Learning crystallographic orientations from electron backscatter diffraction patterns using variational autoencoder

Electron backscatter diffraction (EBSD) is essential for crystallographic orientation analysis and phase identification. While existing EBSD indexing methods—such as Hough-transform-based indexing (HI) and dictionary indexing (DI)—have proven effective, they face trade-offs between accuracy and computational cost. We introduce Latice (latent-space autoencoder for template indexing of crystallographic EBSD), a variational autoencoder (VAE)-based approach that learns compact, physically meaningful representations of EBSD patterns. Once trained, Latice encodes patterns into a 16-dimensional latent space, achieving approximately 99.9% data compression while preserving rotational symmetries and crystallographic features. Applied to experimental patterns from recrystallized 316L stainless steel, Latice achieves a 7.5-fold indexing speedup over DI with a mean disorientation below 1 ∘ compared to DI. Although performance near grain boundaries requires improvement, the method offers significant advantages in indexing efficiency and storage requirements. These results underscore the potential of physics-informed machine learning models for advancing EBSD analysis and broader crystallographic applications.