Physics-based Approaches Research Articles

Advancing Marine Surveillance: A Hybrid Approach of Physics Infused Neural Network for Enhanced Vessel Tracking Using Automatic Identification System Data

In the domain of maritime surveillance, the continuous tracking and monitoring of vessels are imperative for the early detection of potential threats. The Automatic Identification System (AIS) database, which collects vessel movement data over time, including timestamps and other motion details, plays a crucial role in real-time maritime monitoring. However, it frequently exhibits irregular intervals of data collection and intricate, intersecting trajectories, underscoring the importance of analyzing long-term temporal patterns for effective vessel tracking. While Kalman Filters and other physics-based models have been employed to tackle these issues, their effectiveness is limited by their inability to capture long-term dependence and non-linearity in the historical data. This paper introduces a novel approach that leverages Long Short-Term Memory (LSTM), a type of recurrent neural network, renowned for its proficiency in recognizing patterns over extended periods. Recognizing the strengths and limitations of the LSTM model, we propose a hybrid machine-learning algorithm that integrates LSTM with a physics-based model. This combination harnesses the physical laws governing vessel movements alongside data driven pattern mining, thereby enhancing the predictive accuracy of vessel locations. To assess the performance of standalone and hybrid models, various scenarios with different levels of complexity are generated. Furthermore, to simulate real-world data loss conditions often encountered in maritime tracking, temporal data gaps are randomly introduced into the scenarios. The competing approaches are then evaluated using both with time gap and without time gap conditions. Our results show that, although the LSTM model performs better than the physics-based model, the hybrid model consistently outperforms both standalone models across all scenarios. Furthermore, while data gaps negatively impact the accuracy of all models, the performance reduction is minimal for the physics-infused model. In summary, this study not only demonstrates the potential of combining data-driven and physics-based approaches but also sets a new benchmark for maritime vessel tracking.

Physics-Informed Neural Networks for the Reynolds Equation with Transient Cavitation Modeling

Gaining insight into tribological systems is crucial for optimizing efficiency and prolonging operational lifespans in technical systems. Experimental investigations are time-consuming and costly, especially for reciprocating seals in fluid power systems. Elastohydrodynamic lubrication (EHL) simulations offer an alternative but demand significant computational resources. Physics-informed neural networks (PINNs) provide a promising solution using physics-based approaches to solve partial differential equations. While PINNs have successfully modeled hydrodynamics with stationary cavitation, they have yet to address transient cavitation with dynamic geometry changes. This contribution applies a PINN framework to predict pressure build-up and transient cavitation in sealing contacts with dynamic geometry changes. The results demonstrate the potential of PINNs for modeling tribological systems and highlight their significance in enhancing computational efficiency.

Short-Circuit Detection in Lithium-Ion Batteries Using Machine Learning: Analysis and Comparison with Physics-Based Method

Early detection of short circuits in battery-powered systems is critical in preventing potential catastrophic failures. However, nascent short-circuit signatures are extremely weak and challenging to detect using existing algorithms without compromising on prediction accuracy. Traditional physics-based approaches rely on hand-crafted models to establish relationships between battery operating parameters and short resistance, which limits their ability to capture all relevant details, resulting in sub-optimal accuracies. In this study, we present a machine learning-based approach that leverages rest period voltage data to detect short circuits. Our method employs a 1D convolutional neural network (CNN) classifier/estimator that extracts temporal dynamic features relevant to the short circuit prediction problem from both the long and short tails of the rest period voltage profile. The approach is validated using commercial battery data, generated at different conditions including temperatures, and short circuits of varying severities; with prediction accuracies greater than 90% even for soft shorts of 500Ω. The key performance parameters of the 1D CNN model are compared against a physics-based short detection approach, demonstrating its superior performance and cost-effectiveness. Overall, our work represents a significant advancement in the field of short circuit detection in battery-powered systems, offering improved accuracy, efficiency, and cost-effectiveness.

