Physics-informed lithium-ion battery end-of-life prediction model for data-efficient deployment

Abstract

Accurate prediction of the end-of-life (EOL) is essential for safe and efficient deployment of lithium-ion batteries. While numerous models have been proposed for such prediction, they mostly rely on either extensive cycling data or refined health indices which are difficult to obtain during normal operation. Motivated by this, this study introduces a physics-informed EOL prediction framework based on Gaussian process regression. Once trained on a labeled dataset, the model predicts the EOL of a new cell at the deployment stage from its operating conditions (OCs) alone, without any cycling data. To isolate the effects of OCs from cell-to-cell variability, the model defines a nominal EOL and focuses on the impact of three major OCs, namely charging C-rate, ambient temperature, and depth-of-discharge (DOD), under nominal cell specifications. To incorporate electrochemical insights, novel kernel functions are proposed for each OC, including an empirical polynomial formulation for estimating the temperature at which EOL is maximized, as a function of C-rate and DOD. These kernels are combined multiplicatively to capture both individual and interactive degradation effects. The model is validated on two datasets with different chemistries: a synthetic dataset generated using a pseudo-two-dimensional model (LCO/Graphite) and a real-world dataset (NCM/Graphite + SiO). Compared to the conventional radial basis function kernel, the proposed composite kernel reduced the EOL prediction root mean square error by 46.6% and 60.8% and the mean absolute percentage error by 17.1% and 61.6% for the synthetic and real-world datasets, respectively, demonstrating its strong generalization capability and practical applicability for deployment-efficient battery health management.