Research map / Image generation / Autoregressive and latent generation
MOFDiff: Coarse-grained Diffusion for Metal-Organic Framework Design
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 MOFDiff, a coarse-grained equivariant diffusion model that generates building-block-level MOF structures, recovers all-atom structures via an assembly algorithm and UFF relaxation, and optimizes MOFs for CO2 working capacity.
Problem
Template-based MOF generation is limited in chemical diversity, and atomic-resolution diffusion for MOFs (tens to hundreds of atoms per cell) is computationally hard and yields unphysical nodes/linkers.
Approach
Model MOFs as coarse-grained building blocks: a diffusion process over block coordinates and contrastively learned block embeddings (conditioned on a periodic GemNet-OC latent code), followed by a differentiable overlap-based assembly of connection points and force-field relaxation.
Key results
- Of 10,000 generated samples, 2,998 are valid, novel, and unique MOFs.
- MOFDiff-optimized MOFs reach CO2 working capacities up to 4.89 mol/kg; 9 of the top 10 outperform the literature MOFs compared.
- Requires 14.6 GCMC simulations per MOF with working capacity above 2 mol/kg versus 58.1 for naive BW-DB sampling (75% reduction).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Optimize latent vectors for CO2 working capacity | MOFDiff · Carbon capture MOFs | CO2 working capacity [mol/kg]: 3.82 -> 4.89 | Helped |
| Use MOFDiff instead of naive BW-DB sampling to find qualifying MOFs | Carbon capture candidate search · MOFs with working capacity >2 mol/kg | GCMC simulations per qualifying MOF: 58.1 -> 14.6 | Helped |
| Generate larger MOFs with more building blocks | MOFDiff · MOF validity | validity percentage | Hurt |
| Optimize MOFDiff samples for carbon capture | MOFDiff · Top 10 carbon capture candidates | MOFDiff candidates among top 10: 9 out of the top 10 | Helped |
Surprising findings
- Validity percentage declines as the number of building blocks increases.
Limitations
- Validity drops for MOFs with more building blocks.
- Lattice parameters must be very accurate; lattice is fixed during diffusion.
- Relies on GCMC simulations with a re-implemented workflow whose labels differ from BW-DB (~30% underestimation of working capacity).
Experimental setup
Datasets: BW-DB
Benchmarks: BW-DB
Metrics: validity rate novelty uniqueness pore limiting diameter gravimetric surface area density void fraction CO2 working capacity CO2/N2 selectivity GCMC simulations per qualifying MOF
Baselines: BW-DB original labels 18 literature MOFs (e.g., Al-PMOF, UiO-66, MIL-53(Al)) naive sampling from BW-DB
Builds on: DDPM GemNet-OC MOFid UFF MOFChecker
Closest papers by meaning
- MOFFlow: Flow Matching for Structure Prediction of Metal-Organic Frameworks (ICLR 2025 · similarity 0.91)
- Flexible MOF Generation with Torsion-Aware Flow Matching (NeurIPS 2025 · similarity 0.89)
- MolHIT: Advancing Molecular-Graph Generation with Hierarchical Discrete Diffusion Models (NeurIPS 2026 · similarity 0.85)
- SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning (NeurIPS 2024 · similarity 0.84)
- Graph Diffusion that can Insert and Delete (NeurIPS 2025 · similarity 0.84)
- Periodic Materials Generation using Text-Guided Joint Diffusion Model (ICLR 2025 · 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.