primateria/mbgfnet-closed-shell-3d-gw-dataset

Closed-shell 3d transition-metal complexes: PBE0 + G0W0 quasiparticle dataset

2,100 closed-shell mononuclear 3d transition-metal complexes (Sc–Zn), each with a DFT (PBE0/cc-pVDZ) and a one-shot G0W0@PBE0/cc-pVDZ quasiparticle-energy calculation, built to train and benchmark MBGF-Net (Venturella, Li, Hillenbrand, Zhu, arXiv:2407.20384) — a graph neural network that predicts the GW self-energy from cheap DFT quantities — on transition-metal chemistry, which the original paper's architecture had not been shown to handle.

1,889 of the 2,100 rows carry verified, converged labels (see Label quality control below); the rest are retained with an explicit label_status so they can be excluded or used as an open-shell seed set.

What's in it

• tmqm_dblock_2100.csv — one row per complex: identity (CSD refcode, conformer parent), split assignment, label_status, metal, formula, whether it contains a metal–fluorine bond, and ground-truth PBE0 and G0W0 properties: HOMO/LUMO/gap (eV), linearized-density dipole and quadrupole moments, metal-3d IAO population, and the metal-3d quasiparticle renormalization factor (Z). data/<split>.parquet are the same rows split by the split column (browsable via the Hugging Face dataset viewer's split dropdown) — same content, just pre-split into Parquet for browsability; use whichever is more convenient.
• xyz/<label_id>.xyz — the Cartesian geometry for every row (2,100 files). Geometries are GFN2-xTB-optimized structures from the tmQM database (Balcells & Skjelstad, J. Chem. Inf. Model. 2020, 60, 6135, CC BY 4.0); 100 rows are Gaussian-perturbed conformers (σ = 0.03 Å) of 5 tmQM parents, generated for this dataset to test conformational transfer.
• tmqm_splits.json — the fixed train/test assignment used for every result derived from this dataset: a 1,094-complex training pool, a 500-complex random held-out test, a 393-complex unseen-metal test (every Mn and Co complex, held out of training entirely), a 13-complex unseen-ligand test (every metal–fluorine complex), and 100 conformers of 5 training-pool parents.

Level of theory

RKS-PBE0/cc-pVDZ ground state, then one-shot G0W0@PBE0/cc-pVDZ with density fitting on the imaginary frequency axis (100-point integration grid, 30 fit frequencies, full self-energy matrix), continued to the real axis with an 18-point Padé approximant. Reference labels were cross-checked against contour-deformation GW (no analytic continuation) on a 12-molecule subset: agreement ≤ 3.2 meV for occupied frontier states, ≤ 10.8 meV for LUMOs.

Label quality control

An initial batch of 2,100 labels was audited in two passes before this release:

1. SCF convergence. 133 of 2,100 (6.3%) reference SCFs had not converged within the generator's default cycle limit; the GW step had silently run on the unconverged density. 71 were re-converged with a robust protocol (level shift → second-order SCF → stability check) and corrected in place; 62 turned out to be genuinely open-shell at PBE0 (excluded here, label_status = excluded_open_shell_ground_state); 2 have no closed-shell solution at any tested starting point (excluded_no_closed_shell_solution).
2. Spin stability. Every remaining nominally closed-shell label was checked with an unrestricted (broken- symmetry) PBE0 calculation seeded from the closed-shell density; if it converged more than 20 meV below the closed-shell energy, the closed-shell reference is not the true ground state. 139 more complexes failed this test (excluded_open_shell_ground_state) and 10 could not be verified (excluded_unverified_spin_state).

About 10% of tmQM's built-in singlet (S = 0) assignments are wrong at the PBE0/cc-pVDZ level, concentrated in Ni(II) outside square-planar coordination and in Co(III)/Co(II) complexes with antiferromagnetically coupled ligand radicals. The excluded complexes are retained (not deleted) because they are a natural seed set for extending this work to open-shell complexes.

label_status values: ok (1,818), ok_relabeled_after_unconverged_scf (71), excluded_open_shell_ground_state (199), excluded_unverified_spin_state (10), excluded_no_closed_shell_solution (2).

