FAQ & troubleshooting

Performance

Why is my first calculation so slow?

This is normal. The first calculation on a given model downloads its checkpoint (up to ~2 GB across all sizes), which takes roughly 30 seconds. The model is cached afterward, so every subsequent calculation is fast.

I’m getting a CUDA out-of-memory error

Switch Device to CPU in the parameter panel, or choose a smaller model size (small instead of medium or large). Both reduce memory pressure at some cost to speed or accuracy.

Errors

I got an element error with MACE-OFF

MACE-OFF only supports ten organic elements: H, C, N, O, F, P, S, Cl, Br, and I. If your structure contains a metal or any other inorganic element, switch to MACE-MP-0, which covers 89 elements.

A shared link shows “not found”

The result ID in the link may be invalid or mistyped. Shared results are permanent once created and are never deleted, so a valid link should always resolve — double-check the URL was copied in full.

Choosing a model

Two questions settle almost every choice:

1. Is your system purely organic (only H, C, N, O, F, P, S, Cl, Br, I)? If yes, use MACE-OFF. If it contains a metal, a mineral, or any other element, use MACE-MP-0.
2. Do you need van der Waals / dispersion corrections (layered materials, molecular crystals, adsorption)? With MACE-MP-0, enable D3 dispersion. With MACE-OFF, do nothing — dispersion is already included in training.

Within either family, small is fastest and best for quick screening or large systems; medium is the balanced default; large is the most accurate and best suited to production or publication-quality results.

Cost & accounts

Is SimpleAtom free? Do I need an account?

Yes, it’s free, and no account is required. You can run calculations and generate shareable MACE Links entirely anonymously — there is no sign-up or login anywhere in the workflow.

Accuracy

How accurate are the results?

MACE foundation models approach DFT-level accuracy for the chemistries they were trained on. MACE-MP-0 is trained at the PBE+U level and typically overbinds by roughly 0.1–0.5 eV/atom relative to experiment. MACE-OFF is trained on ωB97M-D3BJ data, close to coupled-cluster quality for organic chemistry. Both are general-purpose foundation models, not a replacement for system-specific validation when your application demands the highest confidence — see Validation & reproducibility.