MOF Dataset · Orbital-Materials
MofasaDB
Metal-organic framework dataset from Orbital — large-scale curated MOF structures for materials-discovery ML and synthetic chemistry workflows.
What MofasaDB is
MOFs are porous crystalline materials composed of metal nodes linked by organic ligands, widely studied for gas storage, separations, and catalysis. The database pairs unconditionally generated structures from the Mofasa model with their geometry-relaxed counterparts.
A key challenge in computational materials discovery is the availability of diverse, physically valid synthetic candidate structures at scale. MofasaDB addresses this by supplying large numbers of novel frameworks trained on the QMOF database with up to 170 atoms per structure, annotated with chemical, structural, and energetic features to support machine learning and synthetic chemistry workflows.
How it was made
The dataset contains unconditionally generated MOF structures produced by Mofasa, a generative model trained on structures from the QMOF dataset containing up to 170 atoms. Generated samples are stored in an Atomic Simulation Environment (ASE) database (samples.db) and matched row-for-row with geometry-relaxed counterparts (relaxed.db).
Each entry is annotated through multiple characterisation pipelines. Energy and maximum atomic forces are computed using the ORB machine-learned interatomic potential, along with latent graph representations across five graph neural network layers. Porosity and geometric features (such as cavity diameters, surface areas, and pore volumes) are calculated via Zeo++ using an N₂ probe radius of 1.86 Å. Structural validity, coordination geometry, and chemical consistency are assessed using MOFChecker, SMACT charge balancing, and MOFID framework decomposition into constituent nodes, linkers, and Reticular Chemistry Structure Resource (RCSR) network topologies.
Key facts
- Maximum atoms per structure
- 170
- File formats
- ASE SQLite database (.db), NumPy array (.npy)
- Primary components
- samples.db, relaxed.db, sample_latents/, relaxed_latents/
- Interatomic potential
- ORB machine-learned force field
- Pore calculation probe
- Spherical probe (N₂ radius: 1.86 Å via Zeo++)
What you can use it for
- Materials Screening for Adsorption
- Filter relaxed frameworks by largest cavity diameter, pore limiting diameter, and accessible surface area to identify candidates for gas storage or separation.
- Machine Learning Representation Learning
- Train surrogate models or evaluate similarity metrics using precomputed ORB graph-level and sub-component latent vectors alongside standardized Morgan linker fingerprints.
- Synthetic Chemistry and Reticular Design
- Mine valid organic linkers and coordination environments by inspecting decomposed linker SMILES, coordination site counts, and underlying RCSR network topologies.
Explore the data
No queryable preview
This dataset is not available in a queryable form — usually an unusual file format or a repository too large to convert. The original source may still be readable.Strengths and limits
Good at
- Provides paired, row-aligned raw generated structures and geometry-relaxed counterparts for direct structural and energetic comparisons.
- Includes extensive multi-layer graph neural network latent embeddings (ORB) pooled at framework, node, linker, and solvent levels.
- Features comprehensive automated validation metadata from MOFChecker, SMACT, and Zeo++.
Watch out for
- Structures are restricted to up to 170 atoms, limiting the representation of very large-pore or highly complex unit cells.
- Geometric porosity descriptors are absent, zero, or null for frameworks determined by Zeo++ to lack sufficient porosity for probe access.
- Component breakdown or topology attributes contain error or unknown flags when MOFID fails to deconstruct building blocks.
Questions
Can I look inside MofasaDB without downloading it?
Not here: it is not available in a queryable form. Its description and statistics are still shown.
What licence is MofasaDB under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
How are the generated and relaxed databases indexed?
The ASE databases (samples.db and relaxed.db) are row-aligned and 1-indexed, meaning db.get(i) in samples.db matches db.get(i) in relaxed.db. The corresponding NumPy latent embeddings are 0-indexed, so latent[i - 1] corresponds to database row i.
What happens if a framework contains no solvent molecules?
In structures that lack solvent, the latent embedding vectors for bound_solvent and free_solvent are populated as zero vectors.
How is structural validity determined across the dataset?
Structure quality is evaluated through MOFChecker checks (such as atomic overlaps, over- or under-coordination, and 3D connectivity) combined with SMACT charge balance tests and interatomic distance metrics.
Written from MofasaDB's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.