Reliable state-of-health (SOH) estimation is a key prerequisite for the safe and effective reuse of second-life lithium-ion batteries. However, practical assessment during early-stage screening is often constrained by extremely limited cycling data, where only a few discharge cycles are available due to time and cost limitations. This study investigates SOH estimation under an extreme sparse-cycling scenario in which only three discharge cycles per battery are available, reflecting realistic constraints in early-stage second-life battery screening. Under such severe data limitations, conventional data-driven models become unreliable, motivating the need for data-efficient and interpretable approaches. To address this challenge, a physics-aware and explainable machine learning framework is proposed, integrating physically interpretable feature extraction with lightweight regression models and Shapley Additive exPlanations SHAP-based interpretability analysis. Electrochemically motivated and mathematically derived features are extracted from voltage, current, and capacity measurements to ensure robustness under severe data scarcity. Multiple model classes, including linear regression, support vector regression, tree-based ensembles, and deep learning architectures, are systematically evaluated to assess their suitability in this constrained regime. Experimental results on real second-life battery datasets demonstrate that physics-aware linear models provide stable and interpretable SOH estimates under extreme data sparsity, whereas more complex nonlinear and deep learning models exhibit higher variability due to insufficient training data. These findings highlight that model suitability is strongly dependent on data availability and support the adoption of interpretable, physics-aware approaches for early-stage second-life battery screening rather than long-term degradation modeling.
Hossen et al. (Thu,) studied this question.