Enforcing Electrochemical Consistency: A Hard-Constrained Physics-Informed Neural Network for Battery RUL Prediction

Abstract

Predicting the Remaining Useful Life (RUL) of lithium-ion batteries accurately is imperative in ensuring the safety and reliability of energy storage devices. Although recent physics-aware approaches based on neural networks have successfully shown the importance of extracting parameters related to battery aging from charging curves, their architecture still remains an agnostic and universal approximator without incorporating the underlying physical governing equations. In this work, they propose a Hard-Constrained Physics-Informed Neural Network (HC-PINN) where the governing differential equation of Solid Electrolyte Interphase (SEI) growth and Arrhenius temperature dependency is explicitly integrated within the loss function. Contrary to soft-constrained loss functions, which only penalize the deviation, this approach uses a hard residual loss to penalize violations of the square root time degradation law and Arrhenius law. Numerical results based on the selected battery dataset show that under optimal conditions, HC-PINN reaches an RMSE score of 9.87 cycles, significantly better than the physics-featured FCNN baseline (11.42 cycles). Importantly, when applying 15% Gaussian noise to input data (to represent sensor errors in reality), HC-PINN reaches an RMSE value of 14.23 cycles, whereas FCNN (unconstrained neural network) achieves only 38.56 cycles. Their novel approach brings a new direction in embedding the thermodynamic consistency constraint into neural network optimization.