Bayesian neural networks with physics-based priors for robust battery health assessment

Abstract Accurate and uncertainty-aware health assessment of lithium-ion batteries is essential for ensuring the reliability, safety, and efficiency of energy storage systems. While data-driven methods provide strong predictive performance, they often lack interpretability and reliable uncertainty quantification, whereas physics-based models rely on accurate parameterization and may exhibit limited generalization. To address these challenges, this paper proposes a Bayesian neural network (Bayesian Neural Network (BNN)) framework incorporating physics-based priors derived from electrochemical battery simulations. These priors embed electrochemical domain knowledge directly into the Bayesian inference process via the Kullback–Leibler divergence term of the variational objective, providing a principled regularization mechanism that balances data fidelity with physical consistency. The proposed framework produces a physics-informed posterior predictive distribution and enables end-to-end uncertainty quantification. Its performance is evaluated on the publicly available NASA battery dataset and the Wenzhou Pack Degradation dataset, covering diverse discharge profiles, operating conditions, and battery scales. Across both datasets, the method consistently achieves improvements in predictive accuracy, probabilistic calibration, and robustness compared to standard BNNs with conventional priors and state-of-the-art physics-guided Bayesian approaches. These results highlight the effectiveness of physics-informed priors for reliable battery health prognostics across different chemistries and cell-to-pack validation settings, under controlled nominal operating conditions.