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

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

1,240 open-shell mononuclear 3d transition-metal complexes (Ti, V, Cr, Mn, Fe, Co, Ni; spin multiplicity 1-6, including broken-symmetry open-shell singlets — see note below), each with a spin-unrestricted DFT (UKS-PBE0/cc-pVDZ) and one-shot unrestricted G0W0@PBE0/cc-pVDZ (UGWAC) quasiparticle-energy calculation, built to extend MBGF-Net (Venturella, Li, Hillenbrand, Zhu, arXiv:2407.20384) — a graph neural network that predicts the GW self-energy from cheap DFT quantities — from closed-shell to open-shell transition-metal chemistry. Companion release to primateria/mbgfnet-closed-shell-3d-gw-dataset; see that dataset's card for the closed-shell half of this work.

All 1,240 complexes carry converged labels — no exclusions. A subset (53 complexes, concentrated in the highest spin states) could not be labelled by the standard method and required an extension to it; see Label generation and a known limitation below. Every complex that failed under the standard method was eventually converged by one route or the other, so unlike the closed-shell release, no excluded_* rows are needed here (label_status is ok throughout; label_method records which route each row took).

What's in it

• omol_dblock_full_1240.csv — one row per complex: identity, geometry_source (OMol25 or tmQM, see below), split assignment, label_status, label_method, metal, spin multiplicity, atom count, and ground-truth UKS and G0W0 properties per spin channel (alpha/beta): HOMO/LUMO/gap (eV), <S²> spin contamination. (Dipole, quadrupole, metal-3d population, and the quasiparticle Z-factor — present in the closed-shell CSV — are not yet included here; computing them for open-shell channels needs a SAIAO-basis adaptation not yet built. Planned as a follow-up addition, not required for this release.) 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. For the 1,049 OMol25-sourced rows, source is OMol25's own internal storage path (e.g. tm_react/MOBH15_Cr3_.../orca.tar.zst), carried through unchanged for traceability back to the original OMol25 entry — it reflects OMol25's own data-generation pipeline, not this dataset's; every quantum-chemistry calculation in this dataset (UKS-PBE0 and G0W0/UGWAC) was done with PySCF and fcdmft only, per Level of theory below. For the 191 pilot191 rows, source is the tmQM CSD refcode.
• xyz/<id>.xyz — the Cartesian geometry for every row, both geometry sources together.
  • 1,049 rows (geometry_source = OMol25): taken as-is from OMol25 (Meta FAIR, CC-BY-4.0), no re-optimization performed — this dataset only adds PBE0 and G0W0 labels on top of OMol25's own geometries.
  • 191 rows (geometry_source = tmQM, split = pilot191): the same experimental structures published in primateria/mbgfnet-closed-shell-3d-gw-dataset as closed_shell_release_label_status = excluded_open_shell_ground_state (189) / excluded_no_closed_shell_solution (2) — tmQM's own metadata marked them closed-shell singlets, but a broken-symmetry check found a lower-energy open-shell solution at the PBE0 level, so they carried no valid property values in that release. This dataset supplies the open-shell labels those rows lacked; the closed_shell_release_id / closed_shell_release_label_status columns (empty for the 1,049 OMol25 rows) record the cross-reference. Geometries are duplicated here (not just referenced) for a self-contained release — both copies are the same CC-BY-4.0 tmQM structures with identical provenance. Serves as an out-of-distribution test: experimental geometries vs. the OMol25-generated geometries the rest of the dataset (and any model trained on it) uses.
• splits.json — the fixed train/test assignment used for every result derived from this dataset: a 658-complex training pool, a 165-complex random held-out test, a 226-complex unseen-metal test (every Co and Ni complex, held out of training entirely — chosen because the closed-shell campaign's spin-stability audit found Co and Ni the most open-shell-prone metals), and the 191-complex pilot191 out-of-distribution test described above.

A note on mult = 1 rows: 84 of the 191 pilot191 rows have net spin multiplicity 1 (a singlet) yet are still unrestricted, spin-broken-symmetry calculations with <S²> well above the ideal singlet value of 0 (0.17-1.80 in this dataset) — a lower-energy antiferromagnetically-coupled solution than the closed-shell answer, not a mislabeled closed-shell complex. label_method/twoS and <S²> together identify these; filter on <S²> if you want strictly non-degenerate high-spin systems only.

Level of theory

UKS-PBE0/cc-pVDZ ground state, then one-shot unrestricted G0W0@PBE0/cc-pVDZ (fcdmft UGWAC) 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 — the same recipe as the closed-shell dataset, generalized to two independent spin channels.

