Improved crystal graph networks with periodic invariance from a global perspective

Abstract

In recent decades, the swift progress of machine learning has significantly propelled the advancement of crystal material property prediction and generation. Graph Neural Networks (GNN), as a key tool in AI for Science research, play a crucial role in processing data from various scientific problems. However, most existing GNN models focus on improving the accuracy of crystal material property predictions by exploring the crystal’s geometric structure and altering deep learning model approaches, while neglecting the ratio information of elemental composition, which represents the global information of the crystal. In this study, we propose a new model for crystal material property prediction, Gformer. Gformer combines periodic pattern encoding with self-connecting edges, graph attention convolution, and a global feature extraction and processing module. This allows it to capture repeated patterns in the crystal structure and the macroscopic effects of elemental composition on material properties. The model was trained and evaluated using the JARVIS-DFT and Materials Project databases. The results demonstrate that Gformer achieves outstanding performance in six crystal property prediction tasks. Our code is publicly available at https://github.com/wwzhui/Gformer.