Machine learning-based thermoelectric materials screening with physical consistency and uncertainty propagation

Machine learning (ML) has emerged as a promising strategy for accelerating thermoelectric materials discovery. However, direct prediction of the thermoelectric figure of merit (ZT) often can introduce target leakage and overestimates predictive performance because ZT is intrinsically derived from multiple transport properties. In this work, a physics-guided and uncertainty-aware ML framework is developed for thermoelectric material screening by independently predicting the Seebeck coefficient, electrical conductivity, and thermal conductivity, followed by analytical evaluation of thermoelectric performance. Experimental data from the ESTM database were evaluated using material-wise validation, with exact chemical formulas kept exclusively in either the training or held-out test subset to reduce material-level information leakage. Predictive uncertainty was quantified using ensemble learning and propagated into thermoelectric performance evaluation and uncertainty-aware LCB screening. The proposed framework demonstrates strong predictive capability for individual transport properties while providing physically consistent thermoelectric performance estimation. Although the framework exhibits lower apparent predictive accuracy than direct ZT regression models, it provides improved interpretability and a more physically structured approach to thermoelectric performance evaluation. The results demonstrate that integrating physical consistency, uncertainty propagation, and uncertainty-aware screening provides a robust strategy for data-driven thermoelectric materials screening.