High-Performance stacking ensemble learning for thermoelectric figure-of-merit prediction

Thermoelectric materials, which convert thermal energy directly into electricity, hold promise for sustainable energy applications. However, accurately predicting their efficiency, quantified by the figure of merit (zT), remains challenging, especially for doped materials. Here we present a machine learning (ML) approach, the stacking model, that significantly improves zT prediction accuracy for doped thermoelectric. By combining five regression models through stacking ensemble learning and introducing 100 coordination number features alongside conventional features, our model achieves a coefficient of determination (R2) value of 0.970. This high performance demonstrates unprecedented sensitivity to zT variations due to doping. We validate our model using an expanded dataset of over 230 new materials from recent literature. The model identifies 43 potential high-zT materials, including Pb0.97K0.03Te0.65S0.25Se0.1 with a predicted zT of 1.9. Density functional theory calculations confirm the superior electrical properties of this compound. Our approach offers an efficient strategy for largescale screening of high-performance thermoelectric materials, potentially accelerating their discovery and development for energy applications.