Graph Attention Networks for Microstructural Understanding of Abnormal Grain Growth

Polycrystalline microstructures can be represented as graphs, which capture both grain geometry and connectivity. A number of machine learning (ML) algorithms operate on graph data, raising the question of whether they can be used to predict microstructural evolution in polycrystals. Operating on a data set of Monte Carlo grain growth simulations, we find that a simple graph convolution network outperforms a computer vision approach for predicting the occurrence of abnormal grain growth (AGG) in a model polycrystalline system. Based on this successful proof-of-concept, we extend the data set and enhance the data structure. A graph attention network significantly outperforms simple graph convolution, achieving a 20% reduction in error rate. In addition, feature importance analysis identifies the grain characteristics associated with AGG. Taken together, these results show the promise of ML for both predicting microstructural outcomes and supporting microstructural science.