Machine Learning-Driven Inverse Design and Role of Dopant for Tuning Thermoelectric Efficiency

The synergy between a data-driven machine learning (ML) strategy and thermoelectrics (TE) has emerged as a powerful approach for accelerating the discovery of materials. In this study, we have trained an ML model using experimental datasets encompassing chalcogenides, skutterudite, and clathrates, at varying temperatures. With the trained ML model, we successfully predicted five crucial TE properties for 390 strategically doped, previously unexplored materials across different temperatures. ML explainability elucidated the impact of strategic doping, including lone-paired electrons, polarizability, and electronegativity, which is crucial for optimizing TE properties. We identified Bi-doped Bi 0.1 Sb 1.9 Te 3 material as a top-performing candidate, exhibiting high ZT values ranging from 1.74 to 2.20, ultralow κ L < 0.40 W m –1 K –1, and a decent power factor across a broad temperature range. Co-doping strategies involving Cd/Sb and Sb/In in GeTe, as well as Na/Cd/Se in PbTe, led to high ZT and κ L, enabling the tuning of their electronic band structure to foster an alternative pathway with a high power factor. Leveraging insights from regression analysis, we categorized TE materials based on their performance, particularly emphasizing chalcogenide materials. This fusion of ML and experimental database methodologies presents a promising avenue for the inverse design of high-performance TE materials, achieved through strategic adjustments of dopant and composition ratios.