Modeling and prediction of lithium-ion battery thermal runway via multiphysics-informed neural network

Abstract

In this study, a multiphysics-informed neural network (MPINN) is proposed for the estimation and prediction of thermal runaway (TR) in lithium-ion batteries (LIBs). MPINNs are encoded with the governing laws of physics, including the energy balance equation and Arrhenius law, ensuring accurate estimation of time and space-dependent temperature and dimensionless concentration in comparison to a purely data-driven approach. Spe-cifically, the network is trained using data from a high-fidelity model of an LIB, which describes TR by addressing several coupled partial differential equations. Quantitative analysis reveals that the mean absolute error (MAE) and root mean squared error (RMSE) of the MPINN for TR estimation are less than an artificial neural network (ANN) by 0.71 and 1.57, respectively, when using fully labeled data for training. It outperforms the ANN in terms of MAE and RMSE by 90.56 and 118.64, when only a small portion of labeled data (semi-supervision) are used for TR prediction. Importantly, it predicts TR without any labeled data when the decomposition of reactive species is modeled in the positive electrode. The MPINN exhibits promising results in surrogate modeling, implying it can be successfully implemented in practical scenarios and stimulate further research related to TR modeling using physics-informed deep learning.