3D-structure-attention graph neural network for crystals and materials

Abstract

Machine learning has been widely used in physics and chemistry. As a deep learning method based on graph domain analysis, graph neural networks (GNNs) have natural advantages in predicting material properties. We find that most existing models focus on the topological relationship between atoms without considering the specific positions. However, 3D-spatial distribution is the key to affecting the atomic state and interaction relationship, which has a decisive impact on the material properties. Here, we present a 3D-structure-attention graph neural network (3SAGNN) model, introducing the attention mechanism. The model focuses on the critical areas in the material 3D structure that significantly impact the prediction properties to effectively improve the accuracy of material properties prediction. We prove that the performance of 3SAGNN on a variety of datasets outperforms prior ML models, such as CGCNN. Our proposed model was tested on 36,000 inorganic materials dataset, 20,000 Pt nanocluster dataset, 18,000 porous materials, and 37,000 alloy surface reactions. The experimental results show that 3SAGNN can predict formation energies, total energies, band gaps, and surface catalytic properties more accurately and quickly than density functional theory. Finally, we improve the interpretability of the model through visualisation and show the working mechanism of the network.