Do LSTM memory states reflect the relationships in reduced-complexity sandy shoreline models

Equilibrium-based models are a transparent method of modelling shoreline change, though often too simplistic to capture complex dynamics. Conversely, deep learning methodologies offer greater predictive power at the expense of transparency. In this research we scrutinize the internal workings of an LSTM shoreline model. A regression-based probe is used to show that cell state vectors, responsible for past-to-future information flow, autonomously generate equilibrium-like information akin to the physics-based equilibrium term of the ShoreFor model, Ωeq. The variation in probe skill throughout training is tracked to show that at 5 of 6 transects, the LSTM was able to meaningfully acquire equilibrium information (ΣΔR2 = 0.3–0.6). The results of this work offer evidence that an LSTM may model shoreline change with internal methods that are consistent with the current understanding of coastal shoreline dynamics. These physically meaningful representations emphasize the importance of co-evolution between machine learning and physics-based approaches moving forward.

APPROACH TO CREATION OF GEOLOGICAL AND RESERVOIR SIMULATION MODEL AND METHODS FOR HISTORY MATCHING OF RESERVOIR ENERGY STATE

Under conditions of complex geological structure and substantial share of hard-to-recover oil reserves, reservoir simulation is a key tool in reservoir management. Direct problems solved by reservoir simulation include forecasting of the production performance, control and management of reservoir development, and localization of residual oil reserves. Inverse problem solution is useful for updating the knowledge of geological structure, reservoir quality, and historical production data by means of history matching. The quality of history matching is fundamental to ensure reliable future forecasts and localization of residual oil zones. The main approaches to appropriate history matching rely on analysis and adjustment of well logging and geological parameters, as well as checking and correction of errors, inadequate historical data, application of physics-based and reasonable approaches and tools. The prerequisite for adequate history matching and predictive ability of the resultant reservoir simulation model is the compliance of simulated and actual energy state of the reservoir. In this paper, the initial stage of reservoir simulation is performed, the approach to history matching of reservoir energy state aimed at improving the quality of further predictions is described as exemplified by one of the production areas of Russian oil field. The main causes of poor history match of formation and bottomhole pressures are revealed. These causes are grouped as follows: — geological causes/reservoir quality, which include permeability field distribution, reservoir connectivity, and determination of rock and fluid compressibility;— technical causes (during reservoir simulation), that is model limitations in case of larger areas, the need to use sector models, consideration of buffer region;— technological causes, including inadequate historical data or lack of data, detection of behind-the-casing flows and production casing leaks in wellbore, etc. For each group of factors that have a substantial impact on reservoir pressure, recommendations and methods of energy state history matching on geological and reservoir simulation models are provided, examples are given.

Physics-Based vs Data-Driven 24-Hour Probabilistic Forecasts of Precipitation for Northern Tropical Africa

Abstract Numerical weather prediction (NWP) models struggle to skillfully predict tropical precipitation occurrence and amount, calling for alternative approaches. For instance, it has been shown that fairly simple, purely data-driven logistic regression models for 24-h precipitation occurrence outperform both climatological and NWP forecasts for the West African summer monsoon. More complex neural network–based approaches, however, remain underdeveloped due to the non-Gaussian character of precipitation. In this study, we develop, apply, and evaluate a novel two-stage approach, where we train a U-Net convolutional neural network (CNN) model on gridded rainfall data to obtain a deterministic forecast and then apply the recently developed, nonparametric Easy Uncertainty Quantification (EasyUQ) approach to convert it into a probabilistic forecast. We evaluate CNN+EasyUQ for 1-day-ahead 24-h accumulated precipitation forecasts over northern tropical Africa for 2011–19, with the Integrated Multi-satellitE Retrievals for GPM (IMERG) data serving as ground truth. In the most comprehensive assessment to date, we compare CNN+EasyUQ to state-of-the-art physics-based and data-driven approaches such as monthly probabilistic climatology, raw and postprocessed ensemble forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), and traditional statistical approaches that use up to 25 predictor variables from IMERG and the ERA5 reanalysis. Generally, statistical approaches perform about on par with postprocessed ECMWF ensemble forecasts. The CNN+EasyUQ approach, however, clearly outperforms all competitors in terms of both occurrence and amount. Hybrid methods that merge CNN+EasyUQ and physics-based forecasts show slight further improvement. Thus, the CNN+EasyUQ approach can likely improve operational probabilistic forecasts of rainfall in the tropics and potentially even beyond. Significance Statement Precipitation forecasts in the tropics remain a great challenge despite their enormous potential to create socioeconomic benefits in sectors such as food and energy production. Here, we develop a purely data-driven, machine learning–based prediction model that outperforms traditional, physics-based approaches to 1-day-ahead forecasts of rainfall occurrence and rainfall amount over northern tropical Africa in terms of both forecast skill and computational costs. A combined data-driven and physics-based (hybrid) approach yields further (slight) improvement in terms of forecast skill. These results suggest new avenues to more accurate and more resource-efficient operational precipitation forecasts in the Global South.

