3.45 Å, which is close to the first peak position of monoclinic phase ZrO2 (Fig. 1c). All the positions of main peaks and valleys in our Zr- Zr PDFs obtained from the atomic coordinates of our a- ZrO2 structures agree with those obtained from synchrotron X- ray diffraction<sup>40</sup>. The most populated Zr- Zr bond lengths in c- ZrO2 and a- ZrO2 are 3.6 Å and 3.45 Å, respectively (Supplementary Fig. 8). Interestingly, there is a small Zr metal core inside Zr1, confirmed by polyhedron template matching<sup>41</sup> and atomic concentration analysis (Methods). Fig. 1e shows the atomic structure of the pure Zr metal core viewing from (110) direction. The metal core has a distorted FCC structure with averaged Zr- Zr bond length being 3.3 Å, slightly longer than the standard value in Zr metal (3.2 Å). These observations are different from the bulk behavior, where Zr typically forms hexagonal close- packed (HCP) or body- centered cubic (BCC) structures; and usually forms monoclinic phase ZrO2 after natural oxidation<sup>42</sup>. This discrepancy highlights the distinctive behavior of materials at the nanoscale.  
+
+<|ref|>image<|/ref|><|det|>[[122, 386, 870, 646]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 656, 883, 821]]<|/det|>
+<center>Fig. 1 | Atomic structures of Zr-ZrO2 nanoparticles in 3D. a, Experimental 3D atomic model of the Zr1 NP with a-ZrO2 in blue, c-ZrO2 in green and metal core in red. b, Normalized BOO parameters of all atoms. The red dashed curve is a criterion to distinguish the disordered atoms (32% in total, atoms below the curve) and ordered atoms (68% in total, atoms above the curve). The standard BCC, HCP and FCC parameters are marked as black dots for reference. c, Zr-Zr PDFs of the c-ZrO2 (top panel) and a-ZrO2 (bottom panel), with Zr1 in red, Zr2 in blue and Zr3 in black. The gray peaks show the peak positions of the standard PDF of cubic phase ZrO2 (top panel) and monoclinic phase ZrO2 (bottom panel) for comparison. d, The NP consists of a-ZrO2, c-ZrO2 and a metal core grain. e, Magnified atomic structure of the pure Zr metal core viewing from (110) direction. </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 855, 544, 872]]<|/det|>
+## Atomic concentration and the degree of oxidation  
+
+<|ref|>text<|/ref|><|det|>[[115, 881, 883, 917]]<|/det|>
+To compare the local atomic packing density of Zr in all phases, we obtained the compactness of the NP by determining the Zr atomic concentration ( \(\rho_{\mathrm{N}}\) ) of all the regions present in Zr1 NP
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 66, 883, 622]]<|/det|>
+(Methods). Fig. 2a shows the 3D \(\rho_{\mathrm{N}}\) distribution of Zr1. The low packing density regions are not related to any voids in the NP as we exclude all the voids from consideration when performing calculation (Methods). The averaged \(\rho_{\mathrm{N}}\) of metal core is \(3.85 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3}\) (Fig. 2c), close to the \(\rho_{\mathrm{N}}\) of ideal close-packed metallic Zr \((3.9 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3})\) . The \(\rho_{\mathrm{N}}\) of oxides (both c- ZrO2 and a- ZrO2 phases) are significantly lower than that of pure metal, being \(2.92 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3}\) and \(2.89 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3}\) , respectively. They are comparable to the \(\rho_{\mathrm{N}}\) of ideal c- phase ZrO2 \((3.0 \times 10^{- 2} \mathring{\mathrm{A}}^{- 3})\) . We also observed 3D local \(\rho_{\mathrm{N}}\) heterogeneity in the oxides particularly distributed around the metal- oxide interfaces. Fig. 2d shows the \(\rho_{\mathrm{N}}\) distribution as a function of the distance from the surface of metal core (metal to c- ZrO2). The gradually decrease in \(\rho_{\mathrm{N}}\) suggests the metal- oxides interfaces are atomically smooth interface. The packing density gradient is attributed to the gradual change of the degree of oxidation of the Zr metal. Our PDFs and Zr- Zr bond length analysis suggest that c- ZrO2 is c- phase, and a- ZrO2 mainly forms the tetrahedral structure locally; \(^{40}\) oxygen should locate in tetrahedral sites in both phases (Supplementary Fig. 9). Next, we quantified the degree of oxidation by geometrically filling oxygen into the tetrahedral sites (Methods). Since EDS and EELS measurements in other similar Zr- ZrO2 NPs suggest the oxide grain are almost fully oxidized which is confirmed by our atomic concentration analysis (Fig. 2c), to satisfy the stoichiometric ratio of ZrO2, oxygen can be filled in eight tetrahedral sites (5.5 \(\mathring{\mathrm{A}}^{3}\) ) of the oxide (Supplementary Fig. 9a); but those tetrahedral sites (4.2 \(\mathring{\mathrm{A}}^{3}\) ) in Zr metal are too small (Supplementary Fig. 9b). Fig. 2b and Supplementary Movie 3 shows the 3D oxidation maps of Zr1. The degree of oxidation distribution in all the phases are shown in Fig. 2e and Supplementary Fig. 10, where the degree of oxidation increases along with the decrease of Zr \(\rho_{\mathrm{N}}\) . The c- ZrO2 and a- ZrO2 grains are fully oxidized in their surfaces; and they become less oxidized as closer to their interfaces with the central Zr core (Fig. 2e). The experimentally measured tetrahedral sites in central Zr core are too small to be filled with oxygen, confirming the core is barely oxidized. A cubic cutout of the 3D oxidation maps reveals that the degree of oxidation is strongly correlated to the atomic packing density of Zr; a highly oxidized region always has a lower Zr \(\rho_{\mathrm{N}}\) (Fig. 2f). It’s notable that some other Zr NPs are completely oxidized to cubic phase ZrO2 and/or amorphous ZrO2 (Supplementary Fig. 11) even from the same batch of oxidation. These results indicate the oxidation process is kinetics controlled, in which we observed several intermediate states of oxidized Zr- ZrO2.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[118, 72, 881, 490]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 511, 883, 714]]<|/det|>
+<center>Fig. 2 | Atomic concentration and the degree of oxidation of Zr-ZrO \(_2\) NP. a,b, Atomic concentration \(\rho_{\mathrm{N}}\) distribution (a) and the degree of oxidation (b) of all the atoms in Zr1. Each slice has a thickness of \(5.3\mathrm{\AA}\) . To increase the signal-to-noise ratio, we convolved the degree of oxidation with a 2-Å-wide 3D Gaussian kernel, but this also reduces the 3D spatial resolution of oxidation map to \(\sim 4\mathrm{\AA}\) . c, Distribution of \(\rho_{\mathrm{N}}\) in c-ZrO \(_2\) (blue), a-ZrO \(_2\) (red) and metal core (yellow) phase. c-ZrO \(_2\) has a slightly larger \(\rho_{\mathrm{N}}\) distribution by \(1\%\) (dashed lines) than a-ZrO \(_2\) . The inset figure shows the magnified histogram of metal core. d, The \(\rho_{\mathrm{N}}\) distribution of metal/c-ZrO \(_2\) as a function of the distance from the surface of metal core. The dashed lines show the standard \(\rho_{\mathrm{N}}\) in Zr metal (red) and cubic phase ZrO \(_2\) (green). e, A slice through the Zr1 NP as the red rectangle marked in (b), showing the degree of oxidation at different regions. f, 3D surface rendering of local degree of oxidation and corresponding atomic concentration \(\rho_{\mathrm{N}}\) , showing the strong correlation. The cutout is \(25\times 25\times 25\mathrm{\AA}^3\) . </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 741, 401, 757]]<|/det|>
+## 3D atomic metal-oxide interfaces  
+
+<|ref|>text<|/ref|><|det|>[[115, 766, 883, 914]]<|/det|>
+Coherency of the metal- oxide interface affects many properties including strain, diffusion and band structure \(^{26,43,44}\) . It is extremely difficult to identify the atomic arrangement of semiconductor or incoherent metal- oxide interfaces from 2D projected images. To probe the 3D structure of metal- oxide interface at atomic level, we focus on the atomic Zr- Zr bonding of the interfaces with a range of \(\sim 10\mathrm{\AA}\) based on the packing density between metal core and oxide phases (Fig. 2d). Fig. 3a presents the 3D surface renderings of three major phase, showing the contour of metal core, c- ZrO \(_2\) and a- ZrO \(_2\) phase. Three slices with four atomic layers in thickness through the metal core show the Zr- Zr bonding of metal- oxide interfaces (Figs. 3b-
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 67, 881, 158]]<|/det|>
+3d). We found several types of interfaces including semi-coherent and incoherent interfaces between metal and c- ZrO₂, and incoherent interface between metal and a- ZrO₂. The white rectangles in Figs. 3b- 3d highlight three cutouts from the atomic structures of a semi- coherent (Fig. 3e) and an incoherent interface (Fig. 3k) between metal and c- ZrO₂, and an incoherent interface (Fig. 3l) between metal and a- ZrO₂, respectively.  
+
+<|ref|>text<|/ref|><|det|>[[115, 166, 881, 650]]<|/det|>
+In the semi- coherent interface, four layers of metal Zr atoms (marked in deep red) from the metal \([1\bar{1} 0]\) direction correspond to four layers of Zr atoms (marked as ivory) in the oxide (Fig. 3e). To see the atomic connections in a single corresponding layer, one plane in the cutout is extracted and viewed from \([\bar{1}\bar{1} 1]\) direction of metal (Fig. 3g, Supplementary Fig. 12). Metal \((\bar{1}\bar{1} 1)\) plane is almost coplanar with oxide (002) plane; and the interface is about two atomic layers in thickness (blue atoms in Fig. 3g) and primarily connects metal (111) face with oxide \((11\bar{1})\) face. The Zr- Zr bond lengths increase from metal side \((\sim 3.3 \text{Å})\) to oxide side \((\sim 3.6 \text{Å})\) . The interface has a long Zr- Zr distance which is due to partially oxidation. Moreover, there is an angular mismatch of \(\sim 11^{\circ}\) between metal planes and the interfacial planes in metal [112] direction (oxide [110] direction), making the interface bending towards the oxide (Fig. 3e). To better illustrate the origin of the angular mismatch, we build an ideal model of Zr crystal grain and connect it to a cubic phase ZrO₂ from the same crystal orientation (Fig. 3h). Since the metal and oxide grains have different crystal orientations, there is a \(15^{\circ}\) of wedge through direct connection (angular mismatch in Fig. 3h). To minimize the interfacial energy while maintain the coherency, the oxide has to adopt a bending of \(15^{\circ}\) to fill the wedge (Fig. 3i). At the interface, the maximum numbers of filling oxygen are four instead of eight (Fig. 3j), which means the interface is partially oxidized and the maximum stoichiometric ratio is ZrO. Besides, it is notable that there is a gap angle of \(\sim 8^{\circ}\) between the Zr (100) planes in the oxide and those in the interface (Fig. 3e), alleviating the overall strain in the whole NPs. By rotating this cutout \(120^{\circ}\) counter- clockwise, we observed another angular mismatch of \(\sim 4^{\circ}\) between the metal \((\bar{1}\bar{1} 1)\) planes and oxide (002) planes in metal \([1\bar{1} 0]\) direction (oxide \([1\bar{1} 0]\) direction; Fig. 3f), which is perpendicular to metal [112] direction. It is considerable to have angular mismatch when two adjacent crystal grains having different crystal plane spacing. The (111) spacing of Zr metal is \(2.694 \text{Å}\) while (200) spacing of Zr oxide is \(2.546 \text{Å}\) . To compensate the spacing mismatch and to maintain the coherency, a certain degree (approximately \(4^{\circ}\) ) of twisting between metal and oxide is preferred (Supplementary Fig. 13)⁴⁵.  
+
+<|ref|>text<|/ref|><|det|>[[115, 656, 881, 895]]<|/det|>
+Most of the metal- oxide interfaces are incoherent in the whole particle. Figs. 3k and 3l show the incoherent interfaces of metal/c- ZrO₂ and metal/a- ZrO₂, respectively. Although the atomic bonding become more distorted and disordered in the incoherent interfaces between metal and c- ZrO₂, most of the metal core {111} faces still correspond to oxide {111} faces (Fig. 3k and Supplementary Fig. 14). Zr atoms form an incoherent boundary with lower coordination number and longer bond length than crystalline region (Supplementary Fig. 15 and Supplementary Fig. 16). Those metal- oxide incoherent interfaces introduce a number of defects which are distributed around the metal core. Many Zr defects are found in those incoherent interfaces. These observations indicate that when semi- coherent interface forms, a significant amount of strain could occur due to lattice and/or angular mismatch during oxidation. Once the strain caused by bending or twisting is too large, some of the semi- coherent interfaces could possibly turn to disordered structures through amorphization⁴⁶, where the coherency of interface is completely broken.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[123, 68, 877, 545]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 549, 883, 919]]<|/det|>
+<center>Fig. 3 | 3D atomic metal-oxide interfaces. a-d, 3D surface renderings of three major phase, showing the contour (a) of metal core (red), c-ZrO2 (green) and a-ZrO2 (blue) of Zr1. Three planes are going through the Zr1 in different directions. The sliced atomic models (4-atom-layers in thickness) highlights three different types of interfaces, i.e., semi-coherent interface between metal core and c-ZrO2 (b; in light blue frame), incoherent interface between metal core and c-ZrO2 (c; in green frame) and incoherent interface between metal core and a-ZrO2 (d; in orange frame). e-g, Experimental semi-coherent interface structures specified by the rectangle region in (b) and (h-i) ideal model built with ideal FCC Zr metal and cubic ZrO2. e, The semi-coherent interface viewing from metal [110] direction. There is a bending of \(\sim 11^{\circ}\) between metal and interfacial layers in metal [112] direction (angle between red line and blue line), and a bending of \(\sim 8^{\circ}\) between interfacial layers and c-ZrO2 in oxide [110] direction (angle in blue line and ivory line). The coordination tripods in red and ivory boxes shows the spatial crystal orientation of metal and oxide, respectively. f, The semi-coherent interface viewing from metal [101] direction (by rotating the cutout in (e) \(120^{\circ}\) counter clockwise), showing a twisting of \(\sim 4^{\circ}\) in metal [110] direction (angle between red line and ivory line). g, One atomic plane extracted from the semi-coherent interface (the highlighted area in red in e), viewing from metal [111] direction. In this direction, the oxide shows the (002) plane. The color of the atomic bonding shows the Zr-Zr bond length. The Zr-Zr bond lengths in metal and oxide are close to \(3.3 \mathring{\mathrm{A}}\) and \(3.6 \mathring{\mathrm{A}}\) , respectively. The Zr-Zr bond lengths in the interfacial layers are longer. h, The ideal model of an interface structure between ideal FCC Zr metal and ideal cubic ZrO2, showing a \(15^{\circ}\) of wedge if no bending exists, some of the atoms cannot be bonded this way. To minimize the energy and maintain the coherency, a bending of \(15^{\circ}\) (i) is needed to release the stress. The structure changed from metal to oxide shows in (j). The oxygen atoms are </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 67, 881, 138]]<|/det|>
+colored in red. k- l, Incoherent interface structures specified by the rectangle region in (c) and (d), showing the metal/c- ZrO₂ interface (k) and metal/a- ZrO₂ interface (l). In panel e- l, the metal atoms, interfacial atoms and oxide atoms are colored in deep red, blue and ivory, respectively. The Zr atoms are bonded with their first- nearest Zr neighbors and linked with lines (Methods).  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 167, 419, 184]]<|/det|>
+## Porous structures during oxidation  
+
+<|ref|>text<|/ref|><|det|>[[115, 191, 881, 916]]<|/det|>
+Porous structures of the oxide film formed on the surface of metal are usually associated with metal corrosion<sup>11,17,27–29</sup>. We observed numbers of porous structures in the Zr- ZrO₂ particles. Fig. 4a shows a 2.4- Å- thick slice from the reconstruction volume of Zr1; in which significant number of voids, such as Zr vacancies (triangle), nano- pores (rectangle) and the largest pore (circle) are observed in the particle. From the 3D intensity and surface renderings of three consecutive atomic layers, a single Zr vacancy defect can be clearly located (Supplementary Fig. 17). To determine all the voids and evaluate their occupied volume, we employed Voronoi analysis by measuring the distance of Voronoi vertices to atoms (Methods). Fig. 4b shows the histogram of volume distribution of all the voids, which we define as Zr vacancies, nano- pores and an extremely large nanoscale pore throughout the particle. The porosity is 17% and 14% in Zr1 and Zr2, respectively. Fig. 4c and Supplementary Movie 4 show the distribution of all Zr vacancies in Zr1. No vacancy is found in the metal core. More than 110 vacancies are distributed in the particle and all the Zr vacancies contribute 8.4% of the total porosity. Slightly more vacancies are found in a- ZrO₂ than in c- ZrO₂ (Fig. 4d). We plot the density of Zr vacancies from the boundary of the metal core to the surface of the particle (Fig. 4e). Most of the vacancies are distributed in the range of 15 Å between metal core and oxide, which corresponds to the region where Zr packing density exponentially decreases (Fig. 2d). It's notable that we exclude the vacancies from calculating the Zr packing density, the lower \(\rho_{\mathrm{N}}\) of interface is independent with the rich vacancies surrounding the metal core. We found 41 nano- pores in the volume range between 125 and 4500 ų. They are mostly irregular and with a relatively large length- to- radius ratio. Fig. 4f and Supplementary Movie 4 show the distribution of all nano- pores; they mostly sit at the boundaries between c- ZrO₂ and a- ZrO₂ regions. The largest pore is more than 34000 ų and penetrates throughout the whole particle, providing possible pathway for further oxidation (Fig. 4g and Supplementary Movie 4). This pore predominantly sits in the a- ZrO₂ regions, connecting and separating all three phases; it terminates at the c- ZrO₂ region, releasing a large amount of strain. It's interesting that the Zr- Zr bonds are extremely distorted at the boundary between two c- ZrO₂ domains (Figs. 4h and 4i), where some of the Zr atoms turns to be completely amorphous to release strain. Several nano- pores are coincidentally observed at this region near the small amorphous ZrO₂ domain. These findings indicate that when the strain reaches to a certain point, possibly higher than the fracture point, the Zr- Zr crystal bonding could turn distorted and amorphous first, and then rupture to form defects to release the strain. It is generally believed that a compact layer of amorphous oxide at the micrometers scale can protect the interior of metal from further oxidation in aluminum<sup>47,48</sup>. While our results reveals that in zirconium oxide, the amorphous oxide regions are substantially more porous than those in the crystalline regions, the voids would further advance the oxidation of Zr metal. We observed variety of voids in Zr NPs at atomic scale to nanometer scale, which are highly related to the structure of incoherent interface. The rearrangements of all the atom positions including distortion, amorphization and the rupture of bonding are possibly due to the massive mass transportation during oxidation at the metal- oxide interfaces facilitated by these voids.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[120, 90, 875, 540]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 548, 884, 843]]<|/det|>
+<center>Fig. 4 | Porous structures during oxidation. a, A 2.4-Å-thick slice from the reconstructed volume of Zr1, with vacancy (triangle), nano-pores (rectangle) and the largest pore (circle) highlight. b, Volume distribution of all the voids. We define the voids with volume no larger than filling two Zr atoms (125 Å3, Methods) as vacancies, the voids with volume between 125 and 4500 Å3 as nano-pores. We consider the largest pore with volume of 34000 Å3 independently as it touches and separates all three phases. Dashed lines show the boundaries between three types of voids we define. c, The surface renderings of all vacancies in c-ZrO2 (in green) and a-ZrO2 (in blue). The outline of whole NP is plotted with gray contour. d-e, Statistics of vacancies. (d) The fractions of vacancies in c-ZrO2 and a-ZrO2. (e) The radially normalized density distribution of vacancies as a function of distance from the surface core to the surface. f, The surface renderings of all nano-pores in c-ZrO2 (in green), in a-ZrO2 (in blue) and in between c-ZrO2 and a-ZrO2 (in orange). g, The surface rendering of the largest pore. The boundary atoms composed of amorphous and crystalline atoms are colored by blue and green, respectively. h, One interface between two c-ZrO2 regions with distorted interfacial Zr-Zr bonds, amorphous region and nano-pores. The crystal and amorphous atoms distinguished by BOO analysis are colored as green and blue, respectively. The contour of the nano-pore is colored as orange. i, Three representative slices show a 7.8-Å-thick (approximately five atomic layers) cross section of the nano-pores and surrounding atoms. </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 878, 214, 893]]<|/det|>
+## Conclusion
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 66, 883, 270]]<|/det|>
+In conclusion, we determined the 3D atomic structure of metal- oxide interfaces in Zr- ZrO₂ NP for the first time using atomic resolution electron tomography. We quantitatively measured the atomic packing density and the degree of oxidation from our experimental model of metal- oxide interface. The degree of oxidation from metal to oxide increases gradually, resulting a diffuse interface between FCC Zr core and ZrO₂. The Zr metal connects with its oxide via {111} planes; and the semi- coherent interface has severe distortion including bending and twisting. The significant stress in the interface is relieved through low coordination and defects. Numbers of defects including vacancies and nano- pores together leverage the mass transportation during oxidation. We anticipate that our new findings will fulfill the dearth of 3D atomic structure of metal- oxide interface and advance the study of fundamental problems of metal- oxide interfaces such as oxidation kinetics, diffusion and defect evolution in variety of materials.  
+
+<|ref|>sub_title<|/ref|><|det|>[[116, 308, 194, 324]]<|/det|>
+## Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[116, 339, 434, 356]]<|/det|>
+## Synthesis of zirconium nanoparticles  
+
+<|ref|>text<|/ref|><|det|>[[115, 364, 882, 622]]<|/det|>
+The Zr NPs were synthesized in liquid using laser ablation methods. An all- solid- state ultraviolet laser with a wavelength of 355 nm was employed for laser ablation in ethanol, with a pulse width of 7 ps, a max pulse energy of \(18 \mu \mathrm{J}\) , a repetition frequency of \(800 \mathrm{kHz}\) , and a beam spot diameter of \(20 \mu \mathrm{m}\) . Before being placed in a clean beaker, the bulk Zr target (purity \(>99.95\%\) ) was washed by acetone (99.5%) and ethanol (99.9%). The dissolved oxygen in liquid is eliminated to the minimization by nitrogen flow for 60 min with a flow rate of 4 L/min. Then, the Zr target was submerged into ethanol and the laser beam was accurately focused vertically on the surface of the bulk Zr through the ethanol in a closed chamber. The laser ablation was continued for 1 min, and the produced Zr NPs were dispersed using ultrasonic agitation and subsequently isolated via centrifugation to be collected in the ethanol solution. The freshly- prepared Zr NPs were then placed in air for one month to obtain a naturally oxidized layer several nanometers thick. The detailed methods of synthesis are described elsewhere. The final Zr- ZrO₂ NPs were drop cast onto 7- nm- thick \(\mathrm{Si}_3\mathrm{N}_4\) membranes using atomizer for TEM experiment.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 656, 457, 673]]<|/det|>
+## Atomic-resolution electron tomography  
+
+<|ref|>text<|/ref|><|det|>[[116, 681, 882, 792]]<|/det|>
+EELS maps were collected on an aberration- corrected Thermo Fisher Scientific Spectra 300 microscope operated at \(300 \mathrm{kV}\) using a Gatan Continuum GIF with a K3 direct electron detector in the Bay Area Centre for Electron Microscopy at Songshan Lake Materials Laboratory. EDS maps were collected on an aberration- corrected Thermo Fisher Scientific Themis Z microscope at \(300 \mathrm{kV}\) with a \(50 \mathrm{pA}\) beam current and a total acquisition time of approximately 5 min in Analytical Instrumentation Center at Peking University.  
+
+<|ref|>text<|/ref|><|det|>[[116, 800, 882, 928]]<|/det|>
+Tomographic tilt series are acquired by Thermo Fisher Scientific Titan microscope with spherical aberration correction at Electron Microscopy Laboratory of Peking University. The acceleration voltage was \(300 \mathrm{kV}\) and the imaging mode was HAADF mode. The tomographic tilt series were acquired at very low dose rate \((< 5 \times 10^5 \mathrm{e} / \mathrm{\AA}^2)\) to protect the structure of oxide. For each tilt angle, three sequential images with a dwell time of \(2 \mathrm{to} 4 \mu \mathrm{s}\) were acquired and registered using normalized cross- correlation, and then the averaged to enhance the signal- to- noise ratio.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 66, 882, 196]]<|/det|>
+Acquired images were drift corrected, denoised and aligned before reconstruction. Linear drift from the sample or stage was corrected during the image registration. Block- matching and 3D filtering (BM3D) is employed to denoise the images after drift correction<sup>49</sup>. And then, the background was estimated using the discrete Laplacian function of MATLAB and subtracted. In the direction perpendicular to the tilt axis, the images were aligned by maximizing the cross- correlation between the common lines. Along the tilt axis, the images were aligned using the center- of- mass method.  
+
+<|ref|>text<|/ref|><|det|>[[115, 203, 881, 240]]<|/det|>
+After image processing, the 3D reconstruction was computed from experimental tilt series using Real Space Iterative Reconstruction (RESIRE) algorithm<sup>33</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[115, 247, 882, 432]]<|/det|>
+After reconstruction, atom tracing was performed to determine the 3D atomic coordinates. First, we interpolated reconstructed volume with spline method. All the local maxima in the reconstruction were identified as the rough atomic coordinates. Then, the coordinates were optimized according to the local volume of 1.7 Å \(\times\) 1.7 Å \(\times\) 1.7 Å with a polynomial fitting method. To separate the non- atoms from the potential atoms, K- means clustering method was employed based on the integrated intensity of the local volume (1.7 Å \(\times\) 1.7 Å \(\times\) 1.7 Å). For every potential atom, a minimum distance of 2 Å to its nearest atom should be satisfied. By carefully comparing the individual atom in the potential atomic models with the reconstructed volume, we manually corrected the atomic coordinates of unidentified or misidentified atoms (typically \(< 1\%\) ). The more detailed atom tracing procedure is described elsewhere.<sup>33</sup>  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 465, 285, 483]]<|/det|>
+## Calculation of PDF  
+
+<|ref|>text<|/ref|><|det|>[[117, 491, 679, 509]]<|/det|>
+We calculated the PDF curve from experimental 3D atomic model by  
+
+<|ref|>equation<|/ref|><|det|>[[362, 516, 634, 571]]<|/det|>
+\[g(r) = \frac{1}{N^2}\sum_{i = 1}^{N}\sum_{j = 1}^{N}\langle \delta \big(|\mathbf{r}_{ij}| - r\big)\rangle\]  
+
+<|ref|>text<|/ref|><|det|>[[115, 579, 882, 708]]<|/det|>
+where \(N\) is the total number of atoms; \(\delta\) is the Dirac delta function; \(\langle \cdot \rangle\) is the notation for expectation; \(|\mathbf{r}_{ij}|\) is the distance between the \(i\) - th atom and the \(j\) - th atom. To get a smoother PDF curve, a Gaussian kernel function with a \(\sigma\) of 1.5 Å was applied to convolute with original \(g(r)\). Finally, the PDF was scaled to approach one at the large pair distances. Using this procedure, we calculate the c- ZrO<sub>2</sub> and a- ZrO<sub>2</sub> separately in three NPs. From the PDF, we determined the first valley position as 4.5 Å, corresponding to the first- nearest- neighbor shell distance. The distance was used as a cutoff for BOO and alpha shape calculation (see the sections below).  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 742, 320, 759]]<|/det|>
+## Local BOO parameters  
+
+<|ref|>text<|/ref|><|det|>[[115, 767, 882, 916]]<|/det|>
+We calculated the normalized local BOO parameters to distinguish the order and disorder of all the atoms. The normalized BOO parameter is defined as \(\sqrt{\bar{Q}_4^2 + \bar{Q}_6^2} /\sqrt{\bar{Q}_4^2}_{\mathrm{FCC}} + \bar{Q}_6^2_{\mathrm{FCC}}\) , where the \(\bar{Q}_4\) and \(\bar{Q}_6\) values were computed based on the procedure described elsewhere, using 4.5 Å (the first- nearest- neighbor shell distance) as a constraint<sup>50</sup>. The \(\bar{Q}_4_{\mathrm{FCC}}\) and \(\bar{Q}_6_{\mathrm{FCC}}\) are the reference values of the standard FCC structure. We separated the amorphous part from
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 66, 764, 84]]<|/det|>
+crystalline part according to the criterion of the normalized BOO less than \(0.5^{34}\) .  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 118, 315, 135]]<|/det|>
+## Determination of voids  
+
+<|ref|>text<|/ref|><|det|>[[115, 143, 881, 255]]<|/det|>
+Delaunay triangulation and Voronoi tessellation were performed to determine the voids. Delaunay triangulation, Voronoi tessellation and alpha shape were performed with the built- in functions of MATLAB (namely, 'delaunayTriangulation', 'voronoin', and 'alphaShape'). The initial spatial region of NP was calculated by alpha shape with \(\alpha = 4.5 \mathrm{\AA}\) (the first- nearest- neighbor shell distance). Then, we calculated the space that accommodates at least one Zr atom in the initial particle region with the following steps:  
+
+<|ref|>text<|/ref|><|det|>[[115, 261, 798, 280]]<|/det|>
+(1) The initial particle region was divided into tetrahedra by Delaunay triangulation.  
+
+<|ref|>text<|/ref|><|det|>[[115, 287, 881, 380]]<|/det|>
+(2) We determined whether a tetrahedron is void. We calculated the radius of circumscribed sphere for each tetrahedron. The radius represents the maximized sphere that can fit within the NP without intersecting with the center of any atom. Tetrahedra with a radius larger than \(3.19 \mathrm{\AA}\) were classified as voids. This criterion of radius was obtained based on the standard cubic phase of \(\mathrm{ZrO_2}\) with one vacancy.  
+
+<|ref|>text<|/ref|><|det|>[[115, 387, 881, 461]]<|/det|>
+(3) We grouped neighboring voids together to form larger voids. Two voids that share a common face are considered neighboring and thus combined into a single, larger void. The volume of these larger voids was calculated by summing the volumes of each void of component.  
+
+<|ref|>text<|/ref|><|det|>[[115, 468, 881, 542]]<|/det|>
+(4) We classified the voids into vacancies, nano-pores and the largest pore based on volume. We define the voids with volume filling one or two Zr atoms (45-125 \(\mathrm{\AA}^3\) ) as vacancies, the voids with volume between 125 and 4500 \(\mathrm{\AA}^3\) as nano-pores. The largest pore with volume of 34000 \(\mathrm{\AA}^3\) was considered independently.  
+
+<|ref|>text<|/ref|><|det|>[[115, 550, 880, 585]]<|/det|>
+Finally, the contours of voids were displayed with Laplacian or HC smoothing conducted by GIBBON \(^{51}\) .  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 620, 304, 636]]<|/det|>
+## Atomic concentration  
+
+<|ref|>text<|/ref|><|det|>[[115, 644, 881, 756]]<|/det|>
+Topological bonds were determined based on the Voronoi tessellation. Two atoms are considered topologically bonded if their corresponding Voronoi polygons share a common face. In constructing the Voronoi polygons, we removed those surfaces with area less than \(1\%\) of the total area of the polygon surfaces \(^{52}\) . Additionally, this bond must also be shorter than \(4.5 \mathrm{\AA}\) , corresponding to the first- nearest- neighbor shell distance. The atomic concentration was calculated by \(\rho_{\mathrm{N}} = 1 / V\) , where \(V\) is the volume of Voronoi cell of an atom.  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 790, 320, 807]]<|/det|>
+## The degree of oxidation  
+
+<|ref|>text<|/ref|><|det|>[[115, 815, 883, 925]]<|/det|>
+The oxidation state was determined using Delaunay triangulation. First, the distortion of Delaunay tetrahedra was considered. The distortion parameter was calculated by \(\delta = e_{\mathrm{max}} / e_{\mathrm{avg}} - 1\) , where \(e_{\mathrm{max}}\) and \(e_{\mathrm{avg}}\) are the maximum and average edge lengths of tetrahedron \(^{33}\) . We removed the tetrahedron with a distortion parameter larger than 0.255. Then, the volume of
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[116, 66, 881, 140]]<|/det|>
+remaining Delaunay tetrahedra was calculated. If the volume of a tetrahedron is larger than 4.68 \(\mathring{\mathrm{A}}^3\) (the averaged tetrahedron volume of FCC Zr lattice and c- ZrO2 lattice), an oxygen atom was placed inside. Finally, the degree of oxidation for each Zr atom was calculated by the fraction of its surrounding tetrahedra that accommodate one oxygen atom.  
