CrySTARNet: X‑ray Diffraction Refinement for Disordered Inorganic Crystals via Generative Artificial Intelligence

CrySTARNet (Crystal Structure Autonomous Refinement Network) is a diffusion-based framework for autonomous crystal structure refinement from diffraction data. The framework combines XRD patterns, compositional information, and optional TEM data through a shared multimodal representation space, enabling end-to-end prediction of lattice parameters, atomic coordinates, elemental species, and fractional occupancies. By explicitly modeling continuous occupancy distributions, CrySTARNet naturally handles non-stoichiometric compounds, doped materials, and solid-solution systems. Across six representative crystal structure families, CrySTARNet achieves over 85% match accuracy within the top ten generated candidates, demonstrating strong performance for automated materials characterization and crystal structure determination. Because GitHub does not allow hosting large files, the dataset and model checkpoints are provided here. Please place them into the correct directories after downloading: rename best_xrd.pt by removing the _xrd suffix and put it into ./XRD CLIP/; rename best_xrd_TEM.pt by removing the _xrd_TEM suffix and put it into ./XRD-TEM CLIP/; place epoch=699-step=246399.ckpt into ./Diffusion/output/HYDRA/2025-03-12/perov_CaTiO3/; and after extracting the data archive, move the resulting folder into ./Diffusion/.