Orbital-Materials/MofasaDB

MofasaDB

The MofasaDB is a publicly available dataset containing 200.000+ de novo generated MOF (Metal-Organic Framework) structures from Mofasa trained on QMOF (up to 170 atoms), along with their geometry-relaxed counterparts. The database is released alongside the paper Mofasa: A Step Change in Metal-Organic Framework Generation. A user-friendly web interface for search and discovery can be accessed at https://mofux.ai/.

Table of Contents

1. 1.Database Overview
2. 2.Quick Start
3. 3.Property Reference
4. 4.Structural Properties
5. 5.MOFID Properties
6. 6.Zeo++ Geometric Properties
7. 7.ORB Properties
8. 8.MOFChecker Properties
9. 9.MOF Fragment Properties
10. 10.Linker Properties
11. 11.Validation Metrics

Database Overview

The database contains unconditionally generated MOF structures from Mofasa, along with their geometry-relaxed counterparts.

Files

Data Alignment

The databases are row-aligned: row i in samples.db corresponds to row i in relaxed.db.

Indexing:

• —ASE databases are 1-indexed: first row is db.get(1)
• —NumPy arrays are 0-indexed: first element is array[0]
• —Therefore: latent[i] corresponds to db.get(i + 1)

Quick Start

Load a Structure

from ase.db import connect

db = connect("samples.db")
row_id = 1
row = db.get(row_id)                    # Get first structure (1-indexed)
atoms = row.toatoms()              # Convert to ASE Atoms object
print(atoms.get_chemical_formula())

Access Properties

# Get energy per atom
energy = row.data['properties']['orb_properties']['orb_energy_per_atom']

# Get pore diameter
lcd = row.data['properties']['pyzeo_geometric_properties']['lcd']

# Get topology (top-level property)
topology = row.data['topology']

Load Orb Latent Embeddings

import numpy as np

latents = np.load("sample_latents/orb_latent_4_graph.npy")
latent = latents[row_id - 1]  # Convert 1-indexed row to 0-indexed array

Compare Sample and Relaxed

sample_db = connect("samples.db")
relaxed_db = connect("relaxed.db")

# Row i in both databases correspond to the same structure
row_id = 100
sample_atoms = sample_db.get(row_id).toatoms()
relaxed_atoms = relaxed_db.get(row_id).toatoms()

print(f"Sample formula:  {sample_atoms.get_chemical_formula()}")
print(f"Relaxed formula: {relaxed_atoms.get_chemical_formula()}")

Handle Missing Data

Not all properties are available for every structure. Common causes include:

• —MOFID failure: If MOFID cannot identify the MOF building blocks (nodes, linkers, topology), these properties are set to "UNKNOWN", "ERROR", or empty lists for missing SMILES strings.
• —Zeo++ non-porous: If Zeo++ determines a structure has insufficient porosity for probe access, geometric properties (lcd, pld, accessible volume/surface area) may be missing, zero, or None.
• —Component absence: Latent embeddings for bound_solvent and free_solvent are zero vectors when structures contain no solvent molecules.

Property Reference

Properties are stored in row.data with nested paths. Some examples:

PROPERTY_PATHS = {
    # ORB model properties
    'orb_energy_per_atom': 'properties.orb_properties.orb_energy_per_atom',
    'orb_max_force': 'properties.orb_properties.orb_max_force',
    
    # Zeo++ geometric properties
    'lcd': 'properties.pyzeo_geometric_properties.lcd',
    'pld': 'properties.pyzeo_geometric_properties.pld',
    'dif': 'properties.pyzeo_geometric_properties.dif',
    'av_volume_fraction': 'properties.pyzeo_geometric_properties.av_volume_fraction',
    'av_cm3_per_g': 'properties.pyzeo_geometric_properties.av_cm3_per_g',
    'nav_volume_fraction': 'properties.pyzeo_geometric_properties.nav_volume_fraction',
    'asa_m2_per_g': 'properties.pyzeo_geometric_properties.asa_m2_per_g',
    'number_of_channels': 'properties.pyzeo_geometric_properties.number_of_channels',
    'number_of_pockets': 'properties.pyzeo_geometric_properties.number_of_pockets',
    
