Reexamining Machine Learning Models on Predicting Thermoelectric Properties

Abstract

Thermoelectric materials can generate clean energy by transforming waste heat into electricity. The effectiveness of thermoelectric materials is measured by the dimensionless figure of merit, ZT. The quest for high ZT materials has drawn extensive research experimentally and theoretically. However, due to the vast material space, finding high ZT materials is time-consuming and costly. To improve the efficiency of discovering new thermoelectric materials, recent studies have employed machine learning with databases to search for high ZT candidates. In this work, we examine the effects of adding various physical concepts on the performance of machine learning models in predicting TE properties. The objective is to improve the model ability to capture the underlying physics in designing TE materials. These concepts include short range order and crystal structure class. Results show some improvements in accuracy. However, the current models do not distinguish between dilute alloys and concentrated alloys, rendering them inadequate in predicting doping effects. To better capture the electronic band structure effect from doping, we included various dopant properties as features. This increases the prediction accuracy in doped materials. Furthermore, we used a genetic algorithm to rank features for various thermoelectric properties to provide physical insight into key parameters in designing thermoelectric materials.