Physics-informed learning of health-aware equivalent circuit parameter evolution for lithium-ion batteries

Lithium-ion batteries exhibit nonlinear degradation under complex operating conditions and long-term cycling. Existing studies mainly focus on state-of-health (SOH) estimation, while continuous SOC–SOH-dependent ECM parameter representation and validation of neural-network-inverted parameters remain limited. This study proposes an ECM-PINN framework that integrates a neural network with a differentiable first-order RC model for joint SOH estimation, ECM parameter inversion, and voltage reconstruction. An SOH-guided consistency regularization is introduced to regulate excessive cross-cell parameter dispersion. The framework is evaluated on the Samsung 50E and HUST datasets. On unseen Samsung 50E cells, the Pearson correlation coefficients between the inverted and conventionally identified R 0 , R 1 , and C 1 are 0.996, 0.998, and 0.992, respectively. On the HUST dataset, the SOH estimation MAE and RMSE are 0.53% and 0.74%, while the voltage reconstruction MAE and RMSE are 20.50 mV and 26.00 mV. These results demonstrate that the proposed framework can perform SOC–SOH-dependent ECM parameter inversion under both standardized HPPC and non-standard dynamic operating conditions, while maintaining stable SOH estimation and voltage reconstruction performance.