Track the frontier, from condensed matter to cosmology.
Search papers, labs, and topics across Lattice.
This paper introduces PXRD-Bench, a comprehensive benchmark for evaluating generative models in crystal structure determination from powder X-ray diffraction (PXRD), extending beyond single-phase, known-composition scenarios to include unknown atom counts and two-phase mixtures. The benchmark comprises approximately 4.1 million simulated PXRD patterns and real-world experimental data, alongside Bragglite, a new state-of-the-art PXRD-to-structure model achieving a 44.2% Top-1 match rate on experimental data. Findings reveal that while some models struggle with generalization to novel prototypes and two-phase recovery, the benchmark and baseline model offer critical insights into the current state and future directions of automated crystal structure determination.
Current PXRD-to-structure models achieve a 44.2% Top-1 match rate on experimental data, but still struggle significantly with generalization to unseen structural prototypes and recovery from two-phase mixtures.
Determining crystal structures from powder X-ray diffraction (PXRD) would enable routine characterization of newly synthesized materials, including those absent from reference databases. Generative models offer a promising approach, but existing evaluations remain fragmented and largely assume single-phase samples with known composition. To fill this gap, we introduce PXRD-Bench, a unified benchmark that extends evaluation to unknown per-element atom counts and recovery of both structures in two-phase mixtures. It provides approximately 4.1 million training structures with simulated PXRD, tests generalization to unseen compositions and structural prototypes, and evaluates recovery from experimental PXRD after excluding structurally matching training entries. To provide a strong reference baseline, we also introduce Bragglite, a PXRD-to-structure model developed through a systematic study of PXRD modeling choices. It achieves a 44.2\% Top-1 match rate on experimental RRUFF patterns with known composition, 24.6\% above the strongest prior baseline. Further analysis yields several key findings: some models fail to benefit from experimental PXRD, while generalization to unseen structural prototypes and recovery from two-phase mixtures remain challenging. Together, our benchmark, systematic studies, and Bragglite clarify the capabilities and limitations of current PXRD-to-structure models.
MIT CSAIL