Prediction of Material Properties of Inorganic Compounds Using Self-Attention Network

Abstract

Due to the increase in material databases in recent years, there has been a lot of research regarding deep learning models which use large sizes of datasets and are aimed at the prediction of the material properties of inorganic compounds. Particularly, prediction models with Self-Attention structures, such as Roost and CrabNet, have garnered attention because of two reasons: (1) input variables are confined to the chemical composition of each formula and (2) Self-Attention enables models to learn individual element representations based on their chemical environment. However, the existing Self-Attention model yields low prediction accuracy when predicting structure-dependent material properties, such as the magnetic moment, for lack of structural information of compounds as input. In this research, based on the existing Self-Attention model, we set both elemental and structural information, especially the space group number and lattice constant, as input information and successfully construct a prediction model that is more versatile than existing methods. Furthermore, we visualized lists of promising materials by adopting Bayesian optimization. As a result, we have developed a system to propose desired materials for materials researchers.