+
+<|ref|>text<|/ref|><|det|>[[116, 172, 881, 302]]<|/det|>
+Acknowledgments: We thank the support of High- performance Computing Platform of Peking University. We thank the Electron Microscopy Laboratory at Peking University, Bay Area Centre for Electron Microscopy at Songshan Lake Materials Laboratory and Analytical Instrumentation Center at Peking University for the use of the aberration- corrected electron microscope. This work was supported by the National Natural Science Foundation of China (Grant No. 22172003, 52071222) and Guangdong Major Project of Basic and Applied Basic Research, China (Grant No. 2019B030302010).  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 311, 308, 327]]<|/det|>
+## Author contributions:  
+
+<|ref|>text<|/ref|><|det|>[[147, 336, 881, 446]]<|/det|>
+J.Z. conceived the idea and directed the study. Z.L., Z.X. and Y.Z. performed TEM experiment and acquired the data. Y.Z. and S.H. performed the imaging processing, reconstructions, and atom tracing. Y.Z., Z.L., S.H. conducted/discussed data analysis under the direction of J.Z., T.X. and Y.- E. Z. synthesized Zr NPs under the direction of H.- B. K. and W.- H. W. Y.Z., Z.L. and J.Z. wrote the manuscript. All authors commented on the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[116, 455, 655, 472]]<|/det|>
+Competing interests: The authors declare no competing interests.  
+
+<|ref|>text<|/ref|><|det|>[[116, 480, 649, 498]]<|/det|>
+Data availability: All data are available upon reasonable request.  
+
+<|ref|>sub_title<|/ref|><|det|>[[116, 532, 205, 548]]<|/det|>
+## Reference  
+
+<|ref|>text<|/ref|><|det|>[[113, 551, 884, 920]]<|/det|>
+1. Lee, S. W. et al. Controlling hot electron flux and catalytic selectivity with nanoscale metal-oxide interfaces. Nat. Commun. 12, 40 (2021).  
+2. Han, Z., Yang, F. & Li, Y. Dynamics of metal-support interface revealed by environmental transmission electron microscopy. Matter 5, 2531–2533 (2022).  
+3. Lowe, J. S. & Siegel, D. J. Modeling the interface between lithium metal and its native oxide. ACS Appl. Mater. Interfaces 12, 46015–46026 (2020).  
+4. Li, Q. et al. Extra storage capacity in transition metal oxide lithium-ion batteries revealed by in situ magnetometry. Nat. Mater. 20, 76–83 (2021).  
+5. Foerster, B. et al. Interfacial states cause equal decay of plasmons and hot electrons at gold–metal oxide interfaces. Nano Lett. 20, 3338–3343 (2020).  
+6. Tang, W. et al. The roles of metal oxidation states in perovskite semiconductors. Matter 6, 3782–3802 (2023).  
+7. Smialek, J. L. Invited review paper in commemoration of over 50 years of oxidation of metals: alumina scale adhesion mechanisms: a retrospective assessment. Oxid. Met. 97, 1–50 (2022).  
+8. Hu, J. et al. Microstructure and oxidation behavior of the Y/Ta/Hf co-doped AlCoCrFeNi high-entropy alloys in air at 1100 °C. Corros. Sci. 212, 110930 (2023).  
+9. Backman, L., Gild, J., Luo, J. & Opila, E. J. Part I: Theoretical predictions of preferential oxidation in refractory high entropy materials. Acta Mater. 197, 20–27 (2020).  
+10. Jeurgens, L. P. H., Lyapin, A. & Mittemeijer, E. J. The mechanism of low-temperature oxidation of zirconium. Acta Mater. 53, 4871–4879 (2005).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 68, 884, 111]]<|/det|>
+11. Birks, N., Meier, G. H. & Pettit, F. S. Introduction to the High-Temperature Oxidation of Metals, 2nd Edition. (Cambridge University Press, 2006).  
+
+<|ref|>text<|/ref|><|det|>[[115, 104, 881, 140]]<|/det|>
+12. Sun, X. et al. Dislocation-induced stop-and-go kinetics of interfacial transformations. Nature 607, 708-713 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[115, 141, 880, 177]]<|/det|>
+13. Zhu, L. et al. Visualizing anisotropic oxygen diffusion in ceria under activated conditions. Phys. Rev. Lett. 124, 056002 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[115, 178, 880, 213]]<|/det|>
+14. Morin, F., Beranger, G. & Lacombe, P. Limits of application for Wagner's oxidation theory. Oxid. Met. 4, 51-62 (1972).  
+
+<|ref|>text<|/ref|><|det|>[[115, 215, 881, 250]]<|/det|>
+15. Railsback, J. G., Johnston-Peck, A. C., Wang, J. & Tracy, J. B. Size-dependent nanoscale Kirkendall effect during the oxidation of nickel nanoparticles. ACS Nano 4, 1913-1920 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[115, 252, 881, 288]]<|/det|>
+16. You, R. et al. Revealing temperature-dependent oxidation dynamics of Ni nanoparticles via ambient pressure transmission electron microscopy. Nano Lett. 23, 7260-7266 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 289, 881, 325]]<|/det|>
+17. Vijayakumar, J. et al. Absence of a pressure gap and atomistic mechanism of the oxidation of pure Co nanoparticles. Nat. Commun. 14, 174 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 326, 881, 362]]<|/det|>
+18. Gesmundo, F. & Viani, F. Application of Wagner's theory to the parabolic growth of oxides containing different kinds of defects: I. pure oxides. J. Electrochem. Soc. 128, 460-469 (1981).  
+
+<|ref|>text<|/ref|><|det|>[[115, 363, 881, 399]]<|/det|>
+19. Neyman, K. M. & Kozlov, S. M. Quantifying interactions on interfaces between metal particles and oxide supports in catalytic nanomaterials. NPG Asia Mater. 14, 59 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[115, 400, 881, 436]]<|/det|>
+20. Tea, E., Huang, J., Li, G. & Hin, C. Atomic bonding and electrical potential at metal/oxide interfaces, a first principle study. J. Chem. Phys. 146, 124706 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[115, 437, 881, 473]]<|/det|>
+21. Zou, S. et al. Grafting nanometer metal/oxide interface towards enhanced low-temperature acetylene semi-hydrogenation. Nat. Commun. 12, 5770 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 474, 880, 510]]<|/det|>
+22. Pacchioni, G. Electronic interactions and charge transfers of metal atoms and clusters on oxide surfaces. Phys. Chem. Chem. Phys. 15, 1737 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[115, 511, 881, 547]]<|/det|>
+23. Wilson, R. B., Apgar, B. A., Hsieh, W.-P., Martin, L. W. & Cahill, D. G. Thermal conductance of strongly bonded metal-oxide interfaces. Phys. Rev. B 91, 115414 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[115, 548, 881, 584]]<|/det|>
+24. Greiner, M. T., Chai, L., Helander, M. G., Tang, W. & Lu, Z. Metal/metal-oxide interfaces: How metal contacts affect the work function and band structure of MoO3. Adv. Funct. Mater. 23, 215-226 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[115, 585, 881, 621]]<|/det|>
+25. Nian, Y. et al. Atomic-scale dynamic interaction of H2O molecules with Cu surface. Phys. Rev. Lett. 125, 156101 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[115, 622, 881, 658]]<|/det|>
+26. Dong, Z., Ren, Q., Wang, S., Zhang, L. & Luo, L. Atomic-scale interfacial phase transformation governed Cu oxidation in water vapor. J. Phys. Chem. Lett. 12, 6996-7001 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 659, 861, 676]]<|/det|>
+27. Zhu, Q. et al. Defect-driven selective metal oxidation at atomic scale. Nat. Commun. 12, 558 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 678, 881, 713]]<|/det|>
+28. Sainju, R. et al. In situ studies of single-nanoparticle-level nickel thermal oxidation: from early oxide nucleation to diffusion-balanced oxide thickening. ACS Nano 16, 6468-6479 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[115, 715, 881, 750]]<|/det|>
+29. Xia, W. et al. Bimetallic nanoparticle oxidation in three dimensions by chemically sensitive electron tomography and in situ transmission electron microscopy. ACS Nano 12, 7866-7874 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[115, 751, 881, 787]]<|/det|>
+30. Ni, N. et al. How the crystallography and nanoscale chemistry of the metal/oxide interface develops during the aqueous oxidation of zirconium cladding alloys. Acta Mater. 60, 7132-7149 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[115, 789, 881, 824]]<|/det|>
+31. Dong, Y., Motta, A. T. & Marquis, E. A. Atom probe tomography study of alloying element distributions in Zr alloys and their oxides. J. Nucl. Mater. 442, 270-281 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[115, 825, 881, 860]]<|/det|>
+32. Zhou, J. et al. Observing crystal nucleation in four dimensions using atomic electron tomography. Nature 570, 500-503 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[115, 862, 880, 897]]<|/det|>
+33. Yuan, Y. et al. Three-dimensional atomic packing in amorphous solids with liquid-like structure. Nat. Mater. 21, 95-102 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[115, 900, 881, 917]]<|/det|>
+34. Yang, Y. et al. Determining the three-dimensional atomic structure of an amorphous solid. Nature 592,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[145, 68, 250, 83]]<|/det|>
+60- 64 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 87, 883, 122]]<|/det|>
+35. Li, Z. et al. Probing the atomically diffuse interfaces in Pd@Pt core-shell nanoparticles in three dimensions. Nat. Commun. 14, 2934 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 124, 881, 159]]<|/det|>
+36. Ishikawa, R., Shibata, N., Taniguchi, T. & Ikuhara, Y. Three-dimensional imaging of a single dopant in a crystal. Phys. Rev. Applied 13, 034064 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[115, 161, 881, 196]]<|/det|>
+37. Ishikawa, R. et al. High spatiotemporal-resolution imaging in the scanning transmission electron microscope. Microscopy 69, 240-247 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[115, 198, 881, 250]]<|/det|>
+38. Liang, S.-X., Zhang, L.-C., Reichenberger, S. & Barcikowski, S. Design and perspective of amorphous metal nanoparticles from laser synthesis and processing. Phys. Chem. Chem. Phys. 23, 11121-11154 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 252, 881, 305]]<|/det|>
+39. Liu, Y. et al. Fabrication of amorphous PdNiCuP nanoparticles as efficient bifunctional and highly durable electrocatalyst for methanol and formic acid oxidation. J. Mater. Sci. Technol. 122, 148-155 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[115, 307, 881, 342]]<|/det|>
+40. Zhang, F. et al. In situ study of the crystallization from amorphous to cubic zirconium oxide: riveted and reverse Monte Carlo analyses. Chem. Mater. 19, 3118-3126 (2007).  
+
+<|ref|>text<|/ref|><|det|>[[115, 344, 880, 380]]<|/det|>
+41. Larsen, P. M., Schmidt, S. & Schiotz, J. Robust structural identification via polyhedral template matching. Modelling Simul. Mater. Sci. Eng. 24, 055007 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[115, 381, 880, 417]]<|/det|>
+42. Nicholls, R. J. et al. Crystal structure of the ZrO phase at zirconium/zirconium oxide interfaces. Adv. Eng. Mater. 17, 211-215 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[115, 418, 880, 454]]<|/det|>
+43. Dholabhai, P. P., Martinez, E., Brown, N. T. & Uberuaga, B. P. On the mobility of carriers at semiconductor oxide heterointerfaces. Phys. Chem. Chem. Phys. 19, 23122-23130 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[115, 456, 880, 491]]<|/det|>
+44. Pilania, G., Dholabhai, P. P. & Uberuaga, B. P. Role of symmetry, geometry, and termination chemistry on misfit dislocation patterns at semicoherent heterointerfaces. Matter 2, 1324-1337 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[115, 493, 880, 528]]<|/det|>
+45. Shao, S., Wang, J. & Misra, A. Energy minimization mechanisms of semi-coherent interfaces. J. Appl. Phys. 116, 023508 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[115, 530, 881, 583]]<|/det|>
+46. Sun, Z. et al. Janus Icosahedral Particles: Amorphization Driven by Three-Dimensional Atomic Misfit and Edge Dislocation Compensation. http://arxiv.org/abs/2310.16474 (2023) doi:10.21203/rs.3.rs-3439840/v1.  
+
+<|ref|>text<|/ref|><|det|>[[115, 586, 881, 620]]<|/det|>
+47. McGowan, J. C. Estimation of the approximate ionization constants of inorganic bases. Nature 168, 601-602 (1951).  
+
+<|ref|>text<|/ref|><|det|>[[115, 622, 880, 657]]<|/det|>
+48. Nguyen, L. et al. Atomic-scale insights into the oxidation of aluminum. ACS Appl. Mater. Interfaces 10, 2230-2235 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[115, 660, 881, 712]]<|/det|>
+49. Makinen, Y., Azzari, L. & Foi, A. Collaborative filtering of correlated noise: Exact transform-domain variance for improved shrinkage and patch matching. IEEE Trans. on Image Process. 29, 8339-8354 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[115, 715, 881, 750]]<|/det|>
+50. Lechner, W. & Dellago, C. Accurate determination of crystal structures based on averaged local bond order parameters. J. Chem. Phys. 129, 114707 (2008).  
+
+<|ref|>text<|/ref|><|det|>[[115, 752, 880, 768]]<|/det|>
+51. M Moerman, K. GIBBON: The geometry and image-based bioengineering add-on. JOSS 3, 506 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[115, 770, 880, 806]]<|/det|>
+52. Sheng, H. W., Luo, W. K., Alamgir, F. M., Bai, J. M. & Ma, E. Atomic packing and short-to-medium-range order in metallic glasses. Nature 439, 419-425 (2006).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[42, 43, 312, 71]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[42, 93, 768, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[58, 130, 675, 258]]<|/det|>
+- ZrZrO2interfaceNatureSupplementaryInformation20240219final.pdf- MovieS1.mp4- MovieS2.mp4- MovieS3.mp4- MovieS4.mp4
+
+<--- Page Split --->

preprint/preprint__2bbfafaeea0e549a06406df3ec0de2092667f753880d5d03f95c7b3aace942cb/images_list.json

@@ -0,0 +1,62 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1: Fabrication process flow of the all-fibre heterogeneously-integrated parametric mixer. (a) Tapered SCF with core/cladding diameters of about \\(5\\mu \\mathrm{m} / 125\\mu \\mathrm{m}\\) . (b) Heating and tapering process to make the first void gap in the fibre core. (c) Heating and tapering to make the second void gap in the fibre core. (d) fibre tapering process to scale down core size and collapse the void gap to form the nano-spike coupler. (e) Cleave at the center of the void to remove one side of the taper. (f) Splice the SCF nanospike to a tapered SSMF. (g) Employ a polymer tube to mechanically support the SSMF-SCF connection, keeping the fibre straight and tapering the other end of the SCF. (h) Cleave the other end and (i) splice to another tapered SSMF.",
+    "footnote": [],
+    "bbox": [
+      [
+        113,
+        288,
+        882,
+        586
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2: Fully integrated SCF-based parametric comb generation scheme. An electro-optic frequency comb is generated to act as a seed source for the parametric mixing stage. A pulse re-shaping stage compresses the comb pulse train to maximise the pulse peak power and enhance the parametric mixing efficiency. The optical output is then amplified and launched into a 17 mm sample of fully-integrated SCF, where parametric nonlinearities cause comb broadening. Both ends of the SCF are tapered and spliced to tapered single-mode fibre, with nano-spike couplers to facilitate coupling between the heterogenous fibre cores.",
+    "footnote": [],
+    "bbox": [
+      [
+        123,
+        240,
+        875,
+        592
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3: C-band parametric frequency comb generation using SCF. (a) An electro-optic frequency comb with 26 GHz line spacing (orange) was temporally reshaped and launched into a 17mm sample of SCF at an average power of 32 dBm. The optical spectrum of the broadened parametric comb (blue) was obtained at a resolution of 0.02 nm, achieving 143 tones across a 30.0 nm bandwidth. A 10 MHz resolution BOSA was used to obtain close-in traces at different points within the comb bandwidth (Fig.(b)-(d)). The dashed line shows the instrument noise floor. Results in (c) and (d) are both limited by the instrument noise floor. (e) The linewidth of the parametric comb was measured using the delayed self-heterodyne interferometer (DSHI) method with a 80-km delay line and pseudo-Voigt profile fitting. The pseudo-Voigt fitting curves are shown for the two measured beat notes at the extremities of the SCF comb bandwidth ((f) and (g)).",
+    "footnote": [],
+    "bbox": [
+      [
+        194,
+        140,
+        800,
+        710
+      ]
+    ],
+    "page_idx": 13
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4: Parametric frequency comb simulation including the effects of two-photon absorption and free carriers within the SCF. (a) Nonlinear loss induced by TPA and FCs as a function of average input power. (b) Free carrier density in the silicon core versus time (black), and the optical pulse train power versus time (red). Subsequent pulses in the 26 GHz pulse train increase the free carrier density before complete recombination can occur, reaching a steady state over several nanoseconds. (c) Simulated optical spectrum at the output of the SCF mixer, at 27 dBm and 32 dBm input power (top and middle respectively) and (d) at 32 dBm without free carriers. TPA is included in all three spectra, and insertion loss is neglected.",
+    "footnote": [],
+    "bbox": [
+      [
+        192,
+        188,
+        804,
+        690
+      ]
+    ],
+    "page_idx": 14
+  }
+]

preprint/preprint__2bbfafaeea0e549a06406df3ec0de2092667f753880d5d03f95c7b3aace942cb/preprint__2bbfafaeea0e549a06406df3ec0de2092667f753880d5d03f95c7b3aace942cb.mmd

@@ -0,0 +1,268 @@
+
+# All-fibre heterogeneously-integrated frequency comb generation using silicon core fibre  
+
+Ronit Sohanpal UCL  
+
+Haonan Ren Dalian University of Technology  
+
+Li Shen Huazhong University of Science and Technology  
+
+Callum Deakin UCL  
+
+Alexander Heidt University of Bern  
+
+Thomas Hawkins Clemson University  
+
+John Ballato Clemson University https://orcid.org/0000- 0001- 5910- 3504  
+
+Ursula Gibson Clemson University  
+
+Anna Peacock University of Southampton https://orcid.org/0000- 0002- 1940- 7172  
+
+Zhixin Liu ( zhixin.liu@ucl.ac.uk ) University College London https://orcid.org/0000- 0002- 9681- 7933  
+
+## Article  
+
+Keywords: optical frequency combs, optical signal- to- noise ratio, silicon core fiber  
+
+Posted Date: November 11th, 2021  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 968401/v1  
+
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License
+
+<--- Page Split --->
+
+Version of Record: A version of this preprint was published at Nature Communications on July 9th, 2022. See the published version at https://doi.org/10.1038/s41467-022-31637-1.
+
+<--- Page Split --->
+
+# All-fibre heterogeneously-integrated frequency comb generation using silicon core fibre  
+
+Ronit Sohanpal \(^{1\otimes}\) , Haonan Ren \(^{2,3}\) , Li Shen \(^{4\otimes}\) , Callum Deakin \(^{1}\) , Alexander M. Heidt \(^{5}\) , Thomas W. Hawkins \(^{6}\) , John Ballato \(^{6}\) , Ursula J. Gibson \(^{6}\) , Anna C. Peacock \(^{2\otimes}\) , and Zhixin Liu \(^{1\otimes}\)  
+
+\(^{1}\) Department of Electronic and Electrical Engineering, UCL (University College London), London, United Kingdom  
+
+\(^{2}\) Optoelectronics Research Centre, University of Southampton, Southampton, United Kingdom  
+
+\(^{3}\) now with School of Optoelectronic Engineering and Instrumentation Science, Dalian University of Technology, Dalian, China  
+
+\(^{4}\) Wuhan National Laboratory for Optoelectronics and School of Optical and Electrical Information, Huazhong University of Science and Technology, China  
+
+\(^{5}\) Institute of Applied Physics, University of Bern, Switzerland \(^{6}\) Center for Optical Materials Science and Engineering Technologies, and the Department of Materials Science and Engineering, Clemson University, USA  
+
+\(^{1}\) ronit.sohanpal.14@ucl.ac.uk \(^{4}\) lishen@hust.edu.cn \(^{2}\) acp@orc.soton.ac.uk \(^{1}\) zhixin.liu@ucl.ac.uk  
+
+## Abstract  
+
+Originally developed for metrology, optical frequency combs are becoming increasingly pervasive in a wider range of research topics including optical communications, spectroscopy, and radio or microwave signal processing. However, application demands in these fields can be more challenging as they require compact sources with a high tolerance to temperature variations that are capable of delivering flat comb spectra, high power per tone, narrow linewidth and high optical signal- to- noise ratio (OSNR). To date, there has not been a frequency comb technology that is able to simultaneously achieve all these requirements. This work reports the generation of a flat, high power frequency comb in the telecom band using a 17- mm fully- integrated silicon core fibre (SCF) as a parametric mixer. Our all- fibre, cavity- free source combines the materials benefits of planar waveguide structures with the advantageous properties of fibre platforms to achieve a 30 nm bandwidth comb source containing 143 tones with \(< 3 \mathrm{kHz}\) linewidth, 12 dB flatness, and \(>30 \mathrm{dB}\) OSNR over the entire spectral region. The unique combination of technical features offered by this SCF- based source opens a path towards a new class of high- performance frequency comb generators for communications and signal processing applications.  
+
+## 1 Introduction  
+
+Optical frequency combs (OFCs) generate precisely- spaced phase- coherent spectral tones and have revolutionized frequency control and metrology [1]. Over the past decade, numerous new developments have been focused on comb applications in optical communications, spectroscopy, microwave photonics, and optical- assisted signal processing [2, 3, 4, 5]. Unlike metrology, in which broad 'over- one- octave' spectral bandwidths are the predominate requirement to enable self- referencing, communications and signal processing applications typically demand only tens of nanometers of bandwidth, but require high power, high
+
+<--- Page Split --->
+
+optical signal to noise ratios (OSNR), a flat spectral response and narrow comb linewidths. In addition, compactness and temperature stability are essential prerequisites for most practical applications.  
+
+Significant efforts have been made to develop OFCs using both integrated photonic waveguides and optical fibre platforms [6, 7, 8, 9]. Much of the work using integrated platforms has focused on Kerr soliton comb formation using dispersion- engineered microcavities, which have demonstrated comb spectra spanning more than \(100\mathrm{nm}\) [10, 11]. Although Kerr combs have previously been employed to demonstrate C and L band coherent communications [12], their hyperbolic secant ( \(\mathrm{sech}^2\) ) spectral power envelope means the comb bandwidths must significantly exceed the standard C+L telecom bands to avoid having to use the low power wings, limiting the maximum output power- per- tone and the OSNR. Dark pulse Kerr combs with narrower spectral bandwidths ( \(\sim 35\mathrm{nm}\) ) can offer some improvement to the power and OSNR [13], but the need for a high quality factor cavity implies an inherent sensitivity to temperature variation. Of note is that temperature sensitivity is also an issue with other cavity- based comb sources, such as those generated using integrated mode- locked lasers [14].  
+
+In contrast, OFCs generated via cavity- free configurations offer more flexibility in terms of spectral shaping and are in general much more robust to environmental perturbations. Such OFCs can be realized by using cascaded modulators or by pumping a highly nonlinear waveguide using high peak- power pulses, or a combination of both approaches [15, 16]. For example, planar waveguides based on chalcogenide glasses and III- V semiconductor materials (e.g. AlGaAs) possess large nonlinearities and small dimensions, making them an ideal choice for compact parametric mixing stages [17, 18]. However, such high index planar structures suffer from restrictive coupling requirements when compared to fibre systems as a result of their small dimensions and rectangular cross- section [19].  
+
+In comparison, cavity- free OFCs constructed using fibre- optic platforms as the nonlinear media have the advantage of low losses and high power handling, thus significantly increasing the power- per- tone and OSNR. By integrating fibre amplifiers, fibre pulse compressors, highly nonlinear fibre (HNLF)- based saturable absorbers and parametric mixers, all- fibre OFCs have been demonstrated with bandwidths over \(100\mathrm{nm}\) [20], \(\sim 0\mathrm{dBm}\) per tone and a spectral flatness of \(3\mathrm{dB}\) [8], offering superior performance for data transmission [21, 22] and microwave signal processing [23], as well as being immediately compatible with much of the existing fibre infrastructure. However, current fibre- based OFCs are bulky due to the hundreds of meters of HNLF required to achieve significant parametric gain. Moreover, efficient parametric mixing requires careful dispersion- management over the entire HNLF length, and specifically a stable zero dispersion wavelength [24], which increases fabrication difficulties and cost.  
+
+A long- standing goal has been to combine the materials benefits of the integrated planar waveguide structures with the advantageous waveguiding properties of fibre platforms to achieve all of the desired technical features simultaneously. The emerging class of silicon core fibres (SCF) offer a promising solution to this objective, in which a crystalline silicon core material is embedded within a conventional silica glass cladding. Compared to fibres with glassy core materials, the SCF platform offers a significantly enhanced nonlinear coefficient \(\gamma\) due to the combined effects of the high nonlinear refractive index \(n_2\) of silicon and the high core- cladding index contrast. For example, SCFs with core diameters of only a few \(\mu \mathrm{m}\) can have an effective nonlinear coefficient of more than three orders of magnitude higher than that of commercially available Ge- doped HNLFs, which enables the reduction of the parametric mixer length from hundreds of meters to a few millimeters [25]. Such compact SCFs have already shown great promise for efficient nonlinear processing of signals across the extended telecoms band [26]. However, to date, the complete integration of SCFs with silica fibres has been a significant hurdle and, thus far, reported SCF results have been limited to either free space coupling [27] or partial fibre integration [28], which makes the coupling of dense multi- wavelength systems more challenging.  
+
+In this work, we present the first all- fibre integrated SCF parametric mixer for OFC generation. Specifically, the SCF has been processed to allow for direct splicing to standard single- mode fibre (SSMF) connectors at both ends, allowing for straightforward integration with conventional fibre systems. The resulting comb structure has a bandwidth of \(30\mathrm{nm}\) , with a flatness of \(12\mathrm{dB}\) , \(>30\mathrm{dB}\) OSNR and \(< 3\mathrm{kHz}\) linewidth, properties that are desirable for many practical applications in areas such as phase- coherent communications and micrometer/millimeter wave generation [29]. Moreover, our cavity- free design allows for tunable wavelength and comb spacing, thus enabling high resolution dual- comb spectroscopy [30] as well as dual- comb RF processing [31].
+
+<--- Page Split --->
+
+## 2 Results  
+
+### 2.1 \(\mathrm{SiO_2}\) -Si core fibre integration  
+
+The SCFs used in this work were fabricated via the molten core drawing (MCD) technique [32]. The MCD method is the most practical of the SCF production approaches as it can rapidly produce long lengths of fibre that are compatible with traditional fibre post- processing procedures. The as- drawn MCD SCFs possess a polycrystalline silicon core with typical core/cladding diameters of approximately \(12\mu \mathrm{m} / 125\mu \mathrm{m}\) , as detailed in the Methods. To transform the SCFs into low loss nonlinear parametric mixers that are robust and user- friendly, a multi- step tapering and splicing approach is employed. These post- processing steps are important to enhance the efficiency of the nonlinear processing in the SCFs for a number of key reasons. First, the tapering process is used both to reduce the core size as well as to improve the crystalline quality, and hence, the optical transmission [33]. This results in a SCF with core/cladding diameters of approximately \(5\mu \mathrm{m} / 125\mu \mathrm{m}\) and a transmission loss of \(3\mathrm{dB / cm}\) . Second, a modified taper method is applied to further reduce the core size and fabricate nano- spike couplers on the end facets of the fibre. The role of the couplers is to better match the modes of the SCFs with those of the standard single mode fibres (SSMFs), as well as to suppress reflections at the SCF/SSMF interface when splicing.  
+
+The full process to fabricate the integrated SCF mixer is illustrated in Fig. 1. In the first step, the SCF is heated gently while a small tension is applied along the fibre axis. Owing to the tensile stress in the as- drawn SCF, as the heated core cools and recrystallizes, a void- gap can form around the heat zone (Fig. 1b). Repeating this process produces another void- gap at the other end of the fibre (Fig. 1c). Subsequently, a single sweep tapering process is used to reduce the local core/cladding ratio from \(5 / 125\mu \mathrm{m}\) to \(1.1 / 27\mu \mathrm{m}\) over the left- hand side of the fibre, during which the first void- gap collapses to form a nano- spike with a length of \(\sim 200\mu \mathrm{m}\) at the core facet (Fig. 1d). The SCF is then precisely cleaved in the core- less region and spliced with a pre- prepared tapered SSMF with same cladding diameter (Fig. 1e- f). To allow for a second connection to be made at the other end, we enclose the spliced SSMF- SCF section into a polymer capillary that mechanically supports the nano- spike coupler before applying the same procedure to the right- hand side of the fibre (Fig. 1g- i). The result is a fully integrated SCF fibre device with a total length of \(\sim 17\mathrm{mm}\) , which contains two small core sections of \(1.1\mu \mathrm{m}\) diameter at each end (lengths of \(5.1\mathrm{mm}\) and \(9.1\mathrm{mm}\) at the input and output, respectively), connected by a short \(3\mathrm{mm}\) length with a \(5\mu \mathrm{m}\) diameter in the middle. Importantly, thanks to the silica cladding of the SCFs, all of these processing steps can be conducted using a single glass processing system, enabling high yield production of these heterogeneous fibre devices. The loss in the processed SCF section was determined to be \(\sim 2\mathrm{dB / cm}\) with an effective nonlinear coefficient \((\gamma)\) of approximately \(30\mathrm{W}^{- 1}\mathrm{m}^{- 1}\) . From the transmission losses, the coupling losses for the sample used in this work was estimated to be \(\sim 8\mathrm{dB}\) per facet, resulting in an end- to- end loss of \(\sim 19\mathrm{dB}\) . This fairly substantial insertion loss can be attributed to the losses associated with splicing the tapered fibres ( \(27\mu \mathrm{m}\) diameter cladding) and the mode mismatch between the tapered SSMF and the nanospike. Simulations indicate that these losses could be reduced to \(2\mathrm{dB}\) by optimizing the cladding diameters at the connection point ( \(10\mathrm{um}\) ) to reduce the mismatch [34], though these diameters are currently smaller than what our processing system can handle.  
+
+### 2.2 SCF-based comb generation  
+
+The SCF- based parametric comb generator consists of three main sections (Fig. 2): an electro- optic comb, a fibre pulse compressor, and the all- fibre- integrated SCF device as a parametric mixer. To achieve narrow linewidths, we employed a continuous wave (CW) fibre laser with \(\sim 1.6\mathrm{kHz}\) linewidth as the seed to which a phase and intensity modulation were applied to create an \(11.8\mathrm{nm}\) - wide electro- optic (EO) comb (56 tones with 26- GHz spacing), shown in orange in Fig. 3a. In the time domain, this corresponds to a pulse train with a repetition rate of \(26\mathrm{GHz}\) , with each pulse exhibiting a quasi- linear frequency chirp [35]. To optimise the parametric mixing efficiency, we compressed the pulses linearly using a short length of SSMF to eliminate the frequency chirp across the center of the pulse, resulting in a full- width half- maximum (FWHM) pulse width of \(610\mathrm{fs}\) . As obtaining an OFC with good spectral flatness is a key aim of this work, we implemented a nonlinear optical loop mirror (NOLM) to suppress the low- power pedestals that result from the dispersive chirp compression, which can cause significant spectral rippling after the mixing stage and reduce the mixing efficiency [8]. This also reduces the pulse FWHM down to approximately \(440\mathrm{fs}\) before amplification to an average power of \(32\mathrm{dBm}\) using a dispersion- flattened short- pulse fibre amplifier. The amplifier introduces no observable broadening of the seed pulses. The amplified pulses were subsequently launched into the \(17\mathrm{mm}\) all- fibre- integrated SCF device.
+
+<--- Page Split --->
+
+The blue curve in Fig. 3a shows the spectrum of the SCF mixer output, indicating 143 tones within a bandwidth of \(30~\mathrm{nm}\) with a flatness of \(12\mathrm{~dB}\) , from \(1535\mathrm{~nm}\) to \(1565\mathrm{~nm}\) . For the estimation of the OSNR we measure the close- up spectra at 1545, 1555, and \(1563\mathrm{~nm}\) in Fig. 3b- d with \(10\mathrm{MHz}\) resolution. Considering \(0.1\mathrm{~nm}\) noise bandwidth, the OSNR at the center ( \(1550\mathrm{~nm}\) ) and long wavelength edge ( \(1563\mathrm{~nm}\) ) of the SCF comb is greater than \(35\mathrm{~dB}\) and the OSNR at the short wavelength edge ( \(1536\mathrm{~nm}\) ) is approximately \(30\mathrm{~dB}\) . The increased noise in the \(1535\mathrm{- }1545\mathrm{~nm}\) region is primarily attributed to the amplified spontaneous emission (ASE) noise generated by the Er/Yb- doped fibre amplifier before the parametric amplification stage. This could be eliminated using a bandpass filter to ensure a high OSNR across the whole comb bandwidth. Since the nonlinear spectral broadening of the comb in the SCF is dominated by self- phase modulation, the asymmetric temporal pulse shape after the NOLM leads to the observed asymmetric expansion of the comb towards shorter wavelengths (Supplementary Video 1).  
+
+Fig. 3e shows the measured linewidth of the SCF comb tones (triangle marker), and the seed EO comb (circle marker), measured using the delayed self- heterodyne interferometer method [36] for every second comb line. Typically, the EO comb linewidth increases linearly as a function of the absolute comb tone index \(|n|\) due to the scaling noise contribution of the RF- induced phase noise [37]. Our comb design uses an ultra- low phase noise RF signal generator ( \(< 7\) fs integrated jitter from \(100\mathrm{~Hz}\) to \(100\mathrm{~MHz}\) ), which results in a negligible scaling of the EO comb bandwidth with the tone index (orange circles in Fig. 3e).  
