DiffractGPT-Rietveld

DiffractGPT-Rietveld is an automated pipeline for determining crystal structures from powder X-ray diffraction (XRD) data. It pairs generative machine-learning structure prediction (DiffractGPT) with database pattern matching and computational Rietveld refinement. For a known material it matches the measured pattern against reference structures, and for a novel compound it generates candidate structures directly from the diffraction pattern and then refines the lattice parameters. Benchmarked on mineral samples and crystal-structure datasets, it recovers lattice parameters with success rates of roughly 80 to 98 percent, and it runs as an open, web-accessible tool on AtomGPT.org.

This work is described in Hybrid DiffractGPT-Rietveld Refinement Framework for Automated X-ray Diffraction Analysis (arXiv:2607.08890).

Demo: atomgpt.org/xrd