Previous Predictions Research Articles

Short-Term Marine Wind Speed Forecasting Based on Dynamic Graph Embedding and Spatiotemporal Information

Predicting wind speed over the ocean is difficult due to the unequal distribution of buoy stations and the occasional fluctuations in the wind field. This study proposes a dynamic graph embedding-based graph neural network—long short-term memory joint framework (DGE-GAT-LSTM) to estimate wind speed at numerous stations by considering their spatio-temporal information properties. To begin, the buoys that are pertinent to the target station are chosen based on their geographic position. Then, the local graph structures connecting the stations are represented using cosine similarity at each time interval. Subsequently, the graph neural network captures intricate spatial characteristics, while the LSTM module acquires knowledge of temporal interdependence. The graph neural network and LSTM module are sequentially interconnected to collectively capture spatio-temporal correlations. Ultimately, the multi-step prediction outcomes are produced in a sequential way, where each step relies on the previous predictions. The empirical data are derived from direct measurements made by NDBC buoys. The results indicate that the suggested method achieves a mean absolute error reduction ranging from 1% to 36% when compared to other benchmark methods. This improvement in accuracy is statistically significant. This approach effectively addresses the challenges of inadequate information integration and the complexity of modeling temporal correlations in the forecast of ocean wind speed. It offers valuable insights for optimizing the selection of offshore wind farm locations and enhancing operational and management capabilities.

Mapping of Prion Structures in the Yeast Rnq1.

The Rnq1 protein is one of the best-studied yeast prions. It has a large potentially prionogenic C-terminal region of about 250 residues. However, a previous study indicated that only 40 C-terminal residues form a prion structure. Here, we mapped the actual and potential prion structures formed by Rnq1 and its variants truncated from the C-terminus in two [RNQ+] strains using partial proteinase K digestion. The location of these structures differed in most cases from previous predictions by several computer algorithms. Some aggregation patterns observed microscopically for the Rnq1 hybrid proteins differed significantly from those previously observed for Sup35 prion aggregates. The transfer of a prion from the full-sized Rnq1 to its truncated versions caused substantial alteration of prion structures. In contrast to the Sup35 and Swi1, the terminal prionogenic region of 72 residues was not able to efficiently co-aggregate with the full-sized Rnq1 prion. GFP fusion to the Rnq1 C-terminus blocked formation of the prion structure at the Rnq1 C-terminus. Thus, the Rnq1-GFP fusion mostly used in previous studies cannot be considered a faithful tool for studying Rnq1 prion properties.

PCGIMA: developing the web server for human position-defined CpG islands methylation analysis.

Introduction: CpG island (CGI) methylation is one of the key epigenomic mechanisms for gene expression regulation and chromosomal integrity. However, classical CGI prediction methods are neither easy to locate those short and position-sensitive CGIs (CpG islets), nor investigate genetic and expression pattern for CGIs under different CpG position- and interval- sensitive parameters in a genome-wide perspective. Therefore, it is urgent for us to develop such a bioinformatic algorithm that not only can locate CpG islets, but also provide CGI methylation site annotation and functional analysis to investigate the regulatory mechanisms for CGI methylation. Methods: This study develops Human position-defined CGI prediction method to locate CpG islets using high performance computing, and then builds up a novel human genome annotation and analysis method to investigate the connections among CGI, gene expression and methylation. Finally, we integrate these functions into PCGIMA to provide relevant online computing and visualization service. Results: The main results include: (1) Human position-defined CGI prediction method is more efficient to predict position-defined CGIs with multiple consecutive (d) values and locate more potential short CGIs than previous CGI prediction methods. (2) Our annotation and analysis method not only can investigate the connections between position-defined CGI methylation and gene expression specificity from a genome-wide perspective, but also can analysis the potential association of position-defined CGIs with gene functions. (3) PCGIMA (http://www.combio-lezhang.online/pcgima/home.html) provides an easy-to-use analysis and visualization platform for human CGI prediction and methylation. Discussion: This study not only develops Human position-defined CGI prediction method to locate short and position-sensitive CGIs (CpG islets) using high performance computing to construct MR-CpGCluster algorithm, but also a novel human genome annotation and analysis method to investigate the connections among CGI, gene expression and methylation. Finally, we integrate them into PCGIMA for online computing and visualization.

