Machine learning classification and screening of thermoelectric materials: Interpretable predictions with SHAP analysis

This comprehensive study presents a machine learning framework for multi-task classification of thermoelectric materials using 174 descriptors comprising 173 composition-derived features and measurement temperature. We systematically evaluate thirteen state-of-the-art algorithms on 8014 experimental samples across three classification tasks: ZT performance class, carrier type, and thermal conductivity class. Measurement temperature is included as an explicit input feature alongside the Magpie composition descriptors, physically motivated by the temperature dependence of the ZT figure of merit ( Z T = S 2 σ T / κ ) and essential for the temperature-resolved high-throughput screening. Employing SMOTE (Synthetic Minority Over-sampling Technique) to address severe class imbalance and SHAP (SHapley Additive exPlanations) for model interpretability, we achieve exceptional balanced accuracies of 99.2% for carrier type (RandomForest), 95.7% for thermal conductivity (LightGBM), and 93.3% for ZT classification (LightGBM). Comprehensive SHAP analysis on test samples identifies task-specific primary features: maximum valence electrons and measurement temperature for ZT classification, Cu content (0.874) and valence deviation (0.704) for carrier type, and mean boiling point for thermal conductivity, collectively confirming that electronic structure, thermodynamic descriptors, and operating temperature dominate over traditional mass-based features. This work provides both accurate predictive models and actionable design principles for accelerating thermoelectric materials discovery, enabling faster screening compared to traditional experimental approaches.