Predicting Material Properties by Deep Graph Networks

Abstract Nodes and edges constitute crystal structures, so crystal structures are easy to express as graphs. Using graph networks to predict material properties has been a prevailing view in the past years. Neural networks are used to fit the density functional theory (DFT) calculation results to predict the properties of various materials. Based on this train of thought, the authors develop graph neural network models for predicting material properties with crystal structure data, and the model gets competitive results with Material Project datasets. It is demonstrated that the authors’ model outperforms other ML methods. Like MEGNet, this model makes input data three parts: nodes, edges, and global. Using Graph Attention Network (GAT) can aggregate node data with edge information and get attention to important areas in graphs. Integrating three features retains structural information and is more suitable for mapping the original distribution of crystal structures.