primateria/mbgfnet-closed-shell-4d5d-gw-dataset

Closed-shell 4d/5d transition-metal complexes: PBE0 + G0W0 quasiparticle dataset

2,210 closed-shell mononuclear 4d and 5d transition-metal complexes (17 metals: Y, Zr, Nb, Mo, Ru, Rh, Pd, Ag, Cd from the 4d row; Hf, Ta, W, Re, Os, Ir, Pt, Au, Hg from the 5d row), each with a DFT (PBE0/cc-pVDZ, relativistic small-core pseudopotential on the metal) and a one-shot G0W0@PBE0 quasiparticle-energy calculation. Built to extend MBGF-Net (Venturella, Li, Hillenbrand, Zhu, arXiv:2407.20384) further down the periodic table than either of the companion 3d releases (closed-shell, open-shell) — the first data toward whether a GW-surrogate GNN generalizes past 3d chemistry into heavier, relativistic transition metals.

2,208 of the 2,210 rows carry valid labels; 2 are excluded (see Label quality control below). A subset of the valid labels (334 rows, qpe_ambiguous_frontier = True) required extra care in the quasiparticle-equation solve and are flagged so they can be filtered or weighted differently — see below before treating this as a uniformly clean 2,208-row set.

What's in it

• omol_4d5d_closed_shell_2210.csv — one row per structure: identity, label_status, label_method, metal, row (4d/5d), atom count, OMol25 source path, and ground-truth PBE0 and G0W0 HOMO/LUMO/gap (eV), mean-field total energy, a qpe_ambiguous_frontier quality flag (see below), and wall time. data/train.parquet is the same table in Parquet (browsable via the Hugging Face dataset viewer, and consistent with the other two releases) — same content either way.
• xyz/<id>.xyz — the Cartesian geometry for every row, including the 2 excluded ones (kept for reproducibility, not deleted). Geometries are taken as-is from OMol25 (Meta FAIR, CC-BY-4.0) — no re-optimization performed.
• No fixed train/test split is defined for this release. Unlike the two 3d datasets, this is a first-pass validation campaign (scan feasibility → validate a 4d/5d-appropriate level of theory on test cases → label at scale), not yet a benchmark with designed transferability tests. The raw table has everything (metal, row, atom count) needed to build one for a specific question.

Level of theory

RKS-PBE0/cc-pVDZ ground state, with cc-pVDZ-PP (small-core relativistic effective core potential) basis and ECP on the metal instead of plain cc-pVDZ — required for 4d/5d elements, which cc-pVDZ alone (used throughout the 3d releases) does not support correctly. Ligand atoms keep plain cc-pVDZ. One-shot G0W0@PBE0 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 3d releases, validated on two real OMol25 test structures (AgCl, HfF4) before this campaign was run at scale.

Label quality control

The quasiparticle-equation solver used for the 3d releases has a real failure mode, found and fixed during this campaign. Standard practice finds each orbital's quasiparticle energy by solving w = e_KS + Re[Sigma(w)] + dv with Newton's method started from the KS energy. When the Padé-continued self-energy has more than one nearby real-axis root — found at a frontier orbital in 334 of 2,210 structures (15.1%) here — unconstrained Newton can silently converge to a spurious root far from the true solution rather than failing. Fixed with a bounded, bracket-based solver (methodology/windowed_qpe.py): scans a physically-motivated ±4.5 eV window (sized from this dataset's own shift distribution — max legitimate shift 3.47 eV) for sign changes, and only accepts a root when exactly one candidate exists.

Consequences for this dataset:

• 52 structures had a frontier energy that materially changed (>10 meV, median 929 meV) under the fix; all 52 are relabeled here (label_method = relabeled_bracket_solver).
• 2 structures had no candidate root anywhere in the window at either frontier orbital — excluded (label_status = excluded_unresolvable_qpe). Both are Hg complexes from OMol25's scaled_separations_exp subset (deliberately non-equilibrium geometries), and both geometries look like diffuse solvation-shell clusters rather than tight coordination complexes.
• The other 282 structures flagged qpe_ambiguous_frontier = True had multiple candidate roots in the window, but the one closest to the KS energy matched what plain Newton already found — their energies are unchanged, but the underlying solve was genuinely ambiguous, so they carry the flag rather than being silently treated as equivalent to the other 1,874.
• Open question, not yet resolved: Ag complexes are disproportionately affected (23/407 = 5.7% needed relabeling vs 0.5–1.4% for Pd/Pt/Cd/Au). Mechanism not diagnosed — flagged for anyone using this dataset who wants to dig further.
• The fix does not guarantee a physically sane gap — 7 rows still have GW_gap_eV ≤ 0 after everything above. 3 of these (om4d5d0004_015272, om4d5d0035_036213, om4d5d0076_037179) are flagged qpe_ambiguous_frontier = True (multi-root, but the closest-to-KS root happens to match plain Newton). The other 4 are NOT flagged at all (om4d5d0000_010823, om4d5d0041_000023, om4d5d0061_022141, om4d5d0077_041442) — each has a single, unambiguous root in the search window that is nonetheless large enough to invert or nearly close the gap. Whether this reflects genuine strongly-correlated physics in these specific complexes (5 of the 7 are Ag) or a Padé-continuation artifact the window/uniqueness check can't detect is an open question. Not excluded here (kept label_status = ok, consistent with retaining rather than silently dropping edge cases elsewhere in this project) — filter on GW_gap_eV > 0 if you want a strictly sane subset.

Full writeup: primateria/fcdmft#3 (private repo; methodology/ in this release is the self-contained equivalent).

label_status values: ok (2,208), excluded_unresolvable_qpe (2). label_method values: standard_newton (2,158), relabeled_bracket_solver (50, i.e. the 52 changed minus the 2 that were also excluded).

Provenance and licensing

• Geometries are taken from OMol25 (Meta FAIR, CC-BY-4.0), filtered to: neutral charge, closed-shell (multiplicity 1), mononuclear, ≤ 35 atoms, exactly one 4d or 5d metal (La excluded — not covered by this scan) plus ligand atoms 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 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_4d5d_2026,
  title        = {Closed-shell 4d/5d 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-4d5d-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
Reference labels (2,210 × PBE0 + G0W0)	≈ 269 wall-hours summed (≈ 2,150 CPU-hours at 8 threads/job)	9-machine fleet: 1 workstation (AMD Ryzen Threadripper PRO 9965WX, 24c/48t) + 8 compute nodes (128 cores each), ~5.5 h wall clock at ~67 concurrent workers
QPE fix + relabel (post-processing only, reused stored self-energies, no new SCF/GW)	≈ 2 minutes for the full-population diagnostic scan; seconds per structure for the 52 relabeled	same fleet

Contact

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