Compositionally restricted atomistic line graph neural network for improved thermoelectric transport property predictions

Graph neural networks (GNNs) have evolved many variants for predicting the properties of crystal materials. While most networks within this family focus on improving model structures, the significance of atomistic features has not received adequate attention. In this study, we constructed an atomistic line GNN model using compositionally restricted atomistic representations which are more elaborate set of descriptors compared to previous GNN models, and employing unit graph representations that account for all symmetries. The developed model, named as CraLiGNN, outperforms previous representative GNN models in predicting the Seebeck coefficient, electrical conductivity, and electronic thermal conductivity that are recorded in a widely used thermoelectric properties database, confirming the importance of atomistic representations. The CraLiGNN model allows optional inclusion of additional features. The supplement of bandgap significantly enhances the model performance, for example, more than 35% reduction of mean absolute error in the case of 600 K and 1019 cm−3 concentration. We applied CraLiGNN to predict the unrecorded thermoelectric transport properties of 14 half-Heusler and 52 perovskite compounds, and compared the results with first-principles calculations, showing that the model has extrapolation ability to identify the thermoelectric potential of materials.