Physics-informed state-of-health estimation for LiFePO₄ batteries using relaxation-derived features and gradient boosting

• Battery state of health is estimated using physics-informed relaxation modeling and disorder metrics • Compact static features encode relaxation kinetics, waveform morphology, and equilibrium voltages • The novel LightGBM-Pro MAX outperforms RF, SVR, GBR, DNN, and CNN–BiLSTM baselines • Strong robustness is demonstrated for 60- and 2.5-Ah cells without geometric or scaling bias • β₁, V∞ cluster, and energy variance/skewness dominate predictions, consistent with known electrochemical aging mechanisms This study introduces a physics-informed static-feature framework for estimating the state of health of lithium-ion batteries without additional excitation or intrusive sensing. The framework combines Kohlrausch–Williams–Watts relaxation modeling and heterogeneous energy landscape statistics with an optimized light gradient boosting machine-Pro MAX regressor. Multistage relaxation is divided into mechanism descriptors, namely, relaxation times, dispersion exponents, and equilibrium asymptote that quantify diffusion kinetics, interfacial resistance, and lithium inventory loss, respectively. The energy-landscape-derived variance, skewness, and kurtosis complement these descriptors to capture aging-induced disorder. The framework is trained on 1438 relaxation samples from 60-Ah commercial and 2.5-Ah laboratory LiFePO 4 cells. It achieves high prediction accuracy (R 2 = 0.9985 and mean absolute error = 0.0027) and demonstrates robust performance across varying operating conditions and cell capacities within the scope of the investigated datasets. The lightweight second-level inference offered by the framework suggests its suitability for real-time deployment in battery management systems. KWW relaxation → HEL reconstruction → Feature Set → LightGBM-Pro MAX → SOH