DSTCNN: Deformable spatial-temporal convolutional neural network for pedestrian trajectory prediction

Pedestrian trajectory prediction holds significant research value in service robots, autonomous driving, and intelligent monitoring. Currently, most pedestrian trajectory prediction methods focus on data-driven models based on recurrent neural networks, but there is insufficient research on data-driven models based on convolutional neural networks. In this study, we first analyze the two problems in pedestrian trajectory prediction methods based on convolutional neural networks: 1. Previous trajectory prediction methods based on convolutional neural networks have spatial-temporal entanglement problems; 2. These methods are limited by their fixed convolution kernels and cannot accurately model social and temporal interactions. Furthermore, we propose a deformable spatial-temporal convolutional neural network (DSTCNN) to better adapt to the pedestrian trajectory prediction task. The deformable spatial-temporal convolutional neural network models spatial and temporal interactions separately, overcoming the shortcomings of spatial-temporal entanglement. The deformable spatial-temporal convolution also gets rid of the fixed convolution kernel, making the modeling of spatial-temporal interactions more accurate. On the ETH and UCY datasets, the average displacement error and final displacement error of our method are 0.29 and 0.53 meters, respectively. In kernel density estimation, average Mahalanobis distance, and average maximum eigenvalue metrics, our method still achieves better performance compared to baseline methods. Moreover, the deformable spatial-temporal convolutional neural network is a memory-efficient model with only 4.1K parameters.

Business investment decision-making based on mathematical model and risk analysis

Business investments are prone to market risks, so pre-analysis is mandatory. The type of risk, its period, sustainability, and economic impact are the analyzable features for preventing loss and downfall. In recent years, mathematical models have been used for representing business cycles and analyzing the impacting risks. This article introduces a Decisive Risk Analytical Model (DRAM) for identifying spur defects in business investments. The proposed risk analytical model exploits the investments, returns, and influencing factors over the various market periods. The risk model is tuned for identifying the influencing factors across various small and large investment periods. The model is tuned to adapt to different economic periods split into a single financial year. In the process of tuning and training the mathematical analysis model, deep learning is used. The learning paradigm trains the risks and modifying features from expert opinion and previous predictions. Based on these three factors, the risk for the current investment is forecasted. The forecast aids in improving the new investment feasibilities with minimal risks and model modifications. The frequent market status is identified for preventing unnecessary risk-oriented forecasts using the training performed. Therefore, the proposed model is reliable in identifying risks and providing better investment recommendations.

Homogeneous oxidation of EE2 by ozone: Influencing factors, degradation pathway, and toxicity assessment

17α-Ethinylestradiol(EE2) is a common hormone drug that plays an essential role in endocrine and reproductive health. However, estrogens are degraded only to a limited extent by humans and livestock themselves, and most of them are excreted in undegraded or partially degraded form from human and livestock feces and urine into water bodies, posing a austere ecological risk. Of the known degradation processes in wastewater treatment plants, ozone-induced advanced oxidation shows tremendous potential. At the same time, the limitations of the test methodology makes it extremely difficult to do an in-depth investigation of the degradation mechanism. In this study, the influencing factors, including ozone dose, initial concentration, and pH were investigated. The degradation pathways of EE2 were analyzed by combing with quantum chemical calculations and high performance liquid chromatography coupled with quadrupole-time of flight tandem mass spectrometry (HPLC-Q-TOF-MS). Toxicity Estimation Software Tool (T.E.S.T) evaluated EE2 and its degradation intermediates for a variety of toxicities. And the results showed that reactions such as oxidative hydroxylation of the benzene ring and oxidative ring opening of the aliphatic ring mainly occurred during the degradation of estrogens, and the reaction was initially fast. The degradation effect is lower under neutral conditions, and the acidity and alkalinity of the solution are reduced with the reaction process under strong acid and alkaline conditions. The products of the experimental reactions reflect the active site, e.g., product P1 reflects C(5) as the site of easy reaction, which is consistent with previous theoretical predictions, suggesting that theoretical studies can provide complementary information on oxidation. The toxicity of intermediates obtained from partial ozone degradation improved compared to EE2, and the mutagenicity of EE2 was elevated greatly. However, single ozonation significantly diminished the developmental toxicity and bioaccumulation factor for a large proportion of intermediates. This work may contribute to the theoretical foundation for the practical application of ozone oxidation of environmental estrogens.

