Machine learning for predicting ZT values of high-performance thermoelectric materials in mid-temperature range

Abstract

Machine learning (ML) is increasingly being adopted to accelerate the development of materials research. In this work, we applied the ML approach to predict the figure-of-merit (ZT) of thermoelectric (TE) materials. The experimental datasets were gathered from 150 published articles for five high-performance TE groups in the mid-temperature range, i.e., PbTe, Co4Sb12, Mg2Si, BiCuSeO, and Cu2Se, resulting in 1563 data points in total. The chemical formulas of individual compounds, including the dopant types and concentrations, were extracted as ML features using the Magpie software. The ZT value was set as the target value. The model was built based on different regression algorithms, and its accuracy for predicting ZT was evaluated using the coefficient of determination (R2) and the root mean squared error (RMSE). It was found that the model’s accuracy increased with increasing datasets and by incorporating features from experimental parameters (measurement temperature, sintering temperature, and sintering pressure). The final ML model showed relatively high accuracy, with an R2 of 0.859 and an RMSE of 0.156 for a test set. It means that the model can confidently predict the ZT of specific doped compounds in the selected TE groups. To utilize the model effectively, it is implemented as a webpage application with a user-friendly interface so that researchers without ML expertise can explore the ZT values of the doped TE materials. It will certainly be beneficial for experimentalists as a guideline for designing their experiments.