Predicting the compressive strength of CFRP-confined concrete using deep learning

The application of externally bonded fiber-reinforced polymers (FRP) has revolutionized structural rehabilitation, offering cost-effective and rapid solutions for damaged structures. This study presents a novel deep learning approach for predicting the compressive strength of circular carbon-FRP-confined concrete columns. Leveraging an extensive database of 664 experimental results, the model incorporates monotonicity and smoothness constraints, with hyperparameters optimization using the OPTUNA framework. The proposed model demonstrates exceptional accuracy, achieving R² = 0.93, a20-index = 0.95, and MAPE = 7.89 % on unseen test data, consistently outperforming nine benchmark models including established design guidelines. Scenario-based analysis confirms the model's ability to capture known physical behaviors, such as the effects of concrete strength, column diameter, and FRP thickness on confinement effectiveness. The integration of physical constraints enhances the model's reliability and interpretability, bridging the gap between data-driven and physics-based approaches. This research contributes to the advancement of more accurate, economical, and reliable design guidelines for FRP-strengthened structures, while also demonstrating the potential of constrained deep learning in structural engineering applications.

Quantifying Electrolyte Variation through Coupled Molecular Dynamics and Porous Electrode Modeling of Lithium-Ion Battery Cells

Lithium-ion battery cell modeling using physics-based approaches such as porous electrode theory is a powerful tool for battery design and performance analysis both at the beginning and end of life[1]. Cell metrics such as resistance and thermal performance[2-6] can be quickly calculated in a pseudo-two-dimensional (P2D) framework. For engineering of electric vehicle batteries, speed and fidelity of electrochemical models is paramount in a competitive landscape. Physics-based models[7, 8] allow for high fidelity but require detailed knowledge of the cell component material properties. Acquiring these material characteristics typically requires time-consuming and expensive experiments limiting the ability to quickly screen through cell designs. One approach to circumvent costly experiments is to use molecular dynamics[9, 10] to calculate electrolyte transport properties. We demonstrate how cell modeling using simulated transport properties enables predictions of cell level metrics, allowing for experiment-free component screening. We also show how the variation in transport property predictions from molecular dynamics affects the final cell level performance predictions.

Advancements in Li-Ion Battery Safety: A Multiphysics and Deep Learning Approach for Thermal Runaway Prediction