A single-cell genomic strategy for alternative transcript start sites identification.

Alternative transcription start sites (TSSs) usage plays a critical role in gene transcription regulation in mammals. However, precisely identifying alternative TSSs remains challenging at the genome-wide level. We report a single-cell genomic technology for alternative TSSs annotation and cell heterogeneity detection. In the method, we utilize Fluidigm C1 system to capture individual cells of interest, SMARTer cDNA synthesis kit to recover full-length cDNAs, then dual priming oligonucleotide system to specifically enrich TSSs for genomic analysis. We apply this method to a genome-wide study of alternative TSSs identification in two different IFN-β stimulated mouse embryonic fibroblasts (MEFs). The data clearly discriminate two IFN-β stimulated MEFs. Moreover, our results indicate 81% expressed genes in these two cell types containing multiple TSSs, which is much higher than previous predictions based on Cap-Analysis Gene Expression (CAGE) (58%) or empirical determination (54%) in various cell types. This indicates that alternative TSSs are more pervasive than expected and implies our strategy could position them at an unprecedented sensitivity. It would be helpful for elucidating their biological insights in future.

Multivariable machine learning models for clinical prediction of subsequent hip fractures in older people using the Chinese population database.

This study aimed to develop and validate clinical prediction models using machine learning (ML) algorithms for reliable prediction of subsequent hip fractures in older individuals, who had previously sustained a first hip fracture, and facilitate early prevention and diagnosis, therefore effectively managing rapidly rising healthcare costs in China. Data were obtained from Grade A Tertiary hospitals for older patients (age ≥ 60years) diagnosed with hip fractures in southwest China between 1 January 2009 and 1 April 2020. The database was built by collecting clinical and administrative data from outpatients and inpatients nationwide. Data were randomly split into training (80%) and testing datasets (20%), followed by six ML-based prediction models using 19 variables for hip fracture patients within 2years of the first fracture. A total of 40,237 patients with a median age of 66.0years, who were admitted to acute-care hospitals for hip fractures, were randomly split into a training dataset (32,189 patients) and a testing dataset (8,048 patients). Our results indicated that three of our ML-based models delivered an excellent prediction of subsequent hip fracture outcomes (the area under the receiver operating characteristics curve: 0.92 (0.91-0.92), 0.92 (0·92-0·93), 0.92 (0·92-0·93)), outperforming previous prediction models based on claims and cohort data. Our prediction models identify Chinese older people at high risk of subsequent hip fractures with specific baseline clinical and demographic variables such as length of hospital stay. These models might guide future targeted preventative treatments.

Performance of machine-learning approach for prediction ofpre-eclampsia in a middle-income country.

