Machine Learning Predictions of Thermopower for Thermoelectric Material Screening

Machine or deep learning studies to predict the electronic properties, particularly the Seebeck coefficient, of thermoelectric (TE) materials remain relatively few. Considering the benefits and limitations of these studies, we developed a data-driven, structure-independent machine learning script. The script can produce Seebeck coefficient predictions by using a terminal-executed batch file. Our model within the script was trained with the latest version of the Starry data set, with experimentally derived thermoelectric data. The model was then tested across multiple sets of test sets. Postprediction, we use SHapley Additive exPlanations to analyze the interpretability of our model with chemical insights into features influencing our models’ decisions.