Local and global attention mechanisms synergy for material property prediction

Efficient machine learning (ML) models for predicting and discovering new materials present a significant challenge in materials science. Graph neural networks (GNNs) have garnered considerable attention due to their advantages in material applications. However, many GNNs fall short in representing the topological and geometric structures of crystals, leading to limited accuracy in characterising complex structures and predicting properties. This deficiency can lead to insufficient accuracy in capturing the intricate structural information of crystals, thus limiting the effectiveness of GNNs in characterising complex crystal structures and predicting their properties. To address this challenge, we introduce a dual-level attention crystal graph neural network with enhanced topological-geometric information (TGCGNN) for predicting the properties of crystalline materials. TGCGNN integrates a topologically and geometrically enhanced crystal graph attention layer (GeoGAT) with a global attention layer. The GeoGAT layer learns the topology and spatial geometry of crystals by incorporating the distance vectors between nodes and their neighbours, as well as bond lengths. Additionally, the global attention layer incorporates crystal composition to assess the collective contributions of atoms to the material properties. Through numerous experiments, we demonstrate that TGCGNN outperforms several state-of-the-art models, offering significant improvements in the accuracy of material property prediction.