The increasing adoption of Lithium-ion batteries (LIBs) in various applications necessitates an enhanced focus on safety, particularly concerning the thermal runaway (TR) phenomenon. This study introduces a novel framework combining multiphysics modeling and deep learning (DL) to predict TR in cylindrical Li-ion batteries. The primary objective is to prevent catastrophic battery failures by identifying critical thermal conditions leading to TR.Our methodology integrates state-of-the-art convolutional neural networks (CNNs) and the "You Only Look Once" (YOLO) object detection model to classify and detect the evolution of TR stages in LIBs. The CNNs, trained on simulated thermal images generated through comprehensive multiphysics modeling, classify TR into three distinct stages: safe operation, critical condition, and actual TR occurrence. The YOLO model further pinpoints the location of TR heat sources, enhancing the predictive accuracy.The multiphysics model employed captures the coupled thermal and electrochemical behavior of LIBs, focusing on the solid electrolyte interface (SEI) formation/decomposition on the anode as the primary degradation mechanism. The simulated thermal imagery, reflecting real-time battery conditions, serves as a dataset for training our DL models. This combination of physics-based and data-driven approaches allows for a detailed understanding of TR dynamics.In our study, we also explored the efficiency of other neural network architectures, notably ResNet, EfficientNet, and MobileNet for their potential in TR prediction. ResNet, known for its deep residual learning framework, demonstrated proficiency in handling the complex, layered features of thermal images. EfficientNet, optimized for efficiency and accuracy, showed promise in its scalable architecture and balanced depth, width, and resolution, making it a viable alternative for real-time TR detection. The MobileNet known for its lightweight architecture and efficiency on mobile devices, was evaluated for its capability to process thermal images in a resource-constrained environment. It demonstrated a remarkable balance between computational efficiency and predictive accuracy, making it a viable option for real-time TR monitoring in portable LIB applications.While MobileNet, ResNet, and EfficientNet each showed potential in their respective capacities, our primary CNN-YOLO framework emerged as the most effective for TR prediction in this study. Its proficiency in extracting spatial features and detecting anomalies in real-time was critical for early TR identification in LIBs.Results demonstrate high accuracy in predicting TR stages and heat source locations, emphasizing the ability of the integrated approach. The study highlights the importance of thermal management and real-time monitoring in LIBs, offering insights into the thermal behavior under various load conditions. The predictive framework not only advances the fundamental understanding of TR in LIBs but also provides practical implications for enhancing battery safety in applications like electric vehicles and consumer electronics.In conclusion, this work presents a significant stride in battery research, paving the way for safer and more reliable energy storage solutions. The combined multiphysics modeling and DL framework offer a robust tool for predicting and preventing TR in LIBs, contributing to the broader goal of advancing battery technology in the modern energy landscape.

A Review on Modern Techniques in 3D Character Animation

3D character animation is a pivotal aspect of digital graphics, essential for various applications such as films, games, and simulations. This review paper explores advancements in character animation techniques, including linear blend skinning, dual quaternion skinning, and physics-based approaches. Emphasis is placed on recent innovations aimed at improving the realism and computational efficiency of character deformations, especially non-rigid transformations. By examining key methodologies and their applications, this paper provides a comprehensive overview of the current state and future directions in 3D character animation

The Art and Science of Molecular Docking.

Molecular docking has become an essential part of a structural biologist's and medicinal chemist's toolkits. Given a chemical compound and the three-dimensional structure of a molecular target-for example, a protein-docking methods fit the compound into the target, predicting the compound's bound structure and binding energy. Docking can be used to discover novel ligands for a target by screening large virtual compound libraries. Docking can also provide a useful starting point for structure-based ligand optimization or for investigating a ligand's mechanism of action. Advances in computational methods, including both physics-based and machine learning approaches, as well as in complementary experimental techniques, are making docking an even more powerful tool. We review how docking works and how it can drive drug discovery and biological research. We also describe its current limitations and ongoing efforts to overcome them.

Feature extraction and pattern recognition in time-lapse pressure transient responses

Monitoring of modern wells equipped with permanent downhole gauges and flowmeters provides large datasets of pressure, temperature and flowrate measurements. On a limited scale, selected events from these datasets are traditionally used in well performance analysis and monitoring, reservoir simulation, and many other tasks employing physics-based models. New methods that combine model- and data-driven approaches for full-scale analysis of these large datasets have recently attracted interest, suggesting new metrics for knowledge extraction. Most recent studies have used a combination of Pressure Transient Analysis (PTA), which is a key reservoir engineering tool, with hybrid methods, such as PTA-metrics and AI-powered models. The emergence of hybrid methodologies combining model- and data-driven approaches has opened new perspectives for comprehensive analysis and knowledge extraction from big well monitoring data sets, accumulated in the industry. Despite their potential, these methodologies often face challenges in reliability, effective data processing, interpretation and feature engineering.This paper introduces a novel PTA-feature extraction and pattern recognition methods for analysis of time-lapse pressure transient responses and their Bourdet derivatives. The pattern recognition method builds on an unsupervised classification method to extract PTA-features, associated with different flow regimes commonly used in PTA. Each pressure transient in a time-lapse series is first processed by an autonomously fine-tuned algorithm that measures the signal's distance to an ensemble of physically meaningful responses defined by PTA-feature library, thus breaking the transient into a series of likely PTA-features. A set of hyperparameters is used in the unsupervised classification, where an optimization procedure is employed for automated tuning of the hyperparameters to a particular transient. Subsequently, recognition of underlying patterns governed by sequences of these PTA-features in the time-lapse transient responses is performed. Testing of the combination of these new methods through synthetic and field cases is further carried out with verification via comparison with expert’ interpretation results. Added value to the previously introduced PTA metrics, which provide on-the-fly well and reservoir performance analysis, is finally demonstrated. The proposed pattern recognition method improves reliability and automates the calculation of the PTA metrics.The developed methodology serves as a new tool for knowledge extraction from big well monitoring datasets, available in the companies operating in the oil and gas industry, as well as the emerging industries such as carbon capture and storage and geothermal energy production. The article concludes with a discussion of the main advantages and limitations of the suggested feature extraction and pattern recognition methods. Besides the combined use of the methods described in this article, these methods may also be integrated with conventional physics-based approaches widely used in the industry for well data interpretation and reservoir simulations, improving their performance and efficiency for big data sets.

