Incorporating edge convolution and correlative self-attention into graph neural network for material properties prediction

Abstract

The prediction of material properties is a crucial challenge in the design of new materials. Traditional methods based on either trial-and-error experiments or large-scale density functional theory calculations are known to possess various limitations. Although recent machine learning (ML) methods have shed light on resolving this problem efficiently, the majority of ML models consider only the local atomic environment while ignoring the nonlocal correlations between atoms. Indeed, even the periodic patterns of the crystal structures are not seriously considered. Consequently, these issues lead to an insufficient understanding of the feature information of atoms and bonds. In this study, we propose a crystal graph convolutional neural network based on edge convolution (EdgeConv) and correlative self-attention, namely, EdgeConv-Graph attention neural network (GANN). This network is able to efficiently extract atomic and bonding feature information, while effectively learning the importance weights of all neighbouring nodes. Numerical experiments predicting the electronic structural properties of metal–organic frameworks show that the developed model achieves state-of-the-art performance. Moreover, the proposed model was applied to predict the heat capacity and thermal decomposition temperature of material, demonstrating the ability of this method to effectively generalise multiscale prediction tasks.

Keywords

• material property prediction
• edge convolution
• correlative self-attention
• graph convolutional neural network
• metal–organic frameworks