Combining machine-learning models with first-principles high-throughput calculations to accelerate the search for promising thermoelectric materials

Abstract

Thermoelectric materials can achieve direct energy conversion between electricity and heat, and thus can be applied to waste-heat harvesting and solid-state cooling. The discovery of new thermoelectric materials is mainly based on experiments and first-principles calculations. However, these methods are usually expensive and time-consuming. Recently, the prediction of properties via machine learning has emerged as a popular method in materials science. Herein, we firstly did first-principles high-throughput calculations for a large number of chalcogenides and built a thermoelectric database containing 796 compounds. Many novel and promising thermoelectric materials were discovered. Then, we trained four ensemble learning models and two deep learning models to distinguish the promising thermoelectric materials from the others for n-type and p-type doping, respectively. All the presented models achieve a classification accuracy higher than 85% and area under the curve (AUC) higher than 0.9. In particular, the M3GNet model for n-type data achieves accuracy, precision and recall all higher than 90%. Our work demonstrates a very efficient way of combining machine-learning prediction and first-principles high-throughput calculations to accelerate the discovery of advanced thermoelectric materials.