MAPLE Logo MAPLE

Machine Learning Potentials

MAPLE uses machine learning interatomic potentials (MLIPs) to compute energies and forces instead of traditional quantum mechanical methods. This approach provides near-DFT accuracy at a fraction of the computational cost, enabling rapid geometry optimizations, frequency calculations, and molecular dynamics simulations. The #model command in the settings block selects which potential to use.

Available Models

The table below lists the primary user-facing machine learning potentials in MAPLE, along with their keyword, element coverage, and accuracy characteristics. For periodic, charge/spin, Hessian, solvent, D4, HVP, and development backends such as MACE-Polar, see the calculator backend capability matrix.

Model Name Keyword Supported Elements Accuracy
ANI-2x ani2x H, C, N, O, F, S, Cl Trained on wB97X/6-31G(d); good general-purpose accuracy for organic molecules. Details →
ANI-1x ani1x H, C, N, O Trained on wB97X/6-31G(d); limited to four elements but well-validated
ANI-1ccx ani1ccx H, C, N, O Transfer-learned to CCSD(T)/CBS; high accuracy for small organic molecules
ANI-1xnr ani1xnr H, C, N, O Non-reactive variant of ANI-1x; trained without bond-breaking data
AIMNet2 aimnet2 H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I Broad element coverage with wB97M-D3BJ accuracy; supports charged species. Details →
AIMNet2-NSE aimnet2nse H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I AIMNet2 variant with nuclear spin-orbit effects; useful for heavy-element systems
MACE-OFF23 (Small) maceoff23s H, C, N, O, F, S, Cl Fast equivariant model; suitable for quick screening of organic molecules. Details →
MACE-OFF23 (Medium) maceoff23m H, C, N, O, F, S, Cl Balanced speed and accuracy for organic chemistry applications
MACE-OFF23 (Large) maceoff23l H, C, N, O, F, S, Cl Highest accuracy MACE-OFF variant; recommended for publication-quality results
MACE-OMol maceomol H, Li, B, C, N, O, F, Na, Mg, Si, P, S, Cl, K, Ca, Br, I Extended organic-molecular MACE model; broader element support than OFF23
EGRET egret H, C, N, O, F, S, Cl Graph-based equivariant model; competitive accuracy for drug-like molecules
UMA uma Broad periodic table coverage (83+ elements) Universal model from Meta FAIR; supports 7 task domains, 3 model sizes, and inference=default|turbo. Details →
Your own model module=... Whatever your model supports Plug in your own TorchScript or PyTorch checkpoint (.pt/.jpt/.model) with no source edits. Tutorial →

Model Selection Guide

Choosing the right model depends on your molecular system, required accuracy, and available computational resources. Here are general recommendations:

  • Organic molecules (H, C, N, O only): Start with ani2x for general use or ani1ccx when coupled-cluster accuracy is needed for small molecules.
  • Organic molecules with halogens (F, S, Cl): Use ani2x, maceoff23m, or egret. For the highest accuracy, try maceoff23l.
  • Systems with heavier elements (Br, I, P, Si, etc.): Use aimnet2 or maceomol, which support a broader range of elements.
  • Diverse or unknown systems: Use uma as a reliable general-purpose potential with the widest element coverage. Specify a task domain for best accuracy.
  • Periodic materials / catalysis: Use uma with an explicit periodic-capable task such as omat, oc20, oc22, oc25, omc, or odac. Stress and virial output are not yet validated.
  • Charged or open-shell molecules: Use aimnet2 or uma with the omol task. Molecular wrappers such as ANI and MACE-OFF do not honor charge or multiplicity.
  • Speed-critical workflows: Use maceoff23s or ani2x for the fastest evaluation times. The ANI models are particularly fast on GPUs.
  • Publication-quality results: Use maceoff23l or ani1ccx for the highest available accuracy within their element ranges.

Physical Corrections

ML potentials can be augmented with physical corrections to improve accuracy for specific interactions that are challenging for neural network models. Support is backend-specific; do not assume a correction is available for every model.

DFT-D4 Dispersion Correction

The D4 dispersion correction adds London dispersion interactions (van der Waals forces) to the ML potential energy. In the current MAPLE calculator layer, D4 is supported for ANI backends only.

#model = ani2x
#d4
#opt

Experimental Implicit Solvation

MAPLE's current implicit solvent path is an experimental GB-polar/QEq correction for single-point energies. It is configured through the #solv command with method=gbsa, an implicit solvent name, and experimental=true. See the Solvation page for details.

#model = uma
#solv(method=gbsa, implicit=water, experimental=true)
#sp

QEq Charges

The charge equilibration (QEq) method computes atomic partial charges used internally by the experimental implicit solvent correction. QEq charges are computed automatically when implicit solvation is enabled and do not require a separate command.

Important

Large models such as maceoff23l and maceomol require significantly more GPU memory than smaller models. If you encounter out-of-memory errors, try switching to a smaller model variant (e.g., maceoff23s) or use #device = cpu for systems that exceed your GPU memory.

Usage Examples

Below are complete input file examples demonstrating model selection for various scenarios.

General-purpose optimization with UMA (default: small v1.1 checkpoint, omol task)

# Uses uma-s-1p1 checkpoint and default inference interface
#model=uma
#device=gpu0
#opt(method=lbfgs, level=tight)

0 1
C    0.000000    0.000000    0.000000
O    1.200000    0.000000    0.000000
H   -0.540000    0.940000    0.000000
H   -0.540000   -0.940000    0.000000

UMA for inorganic materials (omat task)

#model=uma(task=omat)
#device=gpu0
#opt(method=lbfgs)

XYZ /path/to/material.xyz

UMA with CUDA turbo inference

# Opt in to the FAIR Chemistry turbo inference interface for repeated GPU evaluations
#model=uma(inference=turbo)
#device=gpu0
#ts(method=neb)

XYZ /path/to/reactant.xyz

XYZ /path/to/product.xyz

Charged-molecule calculation with AIMNet2

#model = aimnet2
#device = gpu0
#opt(method=lbfgs, level=extratight)

-1 1
C    0.000000    0.000000    0.000000
O    1.250000    0.000000    0.000000
O   -1.250000    0.000000    0.000000

Frequency calculation with MACE-OFF23

#model = maceoff23m
#device = gpu0
#freq

XYZ /path/to/optimized_geometry.xyz

CPU-based calculation for large systems

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

XYZ /path/to/large_molecule.xyz