OMol-Descriptors-4M (anonymous review mirror)

Post-DFT electronic-structure descriptors for the 4M-structure wavefunction subset of OMol25: six partial-charge schemes, three stored bond-order schemes plus QTAIM bond paths, full QTAIM critical-point topology, fuzzy-atom integrations, ALIE surface and planarity descriptors, ORCA-derived globals, and the OMol25 energy/force labels aligned to the same keys. Geometries and wavefunctions are inherited from OMol25 (wB97M-V/def2-TZVPD).

• Corpus: 3,986,738 catalogued structures, 34 chemical verticals, 80+ elements.
• Pipeline (anonymized): https://anonymous.4open.science/r/qtaim_generator-8C0C/README.md
• Graph construction (anonymized): https://anonymous.4open.science/r/qtaim_embed-0E05
• License: CC-BY-4.0, inheriting the OMol25 terms.

This mirror is for double-blind review. It hosts the validation split, the test split and the five held-out stress suites. The train split (3,125,019 structures) will be released with the de-anonymized version.

What is in this mirror

Folder	Structures (structure.lmdb)	Format
splits/val/	398,813	gzipped LMDBs split into parts
splits/test/	403,770	gzipped LMDBs split into parts
holdouts/H1/	15,025	plain LMDBs
holdouts/H3/	12,506	plain LMDBs
holdouts/H6/	2,589	plain LMDBs
holdouts/H7/	18,096	plain LMDBs
holdouts/H8/	12,362	plain LMDBs

The five held-out suites contain 60,578 records spanning 58,786 distinct structures (1,792 records are structures that belong to two suites). Held-out structures are removed from train/val/test, so no structure is both an in-distribution test point and a stress-test point.

Other files:

• splits/SHA256SUMS: checksums of every .gz.part_NNN file (paths relative to the repo root).
• splits/SHA256SUMS.lmdb: checksums of the reassembled LMDBs (paths relative to splits/).
• holdouts/manifest_holdout.parquet: held-out membership manifest.
• holdouts/pull_summary.csv: requested vs. found records per suite and LMDB.

Reassembling the split LMDBs

hf download confsubanon/OMol-Descriptors-4M --repo-type dataset --include "splits/*" --local-dir omol_desc
cd omol_desc
sha256sum -c splits/SHA256SUMS          # verify the parts
cd splits
for s in val test; do
  for f in $(ls $s/*.lmdb.gz.part_000 | sed 's/.gz.part_000$//'); do
    cat "$f".gz.part_* | gunzip > "$f"
  done
done
sha256sum -c SHA256SUMS.lmdb            # verify the reassembled LMDBs

Reading an LMDB

Every LMDB is a single file (subdir=False). Keys are ASCII strings and values are pickled Python objects. Each database also holds a length entry that should be skipped when iterating.

import lmdb, pickle

env = lmdb.open("splits/test/charge.lmdb", subdir=False, readonly=True, lock=False)
with env.begin() as txn:
    for k, v in txn.cursor():
        if k == b"length":
            continue
        record = pickle.loads(v)   # {"hirshfeld": {"charge": {"1_N": -0.28, ...}, ...}, "cm5": ...}
        break

LMDB files

Each split and each held-out folder contains the same nine LMDBs, all keyed on

k = relpath(job_folder, root).replace("/", "__")

so any combination of descriptor families joins by key.

File	Contents
structure.lmdb	pymatgen Molecule (positions, elements), total charge and spin, RDKit-derived bond list, source identifiers
charge.lmdb	per-atom charges by scheme: hirshfeld, cm5, adch, becke (ADC-corrected Becke), mulliken_orca, loewdin_orca; total dipole for Hirshfeld/CM5, total and per-atom dipoles for ADCH/Becke
qtaim.lmdb	critical points keyed by atom (nuclear CPs) or atom pair (bond CPs); 26 scalar fields per CP including rho, Laplacian, ellipticity, kinetic and potential energy densities; ring/cage CPs where present
bond.lmdb	bond orders by scheme: fuzzy_bond, mayer_orca, loewdin_orca. QTAIM bond presence is recovered from the bond-CP keys in qtaim.lmdb
fuzzy.lmdb	per-atom Multiwfn fuzzy integrations (becke_fuzzy_density, hirsh_fuzzy_density, spin counterparts on open-shell records); see Known issues
other.lmdb	ALIE surface statistics (volume, area, extrema, skewness) and planarity descriptors (mpp_full, sdp_full, mpp_heavy, sdp_heavy)
orca.lmdb	ORCA-derived globals: SCF energy and components, HOMO/LUMO energies, dipole and quadrupole, gradient RMS, rotational constants, S^2, SCF metadata
energy.lmdb	OMol25 reference labels on the same keys: total energy (eV), per-atom forces (eV/A, atom order of structure.lmdb), OMol25 record identifier and per-record metadata
timings.lmdb	per-job step durations; provenance only