+
+After the SCF, the EO comb linewidth increases to \(1.9\mathrm{- }2.4\mathrm{kHz}\) , which we attribute to the Gordon- Mollenauer effect (nonlinearity- induced amplitude- to- phase noise conversion) [38]. Numerical analysis of our system shows that the spectral coherence of the SCF comb degrades as the amplifier noise figure increases, which results in the increased comb linewidth observed on the short- wavelength edge (Supplementary Figure 3). Nevertheless, it is clear that the SCF comb retains a well- preserved linewidth performance across the whole spectrum. Fig. 3f and Fig. 3g show the measured beat note and their fitting using a pseudo- Voigt profile at both extremities of the comb for the linewidth characterization. We use a pseudo- Voigt profile to account for the fact that the frequency noise of the SCF tones is a mixed contribution of \(1 / \mathrm{f}\) (flicker) frequency noise and white frequency noise from the fibre laser [39, 40].  
+
+To study the role of the integrated SCF device losses on the spectral bandwidth of the SCF- based comb, we simulate the comb generation system to mirror the experimental setup. The numerical simulation uses a modified generalized nonlinear Schrödinger equation (GNLSE) that incorporates two photon absorption (TPA), free carrier absorption (FCA) and free carrier dispersion (FCD) [41] (see Methods). We begin by evaluating the nonlinear loss due to TPA and FCA for different average launch powers within the SCF device used in our system. In these investigations, we assume zero insertion loss so that we can gauge the full potential of our approach. Fig. 4a shows that the nonlinear loss becomes greater than \(1\mathrm{~dB}\) when the average power increases to \(>17\mathrm{~dBm}\) and increases exponentially thereafter. Compared to previous studies of SCF- based nonlinear signal processing [26], our 26- GHz- repetition- rate frequency comb has a significantly shorter interval between the pulses than the free carrier lifetime ( \(38\mathrm{~ps}\) versus \(\sim 1\mathrm{~ns}\) , respectively). As a result, there is only a partial recombination of free carriers, leading to an accumulation of the free carrier density before reaching a steady state after about \(2\mathrm{~ns}\) , as shown in Fig. 4b. The reduction in the pulse train power due to FCA is shown as the red lines in Fig. 4b.  
+
+As expected, TPA, FCA and FCD have negative effects for wide- band comb generation and result in a reduced parametric gain, limiting the comb bandwidth at high average pump powers. As shown in Fig. 4c, increasing the average input power from \(27\mathrm{~dBm}\) to \(32\mathrm{~dBm}\) results in a negligible change of the comb spectral bandwidth due to the increase in nonlinear absorption. Moreover, the results for \(27\mathrm{~dBm}\) of pump power have produced a \(30\mathrm{~nm}\) comb bandwidth with a spectral flatness of \(13\mathrm{~dB}\) at the telecom C- band, similar to our measured results. Thus these findings suggest that the main limitation to the comb bandwidth is the free carrier effects, rather than the insertion loss, and that the 'sweet spot' observed in both our simulations and experiment is a balance between the input power and nonlinear losses. Nevertheless, OFCs with a \(30\mathrm{~nm}\) bandwidth are suitable for many target applications in communications and signal processing, and the SCF comb delivers both high power- per- tone and all- fibre connectivity required for practical systems.  
+
+## 3 Discussion  
+
+The last 20 years of frequency comb development has resulted in an array of comb generation technologies that have been used in numerous electronic and photonic applications. Yet, despite the various platform options that have been considered, there are relatively few OFCs designed specifically with telecommunications and optical signal processing applications in mind. Our SCF- based OFC fills an important gap
+
+<--- Page Split --->
+
+in the frequency comb family to provide a cavity- free, temperature- insensitive, flat and high power OFC source with a compact and portable form. Although the bandwidth of our system is currently limited by the free carrier effects associated with the silicon core, it is possible to mitigate these effects using carrier sweep- out schemes employed in planar silicon systems [42]. This could potentially be achieved by introducing two platinum rods next to the semiconductor core when making the fibre preform [43]. Alternative schemes involving gold- doping of the crystalline Si core can also be implemented to reduce the free- carrier lifetime [44].  
+
+Removing free carrier absorption would result in further increase of the bandwidth to \(>65 \mathrm{nm}\) , as illustrated by the simulated spectrum in Fig. 4d. In the future, the connection losses between the SSMF- SCF could also be reduced to below \(1 \mathrm{dB}\) per facet by optimising the nano- spike coupler design (e.g. employing a thinner silica cladding with outer diameter \(< 10 \mu \mathrm{m}\) ) [34], though this would require a customized mounting rig during the splicing, which is not currently available. With the reduced SSMF- SCF connection losses, we envisage replacing the HNLF in the NOLM stage with another section of SCF to reduce the size of the OFCs and enable a more compact SCF comb solution.  
+
+In summary, we have presented heterogeneous integration of a SCF with SSMFs for compact and efficient all- fibre frequency comb generation. Using our fabricated SCF as a mixer, we obtain 143 tones in a flat, 30- nm bandwidth frequency comb that exhibits narrow linewidths across the whole frequency region. Our approach harnesses the merits of nonlinear silicon waveguides and optical fibre platforms, underpinning comb applications requiring signal generation, processing and detection.  
+
+## 4 Methods  
+
+SCF fabrication The as- drawn SCFs are fabricated using the molten core fibre drawing (MCD) technique. This process uses a standard fibre drawing tower to heat and melt the silicon core that is surrounded by a softened silica cladding (drawing temperature of \(1950^{\circ}\mathrm{C}\) ), which acts as a crucible to retain the fibre profile as it is drawn down, as detailed in [45]. A thin layer of calcium oxide (CaO) is included as an interfacial barrier between the core and cladding during the drawing process, which limits dissolution of silica from the cladding into the silicon core and reduces the thermal strain arising from high- temperature processing. The as- drawn SCFs have a poly- crystalline core material with uniform core/cladding diameters of \(12\mu \mathrm{m} / 125\mu \mathrm{m}\) . To improve the crystalline quality and reduce the losses of the as- drawn fibres, we insert the original SCFs into a silica capillary ( \(400\mu \mathrm{m} / 150\mu \mathrm{m}\) inner/outer diameter) and taper this down to have core/cladding diameters of about \(5\mu \mathrm{m} / 125\mu \mathrm{m}\) . The fabrication is realized using a glass processing system (Vytran GPX- 3400- V4), which is widely accessible for heat- polishing, tapering and splicing.  
+
+Similar to the nano- taper couplers commonly used in planar silicon waveguides [46], nano- spike couplers are fabricated on the SCF facets to improve the coupling to SSMF. The nano- spikes are created by carefully tapering the SCF with the prepared void- gap, which occurs as a result of releasing the tension in the SCFs that is built- in due to the thermal expansion mismatch of the core/cladding materials. Splicing of the tapered SCF with nano- spike couplers on both ends of the tapered SSMFs is achieved by applying a heating power of \(63 \mathrm{W}\) over 7 seconds.  
+
+Seed comb generation Our seed comb begins with modulating a 1555.72- nm CW signal from a fibre laser using a LiNbO \(_3\) Mach- Zehnder modulator and two phase modulators (Supplementary Figure 2). The 1.6- kHz- linewidth CW source was amplified to 33 dBm by a polarization- maintaining fibre amplifier before launching into the modulators. The modulators transform the CW light into a repeated pulse train with the pulse period corresponding to each modulation cycle [47]. The resulting linear chirp yields pulses with relatively flat spectral envelopes for a tone spacing of 26 GHz. A low phase noise RF source was employed to generate the 26- GHz signal that drives the modulators. Generally, an arbitrary frequency can be used to enable a tunable tone spacing that suits DWDM applications. In our system, the RF frequency was tunable between 22- 26.5 GHz, limited by our frequency synthesizer and the electronic devices. Subsequent linear pulse compression was realized by compensating the spectral phase of the pulses using a reel of 65 m of SSMF, which provides second- order dispersion to compress the pulses to their Fourier- transform limit. The pulse full- width half maximum (FWHM) was measured to be approximately 610 fs using an optical autocorrelator (FEMTOCHROME FR- 103XL) with a Gaussian pulse profile assumed. An erbium- doped fibre amplifier (FA2) was used to amplify the pulse train before reshaping via a nonlinear optical loop mirror (NOLM). The NOLM consists of a 3- dB optical coupler connected to a 105 m Ge- doped HNLF with a dispersion of - 0.38 psnm \(^{- 1}\) km \(^{- 1}\) and a nonlinear coefficient of \(>10 \mathrm{W}^{- 1}\mathrm{km}^{- 1}\) at 1550 nm. The fibre was placed in the NOLM loop, along with a polarization controller and a 5- dB attenuator. The NOLM acts as an intensity discriminator, which transmits the
+
+<--- Page Split --->
+
+high- power peak regions at the center of each pulse and reflects the low- power background, providing a pedestal suppression ratio of 17.1 dB and additional pulse compression [48].  
+
+Linewidth characterization The linewidth characterization is performed by using a delayed self- heterodyne interferometer with 80- km ultra- low loss single- mode fibre, providing about \(4\mu \mathrm{m}\) delay or \(\sim 1.2\mathrm{kHz}\) spectral resolution, necessary for characterizing the narrow linewidth tones. As the seed CW source for the comb is a fibre laser, there is a significant \(1 / \mathrm{f}\) - type (flicker) frequency noise contribution to the frequency noise power spectral density [39]. In this case a Lorentzian profile cannot be assumed for the line shape since the frequency noise power spectral density is not dominated by white frequency noise. As such, we use a pseudo- Voigt profile, rather than a Lorentzian profile, to fit the measured beat note to appropriately account for the \(1 / \mathrm{f}\) - induced linewidth broadening [40]. While the RF driving signal also contributes both white phase ( \(\mathrm{f}^{2}\) frequency noise) and coloured phase noise to the comb noise power spectral density, this is difficult to generalise and has been neglected in this analysis.  
+
+Simulation We simulate the comb generation scheme with the same properties as the experiment to ensure a close match to our measured results. Pulse train propagation through the SCF was modelled by the generalized nonlinear Schrödinger equation (GNLSE), including TPA and free carrier effects (FCA and FCD) [49]:  
+
+\[\frac{\partial E}{\partial z} = -\frac{\alpha_l}{2} E + \sum_{m = 2}^{4}i\frac{i^m\beta_m}{2!}\frac{\partial^mE}{\partial t^m} +i\gamma \left(|E|^2 E + \frac{i}{\omega_0}\frac{\partial}{\partial t} (|E|^2 E)\right) - \frac{\sigma}{2} (1 + i\mu)N_cE \quad (1)\]  
+
+where \(E\) is the electric field envelope, \(\alpha_{l}\) is the linear attenuation and \(\beta_{m}\) is the \(m\) - th order dispersion parameter. TPA is included as the imaginary component of the nonlinear coefficient \(\gamma\) :  
+
+\[\gamma = \frac{2\pi n_2}{\lambda A_{\mathrm{eff}}} +\frac{i\beta_{\mathrm{TPA}}}{2A_{\mathrm{eff}}} \quad (2)\]  
+
+where \(n_{2}\) is the Kerr coefficient, \(\beta_{\mathrm{TPA}}\) is the TPA parameter and \(A_{\mathrm{eff}}\) is the effective mode area. FCA and FCD are included in the last term in equation 1, where \(\sigma\) is the FCA coefficient and \(\mu = 2k_{c}k_{0} / \sigma\) , with \(k_{0} = 2\pi /\lambda\) and \(k_{c}\) is the free- carrier- induced refractive index change. The magnitude of the FCA and FCD effects are governed by the rate equation for the free carrier density \(N_{c}\) [50]:  
+
+\[\frac{\partial N_{c}(z,t)}{\partial t} = \frac{\beta_{\mathrm{TPA}}}{2h\nu_{0}}\frac{|E(z,t)|^{4}}{A_{\mathrm{eff}}} -\frac{N_{c}(z,t)}{\tau} \quad (3)\]  
+
+where \(\tau\) is the free carrier lifetime. The GNLSE was solved using the split- step Fourier method (SSFM). Dispersion was included up to fourth- order and Raman scattering was neglected due to the short duration of our pulses. To accurately model the tapered SCF device with a varying core diameter, the SCF was separated into three distinct segments of length \(5.1\mathrm{mm}\) , \(3\mathrm{mm}\) and \(9.1\mathrm{mm}\) (input taper, middle and output taper respectively). These corresponded to core diameters of \(1.1\mu \mathrm{m}\) for the small tapered regions and \(5\mu \mathrm{m}\) for the middle region, and the mode properties of each diameter were estimated from COMSOL Multiphysics software simulations. The parameters used in the simulations are listed in Supplementary Table 1 and Figure 1, and were obtained via a combination of the mode simulations and laboratory experiments. The free carrier density and pulse train shown in Fig. 4b was taken from the final step of the SSFM to illustrate the free carrier density reaching steady state.  
+
+## 5 Data availability  
+
+The data that support the plots within this paper and other findings of this study are available from the corresponding author on reasonable request.  
+
+## 6 Acknowledgements  
+
+We acknowledge financial support from EPSRC grants EP/P000940/1, EP/R041792/1, EP/L015455/1 and EP/V007734/1.  
+
+## 7 Author Contributions  
+
+L.S., A.C.P. and Z.L. conceived the experiment. U.J.G. fabricated the silicon core fibre preforms and T.W.H and J.B. developed the drawing process and drew the preform into a fibre. H.R., L.S. and A.C.P
+
+<--- Page Split --->
+
+developed and implemented the tapering and splicing recipe used to splice the silicon core fibre to the SSMF, and developed simulations to estimate the fibre properties. C.D. and Z.L. built the electro-optic comb generator. R.S. and Z.L. designed the experimental setup, built the nonlinear optical loop mirror and measured the parametrically broadened spectrum. A.M.H. provided the highly nonlinear fibre used in the nonlinear optical loop mirror, and simulated both the statistical coherence analysis and nonlinear pulse evolution of the SCF comb. R.S., H.R., A.M.H. and Z.L. developed the simulation used to estimate the impact of two- photon absorption and free carriers in the system. All authors discussed and collaborated on the manuscript.  
+
+## 8 Competing Interests Statement  
+
+The authors declare no competing interests.  
+
+## References  
+
+[1] Diddams, S. A., Vahala, K. & Udem, T. Optical frequency combs: coherently uniting the electromagnetic spectrum. Science 369 (2020). [2] Lundberg, L. et al. Phase- coherent lightwave communications with frequency combs. Nature Communications 11, 201 (2020). URL https://doi.org/10.1038/s41467- 019- 14010- 7. [3] Picqué, N. & Hänsch, T. W. Frequency comb spectroscopy. Nature Photonics 13, 146- 157 (2019). [4] Fortier, T. M. et al. Generation of ultrastable microwaves via optical frequency division. Nature Photonics 5, 425- 429 (2011). [5] Deakin, C. & Liu, Z. Dual frequency comb assisted analog- to- digital conversion. Optics Letters 45, 173- 176 (2020). [6] Rueda, A., Sedlmeir, F., Kumari, M., Leuchs, G. & Schwefel, H. G. L. Resonant electro- optic frequency comb. Nature 568, 378- 381 (2019). URL https://doi.org/10.1038/s41586- 019- 1110- x. [7] Kippenberg, T. J., Gaeta, A. L., Lipson, M. & Gorodetsky, M. L. Dissipative kerr solitons in optical microresonators. Science 361, eaan8083 (2018). URL https://www.science.org/doi/abs/10.1126/science.aan8083. https://www.science.org/doi/pdf/10.1126/science.aan8083. [8] Ataie, V., Myslivets, E., Kuo, B. P.- P., Alic, N. & Radic, S. Spectrally equalized frequency comb generation in multistage parametric mixer with nonlinear pulse shaping. Journal of Lightwave Technology 32, 840- 846 (2014). [9] Supradeepa, V. R. & Weiner, A. M. Bandwidth scaling and spectral flatness enhancement of optical frequency combs from phase- modulated continuous- wave lasers using cascaded four- wave mixing. Opt. Lett. 37, 3066- 3068 (2012). URL http://www.osapublishing.org/ol/abstract.cfm?URI=ol- 37- 15- 3066. [10] Moss, D. J., Morandotti, R., Gaeta, A. L. & Lipson, M. New cmos- compatible platforms based on silicon nitride and hydex for nonlinear optics. Nature photonics 7, 597- 607 (2013). [11] Gaeta, A. L., Lipson, M. & Kippenberg, T. J. Photonic- chip- based frequency combs. nature photonics 13, 158- 169 (2019). [12] Marin- Palomo, P. et al. Microresonator- based solitons for massively parallel coherent optical communications. Nature 546, 274- 279 (2017). URL https://doi.org/10.1038/nature22387. [13] Torres- Company, V. et al. Laser frequency combs for coherent optical communications. Journal of Lightwave Technology 37, 1663- 1670 (2019). [14] Calo, C. et al. Single- section quantum well mode- locked laser for 400 Gb/s SSB- OFDM transmission. Opt. Express 23, 26442- 26449 (2015). URL http://www.opticsexpress.org/abstract.cfm?URI=oe- 23- 20- 26442. [15] Metcalf, A. J., Torres- Company, V., Leaird, D. E. & Weiner, A. M. High- power broadly tunable electrooptic frequency comb generator. IEEE Journal of Selected Topics in Quantum Electronics 19, 231- 236 (2013). [16] Hu, H. et al. Single- source chip- based frequency comb enabling extreme parallel data transmission. Nature Photonics 12, 469- 473 (2018). URL https://doi.org/10.1038/s41566- 018- 0205- 5.
+
+<--- Page Split --->
+
+[17] Pelusi, M. D. et al. Applications of highly- nonlinear chalcogenide glass devices tailored for high- speed all- optical signal processing. IEEE Journal of Selected Topics in Quantum Electronics 14, 529- 539 (2008).  
+
+[18] Kuyken, B., Billet, M., Leo, F., Yvind, K. & Pu, M. Octave- spanning coherent supercontinuum generation in an algaas- on- insulator waveguide. Opt. Lett. 45, 603- 606 (2020). URL http://ol.osa.org/abstract.cfm?URI=ol- 45- 3- 603.  
+
+[19] Son, G., Han, S., Park, J., Kwon, K. & Yu, K. High- efficiency broadband light coupling between optical fibers and photonic integrated circuits. Nanophotonics 7, 1845- 1864 (2018). URL https://doi.org/10.1515/nanoph- 2018- 0075.  
+
+[20] Kuo, B. P. et al. Wideband parametric frequency comb as coherent optical carrier. Journal of Lightwave Technology 31, 3414- 3419 (2013).  
+
+[21] Puttnam, B. J. et al. 2.15 pb/s transmission using a 22 core homogeneous single- mode multi- core fiber and wideband optical comb. In 2015 European Conference on Optical Communication (ECOC), 1- 3 (IEEE, 2015).  
+
+[22] Temprana, E. et al. Overcoming kerr- induced capacity limit in optical fiber transmission. Science 348, 1445- 1448 (2015). URL https://science.sciencemag.org/content/348/6242/1445. https://science.sciencemag.org/content/348/6242/1445. full.pdf.  
+
+[23] Hu, H., Ataie, V., Myslivets, E. & Radic, S. Optical comb assisted ofdm rf receiver. Journal of Lightwave Technology 37, 1280- 1287 (2019).  
+
+[24] Kuo, B. P.- P., Fini, J. M., Gruner- Nielsen, L. & Radic, S. Dispersion- stabilized highly- nonlinear fiber for wideband parametric mixer synthesis. Optics Express 20, 18611- 18619 (2012).  
+
+[25] Ren, H. et al. Nonlinear optical properties of polycrystalline silicon core fibers from telecom wavelengths into the mid- infrared spectral region. Opt. Mater. Express 9, 1271- 1279 (2019). URL http://www.osapublishing.org/ome/abstract.cfm?URI=ome- 9- 3- 1271.  
+
+[26] Wu, D. et al. Four- wave mixing- based wavelength conversion and parametric amplification in submicron silicon core fibers. IEEE Journal of Selected Topics in Quantum Electronics 27, 1- 11 (2020).  
+
+[27] Wu, D. et al. Net optical parametric gain in a submicron silicon core fiber pumped in the telecom band. APL Photonics 4, 086102 (2019). URL https://doi.org/10.1063/1.5103272. https://doi.org/10.1063/1.5103272.  
+
+[28] Huang, M. et al. Fiber integrated wavelength converter based on a silicon core fiber with a nano- spike coupler. IEEE Photonics Technology Letters 31, 1561- 1564 (2019).  
+
+[29] Tetsumoto, T. et al. Optically referenced 300 ghz millimetre- wave oscillator. Nature Photonics 1- 7 (2021).  
+
+[30] Millot, G. et al. Frequency- agile dual- comb spectroscopy. Nature Photonics 10, 27- 30 (2016).  
+
+[31] Deakin, C. & Liu, Z. Noise and distortion analysis of dual frequency comb photonic rf channelizers. Optics Express 28, 39750- 39769 (2020).  
+
+[32] Gibson, U. J., Wei, L. & Ballato, J. Semiconductor core fibres: materials science in a bottle. Nature Communications 12, 1- 15 (2021).  
+
+[33] Franz, Y. et al. Material properties of tapered crystalline silicon core fibers. Opt. Mater. Express 7, 2055- 2061 (2017). URL http://www.osapublishing.org/ome/abstract.cfm?URI=ome- 7- 6- 2055.  
+
+[34] Ren, H. et al. Tapered silicon core fibers with nano- spikes for optical coupling via spliced silica fibers. Opt. Express 25, 24157- 24163 (2017). URL http://www.opticsexpress.org/abstract.cfm?URI=oe- 25- 20- 24157.  
+
+[35] Torres- Company, V., Lancis, J. & Andrés, P. Lossless equalization of frequency combs. Opt. Lett. 33, 1822- 1824 (2008). URL http://ol.osa.org/abstract.cfm?URI=ol- 33- 16- 1822.  
+
+[36] Okada, K., Hashimoto, K., Shibata, T. & Nagaki, Y. Novel method for high resolution measurement of laser output spectrum. Electronics Letters 16, 630 (1980).  
+
+[37] Skehan, J. C., Naveau, C., Schroder, J. & Andrekson, P. Widely tunable, low linewidth, and high power laser source using an electro- optic comb and injection- locked slave laser array. Opt. Express 29, 17077- 17086 (2021). URL http://www.osapublishing.org/oe/abstract.cfm?URI=oe- 29- 11- 17077.
+
+<--- Page Split --->
+
+[38] Gordon, J. P. & Mollenauer, L. F. Phase noise in photonic communications systems using linear amplifiers. Opt. Lett. 15, 1351- 1353 (1990). URL http://www.osapublishing.org/ol/abstract.cfm?URI=ol- 15- 23- 1351.  
+
+[39] Horak, P. & Loh, W. H. On the delayed self- heterodyne interferometric technique for determining the linewidth of fiber lasers. Opt. Express 14, 3923- 3928 (2006). URL http://www.opticsexpress.org/abstract.cfm?URI=oe- 14- 9- 3923.  
+
+[40] Stephan, G. M., Tam, T. T., Blin, S., Besnard, P. & Tétu, M. Laser line shape and spectral density of frequency noise. Phys. Rev. A 71, 043809 (2005). URL https://link.aps.org/doi/10.1103/PhysRevA.71.043809.  
+
+[41] Yin, L. & Agrawal, G. P. Impact of two- photon absorption on self- phase modulation in silicon waveguides. Opt. Lett. 32, 2031- 2033 (2007). URL http://www.osapublishing.org/ol/abstract.cfm?URI=ol- 32- 14- 2031.  
+
+[42] Griffith, A. G. et al. Silicon- chip mid- infrared frequency comb generation. Nature Communications 6, 6299 (2015). URL https://doi.org/10.1038/ncomms7299.  
+
+[43] Wei, L. et al. Optoelectronic fibers via selective amplification of in- fiber capillary instabilities. Advanced Materials 29, 1603033 (2017). URL https://onlinelibrary.wiley.com/doi/abs/10.1002/adma.201603033. https://onlinelibrary.wiley.com/doi/pdf/10.1002/adma.201603033.  
+
+[44] Sorgard, T., Hawkins, T., Ballato, J., Österberg, U. L. & Gibson, U. J. All- optical high- speed modulation of thz transmission through silicon core optical fibers. Opt. Express 29, 3543- 3552 (2021). URL http://www.osapublishing.org/oe/abstract.cfm?URI=oe- 29- 3- 3543.  
+
+[45] Nordstrand, E. F., Dibbs, A. N., Eraker, A. J. & Gibson, U. J. Alkaline oxide interface modifiers for silicon fiber production. Opt. Mater. Express 3, 651- 657 (2013). URL http://www.osapublishing.org/oe/abstract.cfm?URI=oe- 3- 5- 651.  
+
+[46] Wood, M., Sun, P. & Reano, R. M. Compact cantilever couplers for low- loss fiber coupling to silicon photonic integrated circuits. Opt. Express 20, 164- 172 (2012). URL http://www.opticsexpress.org/abstract.cfm?URI=oe- 20- 1- 164.  
+
+[47] Wu, R., Supradeepa, V., Long, C. M., Leaird, D. E. & Weiner, A. M. Generation of very flat optical frequency combs from continuous- wave lasers using cascaded intensity and phase modulators driven by tailored radio frequency waveforms. Optics letters 35, 3234- 3236 (2010).  
+
+[48] Smith, K., Doran, N. J. & Wigley, P. G. J. Pulse shaping, compression, and pedestal suppression employing a nonlinear- optical loop mirror. Opt. Lett. 15, 1294- 1296 (1990). URL http://ol.osa.org/abstract.cfm?URI=ol- 15- 22- 1294.  
+
+[49] Yin, L., Lin, Q. & Agrawal, G. P. Soliton fission and supercontinuum generation in silicon waveguides. Opt. Lett. 32, 391- 393 (2007). URL http://ol.osa.org/abstract.cfm?URI=ol- 32- 4- 391.  
+
+[50] Yin, L., Lin, Q. & Agrawal, G. P. Dispersion tailoring and soliton propagation in silicon waveguides. Opt. Lett. 31, 1295- 1297 (2006). URL http://ol.osa.org/abstract.cfm?URI=ol- 31- 9- 1295.
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1: Fabrication process flow of the all-fibre heterogeneously-integrated parametric mixer. (a) Tapered SCF with core/cladding diameters of about \(5\mu \mathrm{m} / 125\mu \mathrm{m}\) . (b) Heating and tapering process to make the first void gap in the fibre core. (c) Heating and tapering to make the second void gap in the fibre core. (d) fibre tapering process to scale down core size and collapse the void gap to form the nano-spike coupler. (e) Cleave at the center of the void to remove one side of the taper. (f) Splice the SCF nanospike to a tapered SSMF. (g) Employ a polymer tube to mechanically support the SSMF-SCF connection, keeping the fibre straight and tapering the other end of the SCF. (h) Cleave the other end and (i) splice to another tapered SSMF. </center>
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure 2: Fully integrated SCF-based parametric comb generation scheme. An electro-optic frequency comb is generated to act as a seed source for the parametric mixing stage. A pulse re-shaping stage compresses the comb pulse train to maximise the pulse peak power and enhance the parametric mixing efficiency. The optical output is then amplified and launched into a 17 mm sample of fully-integrated SCF, where parametric nonlinearities cause comb broadening. Both ends of the SCF are tapered and spliced to tapered single-mode fibre, with nano-spike couplers to facilitate coupling between the heterogenous fibre cores. </center>
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Figure 3: C-band parametric frequency comb generation using SCF. (a) An electro-optic frequency comb with 26 GHz line spacing (orange) was temporally reshaped and launched into a 17mm sample of SCF at an average power of 32 dBm. The optical spectrum of the broadened parametric comb (blue) was obtained at a resolution of 0.02 nm, achieving 143 tones across a 30.0 nm bandwidth. A 10 MHz resolution BOSA was used to obtain close-in traces at different points within the comb bandwidth (Fig.(b)-(d)). The dashed line shows the instrument noise floor. Results in (c) and (d) are both limited by the instrument noise floor. (e) The linewidth of the parametric comb was measured using the delayed self-heterodyne interferometer (DSHI) method with a 80-km delay line and pseudo-Voigt profile fitting. The pseudo-Voigt fitting curves are shown for the two measured beat notes at the extremities of the SCF comb bandwidth ((f) and (g)). </center>
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Figure 4: Parametric frequency comb simulation including the effects of two-photon absorption and free carriers within the SCF. (a) Nonlinear loss induced by TPA and FCs as a function of average input power. (b) Free carrier density in the silicon core versus time (black), and the optical pulse train power versus time (red). Subsequent pulses in the 26 GHz pulse train increase the free carrier density before complete recombination can occur, reaching a steady state over several nanoseconds. (c) Simulated optical spectrum at the output of the SCF mixer, at 27 dBm and 32 dBm input power (top and middle respectively) and (d) at 32 dBm without free carriers. TPA is included in all three spectra, and insertion loss is neglected. </center>
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SupplementaryinfoAllfibreSCFcomb.pdf SupplementarySCFSpectrogramvideo.mp4
+
+<--- Page Split --->

preprint/preprint__2bbfafaeea0e549a06406df3ec0de2092667f753880d5d03f95c7b3aace942cb/preprint__2bbfafaeea0e549a06406df3ec0de2092667f753880d5d03f95c7b3aace942cb_det.mmd

@@ -0,0 +1,368 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 858, 177]]<|/det|>
+# All-fibre heterogeneously-integrated frequency comb generation using silicon core fibre  
+
+<|ref|>text<|/ref|><|det|>[[44, 196, 180, 235]]<|/det|>
+Ronit Sohanpal UCL  
+
+<|ref|>text<|/ref|><|det|>[[44, 243, 333, 284]]<|/det|>
+Haonan Ren Dalian University of Technology  
+
+<|ref|>text<|/ref|><|det|>[[44, 290, 477, 331]]<|/det|>
+Li Shen Huazhong University of Science and Technology  
+
+<|ref|>text<|/ref|><|det|>[[44, 336, 172, 374]]<|/det|>
+Callum Deakin UCL  
+
+<|ref|>text<|/ref|><|det|>[[44, 382, 211, 422]]<|/det|>
+Alexander Heidt University of Bern  
+
+<|ref|>text<|/ref|><|det|>[[44, 428, 223, 468]]<|/det|>
+Thomas Hawkins Clemson University  
+
+<|ref|>text<|/ref|><|det|>[[44, 474, 581, 515]]<|/det|>
+John Ballato Clemson University https://orcid.org/0000- 0001- 5910- 3504  
+
+<|ref|>text<|/ref|><|det|>[[44, 520, 223, 560]]<|/det|>
+Ursula Gibson Clemson University  
+
+<|ref|>text<|/ref|><|det|>[[44, 567, 648, 607]]<|/det|>
+Anna Peacock University of Southampton https://orcid.org/0000- 0002- 1940- 7172  
+
+<|ref|>text<|/ref|><|det|>[[44, 612, 640, 654]]<|/det|>
+Zhixin Liu ( zhixin.liu@ucl.ac.uk ) University College London https://orcid.org/0000- 0002- 9681- 7933  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 696, 102, 714]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 733, 747, 754]]<|/det|>
+Keywords: optical frequency combs, optical signal- to- noise ratio, silicon core fiber  
+
+<|ref|>text<|/ref|><|det|>[[44, 772, 345, 791]]<|/det|>
+Posted Date: November 11th, 2021  
+
+<|ref|>text<|/ref|><|det|>[[44, 810, 463, 829]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 968401/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 847, 909, 890]]<|/det|>
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 45, 945, 87]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on July 9th, 2022. See the published version at https://doi.org/10.1038/s41467-022-31637-1.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[137, 135, 857, 187]]<|/det|>
+# All-fibre heterogeneously-integrated frequency comb generation using silicon core fibre  
+
+<|ref|>text<|/ref|><|det|>[[133, 201, 861, 258]]<|/det|>
+Ronit Sohanpal \(^{1\otimes}\) , Haonan Ren \(^{2,3}\) , Li Shen \(^{4\otimes}\) , Callum Deakin \(^{1}\) , Alexander M. Heidt \(^{5}\) , Thomas W. Hawkins \(^{6}\) , John Ballato \(^{6}\) , Ursula J. Gibson \(^{6}\) , Anna C. Peacock \(^{2\otimes}\) , and Zhixin Liu \(^{1\otimes}\)  
+
+<|ref|>text<|/ref|><|det|>[[123, 270, 872, 305]]<|/det|>
+\(^{1}\) Department of Electronic and Electrical Engineering, UCL (University College London), London, United Kingdom  
+
+<|ref|>text<|/ref|><|det|>[[135, 305, 859, 341]]<|/det|>
+\(^{2}\) Optoelectronics Research Centre, University of Southampton, Southampton, United Kingdom  
+
+<|ref|>text<|/ref|><|det|>[[135, 340, 859, 376]]<|/det|>
+\(^{3}\) now with School of Optoelectronic Engineering and Instrumentation Science, Dalian University of Technology, Dalian, China  
+
+<|ref|>text<|/ref|><|det|>[[128, 375, 861, 411]]<|/det|>
+\(^{4}\) Wuhan National Laboratory for Optoelectronics and School of Optical and Electrical Information, Huazhong University of Science and Technology, China  
+
+<|ref|>text<|/ref|><|det|>[[120, 410, 872, 446]]<|/det|>
+\(^{5}\) Institute of Applied Physics, University of Bern, Switzerland \(^{6}\) Center for Optical Materials Science and Engineering Technologies, and the Department of Materials Science and Engineering, Clemson University, USA  
+
+<|ref|>text<|/ref|><|det|>[[377, 465, 620, 536]]<|/det|>
+\(^{1}\) ronit.sohanpal.14@ucl.ac.uk \(^{4}\) lishen@hust.edu.cn \(^{2}\) acp@orc.soton.ac.uk \(^{1}\) zhixin.liu@ucl.ac.uk  
+
+<|ref|>sub_title<|/ref|><|det|>[[465, 578, 531, 592]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[155, 596, 843, 777]]<|/det|>
+Originally developed for metrology, optical frequency combs are becoming increasingly pervasive in a wider range of research topics including optical communications, spectroscopy, and radio or microwave signal processing. However, application demands in these fields can be more challenging as they require compact sources with a high tolerance to temperature variations that are capable of delivering flat comb spectra, high power per tone, narrow linewidth and high optical signal- to- noise ratio (OSNR). To date, there has not been a frequency comb technology that is able to simultaneously achieve all these requirements. This work reports the generation of a flat, high power frequency comb in the telecom band using a 17- mm fully- integrated silicon core fibre (SCF) as a parametric mixer. Our all- fibre, cavity- free source combines the materials benefits of planar waveguide structures with the advantageous properties of fibre platforms to achieve a 30 nm bandwidth comb source containing 143 tones with \(< 3 \mathrm{kHz}\) linewidth, 12 dB flatness, and \(>30 \mathrm{dB}\) OSNR over the entire spectral region. The unique combination of technical features offered by this SCF- based source opens a path towards a new class of high- performance frequency comb generators for communications and signal processing applications.  
