Improving Material Property Prediction by Leveraging the Large-Scale Computational Database and Deep Learning

Abstract

Predicting physical and chemical properties of materials based on structures is critical for bottom-up material design. Many property prediction models and material training databases have been proposed, but accurately predicting properties is still challenging. Here, we report a package of “Matgen + CrystalNet” approach to improve material property prediction. We construct a large-scale material genome database (Matgen) containing 76k materials collected from an experimentally observed database and compute their properties through the density functional theory method with the Perdew–Burke–Ernzerhof (PBE) functional. Our database achieves the same computation accuracy by comparing part of our results with those from the open Material Project and Open Quantum Materials Database, all with PBE computations, and contains more diverse chemical species and big-sized structures. Based on the computed properties of our comprehensive data set, we have developed a new graph neural network (GNN) model, namely, CrystalNet, by strengthening the message passing between atoms and bonds to mimic physical and chemical interactions. The model is shown to outperform other GNN prediction models. The proof-of-concept applications, such as fine-tuning data on experimental values to improve prediction accuracy and bandgap prediction on hypothetical materials, showcase the usability and potential capacity of our package of “database + model” to improve material design.