Open Molecules 2025 (OMol25), Original ASE-LMDB Release from Meta FAIR Chemistry

Provider
Meta FAIR Chemistry, Lawrence Berkeley National Laboratory and collaborators (Levine et al.)
License
CC-BY-4.0 (source)
Stored at
dl.fbaipublicfiles.com
Access
Free and public. No account with the data provider is needed.
DOI
10.48550/arXiv.2505.08762
Category
Chemistry & Materials
Source
https://doi.org/10.48550/arXiv.2505.08762

About Open Molecules 2025 (OMol25), Original ASE-LMDB Release from Meta FAIR Chemistry

The original files of Open Molecules 2025 (OMol25), the molecular DFT dataset from Meta FAIR Chemistry, Lawrence Berkeley National Laboratory and collaborators: more than 100 million single-point calculations at wB97M-V/def2-TZVPD in ORCA 6 over biomolecules, metal complexes, electrolytes and recomputed community datasets, covering 83 elements and systems of up to 350 atoms. Every structure carries total energy (eV) and atomic forces (eV/A), with total charge and spin multiplicity in atoms.info next to per-structure extras such as HOMO-LUMO gap and Mulliken, Lowdin and NBO charges. This row mounts the six tar.gz archives Meta serves from its public file server, exactly as the fairchem documentation and training configs expect them: train_4M (20 GB, the 3,986,754 structure subset), val (21 GB, 2,762,021 structures with labels), test (8 GB, 2,805,046 structures for leaderboard submissions), neutral_train (108 GB), neutral_val (125 MB, 27,697 structures) and the full train split (489 GB). Each archive unpacks to a folder of ASE DB LMDB shards (data0000.aselmdb and so on), which fairchem opens with AseDBDataset. Use these, not a Parquet conversion, to train or fine-tune UMA, eSEN and other machine learning interatomic potentials with fairchem, or to reproduce the published splits. The archives are gzip streams, so copy one to notebook storage and unpack it there; train_4M unpacks to about 22 GB. The data are CC-BY-4.0. The model checkpoints in the gated facebook/OMol25 Hugging Face repository are under the separate FAIR Chemistry License and are not part of this row. Cite Levine et al., arXiv:2505.08762 (2025).

How to load this dataset in Python

# pip install fairchem-core   (brings ase and the aselmdb backend)
import os, tarfile
from fairchem.core.datasets import AseDBDataset
base = os.environ.get("SCIGANTIC_MOUNT_PATH", "/mnt/http-archive/data")
# Start with the smallest split: 125 MB, 27,697 neutral validation structures
with tarfile.open(f"{base}/neutral_val.tar.gz", "r|gz") as t:
    t.extractall("/tmp/omol25", filter="data")
ds = AseDBDataset({"src": "/tmp/omol25/neutral_val"})
print(len(ds), "structures")
atoms = ds.get_atoms(0)
print(atoms.get_chemical_formula(), "charge", atoms.info["charge"], "spin", atoms.info["spin"])
print("energy (eV):", atoms.get_potential_energy(), "forces:", atoms.get_forces().shape)

Written for a Scigantic notebook, where /mnt/archive is the mounted dataset.

Data structure

Fields

positions, numbersase.Atoms
energyfloat
forcesfloat[n_atoms,3]
info.chargeint
info.spinint
info.sourcestr
info.homo_lumo_gap, homo_energyfloat
info.mulliken_charges, lowdin_charges, nbo_chargesfloat[n_atoms]

Files

Example file: neutral_val.tar.gz -> neutral_val/data0000.aselmdb

Path pattern: <split>.tar.gz -> <split>/data<NNNN>.aselmdb (+ .aselmdb-lock)

Analyze this dataset without downloading it

On Scigantic, Open Molecules 2025 (OMol25), Original ASE-LMDB Release from Meta FAIR Chemistry opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.

This page is also available as Markdown and as JSON.

Related Chemistry & Materials datasets