    # Crystal symmetry
    'spacegroup_number': 'properties.crystal_symmetry.symprec_0.01/spacegroup_number',
    'pointgroup': 'properties.crystal_symmetry.symprec_0.01/pointgroup',
    
    # MOFID properties
    'mofid': 'mofid',
    'mofkey': 'mofkey',
    'topology': 'topology',
    'smiles_nodes': 'smiles_nodes',
    'smiles_linkers': 'smiles_linkers',
    'cat': 'cat',
    
    # MOFChecker
    'mofchecker': 'properties.mofchecker',
    'mofchecker_valid': 'properties.mofchecker.mofchecker_valid',
}

Structural Properties

Lattice Parameters

Chemical Composition

MOFID Properties

MOFID is a standardized identifier for MOF structures that encodes topology, nodes, linkers, and catenation information.

Crystal Symmetry

Computed using pymatgen's SpacegroupAnalyzer.

Detailed Crystal Symmetry (nested under properties.crystal_symmetry)

Zeo++ Geometric Properties

Computed using Zeo++ via the pyzeo wrapper. These properties characterize the pore geometry and accessibility using a spherical probe (default: N₂ probe radius of 1.86 Å).

Pore Descriptors

Volume Properties

Surface Area Properties

ORB Properties

Properties computed using the ORB machine-learned interatomic potential.

Energy and Forces

ORB Latent Embeddings

ORB latent embeddings are stored as NumPy files in the sample_latents/ and relaxed_latents/ directories.

File naming: orb_latent_{layer}_{component}.npy

• —Layers 0-4 correspond to different depths in the ORB GNN (layer 4 = final layer)
• —Zero vectors indicate missing data (e.g., structures without solvents)

MOFChecker Properties

Computed using MOFChecker, a tool for validating MOF structures. All keys are prefixed with mofchecker_.

Validity Checks (Binary)

These descriptors are used to determine overall MOF validity. True indicates a problem (except where noted).

Informative Checks (Binary, not used for validity)

Structure Hashes

MOF Fragment Properties

Properties of the decomposed MOF components (nodes, linkers, solvents). Stored under properties.mof_fragments.

Component Types

MOF structures are decomposed into four component types:

• —nodes_and_bridges: Metal nodes and bridging groups
• —linkers: Organic linker molecules
• —bound_solvent: Solvent molecules coordinated to metal centers
• —free_solvent: Unbound solvent molecules in pores

Fragment Formulas

Example: `nodes_and_bridges_formulas = ["Zn4O", "Zn4O"]` for a structure with two identical zinc nodes

Linker SMILES

Linker Properties

Molecular descriptors and fingerprints for organic linker molecules. Stored under properties.linker_properties.

Morgan Fingerprints

Morgan (circular) fingerprints are stored as NumPy files. For similarity search, use the standardized versions.

Scalar metadata:

Molecular Descriptors

Computed on standardized molecules using RDKit.

Coordination Site Descriptors

Counts of metal-coordinating functional groups (computed on as-parsed molecules).

Validation Metrics

Binary metrics used to assess structure quality.

License

CC-BY-4.0

References

• —MOFID: Bucior, B. J., et al. (2019). Identification Schemes for Metal-Organic Frameworks...
• —Zeo++: Willems, T. F., et al. (2012). Algorithms and tools for high-throughput geometry-based analysis...
• —MOFChecker: Ongari, D., et al. (2019). Building a Consistent and Reproducible Database for Adsorption Evaluation...
• —QMOF Andrew S. R., et al. (2021). Paper can be found at Machine learning the quantum-chemical properties of metal–organic frameworks for accelerated materials discovery and corresponding dataset release on GitHub
• —ORB: Orbital ORB v3 Force Field
• —RDKit Morgan Fingerprints: RDKit Documentation