Study of crystal property prediction based on dual attention mechanism and transfer learning

This is the authors' abstract. We don't add key points for this paper.

To avoid the step of manual feature engineering when predicting crystal properties, a graph convolutional neural network based on the dual attention mechanism, named DA-CGCNN, is proposed. It fuses both the channel attention mechanism and self-attention mechanism, named the dual attention mechanism, benefiting from capturing the complex features of each atom and dependencies between atomic nodes better. It is found to have comparable or superior performance to other advanced graph neural network (GNN) models by predicting five properties of the crystal: formation energy, total energy, bandgap, Fermi energy, and density. In addition, cross-property transfer learning is conducted on the computed properties from four small-sample crystal materials. The results show better performance on transferring prediction from these four samples. The proposed model in this study significantly improves the accuracy of crystal property prediction and demonstrates excellent prediction performance by incorporating transfer learning techniques. In summary, this work is important in accelerating the prediction of crystalline material properties and the discovery and design of crystalline materials.