Session: 17-08-01: Renewable Energy Systems

Paper Number: 174039

Physics-Informed Neural Operators for Parameter Estimation in Li-Ion Battery 

Accurate estimation of battery state parameters is key to designing efficient battery management systems (BMS) enabling high-performance, safe, and reliable battery operations. Current lithium-ion battery (LiB) modeling techniques commonly range from equivalent circuit models (ECM) to physics-based modeling approaches with increasing computational complexities and costs. While ECMs can provide low-fidelity estimations due to a lack of consideration for electrochemistry, high-fidelity electrochemical models such as the Pseudo-two-dimensional (P2D) and 3D models can face excessive computational complexity limiting their applications in real-time BMS. Other approaches such as reduced order single particle models (SPM) either ignores the concentration gradients or assume a constant distribution of lithium-ion in the electrolyte phase resulting in inaccurate state estimations, especially at high C-rates and near failure or thermal runaway events. Furthermore, purely data-driven methods are also investigated to predict LiB parameters. While data-driven models can be more efficient than electrochemical models, higher model training time and lack of in-depth physics limits the model’s effectiveness. The current work proposes a novel surrogate modeling approach that leverages Deep Operator Networks (DeepONet) to accurately and efficiently predict lithium-ion concentration gradient and other state parameters over a wide range of LiB operating conditions. The primary focus of this work is to compare DeepONets–based predictions with a simplified electrochemical P2D model with battery ageing. DeepONet is a neural network framework that learns the functional mapping between input conditions (such as current profiles, ambient temperatures, and spatial locations) and the resulting lithium-ion concentration fields within batteries, enabling a fast and accurate prediction of complex electrochemical behavior, especially near LiB failure.

We first generate a comprehensive dataset by systematically solving the simplified P2D models under varying input conditions. This dataset captures the spatiotemporal evolution of lithium-ion concentration and other key state variables. Using this high-fidelity data, we train DeepONet to learn the mapping between input conditions to the target outputs by optimizing different ‘loss’ functions. Once trained, the DeepONet model can provide rapid and accurate predictions of battery parameters, enabling its integration into design optimization, and real-time prediction of battery health and degradation through BMS. Our initial results demonstrate that the DeepONet surrogate achieves high predictive accuracy across the parameter space, with orders-of-magnitude speedup over traditional solvers. This approach opens new possibilities for physics-informed machine learning in battery systems, offering a path toward real-time, scalable, and interpretable battery management systems. The details of this model architecture development and important findings will be communicated in the conference.

Presenting Author: Rishi Roy Sandia National Laboratories

Presenting Author Biography: Rishi Roy is currently a researcher at Sandia National Labs, Livermore, California. His primary interest is in battery safety, laser diagnostics in combustion, hydrogen, and machine learning. He has co-authored several peer-reviewed journals, conference articles, and technical reports in these areas. He regularly serves as a reviewer to several ASME, AIAA, Elsevier journals, and different funding opportunity proposals for the Department of Energy. He is a recipient of different funding awards at Sandia, and various prestigious awards and fellowships by the AIAA and NSF.

Authors: