Aniso-GNN: Physics-Informed Graph Neural Networks Generalizing to Anisotropic Properties of Polycrystals

We present Aniso-GNNs -- graph neural networks (GNNs) that generalize predictions of anisotropic properties of polycrystals to arbitrary loading directions without the need in excessive training data. To this end, we develop GNNs with a physics-inspired combination of node attributes and aggregation function. We further propose a new efficient training strategy leveraging fundamental symmetries of crystallographic orientations and textures as well as tensor properties of individual grains and polycrystals. We demonstrate the predictive power of Aniso-GNNs in modeling anisotropic elastic and inelastic properties of polycrystalline alloys in a wide range of loading directions without training data in those directions.