+
+<|ref|>sub_title<|/ref|><|det|>[[157, 797, 342, 818]]<|/det|>
+## 1 Introduction  
+
+<|ref|>text<|/ref|><|det|>[[156, 825, 842, 911]]<|/det|>
+Optical frequency combs (OFCs) generate precisely- spaced phase- coherent spectral tones and have revolutionized frequency control and metrology [1]. Over the past decade, numerous new developments have been focused on comb applications in optical communications, spectroscopy, microwave photonics, and optical- assisted signal processing [2, 3, 4, 5]. Unlike metrology, in which broad 'over- one- octave' spectral bandwidths are the predominate requirement to enable self- referencing, communications and signal processing applications typically demand only tens of nanometers of bandwidth, but require high power, high
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[155, 91, 840, 120]]<|/det|>
+optical signal to noise ratios (OSNR), a flat spectral response and narrow comb linewidths. In addition, compactness and temperature stability are essential prerequisites for most practical applications.  
+
+<|ref|>text<|/ref|><|det|>[[155, 120, 842, 271]]<|/det|>
+Significant efforts have been made to develop OFCs using both integrated photonic waveguides and optical fibre platforms [6, 7, 8, 9]. Much of the work using integrated platforms has focused on Kerr soliton comb formation using dispersion- engineered microcavities, which have demonstrated comb spectra spanning more than \(100\mathrm{nm}\) [10, 11]. Although Kerr combs have previously been employed to demonstrate C and L band coherent communications [12], their hyperbolic secant ( \(\mathrm{sech}^2\) ) spectral power envelope means the comb bandwidths must significantly exceed the standard C+L telecom bands to avoid having to use the low power wings, limiting the maximum output power- per- tone and the OSNR. Dark pulse Kerr combs with narrower spectral bandwidths ( \(\sim 35\mathrm{nm}\) ) can offer some improvement to the power and OSNR [13], but the need for a high quality factor cavity implies an inherent sensitivity to temperature variation. Of note is that temperature sensitivity is also an issue with other cavity- based comb sources, such as those generated using integrated mode- locked lasers [14].  
+
+<|ref|>text<|/ref|><|det|>[[155, 271, 842, 382]]<|/det|>
+In contrast, OFCs generated via cavity- free configurations offer more flexibility in terms of spectral shaping and are in general much more robust to environmental perturbations. Such OFCs can be realized by using cascaded modulators or by pumping a highly nonlinear waveguide using high peak- power pulses, or a combination of both approaches [15, 16]. For example, planar waveguides based on chalcogenide glasses and III- V semiconductor materials (e.g. AlGaAs) possess large nonlinearities and small dimensions, making them an ideal choice for compact parametric mixing stages [17, 18]. However, such high index planar structures suffer from restrictive coupling requirements when compared to fibre systems as a result of their small dimensions and rectangular cross- section [19].  
+
+<|ref|>text<|/ref|><|det|>[[155, 382, 842, 520]]<|/det|>
+In comparison, cavity- free OFCs constructed using fibre- optic platforms as the nonlinear media have the advantage of low losses and high power handling, thus significantly increasing the power- per- tone and OSNR. By integrating fibre amplifiers, fibre pulse compressors, highly nonlinear fibre (HNLF)- based saturable absorbers and parametric mixers, all- fibre OFCs have been demonstrated with bandwidths over \(100\mathrm{nm}\) [20], \(\sim 0\mathrm{dBm}\) per tone and a spectral flatness of \(3\mathrm{dB}\) [8], offering superior performance for data transmission [21, 22] and microwave signal processing [23], as well as being immediately compatible with much of the existing fibre infrastructure. However, current fibre- based OFCs are bulky due to the hundreds of meters of HNLF required to achieve significant parametric gain. Moreover, efficient parametric mixing requires careful dispersion- management over the entire HNLF length, and specifically a stable zero dispersion wavelength [24], which increases fabrication difficulties and cost.  
+
+<|ref|>text<|/ref|><|det|>[[155, 520, 842, 715]]<|/det|>
+A long- standing goal has been to combine the materials benefits of the integrated planar waveguide structures with the advantageous waveguiding properties of fibre platforms to achieve all of the desired technical features simultaneously. The emerging class of silicon core fibres (SCF) offer a promising solution to this objective, in which a crystalline silicon core material is embedded within a conventional silica glass cladding. Compared to fibres with glassy core materials, the SCF platform offers a significantly enhanced nonlinear coefficient \(\gamma\) due to the combined effects of the high nonlinear refractive index \(n_2\) of silicon and the high core- cladding index contrast. For example, SCFs with core diameters of only a few \(\mu \mathrm{m}\) can have an effective nonlinear coefficient of more than three orders of magnitude higher than that of commercially available Ge- doped HNLFs, which enables the reduction of the parametric mixer length from hundreds of meters to a few millimeters [25]. Such compact SCFs have already shown great promise for efficient nonlinear processing of signals across the extended telecoms band [26]. However, to date, the complete integration of SCFs with silica fibres has been a significant hurdle and, thus far, reported SCF results have been limited to either free space coupling [27] or partial fibre integration [28], which makes the coupling of dense multi- wavelength systems more challenging.  
+
+<|ref|>text<|/ref|><|det|>[[155, 715, 842, 825]]<|/det|>
+In this work, we present the first all- fibre integrated SCF parametric mixer for OFC generation. Specifically, the SCF has been processed to allow for direct splicing to standard single- mode fibre (SSMF) connectors at both ends, allowing for straightforward integration with conventional fibre systems. The resulting comb structure has a bandwidth of \(30\mathrm{nm}\) , with a flatness of \(12\mathrm{dB}\) , \(>30\mathrm{dB}\) OSNR and \(< 3\mathrm{kHz}\) linewidth, properties that are desirable for many practical applications in areas such as phase- coherent communications and micrometer/millimeter wave generation [29]. Moreover, our cavity- free design allows for tunable wavelength and comb spacing, thus enabling high resolution dual- comb spectroscopy [30] as well as dual- comb RF processing [31].
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[156, 86, 282, 105]]<|/det|>
+## 2 Results  
+
+<|ref|>sub_title<|/ref|><|det|>[[156, 117, 486, 135]]<|/det|>
+### 2.1 \(\mathrm{SiO_2}\) -Si core fibre integration  
+
+<|ref|>text<|/ref|><|det|>[[155, 140, 842, 335]]<|/det|>
+The SCFs used in this work were fabricated via the molten core drawing (MCD) technique [32]. The MCD method is the most practical of the SCF production approaches as it can rapidly produce long lengths of fibre that are compatible with traditional fibre post- processing procedures. The as- drawn MCD SCFs possess a polycrystalline silicon core with typical core/cladding diameters of approximately \(12\mu \mathrm{m} / 125\mu \mathrm{m}\) , as detailed in the Methods. To transform the SCFs into low loss nonlinear parametric mixers that are robust and user- friendly, a multi- step tapering and splicing approach is employed. These post- processing steps are important to enhance the efficiency of the nonlinear processing in the SCFs for a number of key reasons. First, the tapering process is used both to reduce the core size as well as to improve the crystalline quality, and hence, the optical transmission [33]. This results in a SCF with core/cladding diameters of approximately \(5\mu \mathrm{m} / 125\mu \mathrm{m}\) and a transmission loss of \(3\mathrm{dB / cm}\) . Second, a modified taper method is applied to further reduce the core size and fabricate nano- spike couplers on the end facets of the fibre. The role of the couplers is to better match the modes of the SCFs with those of the standard single mode fibres (SSMFs), as well as to suppress reflections at the SCF/SSMF interface when splicing.  
+
+<|ref|>text<|/ref|><|det|>[[155, 335, 842, 652]]<|/det|>
+The full process to fabricate the integrated SCF mixer is illustrated in Fig. 1. In the first step, the SCF is heated gently while a small tension is applied along the fibre axis. Owing to the tensile stress in the as- drawn SCF, as the heated core cools and recrystallizes, a void- gap can form around the heat zone (Fig. 1b). Repeating this process produces another void- gap at the other end of the fibre (Fig. 1c). Subsequently, a single sweep tapering process is used to reduce the local core/cladding ratio from \(5 / 125\mu \mathrm{m}\) to \(1.1 / 27\mu \mathrm{m}\) over the left- hand side of the fibre, during which the first void- gap collapses to form a nano- spike with a length of \(\sim 200\mu \mathrm{m}\) at the core facet (Fig. 1d). The SCF is then precisely cleaved in the core- less region and spliced with a pre- prepared tapered SSMF with same cladding diameter (Fig. 1e- f). To allow for a second connection to be made at the other end, we enclose the spliced SSMF- SCF section into a polymer capillary that mechanically supports the nano- spike coupler before applying the same procedure to the right- hand side of the fibre (Fig. 1g- i). The result is a fully integrated SCF fibre device with a total length of \(\sim 17\mathrm{mm}\) , which contains two small core sections of \(1.1\mu \mathrm{m}\) diameter at each end (lengths of \(5.1\mathrm{mm}\) and \(9.1\mathrm{mm}\) at the input and output, respectively), connected by a short \(3\mathrm{mm}\) length with a \(5\mu \mathrm{m}\) diameter in the middle. Importantly, thanks to the silica cladding of the SCFs, all of these processing steps can be conducted using a single glass processing system, enabling high yield production of these heterogeneous fibre devices. The loss in the processed SCF section was determined to be \(\sim 2\mathrm{dB / cm}\) with an effective nonlinear coefficient \((\gamma)\) of approximately \(30\mathrm{W}^{- 1}\mathrm{m}^{- 1}\) . From the transmission losses, the coupling losses for the sample used in this work was estimated to be \(\sim 8\mathrm{dB}\) per facet, resulting in an end- to- end loss of \(\sim 19\mathrm{dB}\) . This fairly substantial insertion loss can be attributed to the losses associated with splicing the tapered fibres ( \(27\mu \mathrm{m}\) diameter cladding) and the mode mismatch between the tapered SSMF and the nanospike. Simulations indicate that these losses could be reduced to \(2\mathrm{dB}\) by optimizing the cladding diameters at the connection point ( \(10\mathrm{um}\) ) to reduce the mismatch [34], though these diameters are currently smaller than what our processing system can handle.  
+
+<|ref|>sub_title<|/ref|><|det|>[[156, 668, 476, 686]]<|/det|>
+### 2.2 SCF-based comb generation  
+
+<|ref|>text<|/ref|><|det|>[[155, 692, 842, 902]]<|/det|>
+The SCF- based parametric comb generator consists of three main sections (Fig. 2): an electro- optic comb, a fibre pulse compressor, and the all- fibre- integrated SCF device as a parametric mixer. To achieve narrow linewidths, we employed a continuous wave (CW) fibre laser with \(\sim 1.6\mathrm{kHz}\) linewidth as the seed to which a phase and intensity modulation were applied to create an \(11.8\mathrm{nm}\) - wide electro- optic (EO) comb (56 tones with 26- GHz spacing), shown in orange in Fig. 3a. In the time domain, this corresponds to a pulse train with a repetition rate of \(26\mathrm{GHz}\) , with each pulse exhibiting a quasi- linear frequency chirp [35]. To optimise the parametric mixing efficiency, we compressed the pulses linearly using a short length of SSMF to eliminate the frequency chirp across the center of the pulse, resulting in a full- width half- maximum (FWHM) pulse width of \(610\mathrm{fs}\) . As obtaining an OFC with good spectral flatness is a key aim of this work, we implemented a nonlinear optical loop mirror (NOLM) to suppress the low- power pedestals that result from the dispersive chirp compression, which can cause significant spectral rippling after the mixing stage and reduce the mixing efficiency [8]. This also reduces the pulse FWHM down to approximately \(440\mathrm{fs}\) before amplification to an average power of \(32\mathrm{dBm}\) using a dispersion- flattened short- pulse fibre amplifier. The amplifier introduces no observable broadening of the seed pulses. The amplified pulses were subsequently launched into the \(17\mathrm{mm}\) all- fibre- integrated SCF device.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[155, 91, 842, 245]]<|/det|>
+The blue curve in Fig. 3a shows the spectrum of the SCF mixer output, indicating 143 tones within a bandwidth of \(30~\mathrm{nm}\) with a flatness of \(12\mathrm{~dB}\) , from \(1535\mathrm{~nm}\) to \(1565\mathrm{~nm}\) . For the estimation of the OSNR we measure the close- up spectra at 1545, 1555, and \(1563\mathrm{~nm}\) in Fig. 3b- d with \(10\mathrm{MHz}\) resolution. Considering \(0.1\mathrm{~nm}\) noise bandwidth, the OSNR at the center ( \(1550\mathrm{~nm}\) ) and long wavelength edge ( \(1563\mathrm{~nm}\) ) of the SCF comb is greater than \(35\mathrm{~dB}\) and the OSNR at the short wavelength edge ( \(1536\mathrm{~nm}\) ) is approximately \(30\mathrm{~dB}\) . The increased noise in the \(1535\mathrm{- }1545\mathrm{~nm}\) region is primarily attributed to the amplified spontaneous emission (ASE) noise generated by the Er/Yb- doped fibre amplifier before the parametric amplification stage. This could be eliminated using a bandpass filter to ensure a high OSNR across the whole comb bandwidth. Since the nonlinear spectral broadening of the comb in the SCF is dominated by self- phase modulation, the asymmetric temporal pulse shape after the NOLM leads to the observed asymmetric expansion of the comb towards shorter wavelengths (Supplementary Video 1).  
+
+<|ref|>text<|/ref|><|det|>[[155, 244, 842, 340]]<|/det|>
+Fig. 3e shows the measured linewidth of the SCF comb tones (triangle marker), and the seed EO comb (circle marker), measured using the delayed self- heterodyne interferometer method [36] for every second comb line. Typically, the EO comb linewidth increases linearly as a function of the absolute comb tone index \(|n|\) due to the scaling noise contribution of the RF- induced phase noise [37]. Our comb design uses an ultra- low phase noise RF signal generator ( \(< 7\) fs integrated jitter from \(100\mathrm{~Hz}\) to \(100\mathrm{~MHz}\) ), which results in a negligible scaling of the EO comb bandwidth with the tone index (orange circles in Fig. 3e).  
+
+<|ref|>text<|/ref|><|det|>[[155, 339, 842, 465]]<|/det|>
+After the SCF, the EO comb linewidth increases to \(1.9\mathrm{- }2.4\mathrm{kHz}\) , which we attribute to the Gordon- Mollenauer effect (nonlinearity- induced amplitude- to- phase noise conversion) [38]. Numerical analysis of our system shows that the spectral coherence of the SCF comb degrades as the amplifier noise figure increases, which results in the increased comb linewidth observed on the short- wavelength edge (Supplementary Figure 3). Nevertheless, it is clear that the SCF comb retains a well- preserved linewidth performance across the whole spectrum. Fig. 3f and Fig. 3g show the measured beat note and their fitting using a pseudo- Voigt profile at both extremities of the comb for the linewidth characterization. We use a pseudo- Voigt profile to account for the fact that the frequency noise of the SCF tones is a mixed contribution of \(1 / \mathrm{f}\) (flicker) frequency noise and white frequency noise from the fibre laser [39, 40].  
+
+<|ref|>text<|/ref|><|det|>[[155, 464, 842, 646]]<|/det|>
+To study the role of the integrated SCF device losses on the spectral bandwidth of the SCF- based comb, we simulate the comb generation system to mirror the experimental setup. The numerical simulation uses a modified generalized nonlinear Schrödinger equation (GNLSE) that incorporates two photon absorption (TPA), free carrier absorption (FCA) and free carrier dispersion (FCD) [41] (see Methods). We begin by evaluating the nonlinear loss due to TPA and FCA for different average launch powers within the SCF device used in our system. In these investigations, we assume zero insertion loss so that we can gauge the full potential of our approach. Fig. 4a shows that the nonlinear loss becomes greater than \(1\mathrm{~dB}\) when the average power increases to \(>17\mathrm{~dBm}\) and increases exponentially thereafter. Compared to previous studies of SCF- based nonlinear signal processing [26], our 26- GHz- repetition- rate frequency comb has a significantly shorter interval between the pulses than the free carrier lifetime ( \(38\mathrm{~ps}\) versus \(\sim 1\mathrm{~ns}\) , respectively). As a result, there is only a partial recombination of free carriers, leading to an accumulation of the free carrier density before reaching a steady state after about \(2\mathrm{~ns}\) , as shown in Fig. 4b. The reduction in the pulse train power due to FCA is shown as the red lines in Fig. 4b.  
+
+<|ref|>text<|/ref|><|det|>[[155, 645, 842, 799]]<|/det|>
+As expected, TPA, FCA and FCD have negative effects for wide- band comb generation and result in a reduced parametric gain, limiting the comb bandwidth at high average pump powers. As shown in Fig. 4c, increasing the average input power from \(27\mathrm{~dBm}\) to \(32\mathrm{~dBm}\) results in a negligible change of the comb spectral bandwidth due to the increase in nonlinear absorption. Moreover, the results for \(27\mathrm{~dBm}\) of pump power have produced a \(30\mathrm{~nm}\) comb bandwidth with a spectral flatness of \(13\mathrm{~dB}\) at the telecom C- band, similar to our measured results. Thus these findings suggest that the main limitation to the comb bandwidth is the free carrier effects, rather than the insertion loss, and that the 'sweet spot' observed in both our simulations and experiment is a balance between the input power and nonlinear losses. Nevertheless, OFCs with a \(30\mathrm{~nm}\) bandwidth are suitable for many target applications in communications and signal processing, and the SCF comb delivers both high power- per- tone and all- fibre connectivity required for practical systems.  
+
+<|ref|>sub_title<|/ref|><|det|>[[156, 818, 316, 838]]<|/det|>
+## 3 Discussion  
+
+<|ref|>text<|/ref|><|det|>[[156, 847, 842, 904]]<|/det|>
+The last 20 years of frequency comb development has resulted in an array of comb generation technologies that have been used in numerous electronic and photonic applications. Yet, despite the various platform options that have been considered, there are relatively few OFCs designed specifically with telecommunications and optical signal processing applications in mind. Our SCF- based OFC fills an important gap
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[155, 91, 842, 188]]<|/det|>
+in the frequency comb family to provide a cavity- free, temperature- insensitive, flat and high power OFC source with a compact and portable form. Although the bandwidth of our system is currently limited by the free carrier effects associated with the silicon core, it is possible to mitigate these effects using carrier sweep- out schemes employed in planar silicon systems [42]. This could potentially be achieved by introducing two platinum rods next to the semiconductor core when making the fibre preform [43]. Alternative schemes involving gold- doping of the crystalline Si core can also be implemented to reduce the free- carrier lifetime [44].  
+
+<|ref|>text<|/ref|><|det|>[[155, 188, 842, 285]]<|/det|>
+Removing free carrier absorption would result in further increase of the bandwidth to \(>65 \mathrm{nm}\) , as illustrated by the simulated spectrum in Fig. 4d. In the future, the connection losses between the SSMF- SCF could also be reduced to below \(1 \mathrm{dB}\) per facet by optimising the nano- spike coupler design (e.g. employing a thinner silica cladding with outer diameter \(< 10 \mu \mathrm{m}\) ) [34], though this would require a customized mounting rig during the splicing, which is not currently available. With the reduced SSMF- SCF connection losses, we envisage replacing the HNLF in the NOLM stage with another section of SCF to reduce the size of the OFCs and enable a more compact SCF comb solution.  
+
+<|ref|>text<|/ref|><|det|>[[155, 285, 842, 354]]<|/det|>
+In summary, we have presented heterogeneous integration of a SCF with SSMFs for compact and efficient all- fibre frequency comb generation. Using our fabricated SCF as a mixer, we obtain 143 tones in a flat, 30- nm bandwidth frequency comb that exhibits narrow linewidths across the whole frequency region. Our approach harnesses the merits of nonlinear silicon waveguides and optical fibre platforms, underpinning comb applications requiring signal generation, processing and detection.  
+
+<|ref|>sub_title<|/ref|><|det|>[[156, 374, 298, 394]]<|/det|>
+## 4 Methods  
+
+<|ref|>text<|/ref|><|det|>[[155, 404, 842, 570]]<|/det|>
+SCF fabrication The as- drawn SCFs are fabricated using the molten core fibre drawing (MCD) technique. This process uses a standard fibre drawing tower to heat and melt the silicon core that is surrounded by a softened silica cladding (drawing temperature of \(1950^{\circ}\mathrm{C}\) ), which acts as a crucible to retain the fibre profile as it is drawn down, as detailed in [45]. A thin layer of calcium oxide (CaO) is included as an interfacial barrier between the core and cladding during the drawing process, which limits dissolution of silica from the cladding into the silicon core and reduces the thermal strain arising from high- temperature processing. The as- drawn SCFs have a poly- crystalline core material with uniform core/cladding diameters of \(12\mu \mathrm{m} / 125\mu \mathrm{m}\) . To improve the crystalline quality and reduce the losses of the as- drawn fibres, we insert the original SCFs into a silica capillary ( \(400\mu \mathrm{m} / 150\mu \mathrm{m}\) inner/outer diameter) and taper this down to have core/cladding diameters of about \(5\mu \mathrm{m} / 125\mu \mathrm{m}\) . The fabrication is realized using a glass processing system (Vytran GPX- 3400- V4), which is widely accessible for heat- polishing, tapering and splicing.  
+
+<|ref|>text<|/ref|><|det|>[[155, 570, 842, 654]]<|/det|>
+Similar to the nano- taper couplers commonly used in planar silicon waveguides [46], nano- spike couplers are fabricated on the SCF facets to improve the coupling to SSMF. The nano- spikes are created by carefully tapering the SCF with the prepared void- gap, which occurs as a result of releasing the tension in the SCFs that is built- in due to the thermal expansion mismatch of the core/cladding materials. Splicing of the tapered SCF with nano- spike couplers on both ends of the tapered SSMFs is achieved by applying a heating power of \(63 \mathrm{W}\) over 7 seconds.  
+
+<|ref|>text<|/ref|><|det|>[[155, 654, 842, 904]]<|/det|>
+Seed comb generation Our seed comb begins with modulating a 1555.72- nm CW signal from a fibre laser using a LiNbO \(_3\) Mach- Zehnder modulator and two phase modulators (Supplementary Figure 2). The 1.6- kHz- linewidth CW source was amplified to 33 dBm by a polarization- maintaining fibre amplifier before launching into the modulators. The modulators transform the CW light into a repeated pulse train with the pulse period corresponding to each modulation cycle [47]. The resulting linear chirp yields pulses with relatively flat spectral envelopes for a tone spacing of 26 GHz. A low phase noise RF source was employed to generate the 26- GHz signal that drives the modulators. Generally, an arbitrary frequency can be used to enable a tunable tone spacing that suits DWDM applications. In our system, the RF frequency was tunable between 22- 26.5 GHz, limited by our frequency synthesizer and the electronic devices. Subsequent linear pulse compression was realized by compensating the spectral phase of the pulses using a reel of 65 m of SSMF, which provides second- order dispersion to compress the pulses to their Fourier- transform limit. The pulse full- width half maximum (FWHM) was measured to be approximately 610 fs using an optical autocorrelator (FEMTOCHROME FR- 103XL) with a Gaussian pulse profile assumed. An erbium- doped fibre amplifier (FA2) was used to amplify the pulse train before reshaping via a nonlinear optical loop mirror (NOLM). The NOLM consists of a 3- dB optical coupler connected to a 105 m Ge- doped HNLF with a dispersion of - 0.38 psnm \(^{- 1}\) km \(^{- 1}\) and a nonlinear coefficient of \(>10 \mathrm{W}^{- 1}\mathrm{km}^{- 1}\) at 1550 nm. The fibre was placed in the NOLM loop, along with a polarization controller and a 5- dB attenuator. The NOLM acts as an intensity discriminator, which transmits the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[155, 91, 841, 120]]<|/det|>
+high- power peak regions at the center of each pulse and reflects the low- power background, providing a pedestal suppression ratio of 17.1 dB and additional pulse compression [48].  
+
+<|ref|>text<|/ref|><|det|>[[155, 119, 842, 258]]<|/det|>
+Linewidth characterization The linewidth characterization is performed by using a delayed self- heterodyne interferometer with 80- km ultra- low loss single- mode fibre, providing about \(4\mu \mathrm{m}\) delay or \(\sim 1.2\mathrm{kHz}\) spectral resolution, necessary for characterizing the narrow linewidth tones. As the seed CW source for the comb is a fibre laser, there is a significant \(1 / \mathrm{f}\) - type (flicker) frequency noise contribution to the frequency noise power spectral density [39]. In this case a Lorentzian profile cannot be assumed for the line shape since the frequency noise power spectral density is not dominated by white frequency noise. As such, we use a pseudo- Voigt profile, rather than a Lorentzian profile, to fit the measured beat note to appropriately account for the \(1 / \mathrm{f}\) - induced linewidth broadening [40]. While the RF driving signal also contributes both white phase ( \(\mathrm{f}^{2}\) frequency noise) and coloured phase noise to the comb noise power spectral density, this is difficult to generalise and has been neglected in this analysis.  
+
+<|ref|>text<|/ref|><|det|>[[155, 258, 842, 315]]<|/det|>
+Simulation We simulate the comb generation scheme with the same properties as the experiment to ensure a close match to our measured results. Pulse train propagation through the SCF was modelled by the generalized nonlinear Schrödinger equation (GNLSE), including TPA and free carrier effects (FCA and FCD) [49]:  
+
+<|ref|>equation<|/ref|><|det|>[[232, 323, 840, 360]]<|/det|>
+\[\frac{\partial E}{\partial z} = -\frac{\alpha_l}{2} E + \sum_{m = 2}^{4}i\frac{i^m\beta_m}{2!}\frac{\partial^mE}{\partial t^m} +i\gamma \left(|E|^2 E + \frac{i}{\omega_0}\frac{\partial}{\partial t} (|E|^2 E)\right) - \frac{\sigma}{2} (1 + i\mu)N_cE \quad (1)\]  
+
+<|ref|>text<|/ref|><|det|>[[155, 362, 840, 391]]<|/det|>
+where \(E\) is the electric field envelope, \(\alpha_{l}\) is the linear attenuation and \(\beta_{m}\) is the \(m\) - th order dispersion parameter. TPA is included as the imaginary component of the nonlinear coefficient \(\gamma\) :  
+
+<|ref|>equation<|/ref|><|det|>[[433, 399, 840, 427]]<|/det|>
+\[\gamma = \frac{2\pi n_2}{\lambda A_{\mathrm{eff}}} +\frac{i\beta_{\mathrm{TPA}}}{2A_{\mathrm{eff}}} \quad (2)\]  
+
+<|ref|>text<|/ref|><|det|>[[155, 429, 842, 486]]<|/det|>
+where \(n_{2}\) is the Kerr coefficient, \(\beta_{\mathrm{TPA}}\) is the TPA parameter and \(A_{\mathrm{eff}}\) is the effective mode area. FCA and FCD are included in the last term in equation 1, where \(\sigma\) is the FCA coefficient and \(\mu = 2k_{c}k_{0} / \sigma\) , with \(k_{0} = 2\pi /\lambda\) and \(k_{c}\) is the free- carrier- induced refractive index change. The magnitude of the FCA and FCD effects are governed by the rate equation for the free carrier density \(N_{c}\) [50]:  
+
+<|ref|>equation<|/ref|><|det|>[[370, 494, 840, 523]]<|/det|>
+\[\frac{\partial N_{c}(z,t)}{\partial t} = \frac{\beta_{\mathrm{TPA}}}{2h\nu_{0}}\frac{|E(z,t)|^{4}}{A_{\mathrm{eff}}} -\frac{N_{c}(z,t)}{\tau} \quad (3)\]  
+
+<|ref|>text<|/ref|><|det|>[[155, 525, 842, 667]]<|/det|>
+where \(\tau\) is the free carrier lifetime. The GNLSE was solved using the split- step Fourier method (SSFM). Dispersion was included up to fourth- order and Raman scattering was neglected due to the short duration of our pulses. To accurately model the tapered SCF device with a varying core diameter, the SCF was separated into three distinct segments of length \(5.1\mathrm{mm}\) , \(3\mathrm{mm}\) and \(9.1\mathrm{mm}\) (input taper, middle and output taper respectively). These corresponded to core diameters of \(1.1\mu \mathrm{m}\) for the small tapered regions and \(5\mu \mathrm{m}\) for the middle region, and the mode properties of each diameter were estimated from COMSOL Multiphysics software simulations. The parameters used in the simulations are listed in Supplementary Table 1 and Figure 1, and were obtained via a combination of the mode simulations and laboratory experiments. The free carrier density and pulse train shown in Fig. 4b was taken from the final step of the SSFM to illustrate the free carrier density reaching steady state.  
+
+<|ref|>sub_title<|/ref|><|det|>[[156, 686, 386, 706]]<|/det|>
+## 5 Data availability  
+
+<|ref|>text<|/ref|><|det|>[[155, 716, 842, 744]]<|/det|>
+The data that support the plots within this paper and other findings of this study are available from the corresponding author on reasonable request.  
+
+<|ref|>sub_title<|/ref|><|det|>[[156, 764, 416, 784]]<|/det|>
+## 6 Acknowledgements  
+
+<|ref|>text<|/ref|><|det|>[[155, 794, 842, 822]]<|/det|>
+We acknowledge financial support from EPSRC grants EP/P000940/1, EP/R041792/1, EP/L015455/1 and EP/V007734/1.  
+
+<|ref|>sub_title<|/ref|><|det|>[[156, 842, 450, 863]]<|/det|>
+## 7 Author Contributions  
+
+<|ref|>text<|/ref|><|det|>[[155, 872, 842, 901]]<|/det|>
+L.S., A.C.P. and Z.L. conceived the experiment. U.J.G. fabricated the silicon core fibre preforms and T.W.H and J.B. developed the drawing process and drew the preform into a fibre. H.R., L.S. and A.C.P
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[155, 91, 842, 202]]<|/det|>
+developed and implemented the tapering and splicing recipe used to splice the silicon core fibre to the SSMF, and developed simulations to estimate the fibre properties. C.D. and Z.L. built the electro-optic comb generator. R.S. and Z.L. designed the experimental setup, built the nonlinear optical loop mirror and measured the parametrically broadened spectrum. A.M.H. provided the highly nonlinear fibre used in the nonlinear optical loop mirror, and simulated both the statistical coherence analysis and nonlinear pulse evolution of the SCF comb. R.S., H.R., A.M.H. and Z.L. developed the simulation used to estimate the impact of two- photon absorption and free carriers in the system. All authors discussed and collaborated on the manuscript.  
+
+<|ref|>sub_title<|/ref|><|det|>[[155, 222, 559, 243]]<|/det|>
+## 8 Competing Interests Statement  
+
+<|ref|>text<|/ref|><|det|>[[156, 253, 447, 267]]<|/det|>
+The authors declare no competing interests.  
