Seanmonami/omol25-hydroformylation-matched

OMol25 transition-metal complex subset with hydroformylation-relevant flags

This is a derived subset of the OMol25 training split (Meta FAIR Chemistry, ωB97M-V/def2-TZVPD). It contains every OMol25 train structure whose composition includes at least one of Rh, Co, Ir, Ru, Pt, Pd, Fe, with per-structure geometry flags marking the ligand motifs that matter for hydroformylation chemistry (carbonyl, phosphorus donor, metal hydride).

Despite the repository name, this is not a curated set of hydroformylation catalysts. It is the full metal-containing pool; the flags let you slice it down.

What is inside

Item	Value
Source	OMol25 train (release 250514), all 80 shards data0000..data0079
Structures kept	5,252,820 (of 101,666,280 scanned)
Files	matched_data0000.db .. matched_data0079.db, one per source shard
Format	ASE SQLite database (ase.db), coordinates + DFT energy + forces + key-value metadata
Total size	about 29 GB
Level of theory	ωB97M-V/def2-TZVPD, unchanged from OMol25

Flag summary over the whole subset:

Flag	Count
has_CO (metal-bound carbonyl)	334,082
has_P (phosphorus present)	1,623,449
has_hydride (metal hydride)	164,614
has_CO and has_P	132,686
has_CO and has_P and has_hydride	12,864

Per metal (a structure with two different metals is counted under both):

Metal	All	CO + P	CO + P + hydride
Rh	367,814	9,677	868
Co	600,437	25,213	1,525
Ir	955,690	23,608	2,519
Ru	760,467	23,768	2,308
Pt	926,351	20,508	2,120
Pd	985,555	17,286	1,675
Fe	662,509	12,810	1,849

Selection and flag rules

1. Composition screen. Parse the OMol25 composition string; keep the structure if any of Rh, Co, Ir, Ru, Pt, Pd, Fe is present. Nothing else is filtered out.
2. has_CO. True if some carbon lies within 2.35 Å of a selected metal atom and that carbon has an oxygen at 1.05 to 1.25 Å.
3. has_hydride. True if some hydrogen lies 1.40 to 1.95 Å from a selected metal atom and is not within 1.20 Å of any C, N, or O atom.
4. has_P. True if phosphorus appears in the composition. This is an element check, not a metal-P distance check.

All distances are plain Euclidean distances on the stored geometry (no periodic images; OMol25 structures are molecular).

Key-value fields per row

Key	Type	Meaning
metals	str	comma-separated selected metals present, e.g. Rh or Co,Fe
has_CO, has_P, has_hydride	bool	flags defined above
charge	float	total charge from OMol25
spin	int	spin multiplicity from OMol25
num_atoms	int	atom count
data_id	str	OMol25 top-level category (e.g. metal_complexes)
source	str	OMol25 source subset path (e.g. omol/metal_organics/..., tm_react, mo_hydrides)
shard	str	source shard file the row came from
shard_index	int	row index inside that source shard; (shard, shard_index) is the unique key

Energy and forces are stored in the ASE calculator slots (row.energy, row.forces).

Usage

from huggingface_hub import hf_hub_download
from ase.db import connect

path = hf_hub_download("Seanmonami/omol25-hydroformylation-matched",
                       "matched_data0000.db", repo_type="dataset")
db = connect(path)
for row in db.select(has_CO=True, has_P=True, has_hydride=True):
    atoms = row.toatoms()
    print(row.metals, row.charge, row.spin, row.energy)

To fetch everything:

from huggingface_hub import snapshot_download
snapshot_download("Seanmonami/omol25-hydroformylation-matched",
                  repo_type="dataset", local_dir="omol25_tm_subset")

Known caveats

• The hydride flag depends on hydrogen positions. OMol25 subsets derived from crystal structures under-represent metal hydrides; the dedicated mo_hydrides and tm_react sources carry most of the hydride examples. Stratify by source if hydride chemistry matters to you.
• has_P does not check that phosphorus is coordinated to the metal.
• Shards are the original OMol25 shards, not a random split. Do not use shard boundaries as a train/test split.

Provenance and license

Derived from the OMol25 dataset released by Meta FAIR under CC-BY-4.0. This subset is redistributed under the same license. Changes made: filtering to the seven metals above and adding the flag and metadata columns described here; geometries, energies, and forces are unmodified.

Please cite OMol25 when using this data:

Levine, D. S. et al. The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models. arXiv:2505.08762 (2025). https://arxiv.org/abs/2505.08762

Subset prepared by Seokhyun Choung (CCEL), 2026. Extraction scripts and the distribution report are available on request.