Pre-eclampsia (PE) is a serious complication of pregnancy associated with maternal and fetal morbidity and mortality. As current prediction models have limitations and may not be applicable in resource-limited settings, we aimed to develop a machine-learning (ML) algorithm that offers a potential solution for developing accurate and efficient first-trimester prediction of PE. We conducted a prospective cohort study in Mexico City, Mexico to develop a first-trimester prediction model for preterm PE (pPE) using ML. Maternal characteristics and locally derived multiples of the median (MoM) values for mean arterial pressure, uterine artery pulsatility index and serum placental growth factor were used for variable selection. The dataset was split into training, validation and test sets. An elastic-net method was employed for predictor selection, and model performance was evaluated using area under the receiver-operating-characteristics curve (AUC) and detection rates (DR) at 10% false-positive rates (FPR). The final analysis included 3050 pregnant women, of whom 124 (4.07%) developed PE. The ML model showed good performance, with AUCs of 0.897, 0.963 and 0.778 for pPE, early-onset PE (ePE) and any type of PE (all-PE), respectively. The DRs at 10% FPR were 76.5%, 88.2% and 50.1% for pPE, ePE and all-PE, respectively. Our ML model demonstrated high accuracy in predicting pPE and ePE using first-trimester maternal characteristics and locally derived MoM. The model may provide an efficient and accessible tool for early prediction of PE, facilitating timely intervention and improved maternal and fetal outcome. © 2023 The Authors. Ultrasound in Obstetrics & Gynecology published by John Wiley & Sons Ltd on behalf of International Society of Ultrasound in Obstetrics and Gynecology.

Network traffic prediction with Attention-based Spatial–Temporal Graph Network

Network traffic prediction plays a significant role in network management. Previous network traffic prediction methods mainly focus on the temporal relationship between network traffic, and used time series models to predict network traffic, ignoring the spatial information contained in traffic data. Therefore, the prediction accuracy is limited, especially in long-term prediction. To improve the prediction accuracy of the dynamic network traffic in the long term, we propose an Attention-based Spatial–Temporal Graph Network (ASTGN) model for network traffic prediction to better capture both the temporal and spatial relations between the network traffic. Specifically, in ASTGN, we exploit an encoder–decoder architecture, where the encoder encodes the input network traffic and the decoder outputs the predicted network traffic sequences, integrating the temporal and spatial information of the network traffic data through the Spatio-Temporal Embedding module. The experimental results demonstrate the superiority of our proposed method ASTGN in long-term prediction.

Enhancing Drug-Drug Interaction Prediction with Auxiliary Drug Similarity Estimation using Convolutional Neural Networks

In the last few years, a rapid development of artificial intelligence has revealed its possible applications in many tasks. One of those areas where artificial intelligence showed great promise was predicting drug-drug interactions, which had an enormous amount of data to be interpreted through. Drug-drug interactions are caused by interactions between drugs; it can decrease the effectiveness, lead to serious emergencies of the patient, or have synergies that increase the effectiveness of both. The previous knowledge-based drug-drug interaction prediction methods were costly and time consuming. Thus, a model that can predict these interactions is highly demanded for patients’ safety, furthermore, providing insights to scientists in medical fields. In this research, I propose a similarity aware drug-drug interaction predicting model. The proposed method consists of three modules: a drug feature extractor, a drug similarity estimator, and a drug-drug interaction predictor. The trained network is designed to be drug similarity-aware, which leads to significantly improved accuracy compared to the naive models. By incorporating drug similarity information, the model can provide more precise and reliable predictions for drug-drug interactions. The experimental result of the proposed deep-learning model achieved state-of-the-art performance compared to other previous methods.

Bayesian Vancomycin Model Selection for Therapeutic Drug Monitoring in Neonates.

