Machine learning in thermoelectric materials identification: Feature selection and analysis

Traditional experimental methods and calculation methods have troublesome steps and long cycles for predicting new thermoelectric materials. Here, a machine learning method used to identify the thermoelectric materials with high efficiency and accuracy is introduced. Furthermore, the relationship between features and thermoelectric performance is discovered by model decomposition and feature combination analysis. The data is extracted from the MRL database to generate a dataset. Then the feature selection is based on the information entropy evaluation of the ExtraTree-based model to exclude dependent and redundant features and obtained the minimum complexity model of 4 features. According to the data set, five different machine learning models are trained and tested, it is found that the Random Forest model is the best choice. The model decomposition and feature combination analysis are attempted to discover the relationship between features and thermoelectric performance. Finally, we use the 4 features that have the most contribution to the thermoelectric performance to reduce a search space of more than 130,000 systems to a set of 6476 candidates. Among the 10 most promising candidates identified, 4 are the existing thermoelectric materials and 6 are ideal candidate materials for future experimental investigation and validation, thus providing important information for the further examination of thermoelectric materials. The approach used in this work is not limited to the search of thermoelectric materials and can be applied in searching for other functional materials.