Exploring Phase Stability and Transport Properties of Emerging Thermoelectric Materials: Machine Learning and Experimental Insights

This study investigates the thermoelectric (TE) properties of newly synthesized materials, including Heusler alloys, magnesium silicides, and tetrahedrites, identified through machine learning (ML) approaches and validated experimentally. Critical descriptors for predicting important properties of TE materials, such as the band gap, formation energy, and Seebeck coefficient ( S ), which strongly influence material stability and TE efficiency, are systematically assessed. Although density functional theory (DFT) remains a foundational approach for predicting TE performance, its computational expense limits its widespread application. To address this challenge, crystal graph convolutional neural networks (CGCNN) are employed to efficiently predict band gap and formation energy from graph-based representations of crystal structures. The experimentally measured band gaps of the synthesized materials show strong agreement with CGCNN predictions. Additionally, experimental measurements of S across different TE classes corroborate computational estimates, demonstrating the robustness of the ML models. For instance, the newly synthesized double half-Heusler compound Ti 2 FeNiSb 2 exhibits an experimental S value of approximately −120 μV/K at 700 K, closely matching the Random Forest model prediction of −143.75 μV/K across the 300–700 K range. The CGCNN model achieves a mean absolute error (MAE) of 0.261 eV for band gap and 0.058 eV/atom for formation energy approximately 1.5 and 2 times lower, respectively, than those reported for earlier pretrained models. Furthermore, for the S prediction, the Random Forest algorithm, optimized using a root-mean-square error loss function, improves accuracy by ∼43.55% compared to previous ML methods. Overall, the integration of ML techniques accelerates TE material discovery, overcoming the limitations of conventional computational approaches.