A Hybrid Approach for Damage Detection and Localization on a Thin Metallic Plate

This paper illustrates a hybrid method for identifying structural damage that includes concepts from Physics-Based approaches within a Data-Driven framework. The lessons learned from the use of damage identification methods based on the analysis of modal curvature allows us to develop effective feature engineering in the development of Machine Learning (ML) algorithms. Moreover, the training of the algorithms exploits data provided from a population of damage cases generated by Finite Element (FE) analysis. The above data-analysis pipeline is applied to the identification of damage in a rectangular metallic plate for which measured modal data are available from a dedicated experimental campaign in CNR-INM laboratory using accelerometers. The damage is modelled as a stiffness reduction over a small area being also representative of material thinning due to corrosion. To have meaningful samples of the training database, the modal convergence of the FE model to the real structure is guaranteed by a structural optimization process. Experimentally identified noise, representative of real-life applications, is then added to the FE results before algorithm training. Damage existence and position are determined by a Novelty Detection approach and a Regression Neural Network, respectively. First, damage is identified on new test cases generated by the same FE procedure used for training. Secondly, the trained algorithms are applied to the experimental dataset. Sensitivity analysis on several parameters (number of samples for training, damage severity and noise levels) is numerically carried out to understand the applicability limits of the present methodology.

Deep-learning viscoelastic seismic inversion for mapping subsea permafrost

Marine seismic surveys can be used to map ice-bearing subsea permafrost on a large scale. However, current seismic processing technologies have limited capacity to image permafrost distribution at depth, mainly due to the low sensitivity of primary reflections and refractions to the velocity inversion found at the base of permafrost. Guided waves and multiples are more sensitive to the velocity variations below the top of permafrost, but they remain challenging to use in physics-based inversion approaches. A deep-learning-based seismic inversion has the potential to improve seismic imaging below the top of permafrost by automatically extracting information from all wave modes. We develop a multi-input neural network (NN) to estimate seismic velocities from marine seismic data. The network is trained on synthetic data generated from the representative distributions of the seismic properties of subsea permafrost. We find that our network can image large velocity contrasts and reversals in depth, typical of subsea permafrost. We use our network to estimate the P- and S-wave velocity and Q-factor models from a seismic line in the Beaufort Sea. The NN indicates highly discontinuous subsea permafrost with variable thickness in the area. Our work shows that deep-learning-based seismic inversion could become a cost-effective technology to map the distribution of subsea permafrost on a large scale and, more generally, high-velocity geologic layers located in shallow waters.

