Chemistry-transferable physics-informed neural network surrogate for lithium-ion cells with online continual learning

This paper presents a physics-informed neural network (PINN) surrogate of the single particle model (SPM) for predicting lithium-ion concentrations and terminal voltage under varying operating and aging conditions. The proposed approach employs a two-stage learning framework. In the offline stage, the network learns the nominal electrochemical dynamics by enforcing the governing partial differential equations along with their associated initial and boundary conditions. During online operation, the physics-informed hidden layers are held fixed, and only the parameters of the output layer are adapted using measured terminal-voltage data via a modified Kalman filter. By explicitly incorporating the applied current as an input, the model can represent both constant and time-varying load profiles. The online adaptation strategy eliminates the need for full-network retraining, mitigates the impact of imbalance between physics-based and data-driven loss terms, and enhances robustness to measurement noise. The stability of the online parameter update is established analytically under standard bounded-noise and persistent-excitation assumptions. The framework is evaluated across multiple cell chemistries, operating profiles, and aging conditions using both simulated and experimental datasets. The results demonstrate accurate concentration and voltage prediction, improved adaptive capability relative to conventional PINNs and gradient-based updating schemes, and an online adaptation time of approximately 2.3 s.