Dynamic graph attention network-based crystal space-enriched representation for improving material property prediction

The machine learning approach applied to materials discovery is a popular research direction. Knowledge of quantum chemistry explains that the structure of a material determines its properties. Graph neural networks (GNNs) provide a unique way of predicting the macroscopic properties of molecules and crystals rather than by solving the computationally expensive Schrodinger equation. Graph neural networks can abundantly transform the structural information of materials into corresponding features, and many models based on graph neural networks have been applied to predict material properties. We developed a new model (DYCGNN) containing a node update module for our designed edge-graph attention network composition. Through the application of the edge-gatv2 module, this module can effectively learn the complex relationship between nodes and neighbouring nodes in the crystal. Based on the calculated weight coefficients of each neighbouring node, the representation of the node is updated more effectively. In addition, we fuse the position information of the nodes into the node eigenvectors to complement the spatial information of the crystal and enrich the complete representation of the crystal. As we investigate the DYCGNN model, we find that our approach can outperform the predictions of previous models and provide insights into material crystallization.