Graph isomorphism network for materials property prediction along with explainability analysis

Deep learning, in particular graph neural network (GNN), has emerged as a powerful tool for exploring the relationship between crystal structures and materials properties. Based on the highly discriminative graph isomorphism network (GIN) algorithm, we present an improved GIN that incorporates edge features during the node feature aggregation and adjusts more weights to important nodes at both local and global levels. We show that our model can achieve comparable or better performance on the Matbench suite compared to some previously reported GNNs and the improvement is significantly greater on small datasets. Although GNNs tend to have good performance in accuracy, the black-box nature of GNNs hinders their application in materials domains to some extent. After training our model, we use t-stochastic neighbour embedding visualization and perturbation-based method to provide a straightforward and comprehensive explanation to increase the transparency and credibility of the model. We demonstrate the superiority of the presented framework in terms of accuracy and explainability. This is crucial for the design of novel materials.