Unlocking Thermoelectric Potential: A Machine Learning Stacking Approach for Half-Heusler Alloys

The thermoelectric properties of half-Heusler (HH) alloys are predicted using an ensemble modeling approach, specifically a stacking model that integrates random forest and XGBoost algorithms. Leveraging a diverse data set encompassing thermal conductivity (κ), electrical conductivity (σ), Seebeck coefficient ( S ), and the figure of merit (ZT), the study demonstrates superior accuracy with R 2 scores of 0.93 for κ, 0.96 for σ, 0.99 for S, and 0.92 for ZT. Feature importance analysis revealed that temperature, mean covalent radius, and atomic volume deviations critically influence thermoelectric behavior. The stacking model demonstrates a clear advantage by substantially reducing both the mean absolute error and root-mean-square error across all predicted thermoelectric properties, indicating enhanced prediction accuracy and reliability compared to individual base models. The model was validated against first-principles density functional theory calculations for representative compounds ErNiBi, YPtBi, and YPdSb, showing strong agreement and capturing real-world experimental effects such as nonmonotonic variations and peak values prior to high-temperature decline─phenomena that simplified density functional theory struggles to encapsulate. This comprehensive approach highlights the effectiveness of combining machine learning with physical descriptors to predict complex thermoelectric properties, offering valuable insights for the strategic design of high-performance HH thermoelectric materials. The ensemble modeling framework, coupled with thoughtful feature selection and meticulous engineering, establishes a robust foundation for future research aimed at developing high-performance thermoelectric materials.