Physics-informed transfer learning by embedding physics into activation functions: an application in battery health management

Accurate prediction of battery remaining useful life (RUL) is critical for ensuring battery safety and reliability. Although physics-informed (PI) machine learning models embed degradation mechanisms to improve accuracy, they often require system-specific knowledge, limiting cross-domain generalization. Transfer learning (TL) enables adaptation across datasets, however, it does not leverage prior physical knowledge or established physics-based models. To combine their strengths, we introduce a dual-branch physics-informed transfer learning framework, where the data-driven branch is pre-trained on the Stanford-MIT-Toyota dataset (source domain) and fine-tuned on the XJTU dataset (target domain), while the PI branch incorporates a solid electrolyte interphase degradation mechanism via an Arrhenius-based activation function. Both branches are co-trained on the target domain to demonstrate the improvement in prediction accuracy. The physics-informed transfer learning (PITL) framework consistently improves prediction accuracy across all tested charge-discharge protocols, achieving the lowest mean absolute percentage error (MAPE) of 9.09 %, compared with 10.53 % for TL model and 14.58 % for the baseline model trained from scratch. The PITL model also achieves the lowest MAPE of 7.03 % in the case of individual-battery prediction. A comparative study shows that replacing the standard activation functions with the Arrhenius-based activation function improves generalization and predictive performance by embedding physics into transfer learning.