Update: open-shell G0W0 labels for these excluded rows are now available in a companion release, primateria/mbgfnet-open-shell-3d-gw-dataset (its pilot191.csv subset, geometry_source = tmQM) — the seed-set use mentioned above.

Splits

tmqm_splits.json and the split column are the split used to produce the results reported for this dataset (designed transferability tests, not a random partition — see Rule below), provided so anyone can reproduce those numbers or compare a new model against them on identical held-out complexes. Every row keeps its full metadata (metal, atom count, elements, metal–fluorine bond), so nothing stops you from ignoring this split and building your own from the raw table for a different question.

Split	Size (labelled)	Rule
train_pool	995	remainder after holding out the tests; excludes all Mn/Co and all M–F complexes
test_random	449	random held-out, same metal distribution as training
test_unseen_metal	332	every Mn and Co complex (metals never in training)
test_unseen_ligand_MF	13	every complex with a metal–fluorine bond
conformers	100	20 perturbed geometries each of 5 training-pool parents

(Sizes above are post-QC; tmqm_splits.json gives the pre-QC split membership, which is what determines label_status = excluded_* rows' original split.)

Provenance and licensing

• Geometries derive from tmQM (Balcells, D.; Skjelstad, B. B. J. Chem. Inf. Model. 2020, 60, 6135–6146. CC BY 4.0), filtered to charge-0, tmQM-singlet, ≤ 35-atom, single-3d-metal complexes with ligand elements in {H, B, C, N, O, F, Si, P, S, Cl}.
• All PBE0 and G0W0 quantities in this dataset were computed by us for this work using PySCF and fcdmft (Zhu group, Yale).
• This dataset (everything except the underlying tmQM geometries, which retain their own CC BY 4.0 licence from the original authors) is released under CC BY 4.0. If you use it, please cite tmQM and this dataset.

Citing

Please cite this dataset directly, the MBGF-Net paper this dataset extends, and tmQM:

@misc{mbgfnet_closedshell_3d_2026,
  title        = {Closed-shell 3d transition-metal complexes: PBE0 + G0W0 quasiparticle dataset},
  author       = {Hsieh, Emerson},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/primateria/mbgfnet-closed-shell-3d-gw-dataset}}
}

@article{venturella2024mbgfnet,
  title={Unified Deep Learning Framework for Many-Body Quantum Chemistry via Green's Functions},
  author={Venturella, Christian and Li, Jiachen and Hillenbrand, Christopher and Leyva Peralta, Ximena and Liu, Jessica and Zhu, Tianyu},
  journal={arXiv preprint arXiv:2407.20384},
  year={2024}
}

@article{tmqm2020,
  title={The tmQM Dataset—Quantum Geometries and Properties of 86k Transition Metal Complexes},
  author={Balcells, David and Skjelstad, Bastian B.},
  journal={Journal of Chemical Information and Modeling},
  volume={60}, number={12}, pages={6135--6146}, year={2020}, doi={10.1021/acs.jcim.0c01041}
}

Compute

Generating this dataset is a CPU-only quantum-chemistry workload — no GPU is used or needed.

Stage	Cost	Hardware
Reference labels (2,100 × PBE0 + G0W0)	≈ 2,400 CPU-hours (≈ 59 h wall, 5 processes × 8 threads each)
Label-quality pass (re-converging 133 unconverged SCFs; spin-stability check of every label)	≈ 1,000 CPU-hours
Total	≈ 3,400 CPU-hours	1 workstation: AMD Ryzen Threadripper PRO 9965WX (24 cores / 48 threads), 251 GB RAM

Contact

Emerson Hsieh (Primateria). Questions and issues: open a discussion on this dataset's Hugging Face page.