Formula Graph Self-Attention Network for Representation-Domain Independent Materials Discovery

Material representation is vital in determining the performance of machine learning (ML) models in predicting the properties of materials. There are two main types of materials descriptors, one that encodes crystal structure details and the other that only uses chemical composition with the hope of discovering new materials. Graph neural networks (GNNs) have dominated materials property prediction tasks outperforming their classical ML counterparts given sufficient data. However, current GNNs are limited to only one of the above two avenues due to the lack of a general graph-based material descriptor. In this study, we propose formula graph that unifies composition-only and structure-based materials descriptors for GNNs. We develop a self-attention integrated GNN that assimilates a formula graph and show that the proposed architecture produces material embeddings transferrable between the two domains, opening further research opportunities in crystal structure prediction. The proposed model can outperform some previously reported structure-agnostic models and their structure-based counterparts while exhibiting better sample efficiency and faster convergence. Finally, the model is applied in a challenging exemplar to predict the complex dielectric function of materials and nominate new substances that potentially exhibit epsilon-near-zero (ENZ) phenomena.