Accurate and explainable machine learning for the power factors of diamond-like thermoelectric materials

Abstract

The application of machine learning (ML)-based methods to the study of thermoelectric (TE) materials is promising. Although conventional ML algorithms can achieve high prediction performance, their lack of interpretability severely obstructs researchers from extracting material-oriented insights from ML models. In this work, high ML-based prediction performance was achieved with respect to TE power factors (PFs), and the results were well understood by the SHapley Additive exPlanations (SHAP), a method to identify the correlations between targets and descriptors. We designed a robust PF prediction model for diamond-like compounds via a stacking technique, and the model achieved a coefficient of determination value above 0.95 on the test set. From the SHAP analysis, the PFs were negatively correlated with electronegativity and positively correlated with the descriptor “volume per atom” based on the previously reported dataset. TE domain knowledge was adopted to understand these correlations. This work shows that ML models can achieve high accuracy while exhibiting good interpretability, making them useful for materials scientists.

Keywords

• Machine learning
• Thermoelectric materials
• Interpretability
• Power factor