Oct 2020· JOM· Vol 72, pp. 4695 - 4705· 79 citations· ⚡ 2 influential· 21 references
Computer ScienceMathematics
Abstract
This article investigates several physics-informed and hybrid machine learning strategies that incorporate physics knowledge in experimental data-driven deep-learning models for predicting the bond quality and porosity of fused filament fabrication (FFF) parts. Three types of strategies are explored to incorporate physics constraints and multi-physics FFF simulation results into a deep neural network (DNN), thus ensuring consistency with physical laws: (1) incorporate physics constraints within the loss function of the DNN, (2) use physics model outputs as additional inputs to the DNN model, and (3) pre-train a DNN model with physics model input-output and then update it with experimental data. These strategies help to enforce a physically consistent relationship between bond quality and tensile strength, thus making porosity predictions physically meaningful. Eight different combinations of the above strategies are investigated. The results show how the combination of multiple strategies produces accurate machine learning models even with limited experimental data.
In this lecture, I review our recent progress within our ARTISTIC research initiative regarding the development of Physics-Informed Deep Learning frameworks to unravel the complex links between manufacturing parameters and battery electrode properties. While physics-based simulations (such as Coarse-Grained Molecular Dynamics or Discrete Element Method) offer deep insights into the formation of electrode microstructures, their high computational cost often hinders their direct use in real-time optimization loops.
Here, I will demonstrate how we bridge this gap by merging AI with physics-based simulations. We utilize high-fidelity 3D physical models to generate synthetic datasets describing electrode formation steps, such as slurry drying and calendering. These datasets are then used to train Deep Learning networks that capture the underlying process-structure correlations. The resulting Deep Learning models act as ultra-fast surrogates capable of predicting the dynamics leading to the 3D microstructures of Lithium-ion battery electrodes, including the spatial location of active material, carbon-binder and pores. These predictions are experimentally validated against key metrics, such as porosity, tortuosity factor, and effective conductivities.
I will illustrate how these fast surrogate models are coupled with macroscale models in order to perform multiscale modeling of manufacturing processes, and to numerical optimizers to solve inverse design problems, predicting the optimal manufacturing process parameters required to achieve specific electrode architectures. In addition to microstructural formation, I will also discuss our work on simulating electrolyte wetting phenomena by combining Lattice Boltzmann Method with AI, a critical step linking the manufactured electrodes to their electrochemical activity and aging processes.
Furthermore, I will discuss the integration of experimental data from our pilot line into this workflow to ensure physical relevance, paving the way towards high-throughput screening and digital twins for battery manufacturing.
Finally, I will also illustrate how we are extending these computational modeling concepts to the structural design of Redox Flow Battery electrodes.
References
[1]
Prof. Alejandro A. Franco's group publications: https://www.modeling-electrochemistry.com/publications
[2]
Galvez-Aranda, D. E., Fernandez, F., & Franco, A. A. (2025). Physics-Assisted Machine Learning for the Simulation of the Slurry Drying in the Manufacturing Process of Battery Electrodes: A Hybrid Time-Dependent VGG16-DEM Model.
ACS Applied Materials & Interfaces
,
17
(22), 32150-32162.
[3]
Galvez‐Aranda, D. E., Dinh, T. L., Vijay, U., Zanotto, F. M., & Franco, A. A. (2024). Time‐Dependent Deep Learning Manufacturing Process Model for Battery Electrode Microstructure Prediction.
Advanced Energy Materials
,
14
(15), 2400376.
[4]
Vijay, U., Fernandez, F., Ben Hadj Ali, S., Asch, M., & Franco, A. A. (2025). Surrogate Modeling of Lithium‐Ion Battery Electrode Manufacturing by Combining Physics‐Based Simulation and Deep Learning.
Batteries & Supercaps
, e202500433.
[5]
Vijay, U., Galvez-Aranda, D. E., Zanotto, F. M., Le-Dinh, T., Alabdali, M., Asch, M., & Franco, A. A. (2025). A hybrid modelling approach coupling physics-based simulation and deep learning for battery electrode manufacturing simulations.
Energy Storage Materials
,
75
, 103883.
[6]
Troncoso, J. F., Zanotto, F. M., Galvez‐Aranda, D. E., Zapata Dominguez, D., Denisart, L., & Franco, A. A. (2025). The ARTISTIC battery manufacturing digitalization initiative: from fundamental research to industrialization.
Batteries & Supercaps
,
8
(1), e202400385.
Figure 1
Alejandro A. Franco· ECS Meeting Abstracts· 0 citations
The results demonstrate that PINN achieves more accurate and stable full-field vibration reconstructions than conventional PINNs, particularly under conditions involving high-frequency modes, and highlights the potential of hybrid data-physics neural frameworks as an efficient and reliable approach for solving complex PDE-governed dynamical systems.
Hailong Liu, S. Hedayatrasa, Yunpeng Zhu et al.· e-Journal of Nondestructive...· 0 citations
Plasma etching plays an increasingly critical role in advanced integrated circuit manufacturing, making physical modeling more essential for mechanism analysis and precise control. However, physical etching models involve multiple empirical parameters that are difficult to extract, limiting their reliability and accuracy. Traditional modeling methods rely heavily on manual calibration through extensive design of experiments. This manual approach is inefficient, highly costly, and struggles to resolve complex, nonlinear physical effects in multidimensional parameter optimization. To address these challenges, this paper proposes an autonomous modeling method based on deep ensemble neural networks to construct a high-fidelity physical etching model. It can automatically extract and calibrate key model parameters using significantly less experimental data, to achieve predicted profiles that are in excellent agreement with experiments. The combined deep ensemble and Bayesian optimization framework balances exploration and exploitation, effectively overcoming local optima traps in complex, high-dimensional parameter spaces during the optimization process, demonstrating high accuracy, transferability, and efficiency. We validated this method through Si substrate etching experiments in Cl2 plasma, where the highly matched profiles confirmed the method’s precision. Furthermore, we extended the validation to a different trench dimension, demonstrating its promising transferability. Compared with the traditional gradient descent method, our approach achieves a 10.2× speedup with a lower loss value. This work establishes a scientific and highly efficient methodology for high-cost modeling of the etching process in advanced semiconductor manufacturing.
