Research map / 3D vision / 3D generation
Flexible MOF Generation with Torsion-Aware Flow Matching
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 MOFFlow-2, a two-stage MOF generator that autoregressively writes SMILES for metal and organic building blocks and then uses a torsion-aware Riemannian flow matching model to assemble 3D frameworks without fixed building blocks or known local coordinates.
Problem
Generating MOFs with novel chemistries and accurate 3D structures is hard because prior generators use a fixed building block library and assume rigid building blocks with known local coordinates.
Approach
Combine a SMILES-based autoregressive building-block generator (with RDKit and a metal template library for 3D initialization) with a flow matching model over SO(3) rotations, translations, SO(2) torsions and lattice parameters, using a non-equivariant Transformer and canonicalization of rotation and torsion targets.
Key results
- Structure prediction at stol=0.5 with 5 samples: MOFFlow-2 reaches 43.95% match rate vs 32.71% for MOFFlow.
- Generation: MOFFlow-2 valid rate 38.84% vs 10.13% for MOFDiff; VNU 31.35% vs 7.95%.
- MOFFlow-2 produces novel building blocks in 10.10% of samples (MOFDiff 0.00%), with 1.82 s sampling time vs 3.19 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MOFFlow-2 instead of MOFFlow | MOF structure prediction · Structure prediction, stol=0.3, 5 samples | match rate (%): 8.68 -> 15.98 | Helped |
| Use MOFFlow-2 instead of MOFFlow | MOF structure prediction · Structure prediction, stol=0.5, 5 samples | match rate (%): 32.71 -> 43.95 | Helped |
| Use MOFFlow-2 instead of MOFFlow | MOF structure prediction · Structure prediction, stol=0.3, 5 samples | RMSE (Å): 0.2039 -> 0.1842 | Helped |
| Use MOFFlow-2 instead of MOFFlow | MOF structure prediction · Structure prediction, stol=0.5, 5 samples | RMSE (Å): 0.3290 -> 0.2925 | Helped |
| Use MOFFlow-2 instead of MOFDiff | MOF generation · MOF generation | valid (%): 10.13 -> 38.84 | Helped |
| Use MOFFlow-2 instead of MOFDiff | MOF generation · MOF generation | VNU (%): 7.95 -> 31.35 | Helped |
| Use MOFFlow-2 instead of MOFDiff | MOF generation · MOF generation | NBB (%): 0.00 -> 10.10 | Helped |
| Use MOFFlow-2 instead of MOFDiff | MOF generation · MOF generation | sampling time (s): 3.19 -> 1.82 | Helped |
Limitations
- Dependence on RDKit for conformer initialization
- Partial conditioning in property-guided generation
- Approximate energy evaluation with machine learning interatomic potentials
Experimental setup
Datasets: hypothetical MOF dataset (filtered with MOFChecker, MOFid metal-oxo decomposition)
Benchmarks: structure prediction on hypothetical MOF test split MOF generation VNU/NBB on 10,000 samples
Metrics: match rate RMSE validity novelty uniqueness novel building blocks sampling time
Baselines: RS EA DiffCSP MOFFlow MOFDiff
Builds on: MOFFlow MOFDiff DiffCSP Riemannian flow matching
Closest papers by meaning
- MOFFlow: Flow Matching for Structure Prediction of Metal-Organic Frameworks (ICLR 2025 · similarity 0.93)
- MOFDiff: Coarse-grained Diffusion for Metal-Organic Framework Design (ICLR 2024 · similarity 0.89)
- Fast Organic Crystal Structure Prediction with Unit Cell Flow Matching (NeurIPS 2026 · similarity 0.85)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.84)
- Unified Generative Modeling of 3D Molecules with Bayesian Flow Networks (ICLR 2024 · similarity 0.84)
- 3D Molecule Generation from Rigid Motifs via $\mathrm{SE}(3)$ Flows (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.