Extracting microstructural descriptors from diffraction contrast x-ray tomography for grain growth modeling

Abstract

Three-dimensional polycrystalline microstructure datasets from non-destructive x-ray characterization techniques such as diffraction contrast tomography (DCT) and high-energy diffraction microscopy (HEDM) encode rich information about grain boundary geometry, crystallography, and topology. Translating this into a reduced-order (e.g., graph) representation requires careful choices about feature engineering. Here, we present a systematic workflow for converting 3D grain maps into graph representations where grain boundaries are encoded as edges carrying physically motivated feature vectors. Fifteen grain boundary descriptors spanning three categories are defined: crystallographic character, topological connectivity, and geometric shape. To arrive at a compact, non-redundant feature set, we apply a two-stage downselection procedure combining Pearson correlation analysis with Kruskal-Wallis testing based on time-resolved grain growth dataset. The retained descriptors reflect the physical mechanisms governing boundary migration. This feature engineering framework also serves as the foundation for a topology-aware graph neural network that learns directly from experimental microstructure data, offering a path toward data-driven prediction of grain boundary behavior during thermomechanical processing.