MACE & EGRET Models

MACE (Multi-ACE) is a family of equivariant message-passing neural network potentials that achieves excellent accuracy through higher-order many-body interactions. EGRET is a related graph-based equivariant model with competitive accuracy for drug-like molecules. The MAPLE hand-wrapped MACE and EGRET backends are molecular no-PBC calculators.

MACE-OFF23 Family

The MACE-OFF23 models are trained on the SPICE dataset for organic molecules containing H, C, N, O, F, S, Cl. Three sizes are available with increasing accuracy and cost:

Model Keyword Parameters Best For
MACE-OFF23 (Small) maceoff23s ~0.5M Fast screening, large systems, speed-critical workflows
MACE-OFF23 (Medium) maceoff23m ~2M Balanced speed/accuracy for production calculations
MACE-OFF23 (Large) maceoff23l ~8M Publication-quality results, final geometries

Important

maceoff23l requires significantly more GPU memory than the smaller variants. If you encounter out-of-memory errors, switch to maceoff23m or maceoff23s, or use #device=cpu.

MACE-OMol

MACE-OMol extends element coverage beyond OFF23, supporting 17 elements including alkali and alkaline earth metals:

Elements: H, Li, B, C, N, O, F, Na, Mg, Si, P, S, Cl, K, Ca, Br, I

#model=maceomol
#device=gpu0
#opt(method=lbfgs)

0 1
Li   0.000000    0.000000    0.000000
F    1.570000    0.000000    0.000000

EGRET

EGRET is a graph-based equivariant model with competitive accuracy for drug-like molecules. It covers the same elements as MACE-OFF23 (H, C, N, O, F, S, Cl).

#model=egret
#device=gpu0
#opt(method=lbfgs)

XYZ /path/to/drug_molecule.xyz

Model Files

MAPLE can auto-download maceoff23m and egret checkpoints from the MAPLE model repository. maceoff23s, maceoff23l, maceomol, and the MACE-Polar development backends require a local checkpoint in maple/function/calculator/model/ or an explicit model_path.

Calculator Capabilities

  • Hessian: analytic and numerical Hessians are available in gas phase. Implicit-solvent Hessians (analytic or numerical) are not supported.
  • Implicit solvent: MAPLE's experimental GB-polar/QEq correction is energy-only (#sp with experimental=true); solvent forces and Hessians are not supported.
  • PBC and stress: the MAPLE MACE wrappers build no-PBC molecular graphs and do not provide validated stress or virial output.
  • HVP: MACE and EGRET do not implement get_hvp().
  • MACE-Polar: development keywords macepols, macepolm, and macepoll support charge/multiplicity through spin = mult - 1, but remain no-PBC and local-checkpoint backends.

Usage Examples

Quick screening with MACE-OFF23 Small

#model=maceoff23s
#device=gpu0
#opt(level=loose)

XYZ /path/to/molecule.xyz

Publication-quality optimization with MACE-OFF23 Large

#model=maceoff23l
#device=gpu0
#opt(method=lbfgs, level=tight)

XYZ /path/to/optimized_guess.xyz

Frequency calculation

#model=maceoff23m
#device=gpu0
#freq

XYZ /path/to/optimized_geometry.xyz

CPU fallback for large systems

#model=maceoff23s
#device=cpu
#opt(level=loose)

XYZ /path/to/large_system.xyz

Limitations

  • No charge/multiplicity support for MACE-OFF, MACE-OMol, or EGRET. Use AIMNet2, UMA omol, or a validated MACE-Polar checkpoint for charged/open-shell molecular inputs.
  • No D4 dispersion. D4 correction is not available for MACE or EGRET models.
  • No periodic systems. These wrappers do not support periodic boundary conditions; periodic atoms fail fast.
  • Limited elements. MACE-OFF23 and EGRET cover 7 elements; MACE-OMol covers 17 elements. For broader coverage, use uma.