A multistage data-driven method for deriving a compact heat-generation correlation from high-precision calorimetry data of lithium-ion batteries: A case study for 18650 NMC cell

Thermal safety remains a critical challenge in lithium-ion battery–powered electric vehicles (EVs), where accurate estimation of heat generation rate (HGR) is essential for reliable operation. Despite advances in battery diagnostics, quantitative HGR prediction using machine learning (ML) remains limited, with existing studies often relying on narrow operating ranges or simulation-based datasets, thereby restricting their reliability. Moreover, compact and interpretable correlations suitable for real-time battery management systems (BMS) are rarely reported. This study develops a comprehensive data-driven approach to estimate HGR of commercial 18650 NMC lithium-ion cells using direct measurements from an isothermal battery calorimeter. The dataset includes discharge rates from 0.5C to 8C and temperatures between 10 and 60 °C, representing diverse and practically relevant battery operating conditions. Four supervised ML models (Support Vector Machine (SVM), K-Nearest-Neighbors (KNN), Random Forest Regressor (RFR), and Gaussian Process Regression (GPR)) are trained using C-rate, current, DOD, and temperature as input features. The selection and contribution of these features are identified using SHAP (SHapley Additive exPlanations) feature-importance analysis. SVM demonstrates the highest predictive accuracy (MSE = 0.0009, RMSE = 0.0301, MAE = 0.0122, R 2 = 0.9993) which is further used to generate an interpretable closed-form HGR correlation via PySR symbolic regression. The resulting expression maintains strong predictive capability (R 2 = 0.9867) and enables rapid estimation of HGR from four measurable parameters. This lightweight and transparent correlation provides a fast and practical tool for thermal management, supporting physics-informed modeling, seamless BMS integration, and improved operational safety of lithium-ion battery systems.