Validation & reproducibility

Automated validation suite

Run the full suite from the backend:

python mace-api/validate_calculation.py --test

This runs five tests, all of which must pass:

You can also validate a single result directly:

# From a JSON string
python mace-api/validate_calculation.py '<result_json>'

# From a file
python mace-api/validate_calculation.py result.json

What the validator checks

• Energy bounds (model-aware): MACE-MP-0 between −20 and +100 eV/atom; MACE-OFF between −800 and +100 eV/atom.
• Force magnitude: max force < 50 eV/Å warns, < 200 eV/Å errors — forces above 50 eV/Å generally indicate overlapping atoms.
• Force conservation: net force should be approximately zero for an isolated molecule.
• Interatomic distances: minimum distance > 0.4 Å (overlap detection).
• Lattice: positive volume and a valid 3×3 matrix for periodic systems.
• Trajectory: no NaN/Inf values, bounded energy fluctuation across an MD run.
• Consistency: the number of symbols, positions, and forces all match.

The validator also flags parameter-level issues: D3 dispersion enabled together with MACE-OFF (double-counting), float32 precision requested alongside a phonon/Hessian calculation, extreme timestep or temperature values, and overly loose fmax thresholds.

Manual verification checklist

Beyond the automated suite, a quick manual sanity check catches most real-world mistakes:

1. Energy in a reasonable range for the model: MACE-MP-0 → −1 to −15 eV/atom; MACE-OFF → −100 to −600 eV/atom.
2. Forces < 10 eV/Å for a reasonable structure — forces above 50 eV/Å usually mean overlapping atoms.
3. RMS force decreases monotonically during optimization.
4. MD energy is conserved in NVE; it fluctuates around the target temperature in NVT.
5. Lattice vectors form a right-handed coordinate system.
6. Volume is positive for periodic systems.

Reproducibility & citing

Every result can be published as a permanent MACE Link — shared results are stored immutably, so a link always reproduces the exact same result it did when it was created. This makes MACE Links suitable for citation in a paper, lab notebook, or supplementary material.

If you use SimpleAtom or the underlying MACE models in published work, please cite the MACE architecture:

Batatia, I. et al. “MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields.” NeurIPS 2022.

Known limitations