Inverse inference of lithium-ion battery degradation parameters via physical modeling and data-driven machine learning

Accurate prediction of lithium-ion battery (LIB) degradation is essential for lifetime assessment and safety management, yet remains challenging because of the complex underlying degradation physics and the high-dimensional parameter space. Existing approaches often struggle to capture coupled, multiscale degradation processes while preserving both computational efficiency and physical interpretability. Here, we present a physics-informed inverse inference framework that leverages the intrinsic two-stage degradation behavior of LIBs to overcome these bottlenecks. By extracting compact, observable physical signatures such as the linear-stage slope, and the knee point for the nonlinear transition, we effectively reduce the dimensionality of the degradation parameter space. We integrate pseudo-two-dimensional (P2D) physical modeling with a lightweight machine learning algorithm (XGBoost) to map these macroscopic trajectory features to internal electrochemical aging parameters, such as solid-electrolyte interphase (SEI) growth and lithium plating. Validated against experimental cycling data, our framework reconstructs degradation trajectories with high fidelity and drastically reduced computational complexity. Furthermore, it uniquely provides uncertainty quantification in accelerated aging regimes. This work bridges the gap between data-driven efficiency and mechanistic interpretability, which could be useful for advanced battery health diagnostics and lifetime optimization.