Mapping recrystallization trajectories in GaAs using latent space diffraction analysis

Abstract

Recrystallization in disordered solids proceeds through a sequence of local structural rearrangements that are difficult to resolve using conventional diffraction analysis. In amorphous and partially ordered materials, subtle variations in diffuse scattering, short-range order, and defect-mediated symmetry emergence encode the pathways through which ordering initiates and propagates. Here, we introduce a latent space framework for mapping these pathways directly from \textit{in situ} 4D-STEM diffraction data. A convolutional autoencoder provides a compact representation of structural motifs, and unsupervised clustering identifies recurring microstructural states, including amorphous, paracrystalline, crystalline, twinned, and hybrid intermediates. By tracking these states across temperature, we construct phase trajectory models that reveal the topology of the recrystallization landscape, including metastable basins, branching pathways, hybrid states, and temperature-dependent reorganizations of accessible states. Applied to ion irradiated GaAs, this approach uncovers two distinct recrystallization regimes separated by a transition near 250\textdegree{}C. At low temperature, recrystallization is growth-dominated and dominated by the persistence of amorphous and crystalline states. At high temperature, the transformation landscape reorganizes: hybrid and faulted states become metastable precursors to twinning, polycrystalline regions stabilize, and twinned structures emerge as dominant end states. The latent space representation also identifies amorphous patterns with weak symmetry signatures that precede recrystallization. This reveals structural precursors to ordering that are not captured by conventional descriptors give new insights into how recrystallization is initiated.