Interpretable physics-guided neural networks for state of health estimation of lithium-ion battery

Accurate prediction of lithium-ion battery state of health (SOH) is essential for safe and reliable operation. Although model-based and data-driven methods have achieved significant progress, their lack of decision transparency leads to limited reliability and poor interpretability in industrial applications. To address this key challenge, we propose a novel SOH estimation method based on a Physics-Guided Kolmogorov-Arnold Network (PGKAN). This dual-module coevolutionary architecture integrates an SOH Estimator (SOHE) and a Degradation Mechanism Surrogate (DMS) to co-evolve SOH estimation and degradation mechanism discovery. Specifically, the SOHE leverages Kolmogorov-Arnold networks (KANs) for high-accuracy and interpretable predictions, while the DMS uncovers hidden degradation patterns and introduces physics-based constraints through an adaptive partial differential equation (PDE) formulation. Experiments on four public datasets encompassing 387 batteries show that PGKAN achieves superior accuracy, reducing the mean absolute percentage error by 8.5% to 19.7% compared to existing methods. More importantly, the DMS successfully reveals the dominant degradation partial differential equations (PDEs) governing each dataset. We further leverage the intrinsic interpretability of the architecture: spline-based analysis yields data-driven feature importance rankings that strongly align with the key terms in the discovered PDEs, and symbolic regression applied to each activation function derives an explicit, symbolic degradation expression. The source code will be available at https://github.com/USTC-AI4EEE/PGKAN .