Design of Deep Learning Model for Predicting Material Properties Using Crystal Structure Represented by Three-Dimensional Mesh

Abstract

To improve the efficiency of the development of new materials, research on predicting material properties using deep learning has been actively conducted. Since material properties are greatly affected by the crystal structure, data representation methods for the crystal structure and the design of deep learning models have been proposed. In the previous method, the crystal structure is converted into a two-dimensional graph, and the properties are predicted using a graph convolutional neural network. However, since the crystal structure is originally three-dimensional, it loses three-dimensional information such as atomic positions by converting it into a graph. Here, we focus on representing the crystal structure as a three-dimensional mesh and propose a deep learning model for predicting properties using the mesh data. The proposed method uses data created by converting the crystal structure into a three-dimensional mesh by Delaunay tetrahedralization. The evaluation of the prediction of the formation energy shows that the mean absolute error of the proposed method is 0.066 eV/atom, which is less than the calculation error by simulation.