Self-adaptable materials structure descriptor based on graph attention network for machine learning

Abstract

Machine learning has shown great potential in bridging materials structures and properties. However, efficient and general descriptors for materials structures are still missing. This work dedicates to establishing an efficient and self-adaptable materials structure descriptor based on the graph attention network (GAT). In the GAT descriptor generation, the concrete structure is represented by graphic atomic clusters (atoms as nodes, edges as bonds) consisting of one central atom and the surroundings. A multi-head-GAT is applied to gather the short- and long-distance topological information from the graphical atomic clusters to generate dimension-flexible descriptors regardless of atom number; the GAT descriptors are contracted and help avoiding overfitting during the training process theoretically. Further verification of the GAT descriptors is carried out by performing the phase classification, energy prediction, and potential energy surface (PES) prediction task for Al-Cu alloys with complex precipitates and defects. The GAT descriptors adapt themselves for cross tasks without manual intervention and produce accurate predictions that agree with the density functional theory (DFT) results. The GAT descriptors are also testified as differentiable, universal, and precise on structural characterization containing point-, line-, and planar- defects. Further applications will be found for the materials design field demanding accurate structure descriptions.

Keywords

• Structural descriptors
• Graph attention network
• Multi-head mechanism
• Self-adaptable
• Mode migration