A Prototype-Guided Solver for Crystal Structure Determination from Powder X-ray Diffraction

Determining crystal structures from powder X-ray diffraction (PXRD) is an important but challenging inverse problem due to the substantial information loss. While existing data-driven approaches, including generative models, typically explore a highly flexible space of possible three-dimensional atomic arrangements, many experimentally observed inorganic crystals fall into a relatively small number of crystallographic prototypes, such as rocksalt and perovskite. Based on this observation, we introduce ProtoGPS (Prototype-Guided PXRD Solver), a framework that reformulates crystal structure determination as discrete prototype identification followed by low-dimensional optimization of prototype-specific structural degrees of freedom. Given a chemical stoichiometry and an observed PXRD pattern, ProtoGPS searches over compatible prototypes, optimizes their free structural parameters, and identifies the structure that best matches the observed diffraction pattern. Experiments on simulated PXRD patterns show that ProtoGPS substantially outperforms the existing data-driven method across crystal structure reconstruction metrics.