Hybrid physics-informed machine learning framework for calibration-free degradation prediction of lithium-ion batteries

Lithium-ion battery degradation prediction traditionally requires chemistry-specific laboratory calibration, limiting scalability across diverse operating conditions and cathode materials. This work proposes a Hybrid Physics-Informed Machine Learning Degradation Model (PIML-DM) that enables calibration-free and physics-guided SoH prediction without chemistry-specific laboratory parameterization using only operational telemetry. The framework integrates a dual-branch architecture in which an LSTM network learns nonlinear aging residuals, while a physics-constrained loss enforces Arrhenius temperature kinetics, Wöhler fatigue stress, and strict monotonicity. To rigorously evaluate cross-chemistry robustness, the framework is trained on the NASA LCO dataset, validated on the Oxford NCA dataset, and benchmarked using a locally acquired LFP dataset. Despite the substantially different voltage signatures and degradation pathways across these chemistries, the PIML-DM achieves sub-0.5% RMSE and maintains physically consistent SoH trajectories without prior calibration. The results demonstrate that shared physics-guided degradation priors enable robust generalization across previously unseen battery chemistries, including LFP systems, while substantially reducing the need for chemistry-specific laboratory characterization and calibration procedures. This establishes the PIML-DM as a scalable, deployment-ready prognostic solution for grid-storage and EV BMS applications.