+
+<|ref|>sub_title<|/ref|><|det|>[[156, 287, 280, 307]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[155, 312, 844, 920]]<|/det|>
+[1] Diddams, S. A., Vahala, K. & Udem, T. Optical frequency combs: coherently uniting the electromagnetic spectrum. Science 369 (2020). [2] Lundberg, L. et al. Phase- coherent lightwave communications with frequency combs. Nature Communications 11, 201 (2020). URL https://doi.org/10.1038/s41467- 019- 14010- 7. [3] Picqué, N. & Hänsch, T. W. Frequency comb spectroscopy. Nature Photonics 13, 146- 157 (2019). [4] Fortier, T. M. et al. Generation of ultrastable microwaves via optical frequency division. Nature Photonics 5, 425- 429 (2011). [5] Deakin, C. & Liu, Z. Dual frequency comb assisted analog- to- digital conversion. Optics Letters 45, 173- 176 (2020). [6] Rueda, A., Sedlmeir, F., Kumari, M., Leuchs, G. & Schwefel, H. G. L. Resonant electro- optic frequency comb. Nature 568, 378- 381 (2019). URL https://doi.org/10.1038/s41586- 019- 1110- x. [7] Kippenberg, T. J., Gaeta, A. L., Lipson, M. & Gorodetsky, M. L. Dissipative kerr solitons in optical microresonators. Science 361, eaan8083 (2018). URL https://www.science.org/doi/abs/10.1126/science.aan8083. https://www.science.org/doi/pdf/10.1126/science.aan8083. [8] Ataie, V., Myslivets, E., Kuo, B. P.- P., Alic, N. & Radic, S. Spectrally equalized frequency comb generation in multistage parametric mixer with nonlinear pulse shaping. Journal of Lightwave Technology 32, 840- 846 (2014). [9] Supradeepa, V. R. & Weiner, A. M. Bandwidth scaling and spectral flatness enhancement of optical frequency combs from phase- modulated continuous- wave lasers using cascaded four- wave mixing. Opt. Lett. 37, 3066- 3068 (2012). URL http://www.osapublishing.org/ol/abstract.cfm?URI=ol- 37- 15- 3066. [10] Moss, D. J., Morandotti, R., Gaeta, A. L. & Lipson, M. New cmos- compatible platforms based on silicon nitride and hydex for nonlinear optics. Nature photonics 7, 597- 607 (2013). [11] Gaeta, A. L., Lipson, M. & Kippenberg, T. J. Photonic- chip- based frequency combs. nature photonics 13, 158- 169 (2019). [12] Marin- Palomo, P. et al. Microresonator- based solitons for massively parallel coherent optical communications. Nature 546, 274- 279 (2017). URL https://doi.org/10.1038/nature22387. [13] Torres- Company, V. et al. Laser frequency combs for coherent optical communications. Journal of Lightwave Technology 37, 1663- 1670 (2019). [14] Calo, C. et al. Single- section quantum well mode- locked laser for 400 Gb/s SSB- OFDM transmission. Opt. Express 23, 26442- 26449 (2015). URL http://www.opticsexpress.org/abstract.cfm?URI=oe- 23- 20- 26442. [15] Metcalf, A. J., Torres- Company, V., Leaird, D. E. & Weiner, A. M. High- power broadly tunable electrooptic frequency comb generator. IEEE Journal of Selected Topics in Quantum Electronics 19, 231- 236 (2013). [16] Hu, H. et al. Single- source chip- based frequency comb enabling extreme parallel data transmission. Nature Photonics 12, 469- 473 (2018). URL https://doi.org/10.1038/s41566- 018- 0205- 5.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[152, 90, 844, 134]]<|/det|>
+[17] Pelusi, M. D. et al. Applications of highly- nonlinear chalcogenide glass devices tailored for high- speed all- optical signal processing. IEEE Journal of Selected Topics in Quantum Electronics 14, 529- 539 (2008).  
+
+<|ref|>text<|/ref|><|det|>[[154, 137, 842, 180]]<|/det|>
+[18] Kuyken, B., Billet, M., Leo, F., Yvind, K. & Pu, M. Octave- spanning coherent supercontinuum generation in an algaas- on- insulator waveguide. Opt. Lett. 45, 603- 606 (2020). URL http://ol.osa.org/abstract.cfm?URI=ol- 45- 3- 603.  
+
+<|ref|>text<|/ref|><|det|>[[154, 183, 842, 226]]<|/det|>
+[19] Son, G., Han, S., Park, J., Kwon, K. & Yu, K. High- efficiency broadband light coupling between optical fibers and photonic integrated circuits. Nanophotonics 7, 1845- 1864 (2018). URL https://doi.org/10.1515/nanoph- 2018- 0075.  
+
+<|ref|>text<|/ref|><|det|>[[154, 230, 842, 259]]<|/det|>
+[20] Kuo, B. P. et al. Wideband parametric frequency comb as coherent optical carrier. Journal of Lightwave Technology 31, 3414- 3419 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[154, 263, 842, 305]]<|/det|>
+[21] Puttnam, B. J. et al. 2.15 pb/s transmission using a 22 core homogeneous single- mode multi- core fiber and wideband optical comb. In 2015 European Conference on Optical Communication (ECOC), 1- 3 (IEEE, 2015).  
+
+<|ref|>text<|/ref|><|det|>[[154, 309, 842, 352]]<|/det|>
+[22] Temprana, E. et al. Overcoming kerr- induced capacity limit in optical fiber transmission. Science 348, 1445- 1448 (2015). URL https://science.sciencemag.org/content/348/6242/1445. https://science.sciencemag.org/content/348/6242/1445. full.pdf.  
+
+<|ref|>text<|/ref|><|det|>[[154, 355, 842, 384]]<|/det|>
+[23] Hu, H., Ataie, V., Myslivets, E. & Radic, S. Optical comb assisted ofdm rf receiver. Journal of Lightwave Technology 37, 1280- 1287 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[154, 388, 842, 416]]<|/det|>
+[24] Kuo, B. P.- P., Fini, J. M., Gruner- Nielsen, L. & Radic, S. Dispersion- stabilized highly- nonlinear fiber for wideband parametric mixer synthesis. Optics Express 20, 18611- 18619 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[154, 420, 842, 463]]<|/det|>
+[25] Ren, H. et al. Nonlinear optical properties of polycrystalline silicon core fibers from telecom wavelengths into the mid- infrared spectral region. Opt. Mater. Express 9, 1271- 1279 (2019). URL http://www.osapublishing.org/ome/abstract.cfm?URI=ome- 9- 3- 1271.  
+
+<|ref|>text<|/ref|><|det|>[[154, 467, 842, 495]]<|/det|>
+[26] Wu, D. et al. Four- wave mixing- based wavelength conversion and parametric amplification in submicron silicon core fibers. IEEE Journal of Selected Topics in Quantum Electronics 27, 1- 11 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[154, 499, 842, 541]]<|/det|>
+[27] Wu, D. et al. Net optical parametric gain in a submicron silicon core fiber pumped in the telecom band. APL Photonics 4, 086102 (2019). URL https://doi.org/10.1063/1.5103272. https://doi.org/10.1063/1.5103272.  
+
+<|ref|>text<|/ref|><|det|>[[154, 545, 842, 574]]<|/det|>
+[28] Huang, M. et al. Fiber integrated wavelength converter based on a silicon core fiber with a nano- spike coupler. IEEE Photonics Technology Letters 31, 1561- 1564 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[154, 578, 842, 606]]<|/det|>
+[29] Tetsumoto, T. et al. Optically referenced 300 ghz millimetre- wave oscillator. Nature Photonics 1- 7 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[154, 611, 808, 626]]<|/det|>
+[30] Millot, G. et al. Frequency- agile dual- comb spectroscopy. Nature Photonics 10, 27- 30 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[154, 630, 840, 658]]<|/det|>
+[31] Deakin, C. & Liu, Z. Noise and distortion analysis of dual frequency comb photonic rf channelizers. Optics Express 28, 39750- 39769 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[154, 663, 842, 691]]<|/det|>
+[32] Gibson, U. J., Wei, L. & Ballato, J. Semiconductor core fibres: materials science in a bottle. Nature Communications 12, 1- 15 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[154, 696, 842, 724]]<|/det|>
+[33] Franz, Y. et al. Material properties of tapered crystalline silicon core fibers. Opt. Mater. Express 7, 2055- 2061 (2017). URL http://www.osapublishing.org/ome/abstract.cfm?URI=ome- 7- 6- 2055.  
+
+<|ref|>text<|/ref|><|det|>[[154, 728, 842, 770]]<|/det|>
+[34] Ren, H. et al. Tapered silicon core fibers with nano- spikes for optical coupling via spliced silica fibers. Opt. Express 25, 24157- 24163 (2017). URL http://www.opticsexpress.org/abstract.cfm?URI=oe- 25- 20- 24157.  
+
+<|ref|>text<|/ref|><|det|>[[154, 775, 842, 803]]<|/det|>
+[35] Torres- Company, V., Lancis, J. & Andrés, P. Lossless equalization of frequency combs. Opt. Lett. 33, 1822- 1824 (2008). URL http://ol.osa.org/abstract.cfm?URI=ol- 33- 16- 1822.  
+
+<|ref|>text<|/ref|><|det|>[[154, 808, 842, 836]]<|/det|>
+[36] Okada, K., Hashimoto, K., Shibata, T. & Nagaki, Y. Novel method for high resolution measurement of laser output spectrum. Electronics Letters 16, 630 (1980).  
+
+<|ref|>text<|/ref|><|det|>[[154, 840, 842, 896]]<|/det|>
+[37] Skehan, J. C., Naveau, C., Schroder, J. & Andrekson, P. Widely tunable, low linewidth, and high power laser source using an electro- optic comb and injection- locked slave laser array. Opt. Express 29, 17077- 17086 (2021). URL http://www.osapublishing.org/oe/abstract.cfm?URI=oe- 29- 11- 17077.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[154, 90, 844, 133]]<|/det|>
+[38] Gordon, J. P. & Mollenauer, L. F. Phase noise in photonic communications systems using linear amplifiers. Opt. Lett. 15, 1351- 1353 (1990). URL http://www.osapublishing.org/ol/abstract.cfm?URI=ol- 15- 23- 1351.  
+
+<|ref|>text<|/ref|><|det|>[[155, 137, 844, 180]]<|/det|>
+[39] Horak, P. & Loh, W. H. On the delayed self- heterodyne interferometric technique for determining the linewidth of fiber lasers. Opt. Express 14, 3923- 3928 (2006). URL http://www.opticsexpress.org/abstract.cfm?URI=oe- 14- 9- 3923.  
+
+<|ref|>text<|/ref|><|det|>[[155, 184, 843, 226]]<|/det|>
+[40] Stephan, G. M., Tam, T. T., Blin, S., Besnard, P. & Tétu, M. Laser line shape and spectral density of frequency noise. Phys. Rev. A 71, 043809 (2005). URL https://link.aps.org/doi/10.1103/PhysRevA.71.043809.  
+
+<|ref|>text<|/ref|><|det|>[[155, 230, 843, 272]]<|/det|>
+[41] Yin, L. & Agrawal, G. P. Impact of two- photon absorption on self- phase modulation in silicon waveguides. Opt. Lett. 32, 2031- 2033 (2007). URL http://www.osapublishing.org/ol/abstract.cfm?URI=ol- 32- 14- 2031.  
+
+<|ref|>text<|/ref|><|det|>[[155, 277, 843, 305]]<|/det|>
+[42] Griffith, A. G. et al. Silicon- chip mid- infrared frequency comb generation. Nature Communications 6, 6299 (2015). URL https://doi.org/10.1038/ncomms7299.  
+
+<|ref|>text<|/ref|><|det|>[[155, 310, 843, 365]]<|/det|>
+[43] Wei, L. et al. Optoelectronic fibers via selective amplification of in- fiber capillary instabilities. Advanced Materials 29, 1603033 (2017). URL https://onlinelibrary.wiley.com/doi/abs/10.1002/adma.201603033. https://onlinelibrary.wiley.com/doi/pdf/10.1002/adma.201603033.  
+
+<|ref|>text<|/ref|><|det|>[[155, 369, 843, 412]]<|/det|>
+[44] Sorgard, T., Hawkins, T., Ballato, J., Österberg, U. L. & Gibson, U. J. All- optical high- speed modulation of thz transmission through silicon core optical fibers. Opt. Express 29, 3543- 3552 (2021). URL http://www.osapublishing.org/oe/abstract.cfm?URI=oe- 29- 3- 3543.  
+
+<|ref|>text<|/ref|><|det|>[[155, 416, 843, 459]]<|/det|>
+[45] Nordstrand, E. F., Dibbs, A. N., Eraker, A. J. & Gibson, U. J. Alkaline oxide interface modifiers for silicon fiber production. Opt. Mater. Express 3, 651- 657 (2013). URL http://www.osapublishing.org/oe/abstract.cfm?URI=oe- 3- 5- 651.  
+
+<|ref|>text<|/ref|><|det|>[[155, 463, 843, 506]]<|/det|>
+[46] Wood, M., Sun, P. & Reano, R. M. Compact cantilever couplers for low- loss fiber coupling to silicon photonic integrated circuits. Opt. Express 20, 164- 172 (2012). URL http://www.opticsexpress.org/abstract.cfm?URI=oe- 20- 1- 164.  
+
+<|ref|>text<|/ref|><|det|>[[155, 510, 843, 552]]<|/det|>
+[47] Wu, R., Supradeepa, V., Long, C. M., Leaird, D. E. & Weiner, A. M. Generation of very flat optical frequency combs from continuous- wave lasers using cascaded intensity and phase modulators driven by tailored radio frequency waveforms. Optics letters 35, 3234- 3236 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[155, 556, 843, 598]]<|/det|>
+[48] Smith, K., Doran, N. J. & Wigley, P. G. J. Pulse shaping, compression, and pedestal suppression employing a nonlinear- optical loop mirror. Opt. Lett. 15, 1294- 1296 (1990). URL http://ol.osa.org/abstract.cfm?URI=ol- 15- 22- 1294.  
+
+<|ref|>text<|/ref|><|det|>[[155, 603, 843, 632]]<|/det|>
+[49] Yin, L., Lin, Q. & Agrawal, G. P. Soliton fission and supercontinuum generation in silicon waveguides. Opt. Lett. 32, 391- 393 (2007). URL http://ol.osa.org/abstract.cfm?URI=ol- 32- 4- 391.  
+
+<|ref|>text<|/ref|><|det|>[[155, 636, 843, 664]]<|/det|>
+[50] Yin, L., Lin, Q. & Agrawal, G. P. Dispersion tailoring and soliton propagation in silicon waveguides. Opt. Lett. 31, 1295- 1297 (2006). URL http://ol.osa.org/abstract.cfm?URI=ol- 31- 9- 1295.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[113, 288, 882, 586]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[114, 599, 882, 706]]<|/det|>
+<center>Figure 1: Fabrication process flow of the all-fibre heterogeneously-integrated parametric mixer. (a) Tapered SCF with core/cladding diameters of about \(5\mu \mathrm{m} / 125\mu \mathrm{m}\) . (b) Heating and tapering process to make the first void gap in the fibre core. (c) Heating and tapering to make the second void gap in the fibre core. (d) fibre tapering process to scale down core size and collapse the void gap to form the nano-spike coupler. (e) Cleave at the center of the void to remove one side of the taper. (f) Splice the SCF nanospike to a tapered SSMF. (g) Employ a polymer tube to mechanically support the SSMF-SCF connection, keeping the fibre straight and tapering the other end of the SCF. (h) Cleave the other end and (i) splice to another tapered SSMF. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[123, 240, 875, 592]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[114, 658, 883, 750]]<|/det|>
+<center>Figure 2: Fully integrated SCF-based parametric comb generation scheme. An electro-optic frequency comb is generated to act as a seed source for the parametric mixing stage. A pulse re-shaping stage compresses the comb pulse train to maximise the pulse peak power and enhance the parametric mixing efficiency. The optical output is then amplified and launched into a 17 mm sample of fully-integrated SCF, where parametric nonlinearities cause comb broadening. Both ends of the SCF are tapered and spliced to tapered single-mode fibre, with nano-spike couplers to facilitate coupling between the heterogenous fibre cores. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[194, 140, 800, 710]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[114, 722, 883, 860]]<|/det|>
+<center>Figure 3: C-band parametric frequency comb generation using SCF. (a) An electro-optic frequency comb with 26 GHz line spacing (orange) was temporally reshaped and launched into a 17mm sample of SCF at an average power of 32 dBm. The optical spectrum of the broadened parametric comb (blue) was obtained at a resolution of 0.02 nm, achieving 143 tones across a 30.0 nm bandwidth. A 10 MHz resolution BOSA was used to obtain close-in traces at different points within the comb bandwidth (Fig.(b)-(d)). The dashed line shows the instrument noise floor. Results in (c) and (d) are both limited by the instrument noise floor. (e) The linewidth of the parametric comb was measured using the delayed self-heterodyne interferometer (DSHI) method with a 80-km delay line and pseudo-Voigt profile fitting. The pseudo-Voigt fitting curves are shown for the two measured beat notes at the extremities of the SCF comb bandwidth ((f) and (g)). </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[192, 188, 804, 690]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[114, 704, 883, 811]]<|/det|>
+<center>Figure 4: Parametric frequency comb simulation including the effects of two-photon absorption and free carriers within the SCF. (a) Nonlinear loss induced by TPA and FCs as a function of average input power. (b) Free carrier density in the silicon core versus time (black), and the optical pulse train power versus time (red). Subsequent pulses in the 26 GHz pulse train increase the free carrier density before complete recombination can occur, reaching a steady state over several nanoseconds. (c) Simulated optical spectrum at the output of the SCF mixer, at 27 dBm and 32 dBm input power (top and middle respectively) and (d) at 32 dBm without free carriers. TPA is included in all three spectra, and insertion loss is neglected. </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[44, 93, 765, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 456, 178]]<|/det|>
+SupplementaryinfoAllfibreSCFcomb.pdf SupplementarySCFSpectrogramvideo.mp4
+
+<--- Page Split --->

preprint/preprint__2be28d29fac1837acb16f3a122ee991594a1cc474dcaae29e51328cc54c6d461/images_list.json

@@ -0,0 +1,92 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1. AILA Framework and Implementation. (A) System Architecture of the Artificially Intelligent Laboratory Assistant (AILA). Dotted lines indicate adaptive information flow governed by AILA’s decision-making, and solid lines represent deterministic information pathways with predefined routing protocols. (B) Image of the atomic force microscope (AFM) experimental setup showing key hardware components and control interfaces. (C) Representative demonstration of AILA’s operation: raw transcript of a user query and AILA’s unedited response sequence, showing the system’s query interpretation, task planning, and execution capabilities.",
+    "footnote": [],
+    "bbox": [
+      [
+        118,
+        80,
+        880,
+        590
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "2.2 AFMBench: tasks for evaluating the AILA framework",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        100,
+        875,
+        576
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3. Comparative Performance Analysis of Language Models on AFMBench. A. Venn diagrams showing accuracy (left) and efficiency (right) metrics for GPT-4o across documentation, analysis, and calculation tasks. Numbers indicate percentage accuracy and processing time in seconds. B. Corresponding performance metrics for GPT-3.5-turbo-0125, highlighting domain-specific competencies and processing speeds. C. A horizontal bar chart comparing tool and agent utilization efficiency between models is expressed as a percentage of successful engagements.",
+    "footnote": [],
+    "bbox": [
+      [
+        135,
+        85,
+        840,
+        585
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4. Error Mode Distribution in Model Performance. Error patterns between GPT-4o (top) and GPT-3.5-turbo-0125 (bottom). Segments represent a proportional distribution of error types: Instruction adherence (dark blue), agent/tool selection (gray), and code generation (light",
+    "footnote": [],
+    "bbox": [
+      [
+        366,
+        464,
+        668,
+        780
+      ]
+    ],
+    "page_idx": 10
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5. Autonomous AFM experiments. A. Evolution of AFM image quality under varying PID parameters. The left panels show topographic images of the calibration grid; the right panels display corresponding line scan profiles (solid: trace, dashed: retrace). SSIM scores quantify trace-retrace correlation, with higher values indicating superior imaging quality. Optimal parameters (P: 249, I: 8957, D: 26) achieve \\(\\mathrm{SSIM} = 0.818\\) . B. Large-area scan demonstrating consistent imaging quality using optimized parameters across multiple grid features. C. Convergence plot showing genetic algorithm optimization efficiency. Red circles: maximum SSIM; black circles: mean SSIM per generation.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        85,
+        875,
+        590
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6. Autonomous AFM experiments. A. High-resolution HOPG imaging demonstrating baseline artifact challenges. Top panels: topographic images at different Z ranges; bottom panels: corresponding line profiles revealing surface features. B. Left: HOPG images obtained using setpoints of 0.2 V and 0.4 V, both manually captured and taken by AILA with GPT-4o. Right: Plot showing the relationship between setpoint and average friction.",
+    "footnote": [],
+    "bbox": [
+      [
+        123,
+        273,
+        866,
+        720
+      ]
+    ],
+    "page_idx": 13
+  }
+]

preprint/preprint__2be28d29fac1837acb16f3a122ee991594a1cc474dcaae29e51328cc54c6d461/preprint__2be28d29fac1837acb16f3a122ee991594a1cc474dcaae29e51328cc54c6d461.mmd

@@ -0,0 +1,416 @@
+
+# Autonomous Microscopy Experiments through Large Language Model Agents  
+
+N M Anoop Krishnan krishnan@iitd.ac.in  
+
+Indian Institute of Technology Delhi https://orcid.org/0000- 0003- 1500- 4947  
+
+Indrajeet Mandal Indian Institute of Technology Delhi  
+
+Jitendra Soni Indian Institute of Technology Delhi  
+
+Mohd Zaki Indian Institute of Technology Delhi https://orcid.org/0000- 0002- 4551- 3470  
+
+Morten Smedskjaer Aalborg University https://orcid.org/0000- 0003- 0476- 2021  
+
+Katrin Wondraczek Leibniz Institute of Photonic Technology  
+
+Lothar Wondraczek University of Jena  
+
+Nitya Nand Gosvami Indian Institute of Technology Delhi  
+
+## Article  
+
+Keywords: self- driving laboratory, atomic force microscopy, benchmarking, AI agents  
+
+Posted Date: December 19th, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 5600537/v1  
+
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.
+
+<--- Page Split --->
+
+Version of Record: A version of this preprint was published at Nature Communications on October 14th, 2025. See the published version at https://doi.org/10.1038/s41467-025-64105-7.
+
+<--- Page Split --->
+
+# Autonomous Microscopy Experiments through Large Language Model Agents  
+
+Indrajeet Mandal \(^{1}\) , Jitendra Soni \(^{2}\) , Mohd Zaki \(^{3}\) , Morten M. Smedskjaer \(^{4}\) , Katrin Wondraczek \(^{5}\) , Lothar Wondraczek \(^{6}\) , Nitya Nand Gosvami \(^{1,2,7*}\) , N. M. Anoop Krishnan \(^{1,3,7,*}\)  
+
+\(^{1}\) School of Interdisciplinary Research, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India. \(^{2}\) Department of Materials Science and Engineering, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India \(^{3}\) Department of Civil Engineering, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016,India. \(^{4}\) Department of Chemistry and Bioscience, Aalborg University, Aalborg, Denmark \(^{5}\) Leibniz Institute of Photonic Technology, 07745 Jena, Germany \(^{6}\) Otto Schott Institute of Materials Research, University of Jena, 07743 Jena, Germany \(^{7}\) Yardi School of Artificial Intelligence, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India  
+
+\(^{*}\) Corresponding Authors: ngosvami@iitd.ac.in (N. N. G.); krishnan@iitd.ac.in (N. M. A. K.) Fax: +91- 11- 2658- 1117; Tel: +91- 11- 2659- 1223  
+
+Keywords: self- driving laboratory, atomic force microscopy, benchmarking, AI agents  
+
+## Abstract  
+
+The emergence of large language models (LLMs) has accelerated the development of self- driving laboratories (SDLs) for materials research. Despite their transformative potential, current SDL implementations rely on rigid, predefined protocols that limit their adaptability to dynamic experimental scenarios across different labs. A significant challenge persists in measuring how effectively AI agents can replicate the adaptive decision- making and experimental intuition of expert scientists. Here, we introduce AILA (Artificially Intelligent Lab Assistant), a framework that automates atomic force microscopy (AFM) through LLM- driven agents. Using AFM as an experimental testbed, we develop AFMBench—a comprehensive evaluation suite that challenges AI agents based on language models like GPT- 4o and GPT- 3.5 to perform tasks spanning the scientific workflow: from experimental design to results analysis. Our systematic assessment shows that state- of- the- art language models struggle even with basic tasks such as documentation retrieval, leading to a significant decline in performance in multi- agent coordination scenarios. Further, we observe that LLMs exhibit a tendency to not adhere to instructions or even divagate to additional tasks beyond the original request, raising serious concerns regarding safety alignment aspects of AI agents for SDLs. Finally, we demonstrate the application of AILA on increasingly complex experiments open- ended experiments: automated AFM calibration, high- resolution feature detection, and mechanical property measurement. Our findings emphasize the necessity for stringent
+
+<--- Page Split --->
+
+benchmarking protocols before deploying AI agents as laboratory assistants across scientific disciplines.  
+
+## 1. Introduction  
+
+Scientific experimentation demands exceptional domain expertise, from exploration or hypothesis- driven experimental design to precision execution and rigorous data analysis. This complexity creates bottlenecks in scientific discovery, particularly as experimental techniques grow increasingly sophisticated. The advent of large language models (LLMs) has propelled the development of self- driving laboratories (SDLs) that integrate diverse information sources for automated planning<sup>1</sup> and experimentation. AI- agents<sup>2,3</sup> and SDLs have already achieved several feats in materials or molecular discovery<sup>4- 6</sup>, chemistry research<sup>7</sup>, and inorganic materials synthesis. The promise of SDLs toward achieving sustainable development<sup>8</sup> has resulted in enormous efforts to harness their potential in high- throughput experimentation and discovery<sup>9</sup>. Efforts to streamline SDLs have resulted in orchestration architectures such as ChemOS<sup>10</sup>. Additionally, it has been demonstrated that the capability of SDLs can be enhanced by a human- in- the- loop framework that handles disambiguation, thereby enabling better planning and execution<sup>11</sup>. While early demonstrations of LLM- based lab assistants showed promise in chemistry and materials science<sup>1- 3</sup>, their operational reliability remains largely uncharacterized beyond specific applications or repetitive use cases with predetermined protocols<sup>15- 19</sup>.  
+
+Current research predominantly addresses well- documented or predefined protocols and single- objective tasks, failing to capture the intricate interplay between experimental planning, multi- tool coordination, and result interpretation or online intervention<sup>10</sup>. While recent investigations incorporating planning elements have demonstrated success in achieving specific experimental objectives, they have not systematically evaluated SDL reliability across the broader spectrum of laboratory automation tasks<sup>15,16</sup>. Although several studies have benchmarked LLMs<sup>17,18,20- 23</sup> and vision language models (VLMs)<sup>13- 16</sup> through question- answer protocols to assess their potential as materials research co- pilots, a crucial knowledge gap persists: understanding how these AI systems handle novel experimental scenarios and their fundamental limitations.  
+
+To address this challenge, we here introduce AILA (Artificially Intelligent Lab Assistant), an LLM- powered framework augmented with specialized tools. We selected scanning probe microscopy<sup>20</sup>, specifically atomic force microscopy (AFM), as our experimental testbed, given its inherent complexity and broad applicability in materials research. There have been several efforts to automate microscopy techniques using AI and human- in- the- loop approaches due to their extensive applications in materials characterization<sup>28- 30,30,31,31- 35</sup>. These efforts focus exclusively on advancing specific operational aspects, such as analysing moving objects or optimizing illumination conditions, with an emphasis on improving individual steps within the broader experimental protocol<sup>30,32,35</sup>. AFM operation demands expertise across multiple domains—from probe calibration to parameter optimization and data interpretation—making
+
+<--- Page Split --->
+
+it an ideal platform for evaluating AI agents' ability to manage sophisticated experimental workflows.  
+
+Using AFM as the model system, we probe AILA's capabilities through AFMBench on five critical aspects of scientific automation: experimental workflow design, multi- tool coordination, decision- making, execution of open- ended experiments, and data analysis. Our systematic evaluation reveals key failure modes and areas requiring enhancement. We demonstrate AILA's practical utility through three increasingly complex case studies: automated microscope calibration, high- resolution graphene step- edge imaging, and load- dependent roughness analysis on highly oriented pyrolytic graphite (HOPG).  
+
+## 2. Results  
+
+### 2.1 AILA framework  
+
+AILA's architecture prioritizes modularity, enabling seamless integration with diverse experimental and analytical platforms. At its core lies an LLM- powered planner—the framework's cognitive centre—which orchestrates user interactions and coordinates specialized agents (Fig. 1a). For AFM operations, AILA deploys two agents: the AFM Handler Agent (AFM- HA) for experimental control and the Data Handler Agent (DHA) for analysis. The AFM- HA interfaces with a document retrieval system comprising AFM software documentation and a code execution engine that translates Python commands into experimental actions. A Python- based API establishes the hardware- software interface, enabling direct control of the AFM system through vendor- specific protocols (Fig. 1b). The DHA manages image optimization and analysis through dedicated tools: an Image Optimizer that fine- tunes PID parameters for high- fidelity imaging and an Image Analyzer that extracts targeted features from experimental data. For queries beyond agent capabilities, the planner generates alternative approaches or recommended actions. The technical specifications and implementation details of each module are explained in the Methods section.  
+
+To demonstrate AILA's operational workflow, we present a multi- step experiment: acquiring an AFM image of HOPG and extracting its friction and roughness parameters (Fig. 1c). This open- ended task exemplifies real- world complexity, offering multiple solution pathways. Upon receiving the query, AILA dissects it into sequential objectives: image acquisition via AFM- HA followed by DHA- led analysis. AFM- HA retrieves relevant documentation, generates executable code, and captures the image. Following successful acquisition, AILA transitions control to DHA, which directs the Image Analyzer to compute the specified parameters. This orchestrated sequence exemplifies AILA's core strengths: the ability to parse complex natural language queries, develop strategic workflows, and coordinate multiple agents toward achieving experimental objectives.
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1. AILA Framework and Implementation. (A) System Architecture of the Artificially Intelligent Laboratory Assistant (AILA). Dotted lines indicate adaptive information flow governed by AILA’s decision-making, and solid lines represent deterministic information pathways with predefined routing protocols. (B) Image of the atomic force microscope (AFM) experimental setup showing key hardware components and control interfaces. (C) Representative demonstration of AILA’s operation: raw transcript of a user query and AILA’s unedited response sequence, showing the system’s query interpretation, task planning, and execution capabilities. </center>
+
+<--- Page Split --->
+![](images/Figure_unknown_0.jpg)
+
+<center>2.2 AFMBench: tasks for evaluating the AILA framework </center>  
+
+Figure 2. Task Distribution and Module Utilization in AFMBench. A. Pie charts showing the distribution of tool requirements (left, single vs. multiple) and agent requirements (right, single vs. multiple) across benchmark tasks. B. Operation complexity categorization showing the proportion of basic versus advanced tasks. C. Horizontal bar chart quantifying module engagement frequency across all tasks, demonstrating utilization patterns of each tool and agent. D. Venn diagram illustrating the overlap between documentation, analysis, and calculation tasks. E. Representative examples of basic (left) and advanced (right) tasks, demonstrating increasing complexity in experimental workflows.  
+
+AFMBench comprises 100 expertly curated experimental tasks (see S3.1 in Supplementary Information for a few examples of tasks; all the tasks are available in the GitHub repo) manually designed to rigorously evaluate autonomous AFM operations across multiple dimensions of complexity. Unlike conventional LLM benchmarks or simulation- based evaluations, each AFMBench task demands physical execution on AFM hardware, introducing real- world temporal constraints and experimental variability. Analysis of the task architecture reveals distinct patterns in resource utilization and operational complexity. In Figure 2A, tool coordination requirements highlight a systematic preference for sophisticated workflows, with
+
+<--- Page Split --->
+
+\(69\%\) of tasks demanding multi- tool integration, while \(31\%\) operate through single- tool protocols. Agent deployment analysis reveals a distribution: \(83\%\) of operations utilize single- agent protocols, while \(17\%\) require multi- agent coordination—enabling evaluation of both targeted expertise and system- wide integration capabilities.  
+
+In Figure 2B, the operational landscape is divided into two primary complexity tiers: basic operations ( \(56\%\) ) encompassing fundamental microscopy tasks and advanced procedures ( \(44\%\) ) requiring more sophisticated experimental workflows (for example questions see Figure 2E). Core system components—the AFM Handler, Document Retriever, and Code Executor—demonstrate maximum engagement, each activating in 66 distinct tasks. The Data Handler Agent and Image Analyzer exhibit selective activation patterns (52 and 48 tasks, respectively), while the Image Optimizer engages exclusively in critical parameter optimization scenarios (4 tasks).  
+
+Task distribution across functional domains reveals three primary clusters: documentation (50 standalone tasks), analysis (14 tasks), and calculation (10 tasks) (see Figure 2D). A significant overlap between these domains emerges through integrated tasks that combine multiple functional requirements, reflecting the interconnected nature of experimental workflows. This carefully constructed distribution enables systematic evaluation of AI systems across a spectrum of experimental complexity—from basic instrument control to advanced multi- step procedures requiring mathematical reasoning and dynamic decision- making—effectively mirroring the cognitive hierarchy of expert atomic force microscopists.  
+
+### 2.3 Performance of AI Agents  
+
+Systematic evaluation of AILA using two advanced closed source language models—GPT- 4o and GPT- 3.5- turbo- 0125—unveils distinctive execution patterns and operational efficacies. GPT- 4o exhibits exceptional proficiency in documentation- centric operations, achieving a \(92\%\) success rate, complemented by robust execution in analysis ( \(71\%\) ) and computational tasks ( \(70\%\) ). The model’s strength lies in its ability to navigate interconnected workflows: \(80\%\) success in merged documentation- analysis procedures and \(60\%\) in documentation- computation sequences. These metrics highlight GPT- 4o’s capacity to replicate the integrative reasoning characteristic of expert microscopists.