Pharmacokinetic models can inform drug dosing of vancomycin in neonates to optimize therapy. However, the model selected needs to describe the intended population to provide appropriate dose recommendations. Our study aims to identify the population pharmacokinetic (PopPK) model(s) with the best performance to predict vancomycin exposure in neonates in our hospital. Relevant published PopPK models for vancomycin in neonates were selected based on demographics and vancomycin dosing strategy. The predictive performance of the models was evaluated in Tucuxi using a local cohort of 69 neonates. Mean absolute error (MAE), relative bias (rBias) and relative root mean square error (rRMSE) were used to quantify the accuracy and precision of the predictive performance of each model for three different approaches: a priori, a posteriori, and Bayesian forecasting for the next course of therapy based on the previous course predictions. A PopPK model was considered clinically acceptable if rBias was between ± 20 and 95% confidence intervals included zero. A total of 25 PopPK models were identified and nine were considered suitable for further evaluation. The model of De Cock et al. 2014 was the only clinically acceptable model based on a priori [MAE 0.35 mg/L, rBias 0.8 % (95% confidence interval (CI) -7.5, 9.1%), and rRMSE 8.9%], a posteriori [MAE 0.037 mg/L, rBias -0.23% (95% CI -1.3, 0.88%), and rRMSE 6.02%] and Bayesian forecasting for the next courses [MAE 0.89 mg/L, rBias 5.45% (95% CI -8.2, 19.1%), and rRMSE 38.3%) approaches. The De Cock model was selected based on a comprehensive approach of model selection to individualize vancomycin dosing in our neonates.

Evolution of delayed dispersal with group size effect and population dynamics

Individuals delay natal dispersal for many reasons. There may be no place to disperse to; immediate dispersal or reproduction may be too costly; immediate dispersal may mean that the individual and their relatives miss the benefits of group living. Understanding the factors that lead to the evolution of delayed dispersal is important because delayed dispersal sets the stage for complex social groups and social behavior. Here, we study the evolution of delayed dispersal when the quality of the local environment is improved by greater numbers of individuals (e.g., safety in numbers). We assume that individuals who delay natal dispersal also expect to delay personal reproduction. In addition, we assume that improved environmental quality benefits manifest as changes to fecundity and survival. We are interested in how do the changes in these life-history features affect delayed dispersal. We use a model that ties evolution to population dynamics. We also aim to understand the relationship between levels of delayed dispersal and the probability of establishing as an independent breeder (a population-level feature) in response to changes in life-history details. Our model emphasizes kin selection and considers a sexual organism, which allows us to study parent–offspring conflict over delayed dispersal. At evolutionary equilibrium, fecundity and survival benefits of group size or quality promote higher levels of delayed dispersal over a larger set of life histories with one exception. The exception is for benefits of increased group size or quality reaped by the individuals who delay dispersal. There, the increased benefit does not change the life histories supporting delay dispersal. Next, in contrast to previous predictions, we find that a low probability of establishing in a new location is not always associated with a higher incidence of delayed dispersal. Finally, we find that increased personal benefits of delayed dispersal exacerbate the conflict between parents and their offspring. We discuss our findings in relation to previous theoretical and empirical work, especially work related to cooperative breeding.

Swin transformer based transfer learning model for predicting porous media permeability from 2D images

Soil and rock, as typical porous media, widely exist in natural slopes and landslides and underground reservoirs. Accurate predicting the permeability of porous media is crucial in preventing geological disasters and extracting underground energy. The conventional methods for obtaining the porous media permeability involve experiments or numerical simulations, both of which are usually time-consuming. In recent years, a new approach to predicting the porous media permeability has emerged, utilizing porous media images and leveraging deep learning algorithms. However, previous image-based permeability prediction frameworks have been dominated by convolutional neural networks (CNNs). CNNs require extensive data and intricate network architectures to achieve reasonably accurate predictive performance. State-of-the-art neural networks and methods are employed to overcome the shortcomings of CNNs. To initiate this research, we generated 1200 two-dimensional porous media images. The permeability labels for the target porous media were acquired through computational fluid dynamics. Subsequently, an end-to-end permeability prediction surrogate model was established based on the Swin Transformer architecture. The proposed model falls under the category of transfer learning models, where pre-trained weights and bias from other datasets serve as the initial weights and bias for our dataset. Our model has achieved State-of-the-art performance, outperforming CNN by a significant margin on small-size datasets. Even with a small dataset size as 200, the proposed model still achieves an R2 score of 0.9717 on the testing set. Both transfer learning and the Swin Transformer architecture have demonstrated robust predictive capabilities in the context of predicting permeability in small-size dataset porous media based on computer vision.