Data-driven met-ocean model for offshore wind energy applications

In recent years, the global transition towards green energy, driven by environmental concerns and increasing electricity demands, has remarkably reshaped the energy landscape. The transformative potential of marine wind energy is particularly critical in securing a sustainable energy future. To achieve this objective, it is essential to have an accurate understanding of wind dynamics and their interactions with ocean waves for the proper design and operation of offshore wind turbines (OWTs). The accuracy of met-ocean models depends critically on their ability to correctly capture sea-surface drag over the multiscale ocean surface—a quantity typically not directly resolved in numerical models and challenging to acquire using either field or laboratory measurements. Although skin friction drag contributes considerably to the total wind stress, especially at moderate wind speeds, it is notoriously challenging to predict using physics-based approaches. The current work introduces a novel approach based on a convolutional neural network (CNN) model to predict the spatial distributions of skin friction drag over wind-generated surface waves using wave profiles, local wave slopes, local wave phases, and the scaled wind speed. The CNN model is trained using a set of high-resolution laboratory measurements of air-side velocity fields and their respective surface viscous stresses obtained over a range of wind-wave conditions. The results demonstrate the capability of our model to accurately estimate both the instantaneous and area-aggregate viscous stresses for unseen wind-wave regimes. The proposed CNN-based wall-layer model offers a viable pathway for estimating the local and averaged skin friction drag in met-ocean simulations.

Physics-Informed Machine Learning of Argon Gas-Driven Melt Pool Dynamics

Abstract Melt pool dynamics in metal additive manufacturing (AM) is critical to process stability, microstructure formation, and final properties of the printed materials. Physics-based simulation, including computational fluid dynamics (CFD), is the dominant approach to predict melt pool dynamics. However, the physics-based simulation approaches suffer from the inherent issue of very high computational cost. This paper provides a physics-informed machine learning method by integrating the conventional neural networks with the governing physical laws to predict the melt pool dynamics, such as temperature, velocity, and pressure, without using any training data on velocity and pressure. This approach avoids solving the nonlinear Navier–Stokes equation numerically, which significantly reduces the computational cost (if including the cost of velocity data generation). The difficult-to-determine parameters' values of the governing equations can also be inferred through data-driven discovery. In addition, the physics-informed neural network (PINN) architecture has been optimized for efficient model training. The data-efficient PINN model is attributed to the extra penalty by incorporating governing PDEs, initial conditions, and boundary conditions in the PINN model.

Accuracy requirements on intrinsic alignments for Stage-IV cosmic shear

In the context of cosmological weak lensing studies, intrinsic alignments (IAs) are one of the most In the context of cosmological weak lensing studies, intrinsic alignments (IAs) are one of the most complicated astrophysical systematic to model, given the poor understanding of the physical processes that cause them. A number of modelling frameworks for IAs have been proposed in the literature, both purely phenomenological or grounded on a perturbative treatment of symmetry-based arguments. However, the accuracy with which any of these approaches is able to describe the impact of IAs on cosmic shear data, particularly on the comparatively small scales () to which this observable is sensitive, is not clear. Here we quantify the level of disagreement between the true underlying intrinsic alignments and the theoretical model used to describe them that can be allowed in the context of cosmic shear analyses with future Stage-IV surveys. We consider various models describing this “IA residual’’, covering both physics-based approaches, as well as completely agnostic prescriptions. The same qualitative results are recovered in all cases explored: for a Stage-IV cosmic shear survey, a mis-modelling of the IA contribution at the ∼10% level produces shifts of ≲0.5σ on the final cosmological parameter constraints. Current and future IA models should therefore aim to achieve this level of accuracy, a prospect that is not unfeasible for models with sufficient flexibility.

Leveraging physics-based and explainable machine learning approaches to quantify the relative contributions of rain and air pollutants to wet deposition

A quantitative understanding of the roles of rainfall and pollutant concentrations in wet deposition is important because they critically influence terrestrial and aquatic ecosystems. However, their relative contributions to wet deposition, which vary across regions, have not yet been identified. We propose two methods that quantitatively separate the contributions of rain and pollutant concentrations to wet deposition: one is based on simplified equations describing the wet scavenging of pollutants and the other is based on random forest models employing SHapley Additive exPlanations. Three-dimensional long-term air quality simulations from 2003 to 2019 are used as inputs for both the physics-based and machine learning models. Remarkably, the results drawn from the explainable machine learning model are consistent with those from the physics-based approach: overall, rain is a more important limiting factor than pollutant concentrations and the relative contribution of rain is larger than that of pollutants by up to a factor of 3–4 in polluted regions. In polluted regions, pollutant concentrations can remain relatively high even in the presence of precipitation owing to continuous and intense emissions; therefore, wet deposition is limited by rainfall. The contribution of rainfall is larger by 1.5–2.5 than that of pollutant concentrations in regions even with low emissions and this considerably large role of rain suggests that regional or transboundary pollutant transport plays a key role in modulating wet deposition. However, in very remote regions, once the rainfall amount exceeds a certain value, rainfall no longer contributes to increasing wet deposition because atmospheric pollutants are readily removed by rain. So, the contributions of the two factors are comparable in pristine regions. Our results can serve as a basis for explaining interannual variations in wet deposition and for future projections of wet deposition under emission control plans and climate change scenarios across regions.

