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

Abstract

Listen

Translate article

Abstract Predicting material properties is a crucial problem for new material designs. Traditional methods either based on trail-and-error experiments or large-scale density function theory calculations have their limitations. The recent machine learning methods shed light on resolving this problem efficiently. However, most machine learning models only consider the local atomic environment, while ignoring the non-local correlations between atoms. Even the periodic patterns of crystal structure are not taken into serious consideration. These issues lead to insufficient grasping of the feature information of atoms and bonds.In this paper, we propose a crystal graph convolutional neural network based on edge convolution and correlative self-attention, namely EdgeConv-GANN. The network is able to better extract atomic and bonding feature information and also effectively learning the importance weights of all neighboring nodes. Numerical experiments on predicting electronic structural properties of metal-organic frameworks show that our model achieves state-of-the-art performance. In addition, we applied the proposed model to predict the heat capacity and thermal decomposition temperature of materials, both results show that our method generalize well in multi-scale prediction tasks.