Application of machine learning regressors in estimating the thermoelectric performance of Bi2Te3-based materials

Abstract

Bi2Te3-based semiconductors are versatile thermoelectric energy harvesters, widely employed for their narrow bandgap and low thermal conductivity. The conventional approach to determine their energy conversion efficiency is through the thermoelectric figure of merit (ZT), which depends on the material's thermal and electrical properties. The measurement of electrical parameters (resistivity and Seebeck coefficient) is often straightforward. However, the thermal properties (lattice and electronic) measurement is generally cumbersome. We propose in this work a pioneering technique based on machine learning algorithms to predict the figure of merit of Bi2Te3-based materials using their structural lattice constants (a and c) and the electrical properties as the model predictors, thus, circumventing the arduous work of thermal property measurements. A total of five weak regression models were initially developed including lasso regression, linear regression, decision tree regression (DTR), and support vector regression (SVR) with RBF and polynomial kernels. The performance of the weak models was evaluated using metrics such as correlation coefficients (CC), mean absolute error (MAE), R2-score, and mean square error (MSE). Our results revealed that DTR and SVR-RBF outperform the rest of the models with CC of 99.0% and 90.8% respectively. Adaboost combines the output of these regressors into a weighted sum that represents the output of the strong regressors. The boosted models performed better predictions than the weak models, with CC of boosted DTR and SVR-RBF models reaching up to 99.5% and 94% respectively. The proposed models were further validated externally by studying the effects of elemental doping and sample growth conditions.