Graph neural networks for an accurate and interpretable prediction of the properties of polycrystalline materials

Various machine learning models have been used to predict the properties of polycrystalline materials, but none them directly consider physical interactions among neighboring grains despite such microscopic critically determining macroscopic material properties. Here, we develop a graph neural network (GNN) model for obtaining an embedding microstructure which incorporates not only features individual also their interactions. The is then linked target property using feed-forward network. Using magnetostriction Tb0.3Dy0.7Fe2 alloys as example, show that single GNN with fixed architecture and hyperparameters allows low prediction error ~10% over group remarkably different microstructures well quantifying importance each feature in grain its magnetostriction. Such microstructure-graph-based therefore enables accurate interpretable materials.