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Figure 3. Comparative Performance Analysis of Language Models on AFMBench. A. Venn diagrams showing accuracy (left) and efficiency (right) metrics for GPT-4o across documentation, analysis, and calculation tasks. Numbers indicate percentage accuracy and processing time in seconds. B. Corresponding performance metrics for GPT-3.5-turbo-0125, highlighting domain-specific competencies and processing speeds. C. A horizontal bar chart comparing tool and agent utilization efficiency between models is expressed as a percentage of successful engagements. </center>  
+
+In marked contrast, GPT- 3.5- turbo- 0125 displays compartmentalized strengths: substantial accuracy in standalone documentation (72%), analytical interpretation (71%), and mathematical operations (40%). However, its performance degrades significantly when confronted with multi- domain challenges, registering null success rates in tasks demanding simultaneous expertise across domains. This limitation suggests insufficient development of cross- functional reasoning capabilities essential for autonomous experimentation.  
+
+Temporal efficiency metrics augment these distinctions. GPT- 4o maintains steady execution speeds: documentation tasks took an average of 12.23 seconds, analytical procedures 9.25
+
+<--- Page Split --->
+
+seconds, and computational operations 11.06 seconds. GPT- 3.5- turbo- 0125 demonstrates irregular processing intervals: 8.88 seconds for documentation, 13.48 seconds for analysis, and 3.26 seconds for calculations. This temporal variability indicates underlying differences in task comprehension and execution strategies between the models.  
+
+Component utilization analysis reinforces these observations. GPT- 4o achieves consistently elevated engagement across system modules, particularly excelling in AFM operation control and document interpretation. These results highlight the fundamental importance of model architecture in autonomous experimental platforms, with GPT- 4o's advanced integrative capabilities positioning it as the superior choice for sophisticated experimental automation.  
+
+### 2.4 Error Analysis Reveals Model-Specific Limitations  
+
+Detailed examination of failure cases revealed distinctive error patterns between the two language models, offering insights into their operational limitations. GPT- 4o exhibits a total error rate of \(20\%\) , with errors distributed across three primary categories: code generation \((60\%)\) , agent/tool selection \((25\%)\) , and instruction adherence \((15\%)\) . The predominance of code generation errors suggests challenges in translating conceptual understanding into executable commands despite the model's strong performance in task comprehension.  
+
+GPT- 3.5- turbo- 0125 demonstrates a markedly higher total error rate of \(45\%\) , with errors concentrated in two categories: code generation \((66.7\%)\) and agent/tool selection \((33.3\%)\) . Notably, the model shows no fundamental query interpretation errors, indicating robust natural language processing capabilities. However, the elevated frequency of code generation errors, coupled with significant agent/tool selection failures, points to underlying deficiencies in translating comprehension into actionable experimental protocols.  
+
+A critical finding emerged regarding GPT4o's instruction adherence. In one of the three recorded errors, GPT- 4o exceeded its designated operational limits, performing actions that were not authorized by the provided guidelines. For instance, it carried out potentially risky tip movements while it was only instructed to scan the surface (see S2.3 in the Supplementary Information). In another case, GPT- 4o was instructed to capture an image and calculate surface friction. While the image was captured correctly, the system failed to analyse it and switches the AFM to the lateral force mode instead of following clear instructions for a specific task (see S2.3 in the Supplementary Information). Instead of staying within the scope of the task, it performed additional actions. Although sometimes the final result may have been correct, the failure to follow instructions highlights concerns about AI- agent behaviour and raises safety risks in automated lab environments. Similar to the observation of "hallucination" in LLMs<sup>36</sup>, these results present a unique challenge—SDLs tend to take arbitrary actions potentially based on memory rather than following the instructions, referred to hereafter as "sleepwalking". These issues are especially critical in sensitive experimental settings, where strict protocol adherence is essential to ensure both equipment safety and the validity of results.  
+
+This error distribution illuminates critical areas for framework enhancement. While GPT- 4o's balanced error profile suggests the need for targeted improvements across multiple domains,
+
+<--- Page Split --->
+
+GPT- 3.5- turbo- 0125's concentrated error patterns indicate fundamental limitations in experimental execution capabilities. These findings underscore the necessity for specialized training in automated experimental systems, particularly focusing on the translation of scientific protocols into executable code sequences.  
+
+### 2.5 Safety Alignment in SDLs  
+
+To understand the safety challenges<sup>37</sup> of AI agents, we evaluate the effectiveness of implementing a safety framework in AILA. First, we establish restricted access protocols for critical AFM functions, coupled with ethical system prompts (see S2.1 in Supplementary Information) that constrain code generation to predefined documentation<sup>38</sup>. Second, we develop strict operational boundaries that permit dynamic code generation solely for image analysis while preventing external software installation or system modifications. Evaluation of the improved protocol demonstrates the effectiveness of these safeguards—AILA appropriately failed when prompted to install external Python libraries. (see S3.3 in Supplementary Information for complete validation logs). These findings underscore the critical importance of robust safety protocols in SDLs, emphasizing the necessity of comprehensive benchmarking and operational guardrails.  
+
+![](images/Figure_4.jpg)
+
+<center>Figure 4. Error Mode Distribution in Model Performance. Error patterns between GPT-4o (top) and GPT-3.5-turbo-0125 (bottom). Segments represent a proportional distribution of error types: Instruction adherence (dark blue), agent/tool selection (gray), and code generation (light </center>
+
+<--- Page Split --->
+
+green). Central numbers indicate total error counts for each model, demonstrating significant differences in overall error rates (20 versus 45 errors, respectively).  
+
+### 2.6 Pushing the Limits of Autonomous Experimentation  
+
+Finally, to demonstrate AILA's capabilities in real- world scenarios, we demonstrate three increasingly complex experimental tasks that typically require expert intervention: automated AFM calibration, high- resolution feature detection, and mechanical property measurement.  
+
+#### 2.6.1 AFM Parameter Optimization  
+
+AFM imaging requires precise calibration of Proportional- Integral- Derivative (PID) gain values, which traditionally demand expert intervention due to the continuous nature of these parameters. This dependency on skilled operators presents a significant barrier to broader AFM adoption. We demonstrate AILA's capability to autonomously optimize these parameters by minimizing the forward- backward scan differential on standard calibration grids. To this end, after loading the calibration sample, AILA was prompted to optimize the imaging parameters (see S4 in Supplementary Information for the complete prompt and output log). A total of 45 images are generated, with 3 images produced in each of the 15 generations. Figure 5A presents experimental AFM data acquired by AILA for the \(1^{\mathrm{st}}\) and \(15^{\mathrm{th}}\) generation of variable PID configurations, with corresponding line scan analyses that quantify trace- retrace symmetry. Initial scans with suboptimal parameters (P: 93- 208, I: 1747- 6623, D: 0- 39) exhibit poor SSIM scores (0.392- 0.768), manifesting as visible distortions in topographic data. Note that a higher SSIM value, closer to 1, indicates a perfect match, while a value of 0 represents no similarity. Through iterative optimization, AILA achieves superior scan quality (SSIM \(>0.81\) ) with optimized parameters (P: 246- 249, I: 8676- 8957, D: 17- 30; see Figure 5A).  
+
+The genetic algorithm's convergence efficiency is demonstrated in Figure 5C, where optimal PID configurations are achieved within 15 generations. Both maximum and mean SSIM values show rapid improvement, stabilizing above 0.8, indicating robust parameter optimization. Figure 5B validates the optimized parameters (P:249, I:8957, D:26) across a larger scan area, maintaining high- quality imaging across multiple grid features.
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Figure 5. Autonomous AFM experiments. A. Evolution of AFM image quality under varying PID parameters. The left panels show topographic images of the calibration grid; the right panels display corresponding line scan profiles (solid: trace, dashed: retrace). SSIM scores quantify trace-retrace correlation, with higher values indicating superior imaging quality. Optimal parameters (P: 249, I: 8957, D: 26) achieve \(\mathrm{SSIM} = 0.818\) . B. Large-area scan demonstrating consistent imaging quality using optimized parameters across multiple grid features. C. Convergence plot showing genetic algorithm optimization efficiency. Red circles: maximum SSIM; black circles: mean SSIM per generation. </center>  
+
+#### 2.6.2 High-Resolution Step-Edge Detection  
+
+Surface characterization through AFM is challenged by noise sources such as thermal drift, mechanical vibrations, and electronic interference<sup>39- 41</sup>, which can obscure subtle topographic features like graphene step edges. In this study, we leverage the advanced analytical capabilities of AILA to address these challenges using highly ordered pyrolytic graphite (HOPG) as a model system. AILA autonomously determines the necessity for baseline correction based on feature size, recognizing that baseline artifacts predominantly affect smaller features. For instance, in the raw image (Figure 6A), the graphene step edge remains indiscernible due to
+
+<--- Page Split --->
+
+baseline distortions. AILA applies a fifth- order polynomial baseline correction to generate the \(1^{\mathrm{st}}\) generation image (Figure 6A), which serves as the foundation for PID gain optimization. Following a process similar to the calibration grid optimization, the image is refined through iterative PID adjustments, resulting in the final optimized image in the \(10^{\mathrm{th}}\) generation, where atomic steps become distinctly visible. The automated optimization process surpasses conventional manual adjustments, offering an enhanced resolution of nanoscale features. Additionally, further analysis of the processed data, including the determination of graphene step height, was facilitated through specific prompts, with the prompts and results detailed in the Supplementary Information S4.  
+
+![](images/Figure_6.jpg)
+
+<center>Figure 6. Autonomous AFM experiments. A. High-resolution HOPG imaging demonstrating baseline artifact challenges. Top panels: topographic images at different Z ranges; bottom panels: corresponding line profiles revealing surface features. B. Left: HOPG images obtained using setpoints of 0.2 V and 0.4 V, both manually captured and taken by AILA with GPT-4o. Right: Plot showing the relationship between setpoint and average friction. </center>  
+
+#### 2.6.3 Load-dependent roughness measurement  
+
+We conduct a comprehensive load- dependent roughness analysis of HOPG. The experiment requires iterative adjustments of AFM parameters, including setting a range of setpoints, capturing images, and analyzing the corresponding friction data. Manually performing this
+
+<--- Page Split --->
+
+procedure is time- intensive, involving parameter modifications, image acquisition, data extraction, and result plotting. To streamline this process, we utilize AILA to automate the experiment.  
+
+AILA was instructed (see S4 in Supplementary materials for the complete prompt and output log) to vary the setpoint voltage from 0.2 V to 1.2 V in increments of 0.2 V. At each setpoint, AILA independently captured the AFM image, calculated the average friction value, and generated the corresponding plot. Figure 6B presents the graph of average friction versus setpoint voltage for both manually obtained and AILA- captured images using the GPT- 4o model. The raw images generated by AILA can be found in the Supplementary Information Figure S2. The entire process was conducted without additional user input regarding figure formatting or parameter settings. Remarkably, AILA autonomously develops the required Python script, executes the experimental protocol, and generates the output, including raw, unedited plots. This automation significantly reduces the time and effort compared to manual execution. The results not only validate the capability of AILA in handling complex AFM experiments but also demonstrate its efficiency in generating reproducible and high- quality outputs for scientific analyses.  
+
+## 3. Discussion  
+
+AILA's evaluation of AFMBench reveals quantifiable distinctions between LLM capabilities in experimental automation. The performance gap between GPT- 4o and GPT- 3.5- turbo- 0125 in cross- domain operations (80% versus 0% success in hybrid tasks) quantifies a key requirement for autonomous laboratories: integrated experimental reasoning. Single- domain competence, while necessary, proves insufficient for replicating expert- level experimental workflows. This disparity in cross- domain integration particularly manifests in complex scenarios, requiring simultaneous optimization of imaging parameters and data analysis—tasks routinely performed by experienced microscopists.  
+
+Error pattern analysis identifies a persistent challenge: protocol- to- execution translation. Code generation errors dominate failure modes (60- 67%) across both architectures, establishing a primary bottleneck in experimental automation. This systematic limitation persists even in advanced models, suggesting architectural constraints beyond mere computational capability. The distribution of error types—particularly the presence of instruction adherence errors (15%) in GPT- 4o versus their absence in GPT- 3.5- turbo- 0125—reveals nuanced differences in model behavior that warrant further investigation regarding the safety and alignment issues associated with the LLMs. It is worth noting that AI- agent based on GPT- 4o, a model known for its reasoning capabilities, exhibit tendencies to perform actions beyond what is instructed, termed as “sleepwalking”. Note that depending on the nature of the LLM, “sleepwalking” could result in potentially harmful outcomes in SDLs when dealing with highly dangerous chemicals or conditions such as high temperature pressure to name a few.  
+
+AILA's modular design establishes quantifiable metrics for autonomous system development. The successful automation of AFM operations validates this architecture for complex
+
+<--- Page Split --->
+
+instrumentation control, demonstrating efficacy in parameter optimization and feature extraction tasks. The measured limitations in tool coordination depending on the LLMs define specific thresholds for improving inter- module communication protocols. Notably, the agent deployment ratio (83:17 single- to- multi- agent) establishes an empirical baseline for balancing specialized and integrated operations—a metric applicable to automation across analytical platforms, from mass spectrometers to X- ray diffractometers.  
+
+These findings suggest specific architectural improvements for next- generation autonomous laboratories. Enhanced integration protocols between specialized agents could address the observed limitations in multi- tool coordination. Similarly, dedicated code generation modules might mitigate the predominant error mode, potentially incorporating specialized scientific programming frameworks.  
+
+## 4. Outlook  
+
+Altogether, AILA demonstrates quantifiable advances in experimental automation through systematic benchmarking. The framework's comprehensive performance metrics in AFM operations establish standards for autonomous laboratory evaluation, while AFMBench introduces reproducible protocols for systematic assessment across experimental domains. Successful execution of complex tasks—from automated image optimization to nanomechanical measurements—validates the framework's capabilities for sophisticated materials characterization. However, the observed tendencies of LLM agents to exceed operational boundaries and perform “sleepwalking” while carrying out the experiments raise significant safety concerns. This behavior, reported for the first time to the best of that authors’ knowledge, emphasize critical areas for development in instruction alignment and operational safety.  
+
+This work’s implications transcend materials characterization. The empirically validated principles—modular architecture (69% multi- tool efficiency), strategic agent deployment (83:17 ratio), and cross- domain integration (80% hybrid task success)—establish design parameters for autonomous laboratories across disciplines. Applications span pharmaceutical screening, environmental monitoring, and process optimization. For instance, the documented success in parameter optimization could translate directly to automated high- throughput drug screening or catalyst discovery platforms. While current limitations in code generation (60% error rate) and tool coordination (31% efficiency) define immediate development targets, these metrics provide clear objectives for advancing autonomous scientific platforms. The path forward requires focused development in three key areas: enhanced cross- domain reasoning capabilities, robust code generation protocols, and sophisticated multi- agent coordination mechanisms. Success in these domains would enable truly autonomous scientific platforms capable of accelerating discovery across the scientific landscape.
+
+<--- Page Split --->
+
+## Methodology  
+
+## AILA  
+
+AILA is constructed utilizing the LangChain software framework, incorporating components such as prompts, LLMs, memory, agents, and tools. AILA uses two categories of prompts: system prompts (see S2.1 in Supplementary Information for the system prompts) and user prompts. System prompts define ethical rules for AILA's interactions and describe the responsibilities assigned to each agent, whereas user prompts are variable inputs provided by end- users. AILA's backbone consists of LLMs, namely GPT- 4o and GPT- 3.5- turbo- 0125, which process user input as strings and provide string- based outputs. These LLMs are stateless, indicating that they do not save conversational context. Here, all interactions and agent states are stored in a Python dictionary and can be accessed by other agents. AILA consists of two specialized agents: the AFM Handler Agent and the Data Handler Agent, both equipped with unique tools to do specific tasks. These agents possess individual prompts, LLMs, and tools; however, they utilize a shared memory to store and access states, facilitating smooth interaction. The system prompts within the agents offer instructions for tool utilization and ethical guidelines, whereas the outputs from other tools or agents serve as user prompts. The framework utilizes LangGraph, a library that allows the construction of an effective multi- agent workflow, integrating all agents and tools seamlessly.  
+
+The architecture for AILA's decision- making process is carefully designed to ensure precise information routing. AILA can dynamically select among three primary options: AFM Handler, Data Handler, or FINISH. When AILA identifies the appropriate agent to handle a query, it routes the information to the selected option. In cases where AILA determines that none of the available agents can sufficiently address the question, it generates a well- structured response and selects the FINISH option to conclude the session effectively. The agents within this system are equipped with three distinct operational choices: utilizing their respective tools, transferring information to the next agent, or terminating the session. A system prompt has been integrated to streamline these decisions. Agents append the prefix NEED HELP to their response when transferring information to another agent. Alternatively, if they believe the query has been adequately addressed, they use the prefix FINAL ANSWER to signal the session's conclusion. By analyzing the output for these keywords, the system seamlessly routes the response to the designated agent or tool or finalizes the session. This structured approach enables efficient multi- agent collaboration, ensuring clarity, accuracy, and optimal performance across tasks while maintaining a robust and adaptive framework.  
+
+## AFM Handler Agent  
+
+Atomic Force Microscopy demands precise sequential execution of multiple experimental stages. Image acquisition requires optimization across three critical parameters: imaging conditions, probe selection, and operational mode configuration (tapping/contact). The experimental sequence encompasses surface approach protocols, scanning procedures, and standardized data acquisition—with procedural deviations potentially resulting in equipment damage or data corruption. Our implementation utilizes the DriveAFM instrument (Nanosurf), which is accessed through a Python- based API architecture and designed for universal compatibility with API- enabled AFM systems. To facilitate AFM imaging experiments, we
+
+<--- Page Split --->
+
+have created the AFM Handler agent, which is integrated with two specialized tools: the Document Retrieval Tool and the Code Executor Tool. Every tool has an individual role, and the AFM Handler agent can dynamically assign tasks to these tools. The agent will assign the responsibility to the Data Handler agent if it finds that neither tool can handle the task.  
+
+## Document Retrieval tool  
+
+The documentation for the instrument offers detailed instructions on how to handle and calibrate it. However, providing full access to the documentation to an LLM entails risks, such as inadvertent alterations to factory settings or calibration data, which could potentially result in damage or malfunction of the instrument. To address this concern, we manually extracted the essential information from the AFM documentation necessary for conducting experiments while safeguarding the instrument's integrity. We consolidated all the crucial codes for regulating each parameter of the instrument into a comprehensive Python script. Since Python code relies heavily on precise indentation and line structure, we utilized the Recursive Character Text Splitter from the LangChain library, specifically designed for Python, to divide the script into manageable chunks. The chunk size was set to a maximum of 1000 characters without overlap, adhering to the token limit for embedding models. Each code chunk comprises three sections: the first includes the necessary Python libraries, the second contains the code required to load the application, and the third section features unique Python code specific to the given task. The first two sections are consistent across all chunks (see S2.2 in the Supplementary Information file for more details). These chunks were then combined to generate a document, embedded using OpenAI's text- embedding- 3- large model. This model, with the capability of producing embeddings of size up to 3072 dimensions, delivers exceptional performance compared to other OpenAI embedding models, especially in multi- language retrieval benchmarks like MIRACL<sup>42</sup>. To store the embeddings, we opted for Chroma, an open- source vector database known for its reliability and efficiency in managing large- scale embedding data. We use a vector store retriever to retrieve the data from the vector store.  
+
+## Code Executor tool  
+
+A code executor tool has been developed to execute Python scripts generated by the AFM Handler Agent to control the AFM software. This tool is intended to run Python code, provided as a text string, directly on the local system to allow for smooth integration with the workflow of the AFM Handler Agent. The utility executes the code and sends back a success message or a detailed description of the error that occurred. If there is an error, the error message is returned to the AFM Handler agent so it can correct the error and retry executing. Otherwise, if the script runs without errors, it is considered the final result. This iterative process ensures precise control of the AFM system while systematically addressing any issues in the script.  
+
+## Data Handler Agent  
+
+Surface tracking optimization in AFM requires precise calibration of three fundamental parameters: Proportional (P), Integral (I), and Derivative (D) gains. Optimal calibration manifests as convergence between trace and retrace signals, indicating stable scanning conditions. The Data Handler agent interfaces with specialized optimization and analysis
+
+<--- Page Split --->
+
+modules; these models can access AFM image data stored in local storage systems. The agent can optimize P, I, and D gains or calculate various surface properties, such as average friction and surface roughness, using the help of modules and image files stored locally.  
+
+## Image Optimization Tool  
+
+The feedback system in an Atomic Force Microscope (AFM) plays a crucial role in maintaining control over the interaction between the cantilever tip and the sample surface. During scanning, variations in surface features alter the interaction forces between the tip and the sample, leading to deflections in the cantilever. These deflections are detected by a photodetector. To ensure that these deflections stay within a specified range, the feedback mechanism continuously adjusts the height of either the tip or the sample stage in real time. This process is managed by a PID (Proportional- Integral- Derivative) controller, which regulates the position of the z- piezo actuator. By moving the cantilever probe up or down, the controller maintains a steady interaction force or adheres to a predefined setpoint, depending on the chosen mode of operation.  
+
+Fine- tuning the P, I, and D gain values of the controller is vital for achieving accurate control of the setpoint in AFM imaging. The integral gain is especially important for enhancing image clarity by mitigating drift and reducing steady- state errors. Once the integral gain is optimized, adjusting the proportional gain can provide further refinement. The derivative gain, on the other hand, is particularly beneficial for imaging samples with pronounced edge features. If the gains are set too low, the PID loop may fail to maintain the setpoint effectively, while excessively high gain values can introduce electrical noise into the image due to amplified feedback or overcompensation for deviations. Properly optimized PID parameters ensure that the feedback loop remains stable and responsive, enabling the AFM to accurately track surface topography, even at higher scanning speeds. This balance is especially critical when imaging delicate, irregular, or soft materials, as it preserves the integrity of tip- sample interactions.  
+
+A genetic algorithm (GA) was employed for PID gain optimization. The GA parameters included a fixed population size of three and a total of 15 generations, enabling efficient tuning of the gains. Although these parameters can be manually adjusted, but excessive image scanning may degrade the AFM tip. The optimized gains ensure effective feedback control, producing comparable forward and backward images. This can be achieved by calculating the mean squared error (MSE) between forward and backward z- axis images for various PID gain settings. However, this method is sensitive to drift during scanning, and this method also depends on previously acquired images. To address this, the structural similarity index (SSIM) was adopted as the fitness function in the genetic algorithm, providing a robust measure of image similarity between the z- axis forward and backward image independent of prior image data.  
+
+This metric offers advantages over traditional Mean Square Error (MSE) approaches by (i) addressing tip degradation challenges in contact- mode AFM by minimizing required scan cycles and enabling optimization using low- resolution images, (ii) maintaining accuracy under drift conditions, (iii) incorporating structural, brightness, and contrast variations in
+
+<--- Page Split --->
+
+optimization, and (iv) providing normalized scores between 0 and 1, where 1 indicates perfect similarity.  
+
+The SSIM is defined as:  
+
+\[S S I M(x,y) = [l(x,y)]^{\alpha} \times [c(x,y)]^{\beta} \times [s(x,y)]^{\gamma}\]  
+
+where, \(l(x,y)\) is the luminance comparison, \(c(x,y)\) is the contrast comparison, and \(s(x,y)\) is the structure comparison with \(\alpha , \beta , \gamma\) being the weighting parameters. Note that the individual components are defined as:  
+
+\[l(x,y) = (2\mu_{x}\mu_{y} + C_{1}) / (\mu x^{2} + \mu y^{2} + C_{1})\] \[c(x,y) = (2\sigma_{x}\sigma_{y} + C_{2}) / (\sigma_{x}{}^{2} + \sigma_{y}{}^{2} + C_{2})\] \[s(x,y) = (\sigma_{xy} + C_{3}) / (\sigma_{x}\sigma_{y} + C_{3})\]  
+
+where, \(\mu_{x}, \mu_{y}\) represent the mean intensities of images \(x\) and \(y\) , \(\sigma_{x}, \sigma_{y}\) is the standard deviations of images \(x\) and \(y\) , \(\sigma_{xy}\) is the cross- covariance between images \(x\) and \(y\) , and \(C_{1}, C_{2}, C_{3}\) are constants to avoid instability with \((C_{1} = (k_{1}L)^{2}, C_{2} = (k_{2}L)^{2}, C_{3} = C_{2} / 2)\) and \(L\) being the dynamic range of pixel values and \(k_{1} = 0.01\) and \(k_{2} = 0.03\) .  
+
+## Baseline correction  
+
+The adaptive baseline correction employed in the step- edge detection of graphene is given by  
+
+\[B(x,y) = \Sigma_{i,j}a_{ij}x^i y^j\]  
+
+Where, \(B(x,y)\) is the baseline function, \(a_{ij}\) are the polynomial coefficients, \(i\) and \(j\) are the polynomial degrees \((0 \leq i,j \leq n)\) with \(n\) being the maximum polynomial degree.  
+
+## Image Analysis tool  
+
+AFM instrument stores the image data as a \*.nid file in the local system. This \*.nid file contains deflection, friction force, and z- axis images for both backward and forward scans. To further process any image from the file, exact data must be extracted from the file. To conduct this, we have used the NSFopen python library in the Image Analysis tool, which takes the query from the data handler agent regarding the specific image data and its location and returns the image data in an array to the data handler tool. To conduct further processing of the images, any Python script generated by the data handler tool can be executed in the Image Analysis tool, and the result can be returned to the data handler agent. Note that there is no database available to guide the LLM model in generating the Python script. It can generate the Python script by itself. There is a total of 6 input parameters for this tool:  
+
+(1) path (str): Directory path to search for the latest file (default: None). 
+(2) filename (str): Specific image file to display (default: None). 
+(3) dynamic_code (str): Python code for processing image data (default: None). 
+(4) calculate_friction (bool): Option to compute average friction (default: False). 
+(5) calculate_mean_roughness (bool): Option to compute mean roughness (default: False). 
+(6) calculate_rms_roughness (bool): Option to compute RMS roughness (default: False). Returns: A dictionary with the status, image data, or error details. Average friction was calculated using the following formula:
+
+<--- Page Split --->
+
+\[F_{ave} = \frac{1}{2}\times (f_{ij} - b_{ij})\]  
+
+Where \(f_{ij}\) and \(b_{ij}\) are the element at position \((i,j)\) in the array of the forward and backward friction image data. We have used the formula in this tool to calculate the mean roughness and RMS roughness values  
+
+\[R_{mean} = \frac{1}{M.N}\sum_{i = 1}^{M}\sum_{j = 1}^{N}|z_{ij} - \bar{z}|\] \[R_{rms} = \frac{1}{M.N}\sum_{i = 1}^{M}\sum_{j = 1}^{N}(z_{ij} - \bar{z})^2\]  
+
+where \(z_{ij}\) is the element at position \((i,j)\) in the array, \(\bar{z}\) is the mean of all elements in the array, \(M\) is the number of rows in the array, \(N\) is the number of columns in the array of the \(z\) - axis forward image data.  
+
+## AFMBench  
+
+Dataset preparation. To evaluate the performance of the AILA, we have manually created a set of 100 questions, carefully categorized into three distinct groups. The first classification is based on whether a question requires one or multiple tools/agents to be solved. The second category assesses the complexity of the questions, distinguishing between basic and advanced levels. Lastly, the questions are grouped by their requirements, such as documentation analysis or calculations. The complexity of each question is determined by the number of agents involved and the steps required to achieve the solution. For instance, modifying a parameter in an AFM system typically requires documentation review and the use of a single agent, categorizing it as a basic task. Conversely, capturing an AFM image and analyzing its surface roughness involves multiple agents, documentation analysis, and calculations, making it an advanced task. A comprehensive JSON file has been created, encapsulating detailed metadata about each question, including its respective category, for streamlined analysis and evaluation. This file serves as a structured resource for further investigations and testing. All questions, along with their relevant classifications and details, have been made accessible on GitHub (https://github.com/M3RG- IITD/AILA) to support transparency and reproducibility in research.  
+
+## Evaluation  
+
+We developed a graphical user interface (GUI) using Streamlit, an open- source Python framework, to streamline user interaction with AILA. The GUI allows users to input text- based queries, select the desired LLM model, and specify a log file name. It then executes AILA in the backend, saving the output log file locally and enabling users to observe results directly within the AFM software. Any output images or figures generated by AILA are also stored in the local system for further analysis. To ensure robustness, we manually evaluated all questions using GPT- 4o and GPT- 3.5- turbo- 0125, verifying the output log files and AFM software results multiple times in collaboration with different researchers to eliminate potential human errors. The evaluation of AILA's performance was categorized into two metrics: accuracy and
+
+<--- Page Split --->
+
+efficiency. For accuracy, questions were divided into categories based on complexity and tool/agent usage, with a percentage of correct answers calculated for each category. For efficiency, uniform parameters were maintained across models in the AFM software, including default settings of 0.1 seconds for time per line and 128 points per line and frame when not specified by the user. To ensure precise efficiency measurements, scanning time for images and the time taken by questions with incorrect answers were excluded from the analysis. Average response times were computed for each category to assess AILA's overall efficiency.  
+
+## Evaluation Metric  
+
+To assess the evaluation of questions in terms of accuracy, we classified the answers provided by AILA into two categories: fully correct answers and incorrect or partially correct answers. A fully correct answer was considered accurate, while any incorrect or incomplete response was deemed incorrect. Given that some questions require manual inspection of the AFM software to verify whether specific parameters are set correctly and whether the AFM image is captured as intended, multiple researchers were involved in verifying the results. They carefully checked the outcomes to ensure error- free results. For measurements of different properties, such as average friction, roughness, and RMS value of roughness, we used the Gwydion software to verify the accuracy of the results. Subsequently, the questions were clustered into appropriate groups, and the corresponding average percentage of correct answers was calculated.  
+
+## Data and Code Availability  
+
+All the tasks in AFMBench, along with the complete log files of the responses for each of the tasks from GPT- 4o and GPT- 3.5 are available at: https://github.com/M3RG- IITD/AILA  
+
+## Acknowledgement  
+
+N.M.A.K. acknowledges the funding support from Google Research Scholar Award, and the Alexander von Humboldt Foundation. I.M. thanks University Grants Commission (UGC), Government of India for the NET- JRF fellowship (221610021768). The authors thank IIT Delhi HPC facility for computational and storage resources.  
+
+## References  
+
+1. Zhao, Z., Lee, W. S. & Hsu, D. Large language models as commonsense knowledge for large-scale task planning. Advances in Neural Information Processing Systems 36, (2024).  
+2. Liu, Y. et al. Machine Learning-Based Reward-Driven Tuning of Scanning Probe Microscopy: Towards Fully Automated Microscopy. Preprint at https://doi.org/10.48550/arXiv.2408.04055 (2024).
+
+<--- Page Split --->
+
+3. Lála, J. et al. Paperqa: Retrieval-augmented generative agent for scientific research. arXiv preprint arXiv:2312.07559 (2023).  
+
+4. Koscher, B. A. et al. Autonomous, multiproperty-driven molecular discovery: From predictions to measurements and back. Science 382, eadi1407 (2023).  
+
+5. Szymanski, N. J. et al. An autonomous laboratory for the accelerated synthesis of novel materials. Nature 624, 86-91 (2023).  
+
+6. Dai, T. et al. Autonomous mobile robots for exploratory synthetic chemistry. Nature 635, 890-897 (2024).  
+
+7. Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570-578 (2023).  
+
+8. Sadeghi, S. et al. Engineering a Sustainable Future: Harnessing Automation, Robotics, and Artificial Intelligence with Self-Driving Laboratories. ACS Sustainable Chemistry & Engineering 12, 12695-12707 (2024).  
+
+9. Delgado-Licona, F. & Abolhasani, M. Research acceleration in self-driving labs: Technological roadmap toward accelerated materials and molecular discovery. Advanced Intelligent Systems 5, 2200331 (2023).  
+
+10. Sim, M. et al. ChemOS 2.0: An orchestration architecture for chemical self-driving laboratories. Matter 7, 2959-2977 (2024).  
+
+11. Darvish, K. et al. ORGANA: A robotic assistant for automated chemistry experimentation and characterization. Matter doi:10.1016/j.matt.2024.10.015.  
+
+12. Abolhasani, M. & Kumacheva, E. The rise of self-driving labs in chemical and materials sciences. Nature Synthesis 2, 483-492 (2023).  
+
+13. Bayley, O., Savino, E., Slattery, A. & Noel, T. Autonomous chemistry: Navigating self-driving labs in chemical and material sciences. Matter 7, 2382-2398 (2024).
+
+<--- Page Split --->
+
+14. Tom, G. et al. Self-driving laboratories for chemistry and materials science. Chemical Reviews 124, 9633–9732 (2024).  
+
+15. Canty, R. B. & Abolhasani, M. Reproducibility in automated chemistry laboratories using computer science abstractions. Nature Synthesis 1–13 (2024).  
+
+16. Volk, A. A. & Abolhasani, M. Performance metrics to unleash the power of self-driving labs in chemistry and materials science. Nature Communications 15, 1378 (2024).  
+
+17. Mirza, A. et al. Are large language models superhuman chemists? Preprint at https://doi.org/10.48550/arXiv.2404.01475 (2024).  
