Skip to content

ml_method_selection.md

Latest commit

 

History

History
46 lines (38 loc) · 2.09 KB

ml_method_selection.md

File metadata and controls

46 lines (38 loc) · 2.09 KB
Raw

MACE method-selection rules

Part of the v1.2 ML-potentials reference set. Companion files: ml_validation_contract.md, ml_failure_modes.md, ml_troubleshooting.md. Index: ml_potentials.md.

The framing: ML potentials are an accelerator on top of trusted methods, not a replacement. The cross-validation contract (ml_validation_contract.md) is what makes that statement honest.

When to reach for MACE

The MACE branch is for systems where GFN2-xTB is too slow, not for systems where GFN2-xTB is fine. Use these in order:

  1. System has < ~500 atoms → use GFN2-xTB. Faster than MACE on CPU, no model-weights download, no validation overhead, fully trusted.
  2. System has ~500–1000 atoms and the task is dynamics → use GFN2-xTB if the run is short (~10 ps), MACE if the run is long (>~50 ps). xTB scales O(N³) so the wall-clock crossover is duration- sensitive.
  3. System has > ~1000 atoms → use MACE. Routing:
    • All elements ∈ {H, C, N, O, P, S, F, Cl, Br, I} → MACE-OFF (the organics foundation model; outperforms GFN2-xTB on torsions and conformers).
    • Otherwise → MACE-MP-0 (the materials foundation model; covers 89 elements, MPTrj+sAlex training).
    • Override with --mace-system-class organic or materials if you want to force a specific model (e.g., a metal–organic framework where the auto-rule picks MACE-MP-0 but you'd rather use MACE-OFF on the organic ligands separately).
  4. System has > ~2000 atoms (40 GB GPU) or > ~1000 atoms (CPU) → you're past the practical MACE-medium ceiling. Drop to --mace-size small first; if that still OOMs, shrink the system or wait for v2.2's CHGNet/Orb integration.

How the bundled scripts wire this in

  • scripts/optimize.py --calculator mace [--mace-system-class ...] [--mace-size small|medium|large]
  • scripts/run_md.py --calculator mace [...]

See ml_validation_contract.md for the validation flags that run_md.py adds when --calculator mace is used.