Rank-optimal neural representation of EBSD resolves subgrains and enables their quantitative analysis

Pattern-level implicit neural representations of electron backscatter diffraction scans open doors for differentiable characterization of materials. This study addresses two limitations that hinder their practical use: manual selection of factorization rank and qualitative-only analysis of diffraction gradient fields. First, we introduce a procedure that determines the microstructure-specific optimal rank prior to any neural training. Second, we demonstrate a hierarchical P-watershed algorithm for segmentation of the gradient field and subsequent quantitative analysis. For a steel with dense substructure, our procedure resolves subgrains without indexing, disorientation thresholds, or manual intervention and provides a mean size matching the reported values from traditional semi-manual workflows.