Data-driven approach to the performance of SnSe-based thermoelectric materials

Abstract

To accelerate the development of high-performance thermoelectric (TE) materials, it is crucial to optimize material composition, temperature conditions, and other key factors. Herein, an interpretable machine learning approach is utilized to analyze the performance of SnSe-based TE materials under varying temperature conditions and chemical compositions. A comprehensive dataset of previous TE studies is constructed, and the atomic feature-based descriptors are generated. Feature selection techniques, including recursive feature elimination (RFE) and Pearson's correlation coefficient, are applied to identify the important features, which are further used to train machine learning models, and the performance is evaluated using cross-validation. The Shapley Additive exPlanation (SHAP) method is employed to rank the influence of these features on the figure-of-merits (ZT). Additionally, density functional theory (DFT) calculations are performed for further electronic band structure analysis. The present findings not only propose high-performance doping strategies for SnSe-based materials but also provide a data-driven framework for the future design of advanced TE materials.