Fast Organic Crystal Structure Prediction with Unit Cell Flow Matching
Title: Accelerating Organic Crystal Structure Prediction via Unit Cell Flow Matching
Abstract:
Computational modeling of organic solids necessitates organic crystal structure prediction (CSP), yet this process has traditionally been prohibitively expensive, demanding several CPU-years for each molecule. While generative frameworks like OXtal have significantly lowered these barriers by directly sampling stable organic crystal structures, they rely on modeling large segments of bulk material using costly triangle layers, foregoing explicit lattice parametrization. This approach still incurs a computational overhead of minutes per molecule.
In this study, we introduce Clari, a large-scale flow matching model that reduces this inference time to mere seconds. Clari generates redundancy-free unit cells by substituting triangle layers with pure pair-bias attention mechanisms. The model’s input requirements are streamlined to atom types and bonds, eliminating the need for RDKit-sanitizable input molecules. This design broadens its utility to handle complex chemical systems, including fullerenes, metal complexes, and atom clusters.
We conduct a comprehensive ablation study on critical architectural decisions, including self-conditioning, noise priors, timestep distributions, and auxiliary losses. Benchmarking against OXtal’s test sets, Clari achieves a superior solve rate while delivering a performance speedup of 15 to 30 times. Furthermore, because Clari explicitly models hydrogen atoms, it allows for inference-time scaling through direct energy ranking, bypassing the need for post-generation decoration or relaxation steps. By generating 150 candidate crystals and filtering for the top 30 based on energy, we observe an additional improvement in solve rates while maintaining a speedup factor of 5 to 8 times.
To facilitate future benchmarking, we also present the CSD Teaching Subset, a new test split comprising diverse and complex molecular structures. These advancements render large-scale virtual screening of organic solids a practical reality by enabling CSP within seconds. The source code is publicly available at https://github.com/aspuru-guzik-group/clari.
Source: arXiv Generated at: 2026-06-03 00:00:00 UTC