+
+18. Zaki, M., Jayadeva, Mausam & Krishnan, N. M. A. MaScQA: investigating materials science knowledge of large language models. Digital Discovery 3, 313–327 (2024).  
+
+19. Alampara, N. et al. Probing the limitations of multimodal language models for chemistry and materials research. arXiv preprint arXiv:2411.16955 (2024).  
+
+20. Miret, S. & Krishnan, N. M. A. Are LLMs Ready for Real-World Materials Discovery? Preprint at http://arxiv.org/abs/2402.05200 (2024).  
+
+21. Krishnan, N. M. A., Kodamana, H. & Bhattoo, R. Machine Learning for Materials Discovery: Numerical Recipes and Practical Applications. (Springer International Publishing, Cham, 2024). doi:10.1007/978-3-031-44622-1.  
+
+22. Krishnan, N. M. A., Kodamana, H. & Bhattoo, R. Natural Language Processing. in Machine Learning for Materials Discovery 263–275 (Springer International Publishing, Cham, 2024). doi:10.1007/978-3-031-44622-1_15.  
+
+23. Mishra, V. et al. LLaMat: Large Language Models for Materials Science Information Extraction. in (2024).  
+
+24. Li, Z. et al. Mmsci: A multimodal multi-discipline dataset for phd-level scientific comprehension. in AI for Accelerated Materials Design-Vienna 2024 (2024).
+
+<--- Page Split --->
+
+25. Alampara, N. et al. MACBENCH: A multimodal chemistry and materials science benchmark.  
+
+26. Skarlinski, M. D. et al. Language agents achieve superhuman synthesis of scientific knowledge. Preprint at https://doi.org/10.48550/arXiv.2409.13740 (2024).  
+
+27. Bian, K. et al. Scanning probe microscopy. Nature Reviews Methods Primers 1, 36 (2021).  
+
+28. Liu, Y. et al. AECroscopy: A Software-Hardware Framework Empowering Microscopy Toward Automated and Autonomous Experimentation. Small Methods 2301740 (2024) doi:10.1002/smtd.202301740.  
+
+29. Kandel, S. et al. Demonstration of an AI-driven workflow for autonomous high-resolution scanning microscopy. Nat Commun 14, 5501 (2023).  
+
+30. Vasudevan, R. K. et al. Autonomous Experiments in Scanning Probe Microscopy and Spectroscopy: Choosing Where to Explore Polarization Dynamics in Ferroelectrics. ACS Nano 15, 11253-11262 (2021).  
+
+31. Kalinin, S. V. et al. Automated and Autonomous Experiments in Electron and Scanning Probe Microscopy. ACS Nano 15, 12604-12627 (2021).  
+
+32. Arias, S., Zhang, Y., Zahl, P. & Hollen, S. Autonomous Molecular Structure Imaging with High-Resolution Atomic Force Microscopy for Molecular Mixture Discovery. J. Phys. Chem. A 127, 6116-6122 (2023).  
+
+33. Kang, S., Park, J. & Lee, M. Machine learning-enabled autonomous operation for atomic force microscopes. Review of Scientific Instruments 94, 123704 (2023).  
+
+34. Zaki, M. et al. Interpretable Machine Learning Approach for Identifying the Tip Sharpness in Atomic Force Microscopy. Scripta Materialia 221, 114965 (2022).  
+
+35. Kalinin, S. V. et al. Probe microscopy is all you need. Machine Learning: Science and Technology 4, 023001 (2023).
+
+<--- Page Split --->
+
+36. Huang, L. et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems (2023).  
+
+37. Anwar, U. et al. Foundational challenges in assuring alignment and safety of large language models. arXiv preprint arXiv:2404.09932 (2024).  
+
+38. Liu, Y. et al. Trustworthy LLMs: A survey and guideline for evaluating large language models' alignment. arXiv preprint arXiv:2308.05374 (2023).  
+
+39. Grobelny, J. et al. Size Measurement of Nanoparticles Using Atomic Force Microscopy. in Characterization of Nanoparticles Intended for Drug Delivery (ed. McNeil, S. E.) 71-82 (Humana Press, Totowa, NJ, 2011). doi:10.1007/978-1-60327-198-1_7.  
+
+40. Moreno-Flores, S. Baseline correction of AFM force curves in the force-time representation. Microscopy Research and Technique 79, 1045-1049 (2016).  
+
+41. Burnham, N. A., Lyu, L. & Poulikakos, L. Towards artefact-free AFM image presentation and interpretation. Journal of Microscopy 291, 163-176 (2023).  
+
+42. Zhang, X. et al. Making a MIRACL: Multilingual Information Retrieval Across a Continuum of Languages. Preprint at https://doi.org/10.48550/arXiv.2210.09984 (2022).
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SupplementaryMaterials.pdf
+
+<--- Page Split --->

preprint/preprint__2be28d29fac1837acb16f3a122ee991594a1cc474dcaae29e51328cc54c6d461/preprint__2be28d29fac1837acb16f3a122ee991594a1cc474dcaae29e51328cc54c6d461_det.mmd

@@ -0,0 +1,561 @@
+<|ref|>title<|/ref|><|det|>[[43, 108, 857, 177]]<|/det|>
+# Autonomous Microscopy Experiments through Large Language Model Agents  
+
+<|ref|>text<|/ref|><|det|>[[43, 196, 268, 240]]<|/det|>
+N M Anoop Krishnan krishnan@iitd.ac.in  
+
+<|ref|>text<|/ref|><|det|>[[43, 268, 725, 307]]<|/det|>
+Indian Institute of Technology Delhi https://orcid.org/0000- 0003- 1500- 4947  
+
+<|ref|>text<|/ref|><|det|>[[43, 300, 365, 340]]<|/det|>
+Indrajeet Mandal Indian Institute of Technology Delhi  
+
+<|ref|>text<|/ref|><|det|>[[43, 344, 365, 383]]<|/det|>
+Jitendra Soni Indian Institute of Technology Delhi  
+
+<|ref|>text<|/ref|><|det|>[[43, 388, 725, 428]]<|/det|>
+Mohd Zaki Indian Institute of Technology Delhi https://orcid.org/0000- 0002- 4551- 3470  
+
+<|ref|>text<|/ref|><|det|>[[43, 432, 575, 472]]<|/det|>
+Morten Smedskjaer Aalborg University https://orcid.org/0000- 0003- 0476- 2021  
+
+<|ref|>text<|/ref|><|det|>[[43, 477, 408, 518]]<|/det|>
+Katrin Wondraczek Leibniz Institute of Photonic Technology  
+
+<|ref|>text<|/ref|><|det|>[[43, 523, 365, 563]]<|/det|>
+Lothar Wondraczek University of Jena  
+
+<|ref|>text<|/ref|><|det|>[[43, 568, 365, 609]]<|/det|>
+Nitya Nand Gosvami Indian Institute of Technology Delhi  
+
+<|ref|>sub_title<|/ref|><|det|>[[43, 653, 105, 671]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[43, 692, 777, 712]]<|/det|>
+Keywords: self- driving laboratory, atomic force microscopy, benchmarking, AI agents  
+
+<|ref|>text<|/ref|><|det|>[[43, 730, 349, 749]]<|/det|>
+Posted Date: December 19th, 2024  
+
+<|ref|>text<|/ref|><|det|>[[43, 768, 475, 787]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 5600537/v1  
+
+<|ref|>text<|/ref|><|det|>[[43, 805, 914, 847]]<|/det|>
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[43, 865, 535, 885]]<|/det|>
+Additional Declarations: There is NO Competing Interest.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[41, 45, 944, 88]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on October 14th, 2025. See the published version at https://doi.org/10.1038/s41467-025-64105-7.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[130, 85, 867, 132]]<|/det|>
+# Autonomous Microscopy Experiments through Large Language Model Agents  
+
+<|ref|>text<|/ref|><|det|>[[118, 153, 877, 191]]<|/det|>
+Indrajeet Mandal \(^{1}\) , Jitendra Soni \(^{2}\) , Mohd Zaki \(^{3}\) , Morten M. Smedskjaer \(^{4}\) , Katrin Wondraczek \(^{5}\) , Lothar Wondraczek \(^{6}\) , Nitya Nand Gosvami \(^{1,2,7*}\) , N. M. Anoop Krishnan \(^{1,3,7,*}\)  
+
+<|ref|>text<|/ref|><|det|>[[120, 216, 876, 410]]<|/det|>
+\(^{1}\) School of Interdisciplinary Research, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India. \(^{2}\) Department of Materials Science and Engineering, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India \(^{3}\) Department of Civil Engineering, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016,India. \(^{4}\) Department of Chemistry and Bioscience, Aalborg University, Aalborg, Denmark \(^{5}\) Leibniz Institute of Photonic Technology, 07745 Jena, Germany \(^{6}\) Otto Schott Institute of Materials Research, University of Jena, 07743 Jena, Germany \(^{7}\) Yardi School of Artificial Intelligence, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India  
+
+<|ref|>text<|/ref|><|det|>[[122, 430, 875, 467]]<|/det|>
+\(^{*}\) Corresponding Authors: ngosvami@iitd.ac.in (N. N. G.); krishnan@iitd.ac.in (N. M. A. K.) Fax: +91- 11- 2658- 1117; Tel: +91- 11- 2659- 1223  
+
+<|ref|>text<|/ref|><|det|>[[145, 487, 850, 505]]<|/det|>
+Keywords: self- driving laboratory, atomic force microscopy, benchmarking, AI agents  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 530, 196, 545]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[117, 549, 880, 885]]<|/det|>
+The emergence of large language models (LLMs) has accelerated the development of self- driving laboratories (SDLs) for materials research. Despite their transformative potential, current SDL implementations rely on rigid, predefined protocols that limit their adaptability to dynamic experimental scenarios across different labs. A significant challenge persists in measuring how effectively AI agents can replicate the adaptive decision- making and experimental intuition of expert scientists. Here, we introduce AILA (Artificially Intelligent Lab Assistant), a framework that automates atomic force microscopy (AFM) through LLM- driven agents. Using AFM as an experimental testbed, we develop AFMBench—a comprehensive evaluation suite that challenges AI agents based on language models like GPT- 4o and GPT- 3.5 to perform tasks spanning the scientific workflow: from experimental design to results analysis. Our systematic assessment shows that state- of- the- art language models struggle even with basic tasks such as documentation retrieval, leading to a significant decline in performance in multi- agent coordination scenarios. Further, we observe that LLMs exhibit a tendency to not adhere to instructions or even divagate to additional tasks beyond the original request, raising serious concerns regarding safety alignment aspects of AI agents for SDLs. Finally, we demonstrate the application of AILA on increasingly complex experiments open- ended experiments: automated AFM calibration, high- resolution feature detection, and mechanical property measurement. Our findings emphasize the necessity for stringent
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 85, 877, 120]]<|/det|>
+benchmarking protocols before deploying AI agents as laboratory assistants across scientific disciplines.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 160, 291, 175]]<|/det|>
+## 1. Introduction  
+
+<|ref|>text<|/ref|><|det|>[[118, 179, 880, 479]]<|/det|>
+Scientific experimentation demands exceptional domain expertise, from exploration or hypothesis- driven experimental design to precision execution and rigorous data analysis. This complexity creates bottlenecks in scientific discovery, particularly as experimental techniques grow increasingly sophisticated. The advent of large language models (LLMs) has propelled the development of self- driving laboratories (SDLs) that integrate diverse information sources for automated planning<sup>1</sup> and experimentation. AI- agents<sup>2,3</sup> and SDLs have already achieved several feats in materials or molecular discovery<sup>4- 6</sup>, chemistry research<sup>7</sup>, and inorganic materials synthesis. The promise of SDLs toward achieving sustainable development<sup>8</sup> has resulted in enormous efforts to harness their potential in high- throughput experimentation and discovery<sup>9</sup>. Efforts to streamline SDLs have resulted in orchestration architectures such as ChemOS<sup>10</sup>. Additionally, it has been demonstrated that the capability of SDLs can be enhanced by a human- in- the- loop framework that handles disambiguation, thereby enabling better planning and execution<sup>11</sup>. While early demonstrations of LLM- based lab assistants showed promise in chemistry and materials science<sup>1- 3</sup>, their operational reliability remains largely uncharacterized beyond specific applications or repetitive use cases with predetermined protocols<sup>15- 19</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[118, 499, 880, 686]]<|/det|>
+Current research predominantly addresses well- documented or predefined protocols and single- objective tasks, failing to capture the intricate interplay between experimental planning, multi- tool coordination, and result interpretation or online intervention<sup>10</sup>. While recent investigations incorporating planning elements have demonstrated success in achieving specific experimental objectives, they have not systematically evaluated SDL reliability across the broader spectrum of laboratory automation tasks<sup>15,16</sup>. Although several studies have benchmarked LLMs<sup>17,18,20- 23</sup> and vision language models (VLMs)<sup>13- 16</sup> through question- answer protocols to assess their potential as materials research co- pilots, a crucial knowledge gap persists: understanding how these AI systems handle novel experimental scenarios and their fundamental limitations.  
+
+<|ref|>text<|/ref|><|det|>[[118, 707, 880, 893]]<|/det|>
+To address this challenge, we here introduce AILA (Artificially Intelligent Lab Assistant), an LLM- powered framework augmented with specialized tools. We selected scanning probe microscopy<sup>20</sup>, specifically atomic force microscopy (AFM), as our experimental testbed, given its inherent complexity and broad applicability in materials research. There have been several efforts to automate microscopy techniques using AI and human- in- the- loop approaches due to their extensive applications in materials characterization<sup>28- 30,30,31,31- 35</sup>. These efforts focus exclusively on advancing specific operational aspects, such as analysing moving objects or optimizing illumination conditions, with an emphasis on improving individual steps within the broader experimental protocol<sup>30,32,35</sup>. AFM operation demands expertise across multiple domains—from probe calibration to parameter optimization and data interpretation—making
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 85, 878, 120]]<|/det|>
+it an ideal platform for evaluating AI agents' ability to manage sophisticated experimental workflows.  
+
+<|ref|>text<|/ref|><|det|>[[118, 140, 880, 271]]<|/det|>
+Using AFM as the model system, we probe AILA's capabilities through AFMBench on five critical aspects of scientific automation: experimental workflow design, multi- tool coordination, decision- making, execution of open- ended experiments, and data analysis. Our systematic evaluation reveals key failure modes and areas requiring enhancement. We demonstrate AILA's practical utility through three increasingly complex case studies: automated microscope calibration, high- resolution graphene step- edge imaging, and load- dependent roughness analysis on highly oriented pyrolytic graphite (HOPG).  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 292, 245, 307]]<|/det|>
+## 2. Results  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 311, 332, 327]]<|/det|>
+### 2.1 AILA framework  
+
+<|ref|>text<|/ref|><|det|>[[117, 330, 880, 591]]<|/det|>
+AILA's architecture prioritizes modularity, enabling seamless integration with diverse experimental and analytical platforms. At its core lies an LLM- powered planner—the framework's cognitive centre—which orchestrates user interactions and coordinates specialized agents (Fig. 1a). For AFM operations, AILA deploys two agents: the AFM Handler Agent (AFM- HA) for experimental control and the Data Handler Agent (DHA) for analysis. The AFM- HA interfaces with a document retrieval system comprising AFM software documentation and a code execution engine that translates Python commands into experimental actions. A Python- based API establishes the hardware- software interface, enabling direct control of the AFM system through vendor- specific protocols (Fig. 1b). The DHA manages image optimization and analysis through dedicated tools: an Image Optimizer that fine- tunes PID parameters for high- fidelity imaging and an Image Analyzer that extracts targeted features from experimental data. For queries beyond agent capabilities, the planner generates alternative approaches or recommended actions. The technical specifications and implementation details of each module are explained in the Methods section.  
+
+<|ref|>text<|/ref|><|det|>[[117, 610, 880, 799]]<|/det|>
+To demonstrate AILA's operational workflow, we present a multi- step experiment: acquiring an AFM image of HOPG and extracting its friction and roughness parameters (Fig. 1c). This open- ended task exemplifies real- world complexity, offering multiple solution pathways. Upon receiving the query, AILA dissects it into sequential objectives: image acquisition via AFM- HA followed by DHA- led analysis. AFM- HA retrieves relevant documentation, generates executable code, and captures the image. Following successful acquisition, AILA transitions control to DHA, which directs the Image Analyzer to compute the specified parameters. This orchestrated sequence exemplifies AILA's core strengths: the ability to parse complex natural language queries, develop strategic workflows, and coordinate multiple agents toward achieving experimental objectives.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[118, 80, 880, 590]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[117, 593, 881, 743]]<|/det|>
+<center>Figure 1. AILA Framework and Implementation. (A) System Architecture of the Artificially Intelligent Laboratory Assistant (AILA). Dotted lines indicate adaptive information flow governed by AILA’s decision-making, and solid lines represent deterministic information pathways with predefined routing protocols. (B) Image of the atomic force microscope (AFM) experimental setup showing key hardware components and control interfaces. (C) Representative demonstration of AILA’s operation: raw transcript of a user query and AILA’s unedited response sequence, showing the system’s query interpretation, task planning, and execution capabilities. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[120, 100, 875, 576]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[147, 85, 653, 101]]<|/det|>
+<center>2.2 AFMBench: tasks for evaluating the AILA framework </center>  
+
+<|ref|>text<|/ref|><|det|>[[117, 581, 881, 733]]<|/det|>
+Figure 2. Task Distribution and Module Utilization in AFMBench. A. Pie charts showing the distribution of tool requirements (left, single vs. multiple) and agent requirements (right, single vs. multiple) across benchmark tasks. B. Operation complexity categorization showing the proportion of basic versus advanced tasks. C. Horizontal bar chart quantifying module engagement frequency across all tasks, demonstrating utilization patterns of each tool and agent. D. Venn diagram illustrating the overlap between documentation, analysis, and calculation tasks. E. Representative examples of basic (left) and advanced (right) tasks, demonstrating increasing complexity in experimental workflows.  
+
+<|ref|>text<|/ref|><|det|>[[117, 752, 881, 902]]<|/det|>
+AFMBench comprises 100 expertly curated experimental tasks (see S3.1 in Supplementary Information for a few examples of tasks; all the tasks are available in the GitHub repo) manually designed to rigorously evaluate autonomous AFM operations across multiple dimensions of complexity. Unlike conventional LLM benchmarks or simulation- based evaluations, each AFMBench task demands physical execution on AFM hardware, introducing real- world temporal constraints and experimental variability. Analysis of the task architecture reveals distinct patterns in resource utilization and operational complexity. In Figure 2A, tool coordination requirements highlight a systematic preference for sophisticated workflows, with
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 879, 158]]<|/det|>
+\(69\%\) of tasks demanding multi- tool integration, while \(31\%\) operate through single- tool protocols. Agent deployment analysis reveals a distribution: \(83\%\) of operations utilize single- agent protocols, while \(17\%\) require multi- agent coordination—enabling evaluation of both targeted expertise and system- wide integration capabilities.  
+
+<|ref|>text<|/ref|><|det|>[[118, 178, 880, 328]]<|/det|>
+In Figure 2B, the operational landscape is divided into two primary complexity tiers: basic operations ( \(56\%\) ) encompassing fundamental microscopy tasks and advanced procedures ( \(44\%\) ) requiring more sophisticated experimental workflows (for example questions see Figure 2E). Core system components—the AFM Handler, Document Retriever, and Code Executor—demonstrate maximum engagement, each activating in 66 distinct tasks. The Data Handler Agent and Image Analyzer exhibit selective activation patterns (52 and 48 tasks, respectively), while the Image Optimizer engages exclusively in critical parameter optimization scenarios (4 tasks).  
+
+<|ref|>text<|/ref|><|det|>[[118, 348, 880, 498]]<|/det|>
+Task distribution across functional domains reveals three primary clusters: documentation (50 standalone tasks), analysis (14 tasks), and calculation (10 tasks) (see Figure 2D). A significant overlap between these domains emerges through integrated tasks that combine multiple functional requirements, reflecting the interconnected nature of experimental workflows. This carefully constructed distribution enables systematic evaluation of AI systems across a spectrum of experimental complexity—from basic instrument control to advanced multi- step procedures requiring mathematical reasoning and dynamic decision- making—effectively mirroring the cognitive hierarchy of expert atomic force microscopists.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 518, 407, 535]]<|/det|>
+### 2.3 Performance of AI Agents  
+
+<|ref|>text<|/ref|><|det|>[[118, 536, 880, 686]]<|/det|>
+Systematic evaluation of AILA using two advanced closed source language models—GPT- 4o and GPT- 3.5- turbo- 0125—unveils distinctive execution patterns and operational efficacies. GPT- 4o exhibits exceptional proficiency in documentation- centric operations, achieving a \(92\%\) success rate, complemented by robust execution in analysis ( \(71\%\) ) and computational tasks ( \(70\%\) ). The model’s strength lies in its ability to navigate interconnected workflows: \(80\%\) success in merged documentation- analysis procedures and \(60\%\) in documentation- computation sequences. These metrics highlight GPT- 4o’s capacity to replicate the integrative reasoning characteristic of expert microscopists.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[135, 85, 840, 585]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[117, 591, 880, 724]]<|/det|>
+<center>Figure 3. Comparative Performance Analysis of Language Models on AFMBench. A. Venn diagrams showing accuracy (left) and efficiency (right) metrics for GPT-4o across documentation, analysis, and calculation tasks. Numbers indicate percentage accuracy and processing time in seconds. B. Corresponding performance metrics for GPT-3.5-turbo-0125, highlighting domain-specific competencies and processing speeds. C. A horizontal bar chart comparing tool and agent utilization efficiency between models is expressed as a percentage of successful engagements. </center>  
+
+<|ref|>text<|/ref|><|det|>[[118, 743, 880, 855]]<|/det|>
+In marked contrast, GPT- 3.5- turbo- 0125 displays compartmentalized strengths: substantial accuracy in standalone documentation (72%), analytical interpretation (71%), and mathematical operations (40%). However, its performance degrades significantly when confronted with multi- domain challenges, registering null success rates in tasks demanding simultaneous expertise across domains. This limitation suggests insufficient development of cross- functional reasoning capabilities essential for autonomous experimentation.  
+
+<|ref|>text<|/ref|><|det|>[[118, 875, 880, 912]]<|/det|>
+Temporal efficiency metrics augment these distinctions. GPT- 4o maintains steady execution speeds: documentation tasks took an average of 12.23 seconds, analytical procedures 9.25
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 879, 158]]<|/det|>
+seconds, and computational operations 11.06 seconds. GPT- 3.5- turbo- 0125 demonstrates irregular processing intervals: 8.88 seconds for documentation, 13.48 seconds for analysis, and 3.26 seconds for calculations. This temporal variability indicates underlying differences in task comprehension and execution strategies between the models.  
+
+<|ref|>text<|/ref|><|det|>[[118, 178, 879, 272]]<|/det|>
+Component utilization analysis reinforces these observations. GPT- 4o achieves consistently elevated engagement across system modules, particularly excelling in AFM operation control and document interpretation. These results highlight the fundamental importance of model architecture in autonomous experimental platforms, with GPT- 4o's advanced integrative capabilities positioning it as the superior choice for sophisticated experimental automation.  
+
+<|ref|>sub_title<|/ref|><|det|>[[147, 292, 620, 309]]<|/det|>
+### 2.4 Error Analysis Reveals Model-Specific Limitations  
+
+<|ref|>text<|/ref|><|det|>[[118, 311, 879, 422]]<|/det|>
+Detailed examination of failure cases revealed distinctive error patterns between the two language models, offering insights into their operational limitations. GPT- 4o exhibits a total error rate of \(20\%\) , with errors distributed across three primary categories: code generation \((60\%)\) , agent/tool selection \((25\%)\) , and instruction adherence \((15\%)\) . The predominance of code generation errors suggests challenges in translating conceptual understanding into executable commands despite the model's strong performance in task comprehension.  
+
+<|ref|>text<|/ref|><|det|>[[118, 442, 879, 554]]<|/det|>
+GPT- 3.5- turbo- 0125 demonstrates a markedly higher total error rate of \(45\%\) , with errors concentrated in two categories: code generation \((66.7\%)\) and agent/tool selection \((33.3\%)\) . Notably, the model shows no fundamental query interpretation errors, indicating robust natural language processing capabilities. However, the elevated frequency of code generation errors, coupled with significant agent/tool selection failures, points to underlying deficiencies in translating comprehension into actionable experimental protocols.  
+
+<|ref|>text<|/ref|><|det|>[[118, 574, 880, 855]]<|/det|>
+A critical finding emerged regarding GPT4o's instruction adherence. In one of the three recorded errors, GPT- 4o exceeded its designated operational limits, performing actions that were not authorized by the provided guidelines. For instance, it carried out potentially risky tip movements while it was only instructed to scan the surface (see S2.3 in the Supplementary Information). In another case, GPT- 4o was instructed to capture an image and calculate surface friction. While the image was captured correctly, the system failed to analyse it and switches the AFM to the lateral force mode instead of following clear instructions for a specific task (see S2.3 in the Supplementary Information). Instead of staying within the scope of the task, it performed additional actions. Although sometimes the final result may have been correct, the failure to follow instructions highlights concerns about AI- agent behaviour and raises safety risks in automated lab environments. Similar to the observation of "hallucination" in LLMs<sup>36</sup>, these results present a unique challenge—SDLs tend to take arbitrary actions potentially based on memory rather than following the instructions, referred to hereafter as "sleepwalking". These issues are especially critical in sensitive experimental settings, where strict protocol adherence is essential to ensure both equipment safety and the validity of results.  
+
+<|ref|>text<|/ref|><|det|>[[118, 876, 879, 912]]<|/det|>
+This error distribution illuminates critical areas for framework enhancement. While GPT- 4o's balanced error profile suggests the need for targeted improvements across multiple domains,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 880, 158]]<|/det|>
+GPT- 3.5- turbo- 0125's concentrated error patterns indicate fundamental limitations in experimental execution capabilities. These findings underscore the necessity for specialized training in automated experimental systems, particularly focusing on the translation of scientific protocols into executable code sequences.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 197, 405, 214]]<|/det|>
+### 2.5 Safety Alignment in SDLs  
+
+<|ref|>text<|/ref|><|det|>[[117, 216, 881, 422]]<|/det|>
+To understand the safety challenges<sup>37</sup> of AI agents, we evaluate the effectiveness of implementing a safety framework in AILA. First, we establish restricted access protocols for critical AFM functions, coupled with ethical system prompts (see S2.1 in Supplementary Information) that constrain code generation to predefined documentation<sup>38</sup>. Second, we develop strict operational boundaries that permit dynamic code generation solely for image analysis while preventing external software installation or system modifications. Evaluation of the improved protocol demonstrates the effectiveness of these safeguards—AILA appropriately failed when prompted to install external Python libraries. (see S3.3 in Supplementary Information for complete validation logs). These findings underscore the critical importance of robust safety protocols in SDLs, emphasizing the necessity of comprehensive benchmarking and operational guardrails.  
+
+<|ref|>image<|/ref|><|det|>[[366, 464, 668, 780]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[117, 826, 880, 882]]<|/det|>
+<center>Figure 4. Error Mode Distribution in Model Performance. Error patterns between GPT-4o (top) and GPT-3.5-turbo-0125 (bottom). Segments represent a proportional distribution of error types: Instruction adherence (dark blue), agent/tool selection (gray), and code generation (light </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 878, 120]]<|/det|>
+green). Central numbers indicate total error counts for each model, demonstrating significant differences in overall error rates (20 versus 45 errors, respectively).  
+
+<|ref|>sub_title<|/ref|><|det|>[[147, 178, 631, 196]]<|/det|>
+### 2.6 Pushing the Limits of Autonomous Experimentation  
+
+<|ref|>text<|/ref|><|det|>[[118, 198, 879, 252]]<|/det|>
+Finally, to demonstrate AILA's capabilities in real- world scenarios, we demonstrate three increasingly complex experimental tasks that typically require expert intervention: automated AFM calibration, high- resolution feature detection, and mechanical property measurement.  
+
+<|ref|>title<|/ref|><|det|>[[177, 273, 504, 290]]<|/det|>
+#### 2.6.1 AFM Parameter Optimization  
+
+<|ref|>text<|/ref|><|det|>[[117, 292, 880, 574]]<|/det|>
+AFM imaging requires precise calibration of Proportional- Integral- Derivative (PID) gain values, which traditionally demand expert intervention due to the continuous nature of these parameters. This dependency on skilled operators presents a significant barrier to broader AFM adoption. We demonstrate AILA's capability to autonomously optimize these parameters by minimizing the forward- backward scan differential on standard calibration grids. To this end, after loading the calibration sample, AILA was prompted to optimize the imaging parameters (see S4 in Supplementary Information for the complete prompt and output log). A total of 45 images are generated, with 3 images produced in each of the 15 generations. Figure 5A presents experimental AFM data acquired by AILA for the \(1^{\mathrm{st}}\) and \(15^{\mathrm{th}}\) generation of variable PID configurations, with corresponding line scan analyses that quantify trace- retrace symmetry. Initial scans with suboptimal parameters (P: 93- 208, I: 1747- 6623, D: 0- 39) exhibit poor SSIM scores (0.392- 0.768), manifesting as visible distortions in topographic data. Note that a higher SSIM value, closer to 1, indicates a perfect match, while a value of 0 represents no similarity. Through iterative optimization, AILA achieves superior scan quality (SSIM \(>0.81\) ) with optimized parameters (P: 246- 249, I: 8676- 8957, D: 17- 30; see Figure 5A).  
+
+<|ref|>text<|/ref|><|det|>[[118, 593, 879, 686]]<|/det|>
+The genetic algorithm's convergence efficiency is demonstrated in Figure 5C, where optimal PID configurations are achieved within 15 generations. Both maximum and mean SSIM values show rapid improvement, stabilizing above 0.8, indicating robust parameter optimization. Figure 5B validates the optimized parameters (P:249, I:8957, D:26) across a larger scan area, maintaining high- quality imaging across multiple grid features.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[120, 85, 875, 590]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[117, 593, 881, 741]]<|/det|>
+<center>Figure 5. Autonomous AFM experiments. A. Evolution of AFM image quality under varying PID parameters. The left panels show topographic images of the calibration grid; the right panels display corresponding line scan profiles (solid: trace, dashed: retrace). SSIM scores quantify trace-retrace correlation, with higher values indicating superior imaging quality. Optimal parameters (P: 249, I: 8957, D: 26) achieve \(\mathrm{SSIM} = 0.818\) . B. Large-area scan demonstrating consistent imaging quality using optimized parameters across multiple grid features. C. Convergence plot showing genetic algorithm optimization efficiency. Red circles: maximum SSIM; black circles: mean SSIM per generation. </center>  
+
+<|ref|>title<|/ref|><|det|>[[177, 762, 563, 780]]<|/det|>
+#### 2.6.2 High-Resolution Step-Edge Detection  
+
+<|ref|>text<|/ref|><|det|>[[117, 781, 880, 912]]<|/det|>
+Surface characterization through AFM is challenged by noise sources such as thermal drift, mechanical vibrations, and electronic interference<sup>39- 41</sup>, which can obscure subtle topographic features like graphene step edges. In this study, we leverage the advanced analytical capabilities of AILA to address these challenges using highly ordered pyrolytic graphite (HOPG) as a model system. AILA autonomously determines the necessity for baseline correction based on feature size, recognizing that baseline artifacts predominantly affect smaller features. For instance, in the raw image (Figure 6A), the graphene step edge remains indiscernible due to
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 84, 880, 252]]<|/det|>
+baseline distortions. AILA applies a fifth- order polynomial baseline correction to generate the \(1^{\mathrm{st}}\) generation image (Figure 6A), which serves as the foundation for PID gain optimization. Following a process similar to the calibration grid optimization, the image is refined through iterative PID adjustments, resulting in the final optimized image in the \(10^{\mathrm{th}}\) generation, where atomic steps become distinctly visible. The automated optimization process surpasses conventional manual adjustments, offering an enhanced resolution of nanoscale features. Additionally, further analysis of the processed data, including the determination of graphene step height, was facilitated through specific prompts, with the prompts and results detailed in the Supplementary Information S4.  
+
+<|ref|>image<|/ref|><|det|>[[123, 273, 866, 720]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[117, 727, 880, 819]]<|/det|>
+<center>Figure 6. Autonomous AFM experiments. A. High-resolution HOPG imaging demonstrating baseline artifact challenges. Top panels: topographic images at different Z ranges; bottom panels: corresponding line profiles revealing surface features. B. Left: HOPG images obtained using setpoints of 0.2 V and 0.4 V, both manually captured and taken by AILA with GPT-4o. Right: Plot showing the relationship between setpoint and average friction. </center>  
+
+<|ref|>title<|/ref|><|det|>[[176, 840, 598, 857]]<|/det|>
+#### 2.6.3 Load-dependent roughness measurement  
+
+<|ref|>text<|/ref|><|det|>[[118, 859, 879, 914]]<|/det|>
+We conduct a comprehensive load- dependent roughness analysis of HOPG. The experiment requires iterative adjustments of AFM parameters, including setting a range of setpoints, capturing images, and analyzing the corresponding friction data. Manually performing this
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 878, 139]]<|/det|>
+procedure is time- intensive, involving parameter modifications, image acquisition, data extraction, and result plotting. To streamline this process, we utilize AILA to automate the experiment.  
