Simulation-based and explainable machine learning analysis of a hybrid PCM–air BTMS at high discharge rates

Reliable thermal control of lithium-ion battery packs at high discharge rates remains a decisive challenge for safety and lifetime. We examine a hybrid phase-change-material (PCM)–air battery thermal management system (BTMS) with a unified, data-driven framework that integrates exploratory analysis, explainable learning, and regime discovery. A Light Gradient Boosting Machine (LightGBM) regressor predicts the average cell temperature () with high fidelity (, ), while SHAP attributions and interaction effects quantify physics-consistent roles: PCM thickness is the dominant cooling lever (negative contribution), discharge rate the primary heating driver (positive contribution), and airflow a secondary convective modifier; the benefit of added thickness amplifies with discharge severity. A companion LightGBM classifier identifies safety-critical exceedances (ROC–AUC 0.99), and isotonic calibration improves probability reliability (lower Brier/ECE) while preserving discrimination, enabling recall-tilted operating points for online monitoring. To uncover latent operating modes, DBSCAN clustering of the time–temperature manifold identifies 24 regimes that cleanly stratify stable, nominal, and rapidly escalating behavior and link these bundles to material choice, thickness, and C-rate. The combined evidence supports practical guidance: high-conductivity composite PCM (PCM–2) at 6– substantially delays critical conditions, with elevated Reynolds number providing additional—but secondary—relief.

Unlike prior PCM–air BTMS studies that focus mainly on moderate discharge rates and black-box surrogate models, this work extends the analysis to 7–9C operation and integrates explainable ML to quantify how PCM thickness, C-rate, and airflow jointly govern thermal behavior. The resulting framework provides a physics-consistent, data-driven basis for hybrid PCM–air BTMS design under aggressive loading.

This constitutes, the first hybrid PCM–air BTMS analysis that jointly leverages explainable gradient boosting, calibrated safety classifiers, and density-based clustering over a 3C–9C operating envelope.