Mace

MACE

What this library is for

MACE is a machine-learning interatomic potential framework based on equivariant message passing. It is used as an ASE calculator for energies, forces, stresses, descriptors, geometry optimization, and molecular dynamics, and it can also train or fine-tune MLIPs from reference energies and forces.

When to use this vs. alternatives

• Use MACE when the task needs a pretrained or trained MLIP for atomistic energies, forces, stresses, MD, relaxation, descriptors, or active-learning loops.
• For ordinary relaxation, MD, descriptor, or screening prompts, start with the documented pretrained/foundation-model calculator path (mace_mp or MACECalculator) before discussing training a new model.
• Use ASE for the orchestration layer: reading structures, attaching the MACE calculator, optimizers, MD, trajectories, and file I/O.
• Use pymatgen for Materials Project queries, phase diagrams, and structure analysis before converting to ASE for MACE inference.
• Use DFT codes when the user needs first-principles reference data; MACE predicts a learned approximation to the training level of theory.
• Do not replace MACE with generic PyTorch/sklearn regression unless the user is explicitly developing a new MLIP model. MACE already encodes equivariance, neighbor cutoffs, training losses, and ASE calculator integration.

Canonical workflow

For inference, pick the appropriate MACE calculator, attach it to an ASE Atoms, and use standard ASE optimizers or MD.

from ase import units
from ase.io import read, write
from ase.md.langevin import Langevin
from ase.md.velocitydistribution import MaxwellBoltzmannDistribution
from ase.optimize import BFGS
from mace.calculators import mace_mp

atoms = read("structure.cif")
calc = mace_mp(model="medium-mpa-0", device="cuda", default_dtype="float64")
atoms.calc = calc

opt = BFGS(atoms, trajectory="mace_relax.traj", logfile="mace_relax.log")
opt.run(fmax=0.03)
print("energy_eV", atoms.get_potential_energy())
print("max_force_eV_per_A", abs(atoms.get_forces()).max())
write("relaxed.xyz", atoms)

MaxwellBoltzmannDistribution(atoms, temperature_K=300)
dyn = Langevin(atoms, timestep=0.5 * units.fs, temperature_K=300, friction=0.001)
dyn.run(200)

For local trained checkpoints:

from mace.calculators import MACECalculator

atoms = read("input.xyz")
atoms.calc = MACECalculator(model_paths="MACE_model.model", device="cuda")
energy = atoms.get_potential_energy()
forces = atoms.get_forces()

For deeper examples, read:

• Foundation models: https://mace-docs.readthedocs.io/en/latest/guide/foundation_models.html
• ASE calculator: https://mace-docs.readthedocs.io/en/latest/guide/ase.html
• Training: https://mace-docs.readthedocs.io/en/latest/guide/training.html
• Troubleshooting: https://mace-docs.readthedocs.io/en/latest/guide/troubleshooting.html

Key conventions and gotchas

• MACE calculators are ASE calculators. Attach them with atoms.calc = calc before asking ASE for energies, forces, stresses, descriptors, optimizations, or MD.
• Make the ASE bridge explicit in examples: read or build Atoms, attach the calculator, run optimization/MD/inference, then inspect energy, forces, and stress or cell behavior when relevant.
• Units follow ASE conventions: energies in eV, forces in eV/Angstrom, and distances in Angstrom. Training data should use eV and eV/Angstrom.
• Foundation-model defaults change. In MACE >=0.3.10, mace_mp() defaults to a newer MPA model; specify model=... when reproducibility matters.
• Choose the pretrained model for the domain: MACE-MP/MPA/OMAT/MATPES for materials, MACE-OFF for organic force fields, MACE-MDP for dipoles and polarizabilities only, not energies/forces.
• Check element coverage and license before using a foundation model. Some models have more restrictive licenses or limited element sets.
• device="cuda" needs a compatible PyTorch/CUDA environment. Fall back to device="cpu" only after acknowledging the speed cost.
• default_dtype="float64" is the conservative accuracy choice; float32 is faster but should be validated for the user's system.
• Training expects energy/force keys in extended XYZ or explicit --energy_key/--forces_key. If the initial output reports zero energies or forces loaded, the data keys are wrong.
• E0 atomic reference energies matter. Use isolated atom energies from the same reference settings, recompute for fine-tuning, or use --E0s=average when appropriate; mismatched E0s are a common source of large errors.

Anti-patterns

• Do not call a MACE model directly on raw coordinate arrays for normal workflows. Use ASE Atoms so species, cell, PBC, units, and calculator properties are handled correctly.
• Do not run MD with a foundation model before checking whether the model covers the elements, chemistry, pressure/temperature regime, charge state, and boundary conditions.
• Do not use MACE-MDP for energies or forces; it is for dipole moments and polarizabilities.
• Do not fine-tune from a foundation model while reusing stale E0s, wrong data keys, or mixed spin-polarization/reference settings.
• Do not trust a trained MLIP because training loss decreased. Evaluate on a held-out test set and inspect energy/force/stress errors by configuration type.
• Do not mix structures from pymatgen/RDKit/ASE without validating atom order, cell, PBC, units, and chemical state before MACE inference.

Diagnostic checks

Before trusting outputs, the agent should:

• Log MACE package version, ASE version, model name/path, model family, device, dtype, cutoff/default settings, dispersion setting, element coverage, and license/provenance.
• Print formula, atom count, cell, PBC, and calculator class before inference.
• Confirm the structure cell and PBC are appropriate for the selected model.
• For relaxations, report final max force and save trajectory/log files.
• For MD, record timestep, thermostat/barostat, temperature, friction, ensemble, number of steps, and energy/temperature drift checks.
• For training, verify counts of loaded configurations, energies, forces, and stresses; inspect initial RMSE ranges; and save the full command/config.
• For fine-tuning, compare initial loss to expected ranges and confirm E0s and data keys match the new reference calculations.

Pointers to deeper material

• Documentation: https://mace-docs.readthedocs.io/
• Foundation models: https://mace-docs.readthedocs.io/en/latest/guide/foundation_models.html
• Training guide: https://mace-docs.readthedocs.io/en/latest/guide/training.html
• Troubleshooting: https://mace-docs.readthedocs.io/en/latest/guide/troubleshooting.html
• Source repository: https://github.com/ACEsuit/mace
• Paper: Batatia et al. (2022), "MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields". https://arxiv.org/abs/2206.07697