Thermoelectric Material Performance (zT) Predictions with Machine Learning

Abstract

Research efforts using the tools in machine- and deep learning models have begun to show success in predicting target properties such as thermoelectric (TE) properties, including the figure of merit (zT). These models were trained on various data sources that used experimental, crystallographic, and density functional theory (DFT) data. We developed an interpretable model on a huge experimental data set of ∼160,000 data points to predict the performance of thermoelectric materials. The model predicts the results of three different test sets with high accuracy, such as the root-mean-square error (RMSE) ranging from 0.15 to 0.20 and the evaluation coefficients (R²) ranging from 0.80 to 0.67. Furthermore, we highlight probable reasons such as literature error, varied synthesis routes for the same material, different forms of crystallinity and morphology, and different particle sizes and densities for the deviation of predicted zT from the experimental zT results of the test sets. Lastly, using an experimental data set, our study is one of the few examples that predict a complex zT property directly across the entire gamut of TE materials.