taradutt007/symqm3k

SymQM-3k: A Symmetry-Resolved Quantum Chemistry Dataset

SymQM-3k is a dataset of 3,967 small organic molecules with DFT-validated point group symmetry and symmetry-resolved molecular orbital spectra. It is the first molecular ML dataset where orbital irreducible representation (irrep) labels are primary outputs of the quantum chemistry calculation rather than post-hoc annotations.

Dataset Summary

Property	Value
Molecules	3,967
Total atoms	73,528
Total orbital entries	548,384
Point groups	21
Unique irrep tokens	20
Atoms types	H, C, N, O, F
Level of theory	B3LYP/6-31G*
Software	ORCA 5.0.4
Source dataset	QM9
Train / Val / Test	3,172 / 395 / 400

Files

Primary Files — used in all experiments

File	Rows	Description
target_summary_final.csv	3,967	One row per molecule: HOMO/LUMO energies (eV), irrep labels, irrep indices, point group, total SCF energy (Eh), dipole moment (Debye)
atom_details_final.csv	73,528	One row per atom: 3D coordinates (ORCA symmetry frame), element, Mulliken charge
orbital_profile_final.csv	548,384	One row per orbital: energy (eV), occupation, irrep label, sequential irrep index, is_homo and is_lumo flags
splits.csv	3,967	Fixed train/val/test split stratified by point group

Supplementary Files — provided for future use

File	Rows	Description
mayer_bond_orders.csv	75,123	Mayer bond orders between atom pairs (threshold > 0.1)
mayer_valence.csv	73,528	Mayer valence components VA and BVA per atom
molecule_metadata.csv	3,967	Rotational constants (MHz), asymmetric unit indices, SCF iteration count, asymmetric unit fraction

Model Weights

File	Description
symmace_best.pt	Pretrained SymMACE weights — reproduces paper results

Quick Start

import pandas as pd
from huggingface_hub import hf_hub_download

# Load molecular summary
summary = pd.read_csv(
    hf_hub_download(
        repo_id="taradutt007/symqm3k",
        filename="target_summary_final.csv",
        repo_type="dataset"))

print(f"Molecules: {len(summary)}")
print(f"Point groups: {summary['orca_pg'].nunique()}")
print(summary[['qm9_id', 'orca_pg', 'homo_ev',
               'lumo_ev', 'homo_irrep', 'lumo_irrep']].head())

Load Pretrained Model

from huggingface_hub import hf_hub_download
import torch

# Download weights
weights_path = hf_hub_download(
    repo_id="taradutt007/symqm3k",
    filename="symmace_best.pt",
    repo_type="dataset")

# Load and evaluate — see GitHub repo for full code
# github.com/staradutt/symqm3k
model.load_state_dict(
    torch.load(weights_path, map_location='cpu'))

Dataset Construction

SymQM-3k was constructed from QM9 via a three-stage pipeline:

Stage 1 — Topological filter (RDKit): Canonical atom ranking symmetry ratio < 0.7, retaining 18,830 candidates from 133,885 QM9 molecules.

Stage 2 — Structural point group (Pymatgen): PointGroupAnalyzer at 0.3 Å tolerance, identifying 3,983 non-C1 candidates.

Stage 3 — DFT calculation (ORCA 5.0.4): RKS B3LYP / 6-31G*, TightSCF, UseSym, SymThresh 1e-2 Bohr. 3,967 successfully converged with a non-trivial electronic point group.

Column Descriptions

target_summary_final.csv

Column	Type	Units	Description
qm9_id	int	—	QM9 molecule identifier
py_pg	str	—	Structural point group (Pymatgen)
orca_pg	str	—	Electronic point group (ORCA DFT)
homo_ev	float	eV	HOMO energy
homo_irrep	str	—	HOMO irrep label. NaN for C1 molecules
homo_irrep_idx	float	—	Sequential index within irrep channel. NaN for C1
lumo_ev	float	eV	LUMO energy
lumo_irrep	str	—	LUMO irrep label. NaN for C1 molecules
lumo_irrep_idx	float	—	Sequential index within irrep channel. NaN for C1
total_energy_eh	float	Eh	Total SCF energy
dipole_debye	float	Debye	Dipole moment magnitude

atom_details_final.csv

Column	Type	Units	Description
qm9_id	int	—	Foreign key
atom_idx	int	—	Atom index within molecule (0-indexed)
element	str	—	Atomic symbol: H, C, N, O, F
x	float	Å	Cartesian x (ORCA symmetry frame)
y	float	Å	Cartesian y
z	float	Å	Cartesian z
charge	float	e	Mulliken atomic charge

orbital_profile_final.csv

Column	Type	Units	Description
qm9_id	int	—	Foreign key
no	int	—	Orbital index (0-indexed, energy ordered)
occ	float	—	Occupation: 2.0 (occupied) or 0.0 (virtual)
ev	float	eV	Orbital energy
irrep	str	—	Mulliken irrep label. NaN for C1 molecules
irrep_idx	float	—	Sequential index within irrep channel
is_homo	int	—	1 if HOMO, 0 otherwise
is_lumo	int	—	1 if LUMO, 0 otherwise

Point Group Distribution

Point Group	Count	%
Cs	2,657	66.98
C2	430	10.84
C2v	399	10.06
C1	297	7.49
C3v	57	1.44
C2h	45	1.13
Others	82	2.07

Known Limitations

• 297 molecules (7.49%) have electronic point group C1 despite non-trivial structural symmetry, these have NaN irrep fields
• High-symmetry groups (Td, Oh, D6h) have very few molecules (1-5 each),per-group statistics are unreliable for these
• is_homo sums to 4,025 and is_lumo to 3,995 (not 3,967) due to degenerate frontier orbitals in high-symmetry molecules, this is physically correct
• ORCA uses reduced subgroups of C2v for orbital labeling, irrep labels correspond to the reduced subgroup, not the nominal point group (see paper Appendix Table A1 for full mapping)
• Free valence (FA) in mayer_valence.csv is zero for all atoms, dataset contains only closed-shell molecules

Reproducing Paper Results

git clone https://github.com/staradutt/symqm3k
cd symqm3k
pip install -r requirements.txt
python data/download.py
python evaluate.py --weights symmace_best.pt

Expected output: HOMO MAE : 0.1528 eV LUMO MAE : 0.1393 eV HOMO Irrep Acc : 76.0% LUMO Irrep Acc : 87.6% HOMO Irrep Idx MAE: 3.40 positions LUMO Irrep Idx MAE: 2.05 positions Transition Acc : 76.5%

Citation

License

Dataset: CC BY 4.0 Code: MIT