Models and Properties

  • Standard MACE, including MACE-MH-0, supports energy, free energy, per-atom energies, forces, and stress through the ordinary direct execution planner.

  • MACEField adds polarization and qualified analytical response properties.

  • MACE-MH-1 supports generated direct execution for admitted nonlinear contracts. Its matching graph-time execution-plan resolver is still pending.

Compact Symmetrix-XL JSON is the preferred interchange format. Loading an ordinary MACE checkpoint can require the optional mace-torch dependency. MACEField checkpoints should be exported to compact JSON because the field-aware upstream modules are not yet present in upstream MACE.

Convert a local ordinary MACE checkpoint after installing the optional converter dependency:

uv pip install mace-torch
symmetrix_extract_mace \
    --model mace-omat-0-medium.model \
    --output srtio3-mace.json

Omitting --chemical-symbols and --atomic-numbers produces the preferred universal export and retains the checkpoint’s complete element domain. Compact format v2 stores the shared radial model once with element-indexed parameters, and format v3 uses the same full-domain principle for nonlinear MACE models. Unlike the original Symmetrix pair-spline format (named v1 here), these formats do not add a separate radial spline table for every element pair, avoiding quadratic pair-table growth. A universal file still contains the checkpoint’s element-indexed learned parameters and can therefore be larger than a deliberately restricted v2 subset, but it can be reused across compositions. Use an explicit element subset only for an intentionally restricted format-v2 deployment or when producing the original Symmetrix format with --radial-format pair-splines; format-v3 models reject subsets.

The command retains all compatible prediction heads. It does not download a model; obtain a checkpoint from its model provider, then use the resulting JSON in the calculator. For a GPU LAMMPS run, add --prepare-jit-device-artifact during conversion on the deployment GPU. mace-torch and its PyTorch dependencies are optional and can be a large download; JSON-only evaluation does not require them.

One-interaction MACE models using the invariant correlation-2 nonlinear readout are supported with canonical equal-channel hidden irreps. This includes both the L_max=1 single-v2 layout and the earlier L_max=2 single-layer layout. Their compact JSON contains only the M0 and R0 execution contracts.

Multi-head JSON models retain every head. Select one with head= and inspect calculator.available_heads. SymmetrixEnsemble provides same-process, sequential committee evaluation; it returns mean values, member values with the _comm suffix, and population variance with the _var suffix.

An ordinary ASE calculation can request the standard properties directly:

atoms.calc = Symmetrix("srtio3-mace.json")  # FP32 by default
energy = atoms.get_potential_energy()
forces = atoms.get_forces()
stress = atoms.get_stress()

Use Symmetrix("srtio3-mace.json", dtype="float64") when higher numerical precision is required.

See Python API for the public calculator and ensemble API.