Multiscale Diagnosis of Lithium-Ion Battery Degradation under Extreme Operating Conditions with Integrated Data-Driven and Post-Mortem Validation

Abstract

Lithium-ion batteries subjected to extreme operating conditions—such as high temperature, high C-rates, and deep overdischarge— exhibit rapid and coupled aging behaviors that are challenging to disentangle using conventional diagnostics. While purely data-driven models often lack interpretability ("black-box"), physics-based methods typically require measurements unavailable in practical applications. To bridge this gap, we propose the SIX-ICA framework, an interpretable machine learning approach that integrates Incremental Capacity Analysis (ICA) features with an XGBoost regressor and SHAP analysis. By extracting mechanism-informed ICA peak features from routine cycling data, the framework achieves robust State-of-Health (SOH) estimation. Crucially, SHAP analysis provides transparent feature attribution, linking statistical inputs directly to degradation pathways. Validated on LiFePO₄/graphite pouch cells cycled at 65 °C and 3 C (comparing 2.5 V vs. 1.0 V cutoffs), the framework identifies Loss of Lithium Inventory (LLI) as the primary driver of capacity fade, noting its significant intensification under deep over-discharge, while Loss of Active Material (LAM) plays a secondary role. These findings are corroborated by OCV fitting and post-mortem characterization. This workflow advances interpretable SOH diagnostics under extreme conditions and offers a scalable route for other battery chemistries.

Keywords

Extreme conditions / Deep over-discharge / Incremental capacity analysis / Interpretable machine learning / State-of-health estimation / Loss of lithium inventory / XGBoost / SHAP analysis

RIGHTS & PERMISSIONS

2026 SIOC, CAS, Shanghai, & WILEY-VCH GmbH