Prediction of survival in out-of-hospital cardiac arrest: the updated Swedish cardiac arrest risk score (SCARS) model.

Out-of-hospital cardiac arrest (OHCA) is a major health concern worldwide. Although one-third of all patients achieve a return of spontaneous circulation and may undergo a difficult period in the intensive care unit, only 1 in 10 survive. This study aims to improve our previously developed machine learning model for early prognostication of survival in OHCA. We studied all cases registered in the Swedish Cardiopulmonary Resuscitation Registry during 2010 and 2020 (n = 55 615). We compared the predictive performance of extreme gradient boosting (XGB), light gradient boosting machine (LightGBM), logistic regression, CatBoost, random forest, and TabNet. For each framework, we developed models that optimized (i) a weighted F1 score to penalize models that yielded more false negatives and (ii) a precision-recall area under the curve (PR AUC). LightGBM assigned higher importance values to a larger set of variables, while XGB made predictions using fewer predictors. The area under the curve receiver operating characteristic (AUC ROC) scores for LightGBM were 0.958 (optimized for weighted F1) and 0.961 (optimized for a PR AUC), while for XGB, the scores were 0.958 and 0.960, respectively. The calibration plots showed a subtle underestimation of survival for LightGBM, contrasting with a mild overestimation for XGB models. In the crucial range of 0-10% likelihood of survival, the XGB model, optimized with the PR AUC, emerged as a clinically safe model. We improved our previous prediction model by creating a parsimonious model with an AUC ROC at 0.96, with excellent calibration and no apparent risk of underestimating survival in the critical probability range (0-10%). The model is available at www.gocares.se.

Machine learning models for predicting unscheduled return visits of patients with abdominal pain at emergency department and validation during COVID-19 pandemic: A retrospective cohort study.

Machine learning (ML) models for predicting 72-hour unscheduled return visits (URVs) for patients with abdominal pain in the emergency department (ED) were developed in a previous study. This study refined the data to adjust previous prediction models and evaluated the model performance in future data validation during the COVID-19 era. We aimed to evaluate the practicality of the ML models and compare the URVs before and during the COVID-19 pandemic. We used electronic health records from Chang Gung Memorial Hospital from 2018 to 2019 as a training dataset, and various machine learning models, including logistic regression (LR), random forest (RF), extreme gradient boosting (XGB), and voting classifier (VC) were developed and subsequently used to validate against the 2020 to 2021 data. The models highlighted several determinants for 72-hour URVs, including patient age, prior ER visits, specific vital signs, and medical interventions. The LR, XGB, and VC models exhibited the same AUC of 0.71 in the testing set, whereas the VC model displayed a higher F1 score (0.21). The XGB model demonstrated the highest specificity (0.99) and precision (0.64) but the lowest sensitivity (0.01). Among these models, the VC model showed the most favorable, balanced, and comprehensive performance. Despite the promising results, the study illuminated challenges in predictive modeling, such as the unforeseen influences of global events, such as the COVID-19 pandemic. These findings not only highlight the significant potential of machine learning in augmenting emergency care but also underline the importance of iterative refinement in response to changing real-world conditions.

PrecisionLymphoNet: Advancing Malignant Lymphoma Diagnosis via Ensemble Transfer Learning with CNNs.

Malignant lymphoma, which impacts the lymphatic system, presents diverse challenges in accurate diagnosis due to its varied subtypes-chronic lymphocytic leukemia (CLL), follicular lymphoma (FL), and mantle cell lymphoma (MCL). Lymphoma is a form of cancer that begins in the lymphatic system, impacting lymphocytes, which are a specific type of white blood cell. This research addresses these challenges by proposing ensemble and non-ensemble transfer learning models employing pre-trained weights from VGG16, VGG19, DenseNet201, InceptionV3, and Xception. For the ensemble technique, this paper adopts a stack-based ensemble approach. It is a two-level classification approach and best suited for accuracy improvement. Testing on a multiclass dataset of CLL, FL, and MCL reveals exceptional diagnostic accuracy, with DenseNet201, InceptionV3, and Xception exceeding 90% accuracy. The proposed ensemble model, leveraging InceptionV3 and Xception, achieves an outstanding 99% accuracy over 300 epochs, surpassing previous prediction methods. This study demonstrates the feasibility and efficiency of the proposed approach, showcasing its potential in real-world medical applications for precise lymphoma diagnosis.

