OMol-Descriptors-4M

Post-DFT interpretive descriptors (partial charges, bond orders, QTAIM topology, fuzzy and surface integrations, ORCA globals) for the 4M-structure wavefunction subset of OMol25. Built with the open generator pipeline by wrapping Multiwfn and ORCA post-analysis on geometries and wavefunctions inherited from the OMol25 release (wB97M-V / def2-TZVPD).

• Corpus size: 3,986,738 structures across 80+ elements and 34 verticals.
• License: CC-BY-4.0, inheriting OMol25 terms.
• Pipeline: https://github.com/santi921/qtaim_generator

Repository Layout

OMol-Descriptors-4M/
  train/        # 3,189,390 structures
  val/          #   399,690 structures
  test/         #   415,913 structures
  holdouts/
    h1_metal_ligand/      # 15,030 structures
    h3_reactivity/        # 12,507 structures
    h6_lanthanide_ligand/ #  2,589 structures
    h7_large_systems/     # 18,200 structures
    h8_large_net_charge/  # 12,393 structures

Each split folder (and each holdout folder) contains the same eight LMDB files described below. Held-out structures are removed from train/val/test so the same structure is never both in-distribution and stress-test. Overlaps between held-out sets are allowed.

LMDB Files

Every LMDB inside a split is keyed on the same string identifier

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

so any combination of descriptors joins by key on the full corpus. Cross-LMDB key alignment holds to within 4 records on the 4M-key index.

File	What it stores	What you get from it
structure.lmdb	pymatgen Molecule (positions, elements), total spin and charge, RDKit-derived bond list, external identifiers	Geometry, connectivity, spin/charge state, source IDs
charge.lmdb	per-atom charges nested by scheme	Hirshfeld, CM5, ADCH, Becke, Mulliken (ORCA), Loewdin (ORCA); plus total dipole on Hirshfeld/CM5 and total + per-atom dipoles on ADCH/Becke
qtaim.lmdb	per-structure dict of critical points keyed by atom ID (nuclear CP) or atom-pair (bond CP)	26 scalar fields per CP including rho, laplacian, ellipticity, kinetic and potential energy densities; ring/cage CPs and non-nuclear attractors where present; QTAIM bond presence is recoverable from BCP keys
bond.lmdb	bond-order schemes keyed by directed atom-pair string "{i}_{El_i}_to_{j}_{El_j}"	fuzzy_bond, mayer_orca, loewdin_orca
fuzzy.lmdb	per-atom Multiwfn fuzzy integrations	becke_fuzzy_density, hirsh_fuzzy_density; spin counterparts on open-shell records
other.lmdb	molecular descriptors from ESP and ALIE surface analysis	surface volume and area, signed mean and variance, polarity index, internal charge separation; plus mpp_full, sdp_full, mpp_heavy, sdp_heavy
orca.lmdb	ORCA-derived globals	total SCF energy, HOMO/LUMO orbital energies, dipole vector and magnitude, gradient RMS, energy components, rotational constants, quadrupole, S^2; filtered via DEFAULT_ORCA_FILTER when consumed by the converter
timings.lmdb	per-job step durations	Provenance only, not used by the converter

Coverage

Five of the seven analytic LMDBs show no corrupt or incomplete records. The two with failures are qtaim (25 missing, 0.0006%) and orca (1,607 missing, 0.0403%). ORCA failures concentrate on physically hard cases (solvated biomolecules, scaled-separation electrolytes, open-shell TM redox); geometry and Multiwfn-side descriptors stay valid on those records. Every charge scheme, bond-order scheme, full QTAIM topology, and ORCA globals reaches >= 99% on each of the 34 verticals.

Held-Out Evaluation Sets

Each holdout folder is a standalone copy of the eight LMDBs above, restricted to the structures selected for that stress test. Headline metric per set is Delta = metric_holdout - metric_main_test.

• H1 - Metal-ligand pairs (15,030): eleven (TM, partner) bond pairs sampled stratified across log-spaced frequency bands. Selection uses true bonds (Mayer or fuzzy bond order >= 0.3) from bond.lmdb, not element co-occurrence. Probes rare metal-ligand chemistry transfer while keeping every element represented in training.

• H3 - Reactivity (12,507): composition-stratified subsample of tm_react (5k), electrolytes_reactivity (5k), and pmechdb (2.5k). Tests non-equilibrium and transition-state-adjacent geometries.

• H6 - Lanthanide-ligand pairs (2,589): same construction as H1 applied to (Ln, partner) bonds across five Ln-bearing verticals. Probes rare f-block bonding while keeping every lanthanide represented in training.

• H7 - Large systems (18,200): structures with more than 250 atoms. Size extrapolation regime where fuzzy integration and grid-based QTAIM are most expensive and per-atom DFT signal-to-noise is hardest to maintain.

• H8 - Large net charges (12,393): structures with |net_charge| > 4. The high-charge tail where partial-charge schemes disagree most and SCF is most sensitive to functional and basis.

---

license: cc-by-4.0