Predict transition-metal complex HOMO-LUMO gaps from the open tmQM dataset on a fixed split

Transition-metal chemistry is a known blind spot for ML property models because of open-shell/d-electron and spin-state complexity, and most large QM ML datasets (QM9, ANI-1x, GDB) contain no transition metals. The tmQM dataset provides DFT (TPSSh-D3BJ/def2 with a Cartesian-basis single point) geometries and properties for 86665 mononuclear transition-metal complexes spanning the 3d/4d/5d series, extracted from the Cambridge Structural Database, including the HOMO-LUMO gap, dipole moment, polarizability, and metal-centered partial charge. Using a fixed, disclosed train/validation/test split (for example an 80/10/10 random split with a stated seed, or the split from the tmQMg graph release), train a model to predict the HOMO-LUMO gap (in eV) from structure and report the test-set MAE. Can an agent-built graph model beat the published baseline error for the gap, and does accuracy differ systematically across the 3d vs 4d vs 5d rows?

Acceptance. FULLY RESOLVES: a model predicting the tmQM HOMO-LUMO gap with test-set MAE that improves on the published baseline (target order of magnitude $\le 0.3$ eV with a modern GNN), evaluated on a clearly named fixed split (exact indices or seed+recipe), reported with architecture, hyperparameters, the per-row (3d/4d/5d) MAE breakdown, and a runnable training+evaluation script and weights. PARTIAL: any reproducible tmQM gap-prediction result on a disclosed split with its MAE and error analysis, or a documented negative result showing a named architecture fails to transfer across the metal rows. Metric: MAE of the HOMO-LUMO gap in eV over the held-out test set; the split must be stated exactly for comparability.