Research map / Probabilistic and scientific modeling / Molecules and proteins

Fast Organic Crystal Structure Prediction with Unit Cell Flow Matching

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CLARI, a flow-matching model that generates single organic unit cells (atoms plus lattice vectors) with a pair-bias DiT instead of triangle updates, sampling 15-30x faster than OXtal while beating it on CSP benchmarks.

Problem

Organic crystal structure prediction is accurate with traditional search but costs CPU-years per molecule; existing generative models like OXtal are computationally heavy because they model bulk crops of symmetry copies with triangle-update layers.

Approach

Model a single unit cell by treating lattice vectors as three virtual atoms under a flow-matching objective, replace triangle updates with pair-bias attention, use a data-fitted lattice source prior, and add volume and pairwise-distance auxiliary losses plus self-conditioning.

Key results

• CLARI-L with 400 samples and top-200 by UMA energy reaches Sol 0.919 on Rigid and 0.596 on Flexible, versus OXtal 0.300 and 0.220 at n_s=k=30.
• CLARI-L at n_s=k=1000 attains Sol@1000 of 0.763 on the CSD Teaching Subset.
• CLARI-L generates 150 structures per molecule in 2.2 s on average, or 6.0 s with UMA top-30 selection.

Surprising findings

• Log time discretization is uniformly harmful, contrary to protein and small-molecule practice that skews toward late timesteps.
• A standard Gaussian lattice prior degrades reconstruction metrics substantially compared with a data-fitted prior.

Limitations

• Conditions on a 2D molecular graph that is stereochemistry-agnostic and requires the number of molecules per unit cell (Z) as input.
• Occasionally generates crystals with steric clashes or unphysical voids.
• Sol is an approximate metric needing only 8/15 molecule matches.

Experimental setup

Datasets: Cambridge Structural Database (CSD) OXtal test set (rigid, flexible, CSP5-7) CSD Teaching Subset

Benchmarks: OXtal test set (Rigid, Flexible) CSP Blind Tests CSP5-7 CSD Teaching Subset

Metrics: Sol@k (>=8/15 COMPACK match) Sol@k 15/15 match clash rate PoseBusters pass rate relative volume error EMD PDD

Baselines: OXtal DFT participants (CSP Blind Tests average)

Builds on: OXtal AlphaFold3 Diffusion Transformer flow matching self-conditioning

Closest papers by meaning

• EP-Flow: Disordered Crystal Structure Prediction without Site-level Annotation (NeurIPS 2026 · similarity 0.87)
• MOFFlow: Flow Matching for Structure Prediction of Metal-Organic Frameworks (ICLR 2025 · similarity 0.87)
• All that structure matches does not glitter (NeurIPS 2025 · similarity 0.86)
• Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.85)
• Flexible MOF Generation with Torsion-Aware Flow Matching (NeurIPS 2025 · similarity 0.85)
• CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models (NeurIPS 2026 · similarity 0.85)