Prediction of superior thermoelectric performance in unexplored doped-BiCuSeO via machine learning

Abstract

BiCuSeO compound is a promising thermoelectric material, which has attracted many experimental studies through trial-and-error approaches to improve its thermoelectric performance by element doping, such that a fast and efficient prediction of thermoelectric property for unexplored and rarely explored doped-BiCuSeO is highly desired. In this work, a machine learning (ML) model for predicting the ZT value of M element doped-BiCuSeO (Bi1-xMxCuSeO) has been established via the correlation analysis for descriptors and the comparison among different ML approaches. The results show that Gradient Boosting Regressor is the most appropriate approach for our ML model, which is well validated by comparing the predicted and experimental ZT values for the cases in the dataset. The ML model is also used to predict the ZT values of Bi1-xMxCuSeO with unexplored and rarely explored doping element M, and the optimal doping elements as well as their doping contents are screened out. The results indicate that the ZT of Bi0.86Po0.14CuSeO (Po-doped) and Bi0.88Cs0.12CuSeO (Cs-doped) are higher than that of pure BiCuSeO, and are improved by 104 % and 98 % at the 923 K, respectively. The enhancement is well explained by the first-principles calculations. The findings offer a guideline for exploring superior thermoelectric performance in BiCuSeO.

Keywords

• BiCuSeO
• Thermoelectric
• Machine learning
• Doping
• Prediction