Graph Neural Networks for an Accurate and Interpretable Prediction of the Properties of Polycrystalline Materials.

Machine learning models have been used to predict the properties of polycrystalline materials, but none of the existing models directly consider the physical interactions among the grains despite such interactions critically governing materials properties. Here, we use a graph, which comprises a set of interacting nodes, to represent a polycrystalline microstructure comprising a set of interacting grains. Built upon such microstructure graph, a graph neural network (GNN) model is developed for obtaining a microstructure embedding which incorporates not only the physical features of individual grains but also their interactions. The embedding is then linked to the target property of the microstructure using a feed-forward neural network. Using the magnetostriction of polycrystalline TbxDy1-xFe2 alloys as an example, a low prediction error of <10% is achieved using a dataset of ~500 microstructures. The relative importance of each grain in the microstructure to the predicted effective magnetostriction is quantified through the Integrated Gradient method. Our GNN model, with simultaneously low prediction error and high interpretability, is promising for realizing high-throughput prediction of the properties of polycrystalline materials.