Conventions

• Atom indices in charge.lmdb, bond.lmdb and qtaim.lmdb keys are 1-based (as printed by ORCA and Multiwfn) and follow the atom order of structure.lmdb. Per-atom entries are keyed "{i}_{El}".
• bond.lmdb keys are "{i}_{El_i}_to_{j}_{El_j}", stored once per pair with i < j; read them as undirected.
• Each bond-order scheme lists only pairs above its print threshold (Mayer > 0.10 and Loewdin > 0.05 in ORCA; total fuzzy bond order >= 0.05 in Multiwfn). A pair missing from a scheme was below the threshold, not lost.
• energy.lmdb energies were converted from Hartree with ASE's constant (CODATA 2014), as OMol25 does. They agree with our independently parsed orca.lmdb SCF energies to within 3e-11 eV. Using CODATA 2018 instead shifts the largest systems by up to 0.023 eV.

Coverage

The 1,631 jobs whose post-processing failed (25 QTAIM, 1,607 ORCA, 1 in both) are removed from all eight descriptor LMDBs. energy.lmdb keeps every catalogued structure. Per-LMDB record counts in this mirror:

LMDB	Val	Test	Held-out (distinct)
structure	398,813	403,770	58,786
charge, bond, fuzzy, other, timings	398,813	403,770	58,785
orca	398,812	403,770	58,785
qtaim	398,813	403,769	58,785
energy	398,982	403,951	58,786

Join on the key intersection of the LMDBs you need.

Splits

The main split is by molecular composition. Each structure's electronegativity-ordered formula (pymatgen Composition.formula) is hashed with SHA-256, mapped to [0, 1) and assigned to train / val / test at 0.8 / 0.1 / 0.1. Every structure that shares a composition gets the same assignment. Held-out structures are removed before hashing.

Held-out stress suites

The headline metric per suite is Delta = MAE_holdout - MAE_test, reported next to the absolute errors.

• H1 - Metal-ligand pairs (15,025): eleven (transition metal, partner) bond pairs, sampled across log-spaced frequency bands. Selection uses bonds (Mayer or fuzzy bond order >= 0.3), not element co-occurrence.
• H3 - Reactivity (12,506): composition-stratified subsample of tm_react, electrolytes_reactivity and pmechdb. Non-equilibrium and transition-state-adjacent geometries.
• H6 - Lanthanide-ligand pairs (2,589): the H1 construction applied to (lanthanide, partner) bonds.
• H7 - Large systems (18,096): more than 250 atoms.
• H8 - Large net charges (12,362): |net charge| > 4.

Counts are records present in the released LMDBs. They fall a few records short of the selection targets where a selected structure lacked a complete descriptor set. Membership is per structure while the main split is per composition, so some held-out structures share a composition with a training structure. We report results on the full suites and on their formula-disjoint subsets.

Known issues

• Non-finite charges. 21 records, all from the 5A_elytes vertical, have non-finite Multiwfn charges (Hirshfeld, CM5, ADCH, Becke). 8 of them are in this mirror:

  • val: 5A_elytes__2634_C3H8O3S_5_shell_559_0_1, 5A_elytes__604_C2H6O2_5_shell_631_0_1
  • H7: 5A_elytes__1004_CH4O_5_shell_397_0_1, 5A_elytes__2813_C2H3O2-1_5_shell_2047_1_1, 5A_elytes__2832_C4H11NO_5_shell_58_0_1, 5A_elytes__2864_C9H19NO_5_shell_414_0_1, 5A_elytes__370_C4H11NO_5_shell_31_0_1
  • H8: 5A_elytes__1472_F6P-1_5_shell_2279_5_1

  Drop these before fitting scalers.

• Open-shell fuzzy and spin quantities. A Multiwfn input/parsing issue affects this release:

  • hirsh_fuzzy_density is zero.
  • hirsh_fuzzy_spin holds Hirshfeld electron populations, not spin.
  • On open-shell records, fuzzy_bond and becke_fuzzy_spin are unreliable, as are the QTAIM spin-resolved fields of affected records.

  Closed-shell fuzzy_bond, becke_fuzzy_density, all six charge schemes, and the total-density QTAIM fields (rho, Laplacian, ellipticity, energy densities) are not affected. The pipeline is fixed, and the affected fields will be regenerated in a versioned update.

• ECP elements. Heavy atoms treated with effective core potentials rely on Multiwfn's EDF library to restore the core density. The pipeline re-queues jobs where the EDF load failed. Charges on ECP atoms can still show larger magnitudes than on all-electron atoms.

Intended use

Surrogate models for real-space charges, bond orders and QTAIM topology; charge-aware MLIPs (joint charge and energy/force heads via energy.lmdb); descriptor-augmented property prediction. Out of scope: drawing chemistry conclusions beyond the OMol25 level of theory, and excited states.

Citation

Citation withheld for double-blind review. Please also cite OMol25.