Superconductivity at 215 K in H3SM (M=Ne, Ar, Kr, Xe, Rn) ternary hydrides

The recent discovery of H3S attracts great attention due to the confirmed superconducting critical temperature (Tc) of 203 K. Following the previous structure prediction of H3SXe superconductors, the anisotropic Migdal-Eliashberg (ME) theory is carried out to study the superconducting properties of H3SM (M=Ne, Ar, Kr, Xe, Rn) under pressure. We identify a new H3SRn ternary hydride, with Tc of about 215 K under 240 GPa, belonging to two-gap superconductors with multi-band. H3SNe exhibits remarkable superconductivity with an anomalously low stabilization pressure (Ps) of about 60 GPa. With the motivation of discovering superconductors at the lowest possible pressure, the chemical pressure is quantified based on the same volume-strain effect of physical pressure and chemical pressure. In H3SM, the larger the chemical pressure of the central atom, the larger the physical pressure required for its stabilization. Away from the phase transition, decreasing pressure is beneficial to the superconductivity of the H3SM system. These findings open new avenues for designing and optimizing new high-Tc superconductors with low-Ps.

Physics‐informed surrogates for electromagnetic dynamics using Transformers and graph neural networks

AbstractA novel use case for two data‐driven models, namely, a Transformer and a convolutional graph neural network (CGNN) is proposed. The authors propose to use these models for emulating the dynamics of electromagnetic (EM) propagation and scattering. The Transformer translates a past sequence into a future sequence by constructing representations from the past and using it to predict the future, taking all of its own previous predictions as input at each step of prediction. The CGNN updates the current state of attribute vectors of each node by passing it information (messages) from all of its neighbouring nodes. We train these models with FDTD simulations of plane waves propagating and scattering from PEC objects. The authors demonstrate that, within the bounds of computational resources, the Transformer can be utilised as a surrogate for EM dynamics, providing 14× speed‐up, while the CGNN can be utilised as a next‐frame predictor, providing 9× speed‐up. When comparing the accuracy of these two models with the authors’ previously developed Encoder‐Recurrent‐Decoder (ERD) model, it is observed that the error for both the Transformer and the CGNN remains within the same bound for the ERD model. To the best of the authors’ knowledge, this work is the first to utilise the Transformer as a surrogate for EM dynamics.

Machine Learning for Polaritonic Chemistry: Accessing Chemical Kinetics.

Altering chemical reactivity and material structure in confined optical environments is on the rise, and yet, a conclusive understanding of the microscopic mechanisms remains elusive. This originates mostly from the fact that accurately predicting vibrational and reactive dynamics for soluted ensembles of realistic molecules is no small endeavor, and adding (collective) strong light-matter interaction does not simplify matters. Here, we establish a framework based on a combination of machine learning (ML) models, trained using density-functional theory calculations and molecular dynamics to accelerate such simulations. We then apply this approach to evaluate strong coupling, changes in reaction rate constant, and their influence on enthalpy and entropy for the deprotection reaction of 1-phenyl-2-trimethylsilylacetylene, which has been studied previously both experimentally and using ab initio simulations. While we find qualitative agreement with critical experimental observations, especially with regard to the changes in kinetics, we also find differences in comparison with previous theoretical predictions. The features for which the ML-accelerated and ab initio simulations agree show the experimentally estimated kinetic behavior. Conflicting features indicate that a contribution of dynamic electronic polarization to the reaction process is more relevant than currently believed. Our work demonstrates the practical use of ML for polaritonic chemistry, discusses limitations of common approximations, and paves the way for a more holistic description of polaritonic chemistry.