+
+<|ref|>text<|/ref|><|det|>[[117, 160, 880, 404]]<|/det|>
+AILA was instructed (see S4 in Supplementary materials for the complete prompt and output log) to vary the setpoint voltage from 0.2 V to 1.2 V in increments of 0.2 V. At each setpoint, AILA independently captured the AFM image, calculated the average friction value, and generated the corresponding plot. Figure 6B presents the graph of average friction versus setpoint voltage for both manually obtained and AILA- captured images using the GPT- 4o model. The raw images generated by AILA can be found in the Supplementary Information Figure S2. The entire process was conducted without additional user input regarding figure formatting or parameter settings. Remarkably, AILA autonomously develops the required Python script, executes the experimental protocol, and generates the output, including raw, unedited plots. This automation significantly reduces the time and effort compared to manual execution. The results not only validate the capability of AILA in handling complex AFM experiments but also demonstrate its efficiency in generating reproducible and high- quality outputs for scientific analyses.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 427, 272, 443]]<|/det|>
+## 3. Discussion  
+
+<|ref|>text<|/ref|><|det|>[[118, 446, 880, 594]]<|/det|>
+AILA's evaluation of AFMBench reveals quantifiable distinctions between LLM capabilities in experimental automation. The performance gap between GPT- 4o and GPT- 3.5- turbo- 0125 in cross- domain operations (80% versus 0% success in hybrid tasks) quantifies a key requirement for autonomous laboratories: integrated experimental reasoning. Single- domain competence, while necessary, proves insufficient for replicating expert- level experimental workflows. This disparity in cross- domain integration particularly manifests in complex scenarios, requiring simultaneous optimization of imaging parameters and data analysis—tasks routinely performed by experienced microscopists.  
+
+<|ref|>text<|/ref|><|det|>[[118, 614, 880, 840]]<|/det|>
+Error pattern analysis identifies a persistent challenge: protocol- to- execution translation. Code generation errors dominate failure modes (60- 67%) across both architectures, establishing a primary bottleneck in experimental automation. This systematic limitation persists even in advanced models, suggesting architectural constraints beyond mere computational capability. The distribution of error types—particularly the presence of instruction adherence errors (15%) in GPT- 4o versus their absence in GPT- 3.5- turbo- 0125—reveals nuanced differences in model behavior that warrant further investigation regarding the safety and alignment issues associated with the LLMs. It is worth noting that AI- agent based on GPT- 4o, a model known for its reasoning capabilities, exhibit tendencies to perform actions beyond what is instructed, termed as “sleepwalking”. Note that depending on the nature of the LLM, “sleepwalking” could result in potentially harmful outcomes in SDLs when dealing with highly dangerous chemicals or conditions such as high temperature pressure to name a few.  
+
+<|ref|>text<|/ref|><|det|>[[118, 860, 876, 896]]<|/det|>
+AILA's modular design establishes quantifiable metrics for autonomous system development. The successful automation of AFM operations validates this architecture for complex
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 879, 196]]<|/det|>
+instrumentation control, demonstrating efficacy in parameter optimization and feature extraction tasks. The measured limitations in tool coordination depending on the LLMs define specific thresholds for improving inter- module communication protocols. Notably, the agent deployment ratio (83:17 single- to- multi- agent) establishes an empirical baseline for balancing specialized and integrated operations—a metric applicable to automation across analytical platforms, from mass spectrometers to X- ray diffractometers.  
+
+<|ref|>text<|/ref|><|det|>[[118, 216, 879, 310]]<|/det|>
+These findings suggest specific architectural improvements for next- generation autonomous laboratories. Enhanced integration protocols between specialized agents could address the observed limitations in multi- tool coordination. Similarly, dedicated code generation modules might mitigate the predominant error mode, potentially incorporating specialized scientific programming frameworks.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 348, 252, 364]]<|/det|>
+## 4. Outlook  
+
+<|ref|>text<|/ref|><|det|>[[118, 367, 880, 573]]<|/det|>
+Altogether, AILA demonstrates quantifiable advances in experimental automation through systematic benchmarking. The framework's comprehensive performance metrics in AFM operations establish standards for autonomous laboratory evaluation, while AFMBench introduces reproducible protocols for systematic assessment across experimental domains. Successful execution of complex tasks—from automated image optimization to nanomechanical measurements—validates the framework's capabilities for sophisticated materials characterization. However, the observed tendencies of LLM agents to exceed operational boundaries and perform “sleepwalking” while carrying out the experiments raise significant safety concerns. This behavior, reported for the first time to the best of that authors’ knowledge, emphasize critical areas for development in instruction alignment and operational safety.  
+
+<|ref|>text<|/ref|><|det|>[[118, 593, 880, 837]]<|/det|>
+This work’s implications transcend materials characterization. The empirically validated principles—modular architecture (69% multi- tool efficiency), strategic agent deployment (83:17 ratio), and cross- domain integration (80% hybrid task success)—establish design parameters for autonomous laboratories across disciplines. Applications span pharmaceutical screening, environmental monitoring, and process optimization. For instance, the documented success in parameter optimization could translate directly to automated high- throughput drug screening or catalyst discovery platforms. While current limitations in code generation (60% error rate) and tool coordination (31% efficiency) define immediate development targets, these metrics provide clear objectives for advancing autonomous scientific platforms. The path forward requires focused development in three key areas: enhanced cross- domain reasoning capabilities, robust code generation protocols, and sophisticated multi- agent coordination mechanisms. Success in these domains would enable truly autonomous scientific platforms capable of accelerating discovery across the scientific landscape.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[118, 85, 232, 101]]<|/det|>
+## Methodology  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 105, 170, 119]]<|/det|>
+## AILA  
+
+<|ref|>text<|/ref|><|det|>[[117, 123, 880, 422]]<|/det|>
+AILA is constructed utilizing the LangChain software framework, incorporating components such as prompts, LLMs, memory, agents, and tools. AILA uses two categories of prompts: system prompts (see S2.1 in Supplementary Information for the system prompts) and user prompts. System prompts define ethical rules for AILA's interactions and describe the responsibilities assigned to each agent, whereas user prompts are variable inputs provided by end- users. AILA's backbone consists of LLMs, namely GPT- 4o and GPT- 3.5- turbo- 0125, which process user input as strings and provide string- based outputs. These LLMs are stateless, indicating that they do not save conversational context. Here, all interactions and agent states are stored in a Python dictionary and can be accessed by other agents. AILA consists of two specialized agents: the AFM Handler Agent and the Data Handler Agent, both equipped with unique tools to do specific tasks. These agents possess individual prompts, LLMs, and tools; however, they utilize a shared memory to store and access states, facilitating smooth interaction. The system prompts within the agents offer instructions for tool utilization and ethical guidelines, whereas the outputs from other tools or agents serve as user prompts. The framework utilizes LangGraph, a library that allows the construction of an effective multi- agent workflow, integrating all agents and tools seamlessly.  
+
+<|ref|>text<|/ref|><|det|>[[117, 442, 880, 723]]<|/det|>
+The architecture for AILA's decision- making process is carefully designed to ensure precise information routing. AILA can dynamically select among three primary options: AFM Handler, Data Handler, or FINISH. When AILA identifies the appropriate agent to handle a query, it routes the information to the selected option. In cases where AILA determines that none of the available agents can sufficiently address the question, it generates a well- structured response and selects the FINISH option to conclude the session effectively. The agents within this system are equipped with three distinct operational choices: utilizing their respective tools, transferring information to the next agent, or terminating the session. A system prompt has been integrated to streamline these decisions. Agents append the prefix NEED HELP to their response when transferring information to another agent. Alternatively, if they believe the query has been adequately addressed, they use the prefix FINAL ANSWER to signal the session's conclusion. By analyzing the output for these keywords, the system seamlessly routes the response to the designated agent or tool or finalizes the session. This structured approach enables efficient multi- agent collaboration, ensuring clarity, accuracy, and optimal performance across tasks while maintaining a robust and adaptive framework.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 744, 299, 761]]<|/det|>
+## AFM Handler Agent  
+
+<|ref|>text<|/ref|><|det|>[[118, 763, 880, 911]]<|/det|>
+Atomic Force Microscopy demands precise sequential execution of multiple experimental stages. Image acquisition requires optimization across three critical parameters: imaging conditions, probe selection, and operational mode configuration (tapping/contact). The experimental sequence encompasses surface approach protocols, scanning procedures, and standardized data acquisition—with procedural deviations potentially resulting in equipment damage or data corruption. Our implementation utilizes the DriveAFM instrument (Nanosurf), which is accessed through a Python- based API architecture and designed for universal compatibility with API- enabled AFM systems. To facilitate AFM imaging experiments, we
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 880, 158]]<|/det|>
+have created the AFM Handler agent, which is integrated with two specialized tools: the Document Retrieval Tool and the Code Executor Tool. Every tool has an individual role, and the AFM Handler agent can dynamically assign tasks to these tools. The agent will assign the responsibility to the Data Handler agent if it finds that neither tool can handle the task.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 179, 330, 195]]<|/det|>
+## Document Retrieval tool  
+
+<|ref|>text<|/ref|><|det|>[[117, 199, 880, 610]]<|/det|>
+The documentation for the instrument offers detailed instructions on how to handle and calibrate it. However, providing full access to the documentation to an LLM entails risks, such as inadvertent alterations to factory settings or calibration data, which could potentially result in damage or malfunction of the instrument. To address this concern, we manually extracted the essential information from the AFM documentation necessary for conducting experiments while safeguarding the instrument's integrity. We consolidated all the crucial codes for regulating each parameter of the instrument into a comprehensive Python script. Since Python code relies heavily on precise indentation and line structure, we utilized the Recursive Character Text Splitter from the LangChain library, specifically designed for Python, to divide the script into manageable chunks. The chunk size was set to a maximum of 1000 characters without overlap, adhering to the token limit for embedding models. Each code chunk comprises three sections: the first includes the necessary Python libraries, the second contains the code required to load the application, and the third section features unique Python code specific to the given task. The first two sections are consistent across all chunks (see S2.2 in the Supplementary Information file for more details). These chunks were then combined to generate a document, embedded using OpenAI's text- embedding- 3- large model. This model, with the capability of producing embeddings of size up to 3072 dimensions, delivers exceptional performance compared to other OpenAI embedding models, especially in multi- language retrieval benchmarks like MIRACL<sup>42</sup>. To store the embeddings, we opted for Chroma, an open- source vector database known for its reliability and efficiency in managing large- scale embedding data. We use a vector store retriever to retrieve the data from the vector store.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 631, 286, 647]]<|/det|>
+## Code Executor tool  
+
+<|ref|>text<|/ref|><|det|>[[118, 650, 880, 799]]<|/det|>
+A code executor tool has been developed to execute Python scripts generated by the AFM Handler Agent to control the AFM software. This tool is intended to run Python code, provided as a text string, directly on the local system to allow for smooth integration with the workflow of the AFM Handler Agent. The utility executes the code and sends back a success message or a detailed description of the error that occurred. If there is an error, the error message is returned to the AFM Handler agent so it can correct the error and retry executing. Otherwise, if the script runs without errors, it is considered the final result. This iterative process ensures precise control of the AFM system while systematically addressing any issues in the script.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 820, 295, 836]]<|/det|>
+## Data Handler Agent  
+
+<|ref|>text<|/ref|><|det|>[[118, 839, 879, 911]]<|/det|>
+Surface tracking optimization in AFM requires precise calibration of three fundamental parameters: Proportional (P), Integral (I), and Derivative (D) gains. Optimal calibration manifests as convergence between trace and retrace signals, indicating stable scanning conditions. The Data Handler agent interfaces with specialized optimization and analysis
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 879, 139]]<|/det|>
+modules; these models can access AFM image data stored in local storage systems. The agent can optimize P, I, and D gains or calculate various surface properties, such as average friction and surface roughness, using the help of modules and image files stored locally.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 160, 338, 177]]<|/det|>
+## Image Optimization Tool  
+
+<|ref|>text<|/ref|><|det|>[[118, 179, 880, 366]]<|/det|>
+The feedback system in an Atomic Force Microscope (AFM) plays a crucial role in maintaining control over the interaction between the cantilever tip and the sample surface. During scanning, variations in surface features alter the interaction forces between the tip and the sample, leading to deflections in the cantilever. These deflections are detected by a photodetector. To ensure that these deflections stay within a specified range, the feedback mechanism continuously adjusts the height of either the tip or the sample stage in real time. This process is managed by a PID (Proportional- Integral- Derivative) controller, which regulates the position of the z- piezo actuator. By moving the cantilever probe up or down, the controller maintains a steady interaction force or adheres to a predefined setpoint, depending on the chosen mode of operation.  
+
+<|ref|>text<|/ref|><|det|>[[117, 386, 880, 592]]<|/det|>
+Fine- tuning the P, I, and D gain values of the controller is vital for achieving accurate control of the setpoint in AFM imaging. The integral gain is especially important for enhancing image clarity by mitigating drift and reducing steady- state errors. Once the integral gain is optimized, adjusting the proportional gain can provide further refinement. The derivative gain, on the other hand, is particularly beneficial for imaging samples with pronounced edge features. If the gains are set too low, the PID loop may fail to maintain the setpoint effectively, while excessively high gain values can introduce electrical noise into the image due to amplified feedback or overcompensation for deviations. Properly optimized PID parameters ensure that the feedback loop remains stable and responsive, enabling the AFM to accurately track surface topography, even at higher scanning speeds. This balance is especially critical when imaging delicate, irregular, or soft materials, as it preserves the integrity of tip- sample interactions.  
+
+<|ref|>text<|/ref|><|det|>[[117, 611, 880, 817]]<|/det|>
+A genetic algorithm (GA) was employed for PID gain optimization. The GA parameters included a fixed population size of three and a total of 15 generations, enabling efficient tuning of the gains. Although these parameters can be manually adjusted, but excessive image scanning may degrade the AFM tip. The optimized gains ensure effective feedback control, producing comparable forward and backward images. This can be achieved by calculating the mean squared error (MSE) between forward and backward z- axis images for various PID gain settings. However, this method is sensitive to drift during scanning, and this method also depends on previously acquired images. To address this, the structural similarity index (SSIM) was adopted as the fitness function in the genetic algorithm, providing a robust measure of image similarity between the z- axis forward and backward image independent of prior image data.  
+
+<|ref|>text<|/ref|><|det|>[[118, 838, 879, 911]]<|/det|>
+This metric offers advantages over traditional Mean Square Error (MSE) approaches by (i) addressing tip degradation challenges in contact- mode AFM by minimizing required scan cycles and enabling optimization using low- resolution images, (ii) maintaining accuracy under drift conditions, (iii) incorporating structural, brightness, and contrast variations in
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 84, 880, 120]]<|/det|>
+optimization, and (iv) providing normalized scores between 0 and 1, where 1 indicates perfect similarity.  
+
+<|ref|>text<|/ref|><|det|>[[117, 141, 315, 157]]<|/det|>
+The SSIM is defined as:  
+
+<|ref|>equation<|/ref|><|det|>[[276, 160, 718, 181]]<|/det|>
+\[S S I M(x,y) = [l(x,y)]^{\alpha} \times [c(x,y)]^{\beta} \times [s(x,y)]^{\gamma}\]  
+
+<|ref|>text<|/ref|><|det|>[[117, 199, 880, 254]]<|/det|>
+where, \(l(x,y)\) is the luminance comparison, \(c(x,y)\) is the contrast comparison, and \(s(x,y)\) is the structure comparison with \(\alpha , \beta , \gamma\) being the weighting parameters. Note that the individual components are defined as:  
+
+<|ref|>equation<|/ref|><|det|>[[300, 255, 696, 320]]<|/det|>
+\[l(x,y) = (2\mu_{x}\mu_{y} + C_{1}) / (\mu x^{2} + \mu y^{2} + C_{1})\] \[c(x,y) = (2\sigma_{x}\sigma_{y} + C_{2}) / (\sigma_{x}{}^{2} + \sigma_{y}{}^{2} + C_{2})\] \[s(x,y) = (\sigma_{xy} + C_{3}) / (\sigma_{x}\sigma_{y} + C_{3})\]  
+
+<|ref|>text<|/ref|><|det|>[[117, 321, 880, 398]]<|/det|>
+where, \(\mu_{x}, \mu_{y}\) represent the mean intensities of images \(x\) and \(y\) , \(\sigma_{x}, \sigma_{y}\) is the standard deviations of images \(x\) and \(y\) , \(\sigma_{xy}\) is the cross- covariance between images \(x\) and \(y\) , and \(C_{1}, C_{2}, C_{3}\) are constants to avoid instability with \((C_{1} = (k_{1}L)^{2}, C_{2} = (k_{2}L)^{2}, C_{3} = C_{2} / 2)\) and \(L\) being the dynamic range of pixel values and \(k_{1} = 0.01\) and \(k_{2} = 0.03\) .  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 418, 285, 434]]<|/det|>
+## Baseline correction  
+
+<|ref|>text<|/ref|><|det|>[[118, 437, 875, 455]]<|/det|>
+The adaptive baseline correction employed in the step- edge detection of graphene is given by  
+
+<|ref|>equation<|/ref|><|det|>[[395, 455, 600, 476]]<|/det|>
+\[B(x,y) = \Sigma_{i,j}a_{ij}x^i y^j\]  
+
+<|ref|>text<|/ref|><|det|>[[117, 478, 880, 516]]<|/det|>
+Where, \(B(x,y)\) is the baseline function, \(a_{ij}\) are the polynomial coefficients, \(i\) and \(j\) are the polynomial degrees \((0 \leq i,j \leq n)\) with \(n\) being the maximum polynomial degree.  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 536, 290, 552]]<|/det|>
+## Image Analysis tool  
+
+<|ref|>text<|/ref|><|det|>[[117, 555, 881, 741]]<|/det|>
+AFM instrument stores the image data as a \*.nid file in the local system. This \*.nid file contains deflection, friction force, and z- axis images for both backward and forward scans. To further process any image from the file, exact data must be extracted from the file. To conduct this, we have used the NSFopen python library in the Image Analysis tool, which takes the query from the data handler agent regarding the specific image data and its location and returns the image data in an array to the data handler tool. To conduct further processing of the images, any Python script generated by the data handler tool can be executed in the Image Analysis tool, and the result can be returned to the data handler agent. Note that there is no database available to guide the LLM model in generating the Python script. It can generate the Python script by itself. There is a total of 6 input parameters for this tool:  
+
+<|ref|>text<|/ref|><|det|>[[117, 744, 850, 890]]<|/det|>
+(1) path (str): Directory path to search for the latest file (default: None). 
+(2) filename (str): Specific image file to display (default: None). 
+(3) dynamic_code (str): Python code for processing image data (default: None). 
+(4) calculate_friction (bool): Option to compute average friction (default: False). 
+(5) calculate_mean_roughness (bool): Option to compute mean roughness (default: False). 
+(6) calculate_rms_roughness (bool): Option to compute RMS roughness (default: False). Returns: A dictionary with the status, image data, or error details. Average friction was calculated using the following formula:
+
+<--- Page Split --->
+<|ref|>equation<|/ref|><|det|>[[403, 82, 593, 117]]<|/det|>
+\[F_{ave} = \frac{1}{2}\times (f_{ij} - b_{ij})\]  
+
+<|ref|>text<|/ref|><|det|>[[117, 118, 880, 176]]<|/det|>
+Where \(f_{ij}\) and \(b_{ij}\) are the element at position \((i,j)\) in the array of the forward and backward friction image data. We have used the formula in this tool to calculate the mean roughness and RMS roughness values  
+
+<|ref|>equation<|/ref|><|det|>[[366, 174, 627, 285]]<|/det|>
+\[R_{mean} = \frac{1}{M.N}\sum_{i = 1}^{M}\sum_{j = 1}^{N}|z_{ij} - \bar{z}|\] \[R_{rms} = \frac{1}{M.N}\sum_{i = 1}^{M}\sum_{j = 1}^{N}(z_{ij} - \bar{z})^2\]  
+
+<|ref|>text<|/ref|><|det|>[[117, 300, 880, 358]]<|/det|>
+where \(z_{ij}\) is the element at position \((i,j)\) in the array, \(\bar{z}\) is the mean of all elements in the array, \(M\) is the number of rows in the array, \(N\) is the number of columns in the array of the \(z\) - axis forward image data.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 379, 220, 395]]<|/det|>
+## AFMBench  
+
+<|ref|>text<|/ref|><|det|>[[117, 397, 880, 699]]<|/det|>
+Dataset preparation. To evaluate the performance of the AILA, we have manually created a set of 100 questions, carefully categorized into three distinct groups. The first classification is based on whether a question requires one or multiple tools/agents to be solved. The second category assesses the complexity of the questions, distinguishing between basic and advanced levels. Lastly, the questions are grouped by their requirements, such as documentation analysis or calculations. The complexity of each question is determined by the number of agents involved and the steps required to achieve the solution. For instance, modifying a parameter in an AFM system typically requires documentation review and the use of a single agent, categorizing it as a basic task. Conversely, capturing an AFM image and analyzing its surface roughness involves multiple agents, documentation analysis, and calculations, making it an advanced task. A comprehensive JSON file has been created, encapsulating detailed metadata about each question, including its respective category, for streamlined analysis and evaluation. This file serves as a structured resource for further investigations and testing. All questions, along with their relevant classifications and details, have been made accessible on GitHub (https://github.com/M3RG- IITD/AILA) to support transparency and reproducibility in research.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 720, 215, 735]]<|/det|>
+## Evaluation  
+
+<|ref|>text<|/ref|><|det|>[[118, 737, 880, 904]]<|/det|>
+We developed a graphical user interface (GUI) using Streamlit, an open- source Python framework, to streamline user interaction with AILA. The GUI allows users to input text- based queries, select the desired LLM model, and specify a log file name. It then executes AILA in the backend, saving the output log file locally and enabling users to observe results directly within the AFM software. Any output images or figures generated by AILA are also stored in the local system for further analysis. To ensure robustness, we manually evaluated all questions using GPT- 4o and GPT- 3.5- turbo- 0125, verifying the output log files and AFM software results multiple times in collaboration with different researchers to eliminate potential human errors. The evaluation of AILA's performance was categorized into two metrics: accuracy and
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 880, 215]]<|/det|>
+efficiency. For accuracy, questions were divided into categories based on complexity and tool/agent usage, with a percentage of correct answers calculated for each category. For efficiency, uniform parameters were maintained across models in the AFM software, including default settings of 0.1 seconds for time per line and 128 points per line and frame when not specified by the user. To ensure precise efficiency measurements, scanning time for images and the time taken by questions with incorrect answers were excluded from the analysis. Average response times were computed for each category to assess AILA's overall efficiency.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 236, 278, 252]]<|/det|>
+## Evaluation Metric  
+
+<|ref|>text<|/ref|><|det|>[[117, 254, 880, 459]]<|/det|>
+To assess the evaluation of questions in terms of accuracy, we classified the answers provided by AILA into two categories: fully correct answers and incorrect or partially correct answers. A fully correct answer was considered accurate, while any incorrect or incomplete response was deemed incorrect. Given that some questions require manual inspection of the AFM software to verify whether specific parameters are set correctly and whether the AFM image is captured as intended, multiple researchers were involved in verifying the results. They carefully checked the outcomes to ensure error- free results. For measurements of different properties, such as average friction, roughness, and RMS value of roughness, we used the Gwydion software to verify the accuracy of the results. Subsequently, the questions were clustered into appropriate groups, and the corresponding average percentage of correct answers was calculated.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 480, 355, 496]]<|/det|>
+## Data and Code Availability  
+
+<|ref|>text<|/ref|><|det|>[[118, 499, 880, 535]]<|/det|>
+All the tasks in AFMBench, along with the complete log files of the responses for each of the tasks from GPT- 4o and GPT- 3.5 are available at: https://github.com/M3RG- IITD/AILA  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 556, 279, 572]]<|/det|>
+## Acknowledgement  
+
+<|ref|>text<|/ref|><|det|>[[118, 574, 880, 648]]<|/det|>
+N.M.A.K. acknowledges the funding support from Google Research Scholar Award, and the Alexander von Humboldt Foundation. I.M. thanks University Grants Commission (UGC), Government of India for the NET- JRF fellowship (221610021768). The authors thank IIT Delhi HPC facility for computational and storage resources.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 688, 214, 704]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[115, 725, 874, 874]]<|/det|>
+1. Zhao, Z., Lee, W. S. & Hsu, D. Large language models as commonsense knowledge for large-scale task planning. Advances in Neural Information Processing Systems 36, (2024).  
+2. Liu, Y. et al. Machine Learning-Based Reward-Driven Tuning of Scanning Probe Microscopy: Towards Fully Automated Microscopy. Preprint at https://doi.org/10.48550/arXiv.2408.04055 (2024).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 84, 870, 135]]<|/det|>
+3. Lála, J. et al. Paperqa: Retrieval-augmented generative agent for scientific research. arXiv preprint arXiv:2312.07559 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 149, 822, 200]]<|/det|>
+4. Koscher, B. A. et al. Autonomous, multiproperty-driven molecular discovery: From predictions to measurements and back. Science 382, eadi1407 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 215, 852, 266]]<|/det|>
+5. Szymanski, N. J. et al. An autonomous laboratory for the accelerated synthesis of novel materials. Nature 624, 86-91 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 281, 866, 332]]<|/det|>
+6. Dai, T. et al. Autonomous mobile robots for exploratory synthetic chemistry. Nature 635, 890-897 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[115, 346, 873, 398]]<|/det|>
+7. Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570-578 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 411, 876, 496]]<|/det|>
+8. Sadeghi, S. et al. Engineering a Sustainable Future: Harnessing Automation, Robotics, and Artificial Intelligence with Self-Driving Laboratories. ACS Sustainable Chemistry & Engineering 12, 12695-12707 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[115, 510, 866, 595]]<|/det|>
+9. Delgado-Licona, F. & Abolhasani, M. Research acceleration in self-driving labs: Technological roadmap toward accelerated materials and molecular discovery. Advanced Intelligent Systems 5, 2200331 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[117, 608, 857, 659]]<|/det|>
+10. Sim, M. et al. ChemOS 2.0: An orchestration architecture for chemical self-driving laboratories. Matter 7, 2959-2977 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 673, 775, 725]]<|/det|>
+11. Darvish, K. et al. ORGANA: A robotic assistant for automated chemistry experimentation and characterization. Matter doi:10.1016/j.matt.2024.10.015.  
+
+<|ref|>text<|/ref|><|det|>[[117, 739, 819, 790]]<|/det|>
+12. Abolhasani, M. & Kumacheva, E. The rise of self-driving labs in chemical and materials sciences. Nature Synthesis 2, 483-492 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[117, 805, 852, 856]]<|/det|>
+13. Bayley, O., Savino, E., Slattery, A. & Noel, T. Autonomous chemistry: Navigating self-driving labs in chemical and material sciences. Matter 7, 2382-2398 (2024).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 876, 135]]<|/det|>
+14. Tom, G. et al. Self-driving laboratories for chemistry and materials science. Chemical Reviews 124, 9633–9732 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 148, 861, 201]]<|/det|>
+15. Canty, R. B. & Abolhasani, M. Reproducibility in automated chemistry laboratories using computer science abstractions. Nature Synthesis 1–13 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 214, 870, 267]]<|/det|>
+16. Volk, A. A. & Abolhasani, M. Performance metrics to unleash the power of self-driving labs in chemistry and materials science. Nature Communications 15, 1378 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 280, 806, 333]]<|/det|>
+17. Mirza, A. et al. Are large language models superhuman chemists? Preprint at https://doi.org/10.48550/arXiv.2404.01475 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 345, 872, 399]]<|/det|>
+18. Zaki, M., Jayadeva, Mausam & Krishnan, N. M. A. MaScQA: investigating materials science knowledge of large language models. Digital Discovery 3, 313–327 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 411, 816, 464]]<|/det|>
+19. Alampara, N. et al. Probing the limitations of multimodal language models for chemistry and materials research. arXiv preprint arXiv:2411.16955 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 477, 793, 530]]<|/det|>
+20. Miret, S. & Krishnan, N. M. A. Are LLMs Ready for Real-World Materials Discovery? Preprint at http://arxiv.org/abs/2402.05200 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 542, 850, 627]]<|/det|>
+21. Krishnan, N. M. A., Kodamana, H. & Bhattoo, R. Machine Learning for Materials Discovery: Numerical Recipes and Practical Applications. (Springer International Publishing, Cham, 2024). doi:10.1007/978-3-031-44622-1.  
+
+<|ref|>text<|/ref|><|det|>[[117, 640, 856, 725]]<|/det|>
+22. Krishnan, N. M. A., Kodamana, H. & Bhattoo, R. Natural Language Processing. in Machine Learning for Materials Discovery 263–275 (Springer International Publishing, Cham, 2024). doi:10.1007/978-3-031-44622-1_15.  
+
+<|ref|>text<|/ref|><|det|>[[117, 739, 870, 791]]<|/det|>
+23. Mishra, V. et al. LLaMat: Large Language Models for Materials Science Information Extraction. in (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 805, 840, 857]]<|/det|>
+24. Li, Z. et al. Mmsci: A multimodal multi-discipline dataset for phd-level scientific comprehension. in AI for Accelerated Materials Design-Vienna 2024 (2024).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 83, 844, 135]]<|/det|>
+25. Alampara, N. et al. MACBENCH: A multimodal chemistry and materials science benchmark.  
+
+<|ref|>text<|/ref|><|det|>[[115, 148, 857, 202]]<|/det|>
+26. Skarlinski, M. D. et al. Language agents achieve superhuman synthesis of scientific knowledge. Preprint at https://doi.org/10.48550/arXiv.2409.13740 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[115, 214, 855, 267]]<|/det|>
+27. Bian, K. et al. Scanning probe microscopy. Nature Reviews Methods Primers 1, 36 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 279, 818, 365]]<|/det|>
+28. Liu, Y. et al. AECroscopy: A Software-Hardware Framework Empowering Microscopy Toward Automated and Autonomous Experimentation. Small Methods 2301740 (2024) doi:10.1002/smtd.202301740.  
+
+<|ref|>text<|/ref|><|det|>[[115, 378, 832, 431]]<|/det|>
+29. Kandel, S. et al. Demonstration of an AI-driven workflow for autonomous high-resolution scanning microscopy. Nat Commun 14, 5501 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 443, 875, 528]]<|/det|>
+30. Vasudevan, R. K. et al. Autonomous Experiments in Scanning Probe Microscopy and Spectroscopy: Choosing Where to Explore Polarization Dynamics in Ferroelectrics. ACS Nano 15, 11253-11262 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 541, 814, 594]]<|/det|>
+31. Kalinin, S. V. et al. Automated and Autonomous Experiments in Electron and Scanning Probe Microscopy. ACS Nano 15, 12604-12627 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 606, 872, 692]]<|/det|>
+32. Arias, S., Zhang, Y., Zahl, P. & Hollen, S. Autonomous Molecular Structure Imaging with High-Resolution Atomic Force Microscopy for Molecular Mixture Discovery. J. Phys. Chem. A 127, 6116-6122 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 705, 841, 758]]<|/det|>
+33. Kang, S., Park, J. & Lee, M. Machine learning-enabled autonomous operation for atomic force microscopes. Review of Scientific Instruments 94, 123704 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[115, 771, 832, 825]]<|/det|>
+34. Zaki, M. et al. Interpretable Machine Learning Approach for Identifying the Tip Sharpness in Atomic Force Microscopy. Scripta Materialia 221, 114965 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[115, 837, 844, 890]]<|/det|>
+35. Kalinin, S. V. et al. Probe microscopy is all you need. Machine Learning: Science and Technology 4, 023001 (2023).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 844, 167]]<|/det|>
+36. Huang, L. et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems (2023).  
+
+<|ref|>text<|/ref|><|det|>[[117, 181, 847, 234]]<|/det|>
+37. Anwar, U. et al. Foundational challenges in assuring alignment and safety of large language models. arXiv preprint arXiv:2404.09932 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 247, 808, 300]]<|/det|>
+38. Liu, Y. et al. Trustworthy LLMs: A survey and guideline for evaluating large language models' alignment. arXiv preprint arXiv:2308.05374 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[117, 313, 866, 430]]<|/det|>
+39. Grobelny, J. et al. Size Measurement of Nanoparticles Using Atomic Force Microscopy. in Characterization of Nanoparticles Intended for Drug Delivery (ed. McNeil, S. E.) 71-82 (Humana Press, Totowa, NJ, 2011). doi:10.1007/978-1-60327-198-1_7.  
+
+<|ref|>text<|/ref|><|det|>[[117, 444, 809, 496]]<|/det|>
+40. Moreno-Flores, S. Baseline correction of AFM force curves in the force-time representation. Microscopy Research and Technique 79, 1045-1049 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[117, 510, 808, 562]]<|/det|>
+41. Burnham, N. A., Lyu, L. & Poulikakos, L. Towards artefact-free AFM image presentation and interpretation. Journal of Microscopy 291, 163-176 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[117, 576, 858, 628]]<|/det|>
+42. Zhang, X. et al. Making a MIRACL: Multilingual Information Retrieval Across a Continuum of Languages. Preprint at https://doi.org/10.48550/arXiv.2210.09984 (2022).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[43, 93, 768, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 333, 150]]<|/det|>
+SupplementaryMaterials.pdf
+
+<--- Page Split --->