Attention-based graph neural network for materials property prediction

What is it about?

We propose a new concept of formula graph as a general representation of crystal structure and chemical composition for graph neural networks (GNNs). This enables implementing a novel self-attention integrated GNN that is applicable in both structure-based and structure-agnostic materials property prediction schemes, solving a long-standing problem of maintaining separate machine learning architectures for each domain.

Featured Image

Photo by Hitesh Choudhary on Unsplash

Photo by Hitesh Choudhary on Unsplash

Why is it important?

Current graph neural network (GNN) architectures can only process either crystal structure or chemical composition for materials property prediction. Therefore such models are not transferable between the two domains. The formula graph concept that we propose can represent crystal structure as well as chemical composition for GNNs. Furthermore, our GNN model achieves state-of-the-art results in both domains in predicting diverse materials properties. Due to the transferability between composition-only and structure-based schemes, our model opens up further research opportunities in crystal structure prediction and structure prototype selection for DFT.