Diffraction‐Native Representation Learning for Automatic Symmetry Recognition and Structural Retrieval

ABSTRACT Selected area electron diffraction (SAED) is an indispensable tool for local crystallographic analysis, yet its interpretation remains a time‐consuming bottleneck, particularly when multiple compounds exhibit comparable crystal structures. Conventional coding‐based approaches struggle with complex diffraction geometries, missing reflections, and orientation ambiguities, while existing deep learning (DL) methods are largely constrained to single‐view inputs and fixed classification categories rather than explicit structure identification. A multi‐view convolutional neural network with task‐specific branches is presented, in which diffraction information from two zone axes is fused into a unified diffraction‐native descriptor to enable both symmetry inference and database‐scale structural identification. Trained on ~10 000 materials and evaluated on ~2000 unseen materials from the Materials Project database, the proposed method achieves 95.09% accuracy in crystal system classification and 80.12% accuracy in space group prediction. Under chemical formula constraints, it further attains a Top‐1 recall of 93.9% for structure retrieval across ∼13 000 candidate materials. The model further demonstrates strong robustness across simulated zeolite structures, in situ phase‐transition datasets capturing dynamic symmetry evolution, and experimentally obtained polymorph datasets. This descriptor‐driven paradigm enables rapid in situ analysis (30–50 s) with interpretable and database‐compatible outputs, offering a scalable framework for automated crystallographic analysis and materials discovery.