Ho-dal Song, Yuxuan Zhai, Ziyi Hu et al.· Journal of Vacuum Science &a...· 0 citations
Accurate constitutive modeling of hot deformation behavior is essential for designing thermomechanical processes in advanced structural alloys. Conventional Arrhenius-type and empirical models do not adequately capture the combined effects of strain hardening, dynamic recovery (DRV), and dynamic recrystallization (DRX) across broad processing conditions. In this study, two Stacked Residual Physics-Informed Neural Networks (STAR-PINNs) were developed to simulate the hot deformation response of a Mo-rich $\alpha+\beta$ titanium alloy (Ti-6Al-4Mo-1V-0.1Si). The Enhanced STAR-PINN incorporated thermomechanical constitutive constraints, while the DRX-Aware STAR-PINN employed a dual-output architecture to account for recrystallization kinetics. Both models used a shared residual encoder trained on experimental flow stress data collected at temperatures from 800 to 1050 degrees C and strain rates between 0.01 and 10 per second. Physics-informed constraints, including thermal softening, strain-rate sensitivity, strain hardening, and post-peak softening, were enforced through automatic differentiation. The DRX-Aware model further integrated JMAK-Avrami regularization, DRX saturation constraints, and Arrhenius-based consistency with tunable parameters, directly linking the predicted DRX fraction to stress output via latent-feature fusion. The DRX-Aware STAR-PINN achieved RMSE = 11.69 MPa, MAE = 4.83 MPa, R^2 = 0.9850, and a cross-validated RMSE of 12.47 +/- 0.26 MPa. This model accurately reproduced temperature-dependent flow curves, DRX kinetics, and Zener-Hollomon relationships, while maintaining physically consistent constitutive behavior. These results demonstrate that physics-informed deep learning provides a robust and interpretable framework for constitutive modeling, offering a practical approach for advanced process modeling of titanium alloys.
Prashil S. Joshi, Diksha Mahadule, Rajesh K.Khatirkar· 0 citations
Machine learning (ML) is accelerating the advancement of additive manufacturing (AM) into an intelligent, autonomous technology, spanning basic process optimization to specialized biomedical implant fabrication. Based on a systematic literature review encompassing 103 screened studies from 2014 to 2025 in accordance with PRISMA-P guidelines, this research maps how ML augments AM workflows in design, in-situ monitoring, and material performance. Supervised techniques like support vector machines, artificial neural networks, and convolutional neural networks are used extensively to forecast melt pool behavior, identify defects, and optimize parameters. Analysis of comparative studies within the reviewed cohort shows that while traditional data-driven models are foundational, physics-informed neural network (PINN) architectures pro-vide a 10–12% increase in microstructural and thermal history prediction accuracy by embedding explicit physical conservation laws relative to purely empirical black-box configurations. Emerging frameworks, including large-scale generative models and federated learning, are evaluated not as immediate clinical solutions, but as experimental methodologies that facilitate advanced inverse design, biomimetic lattice synthesis, and decentralized collaborative manufacturing protocols. Ti-based alloys, particularly Ti–6Al–4V, lead the clinical domain due to structural performance, fine-tuned through ML for site-specific characteristics. The findings conclude that transitioning from 'black-box' models to explainable AI represents the definitive path forward for meeting FDA/CE regulatory standards in clinical validation.
V. Harikrishnan, Sathiyamoorthy Margabandu· Acta Mechanica et Automatica· 0 citations
Physics-informed machine learning, digital twins, and additive manufacturing are a new direction for the creation of intelligent, adaptive, and high-performance engineering systems that are being integrated into a smart industrial framework. The strategy combines multimodal sensing, data fusion, physics-based modeling, machine learning, process optimization, and closed-loop control, and addresses the challenges of enhancing manufacturing performance across the product life cycle. In physics-informed machine learning, physics principles are incorporated into the data-driven models, which enhances prediction accuracy, decreases the need for large data sets, and facilitates generalization from model to model for different process conditions. Digital twins are virtual models of AM systems that support real-time monitoring and anomaly detection, predictive analysis, virtual experiments, and adaptive process control. The combination of edge computing and intelligent controllers enhances quick decision-making processes during the fabrication process. Aerospace, defense, biomedical engineering, and advanced composite manufacturing are just a few of the applications that show promise for achieving better dimensional accuracy, defect reduction, lightweight design, energy efficiency, material utilization, and process traceability. Industrial deployment is, however, hindered by the lack of high-quality datasets, class imbalance, limited model transferability, interoperability, high computational requirements, cybersecurity, certification, and lifecycle governance. For scalable implementation, standardized data formats, open architectures, benchmark datasets, federated learning, hybrid modeling, and rigorous validation procedures are all important. In general, physics-based informed intelligence and adaptive AM create a promising basis for reliable, efficient, traceable, and environmentally friendly high-performance engineering.
Fahmina Afrin· Journal of Artificial Intell...· 0 citations
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
MIT News · Artificial Intelligence· news.mit.eduAug 24, 2026