Label generation and a known limitation

Standard unrestricted G0W0+Padé continuation requires one Fermi level shared between both spin channels (the code averages majority-spin HOMO and minority-spin LUMO). For strongly spin-polarized complexes — concentrated in the highest multiplicities and in reactivity/transition-state geometries — majority-spin HOMO can sit above minority-spin LUMO, so no such shared reference exists, and the standard method cannot produce a label at all (not a convergence failure — the calculation is never attempted). We traced this to the analytic-continuation step specifically (a Padé fit evaluated far from its sampled region becomes unreliable), not a limitation of G0W0 itself, and extended the method with an independent per-spin Fermi level and Padé fit for these cases. The exact code used is included in this release, methodology/ugw_ac_perspin.py — an additive extension to fcdmft (Zhu group, Yale) that calls fcdmft's existing, unmodified self-energy routines twice (once per spin's own Fermi level) rather than reimplementing them; see methodology/README.md for how it fits into fcdmft's module layout.

label_method records which route each row took:

• standard (996 rows): the shared-Fermi-level method.
• per_spin_ef (53 rows): the extended method, needed only when no shared Fermi level exists. Validated by (1) reproducing standard-method results to within 14 meV on complexes where both methods are applicable, and (2) confirming resulting quasiparticle shifts on previously-unlabelable complexes match the same magnitude as standard-method results for the same metal and spin state, not an inflated artifact.

(A further ~46 complexes initially failed under the standard method but converged on a fresh attempt without needing the extension at all — ordinary re-convergence noise, not a fix-dependent result. Only the 53 per_spin_ef rows required the new method.)

Splits

splits.json and the split column are the split used to produce the results reported for this dataset (designed transferability tests, mirroring the closed-shell release's structure), provided so anyone can reproduce those numbers or compare a new model against them on identical held-out complexes.

Split	Size	Rule
train_pool	658	remainder after holding out the tests below
test_random	165	random held-out, stratified by metal x multiplicity
test_unseen_metal	226	every Co and Ni complex (metals never in training)
pilot191	191	tmQM experimental geometries — out-of-distribution vs. OMol25-generated training geometries

Only the priority-1 (metal_complexes, ≤35 atoms) tier of the underlying OMol25 scan is labelled in this release. Two further tiers were identified but deliberately not labelled here: a 1,990-complex priority-2 tier (larger metal complexes and reactivity structures) and a 2,542-complex priority-3 electrolytes tier — reserved for a possible future extension, not part of this release.

Provenance and licensing

• Geometries are taken from OMol25 (Meta FAIR, CC-BY-4.0), filtered to the metal_complexes domain: neutral charge, open-shell (multiplicity ≥ 3), mononuclear, ≤ 35 atoms.
• All UKS and G0W0 quantities in this dataset were computed by us for this work using PySCF and fcdmft (Zhu group, Yale), including the per-spin-Fermi-level extension linked above.
• This dataset (everything except the underlying OMol25 geometries, which retain their own CC BY 4.0 licence from Meta FAIR) is released under CC BY 4.0. If you use it, please cite OMol25 and this dataset.

Citing

Please cite this dataset directly, the MBGF-Net paper this dataset extends, and OMol25:

@misc{mbgfnet_openshell_3d_2026,
  title        = {Open-shell 3d transition-metal complexes: PBE0 + unrestricted G0W0 quasiparticle dataset},
  author       = {Hsieh, Emerson},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/primateria/mbgfnet-open-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{omol25_2025,
  title={OMol25: The Open Molecules 2025 Dataset},
  author={{Meta FAIR Chemistry Team}},
  year={2025},
  note={\url{https://huggingface.co/facebook/OMol25}}
}

Compute

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

Stage	Cost	Hardware
UKS-PBE0 ground states (1,049 OMol25-sourced x, incl. re-attempts on initially-failing complexes)	≈ 6,100 CPU-hours	9-machine fleet: 1 workstation (AMD Ryzen Threadripper PRO 9965WX, 24c/48t) + 8 compute nodes (128 cores each), 8 threads/complex
One-shot UGWAC + Padé AC (1,049 OMol25-sourced x, standard + per-spin-ef)	≈ 2,100 CPU-hours	same fleet
Subtotal (OMol25-sourced rows)	≈ 8,300 CPU-hours

The 191 pilot191 (tmQM-sourced) rows were computed in an earlier phase of this project on different hardware; a directly comparable CPU-hour figure for them isn't available.

Contact

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