[Paper Review] OrgFlow: Generative Modeling of Organic Crystal Structures from Molecular Graphs

Mohammadmahdi Vahediahmar, Matthew A. McDonald+1 more|arXiv (Cornell University)|Feb 22, 2026

Machine Learning in Materials Science0 citations

TL;DR

OrgFlow is a conditional flow-matching model that generates full periodic organic crystal structures from molecular graphs, achieving significantly higher match rates than baselines with fewer inference steps, via bond-aware priors and periodic boundary conditioning.

ABSTRACT

Crystal structure prediction is a long-standing challenge in materials science, with most data-driven methods developed for inorganic systems. This leaves an important gap for organic crystals, which are central to pharmaceuticals, polymers, and functional materials, but present unique challenges, such as larger unit cells and strict chemical connectivity. We introduce a flow-matching model for predicting organic crystal structures directly from molecular graphs. The architecture integrates molecular connectivity with periodic boundary conditions while preserving the symmetries of crystalline systems. A bond-aware loss guides the model toward realistic local chemistry by enforcing distributions of bond lengths and connectivity. To support reliable and efficient training, we built a curated dataset of organic crystals, along with a preprocessing pipeline that precomputes bonds and edges, substantially reducing computational overhead during both training and inference. Experiments show that our method achieves a Match Rate more than 10 times higher than existing baselines while requiring fewer sampling steps for inference. These results establish generative modeling as a practical and scalable framework for organic crystal structure prediction.

Motivation & Objective

• Motivate the need for organic crystal structure prediction and bridge the gap left by inorganic-focused CSP methods.
• Propose a conditional flow-matching framework conditioned on molecular graphs to generate full periodic organic crystals.
• Preserve covalent connectivity and crystallographic symmetry through bond-aware regularization and periodic graph representations.
• Build and curate a large organic crystal dataset with bond-length statistics to support training and evaluation.

Proposed method

• Formulate CSP as transport from a Gaussian prior to crystal distributions conditioned on a molecular graph G.
• Use an E(3)-equivariant message-passing architecture on periodic graphs to learn velocity fields for fractional coordinates and lattice parameters.
• Incorporate bond-aware regularization using empirical bond-length statistics to enforce chemically valid geometries.
• Preprocess CIFs into symmetry-consistent periodic molecular graphs with integer image shifts for PBC handling.
• Train with rectified flow principles to achieve near-linear interpolation and reduced ODE steps (20 steps).
• Evaluate using structure-matcher based metrics (pymatgen StructureMatcher and spglib) to assess geometric and symmetry fidelity.

Experimental results

Research questions

• RQ1 Can a conditional flow-matching model conditioned on molecular graphs accurately generate full periodic organic crystal structures?
• RQ2 How does incorporating molecular connectivity and bond-length priors affect chemical validity and crystallographic accuracy of generated structures?
• RQ3 What is the efficiency (inference steps) of OrgFlow compared to baselines when generating organic crystals?
• RQ4 How does OrgFlow perform across diverse organic subsets (drug-like, small molecules, etc.)?
• RQ5 What dataset and preprocessing enable scalable training for organic CSP?

Key findings

• OrgFlow achieves substantially higher match rates than FlowMM across all dataset splits (e.g., 21.94% vs 0.1% on drug-like molecules).
• OrgFlow requires far fewer sampling steps than baselines (near 20 steps vs ~500 steps for FlowMM on small molecules).
• Bond-aware regularization improves match rate (0.0% to 10.1% as λb increases to 5e-3).
• Fractional coordinate and lattice losses further enhance accuracy, with lattice weight yielding up to 13.6% match rate in ablations.
• OrgFlow demonstrates strong qualitative alignment with ground-truth organic crystals, recovering ring systems and packing while errors mostly occur in flexible substituents.
• A large organic-crystal benchmark of 177k structures from the CSD was compiled to support training and evaluation.