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 directly consider the physical interactions among neighboring grains despite such microscopic interactions critically determining macroscopic properties. The authors develop a graph neural network (GNN) model to embed polycrystalline microstructures as graphs that incorporate both individual grain features and their interactions, and then link this embedding to target properties via a feed-forward neural network. Using magnetostriction of polycrystalline Tb0.3Dy0.7Fe2 as an example, a single GNN with fixed architecture and hyperparameters achieves ~10% prediction error across diverse microstructures and quantifies feature importance at the grain level. This graph-based GNN enables accurate and interpretable prediction of polycrystalline materials’ properties.