From partial charging to physics-aligned health prediction: A dual-path neural framework bridging features and degradation signatures

Reliable estimation of lithium-ion battery state of health (SOH) across diverse chemistries and operating conditions remains a central challenge for intelligent battery management. This study presents the Electrochemical Feature–Physics Informed Neural Network (EFPINN), a unified two-stage framework that integrates feature-based electrochemical parameter inference with physics-guided SOH prediction. An extended electrochemical model is first applied to laboratory cycling data across NCM, NCA, and LFP chemistries to extract interpretable electrochemical parameters. For each chemistry, the four parameters most strongly correlated with SOH are selected as inference targets, and a lightweight neural network is trained to map statistical and dynamic features extracted from partial charging segments to these parameters. The trained inference models are subsequently transferred to public datasets of matching chemistries to estimate electrochemical parameters directly from charging features. These inferred parameters, combined with the original features, are then fed into the EFPINN model, where gradient-based monotonicity constraints enforce physically consistent relationships between degradation behavior and SOH evolution. Extensive validation on public datasets demonstrates that EFPINN achieves high SOH prediction accuracy (R² > 0.97), interpretable parameter trajectories, and consistent performance across NCM, NCA, and LFP chemistries. By integrating physically meaningful parameter inference with constrained SOH modeling, the proposed framework offers a scalable and deployment-ready solution for real-world battery health diagnostics.