FEP Augmentation as a Means to Solve Data Paucity Problems for Machine Learning in Chemical Biology.

In the realm of medicinal chemistry, the primary objective is to swiftly optimize a multitude of chemical properties of a set of compounds to yield a clinical candidate poised for clinical trials. In recent years, two computational techniques, machine learning (ML) and physics-based methods, have evolved substantially and are now frequently incorporated into the medicinal chemist's toolbox to enhance the efficiency of both hit optimization and candidate design. Both computational methods come with their own set of limitations, and they are often used independently of each other. ML's capability to screen extensive compound libraries expediently is tempered by its reliance on quality data, which can be scarce especially during early-stage optimization. Contrarily, physics-based approaches like free energy perturbation (FEP) are frequently constrained by low throughput and high cost by comparison; however, physics-based methods are capable of making highly accurate binding affinity predictions. In this study, we harnessed the strength of FEP to overcome data paucity in ML by generating virtual activity data sets which then inform the training of algorithms. Here, we show that ML algorithms trained with an FEP-augmented data set could achieve comparable predictive accuracy to data sets trained on experimental data from biological assays. Throughout the paper, we emphasize key mechanistic considerations that must be taken into account when aiming to augment data sets and lay the groundwork for successful implementation. Ultimately, the study advocates for the synergy of physics-based methods and ML to expedite the lead optimization process. We believe that the physics-based augmentation of ML will significantly benefit drug discovery, as these techniques continue to evolve.

Transient modeling of a solid oxide fuel cell using an efficient deep learning HY-CNN-NARX paradigm

Control and monitoring systems are crucial for ensuring optimal performance, efficiency, and longevity of Solid Oxide Fuel Cells (SOFC). Developing a model to accurately capture the transient behavior of the SOFC under dynamic operation is essential for these applications. As electrochemical, mass-transfer, and thermal equations are involved in SOFCs’ dynamic, physics-based approaches to capture all the complexity of SOFC systems are often too computationally expensive for real-time implementation. In this paper, an innovative multi-input neural network is developed to build an accurate model of SOFC that is computationally efficient. The proposed algorithm uses experimental data and is a modified nonlinear autoregressive exogenous (NARX) network. An optimal sequence of the most recent observations (output voltages) that characterize the SOFC dynamics is passed through stacked 1D convolutional layers (CNN) to extract latent spatial/temporal information. Then the output is concatenated with current and past exogenous inputs. The fusion of learned features, representing the internal dynamics and the effect of exogenous inputs, is then fed to a fully connected network for one-step ahead performance prediction. This hybrid structure, called HY-CNN-NARX, is capable of identifying the transient dynamics of the SOFCs by unrolling information from historical outputs and inputs within a feedforward framework. To evaluate the model performance, lab-scale tubular SOFCs are experimentally tested at 650–750 °C under various dynamic operations and initial conditions, and the transient and steady-state responses of the SOFC are collected. Once the model is validated using a wide operating range of data, the generalization and prediction performance of the proposed network is compared with a conventional NARX and a recurrent stacked Long short-term memory (LSTM) model. The comparative results demonstrate that the proposed HY-CNN-NARX model outperforms the NARX model on unseen data with an accuracy improvement of 59.2% for root mean square error and 53.7% for mean absolute percentage error. In addition, the prediction performance of the HY-CNN-NARX is comparable to its recurrent counterpart, the LSTM model, but with a 39.3% faster execution time.