A systematic evaluation of ML-based crystal structure determination from powder diffraction

Deep generative models have shown promise for crystal structure determination from powder X-ray diffraction (PXRD), addressing the ambiguity of the PXRD inverse problem. However, comparisons across methods are hindered by differences in datasets and evaluation protocols. Moreover, existing methods are evaluated largely on simulated, single-phase data with known primitive-cell composition, leaving important practical settings underexplored. We introduce a unified benchmark spanning three settings: known primitive-cell composition, element-set conditioning, and multiphase recovery. We evaluate prior methods under a common protocol, with particular emphasis on experimental PXRD, and introduce Crystalite-PXRD, a strong baseline built on a recent crystal structure prediction backbone. On experimental RRUFF patterns, Crystalite-PXRD reaches a 50\% top-1 match rate, improving by 18.92 percentage points over the strongest prior baseline. On the two-phase recovery task, it achieves a 40.17\% top-20 all-phase match rate. Our benchmark provides a common basis for comparison and identifies sim-to-real transfer and multiphase recovery as key remaining challenges toward practical PXRD-based structure determination.