Train a transferable MLIP on SPICE and reach released-foundation-model force accuracy on the held-out test set

SPICE is an open dataset built to train general-purpose ML potentials for biomolecular simulation: it contains on the order of 1.1 million conformations of drug-like small molecules, dipeptides, solvated systems, and ion pairs, with energies and forces computed at the wB97M-D3(BJ)/def2-TZVPPD level, covering the elements H, C, N, O, F, Na, Mg, P, S, Cl, K, Ca, Br, I. Using SPICE's provided train/validation/test partition (or a clearly disclosed split), train a machine-learning interatomic potential and report the test-set energy MAE (in meV per atom) and per-atom force MAE (in meV/$\text{\AA}$). The openly released MACE-OFF potentials were trained on SPICE and define the current accuracy reference. Can an agent-built model match or beat the released foundation-model test force MAE on SPICE while training within a single-workstation compute budget, and how does force accuracy vary between the small-molecule and peptide subsets?

Acceptance. FULLY RESOLVES: an MLIP trained on the SPICE training partition achieving test force MAE at or below the released MACE-OFF SPICE test error (report the specific reference number and version used for comparison), together with energy MAE per atom, the exact split, model/hyperparameters, training compute, and a runnable training+evaluation script and weights. PARTIAL: any reproducible SPICE-trained model reporting test energy and force MAE with the split disclosed (even if above the reference), or a documented negative result showing a named architecture cannot match the reference within a stated compute budget. Metrics: energy MAE in meV/atom and force MAE in meV/$\text{\AA}$ on the held-out test set, with the small-molecule vs peptide breakdown.