Models & Backends

Included (pre-parameterized) models

A model is the combination of a functional form with a set of parameters. A number of pre-parameterized models can be selected in AMS.

Model	Primary Domain	Backend	Finetunable with ParAMS a
AIMNet2-B973c	(Charged) Organic Molecules, Organohalides	AIMNet2	No
AIMNet2-NSE	Open Shell Organic Systems	AIMNet2	No
AIMNet2-Pd	Organometallics Involving Palladium	AIMNet2	No
AIMNet2-Rxn	Reactive Organic Molecules, Transition States	AIMNet2	No
AIMNet2-wB97MD3	(Charged) Organic Molecules, Organohalides	AIMNet2	No
ANI-1ccx	Organic Molecules	TorchANI	No
ANI-1x	Organic Molecules	TorchANI	No
ANI-2x	Organic Molecules	TorchANI	No
eSEN-S-Con-OMol	Organic Molecules, Biomolecules, Metal Complexes, Electrolytes	FAIRChem	No
M3GNet-UP-2022	Materials	M3GNet	Yes
MACE-MP-0-Large	Materials	MACE	Yes
MACE-MP-0-Medium	Materials	MACE	Yes
MACE-MP-0-Small	Materials	MACE	Yes
MACE-MPA-0	Materials	MACE	Yes
NEP89	Materials and Condensed Phase	NEP	No
QET-PBE-2025	Materials, Charge Transfer	MatGL	Yes
QET-r2SCAN-2025	Materials, Charge Transfer	MatGL	Yes
TensorNet-PBE-M-2025	Materials	MatGL	Yes
TensorNet-r2SCAN-M-2025	Materials	MatGL	Yes
UMA-S-1.2-OC20	Adsorbates on Catalytic Surfaces	FAIRChem	No
UMA-S-1.2-OC22	Adsorbates on Catalytic Oxide Surfaces	FAIRChem	No
UMA-S-1.2-OC25	Solvents, Ions and Electrolytes with Catalytic Surfaces	FAIRChem	No
UMA-S-1.2-ODAC	CO2/H2O adsorbed in MOFs	FAIRChem	No
UMA-S-1.2-OMat	Inorganic Materials	FAIRChem	No
UMA-S-1.2-OMC	Organic Molecular Crystals	FAIRChem	No
UMA-S-1.2-OMol	Organic Molecules, Biomolecules, Metal Complexes, Electrolytes	FAIRChem	No

^a ) Most of these models can be finetuned using external tools. The table indicates whether finetuning is supported in ParAMS.

AIMNet2 Models

AIMNet2 (Atoms In Molecules Network) is a neural network potential designed for accurate predictions of molecular geometries and reaction energies for both neutral and gas-phase charged organic molecules. It incorporates explicit long-range electrostatics and dispersion contributions and uses charge equilibration within the message passing framework, enabling improved performance for systems where long-range interactions are important.

A series of pre-trained models are available in AMS: AIMNet2-B973c [1], AIMNet2-NSE [2], AIMNet2-Pd [3], AIMNet2-Rxn, AIMNet2-wB97MD3. For general purpose organic chemistry, the models AIMNet2-wB97MD3 and AIMNet2-B973c are available. The difference between them is that AIMNet2-wB97MD3 is trained to more expensive and accurate ωB97M-D3/def2-TZVPP DFT reference data, whereas AIMNet2-B973c uses only B97-3c reference data. For open-shell systems (e.g. radicals, triplet states) the model AIMNet2-NSE gives more accurate results, as it also takes the spin state into consideration. For reactive organic chemistry, where bonds are breaking and forming or when evaluating transition states, the model AIMNet2-Rxn is most appropriate. Finally, for organometallic systems involving Palladium, the AIMNet2-Pd model is available. Note that this model also includes implicit THF solvation.

For all AIMNet2 models, an ensemble of models can also be selected with 2-4 members. These members are independently trained models which can be used together during calculations to get estimates of the uncertainty of calculated energies and forces.

Best for

• Fast calculations of small, drug-like molecules to routine DFT level accuracy

• Predictions of atomic charges and dipole moments

• Uncertainty estimations of energies and forces

Limitations

• Restricted to 14 elements (H, B, C, N, O, F, Si, P, S, Cl, As/Pd, Se, Br, I) or 4 elements (H, C, N, O) for AIMNet2-Rxn

Training data

Models	“AIMNet2-B973c, AIMNet-NSE, AIMNet2-wB97MD3”
Dataset	20 million conformers, including charged species, distilled from an initial pool of 120 million
Reference methods	B97-3c (AIMNet2-B973c), ωB97M-D3/def2-TZVPP (AIMNet2-wB97MD3/AIMNet-NSE)
Included elements	H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I

Model	AIMNet2-Pd
Dataset	20 million conformers, including charged species, distilled from an initial pool of 120 million
Reference methods	ωB97M-D3/def2-TZVPP with implicit CPCM solvation for THF
Included elements	H, B, C, N, O, F, Si, P, S, Cl, Pd, Se, Br, I

Model	AIMNet2-Rxn
Dataset	4.7 million reaction relevant conformers including equilibrium, near-equilibrium, minimum energy pathways, off-pathways, and transition states geometries
Reference methods	ωB97M-D3/def2-TZVPP
Included elements	H, C, N, O

Notes

• Predictions from all AIMNet2 models can be calculated using ensembles, meaning that the final prediction is an average over multiple independently trained neural networks.

Examples

An example AMS input file for the geometry optimization of chloromethane with AIMNet2-wB97MD3 is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" --delete-old-results << eor
Task GeometryOptimization

System
    Atoms
        C 1.1917 0.0037 0.0128 
        C 2.1539 1.0167 -0.0103 
        C -0.168 0.3408 0.0612 
        C -0.5446 1.6751 0.0858 
        C 1.7616 2.3374 0.0146 
        C 0.4166 2.6663 0.0625 
        C 1.604 -1.4116 -0.0198 
        O 0.7642 -2.2896 -0.0047 
        O 2.9122 -1.7313 -0.0664 
        O -1.1117 -0.6342 0.08359999999999999 
        C -2.4107 -0.3144999999999999 -0.0356 
        O -2.7326 0.8244 -0.2789 
        C -3.466699999999999 -1.3761 0.1341 
        H 3.2034 0.7649 -0.0473 
        H -1.591 1.9396 0.1227 
        H 2.506 3.119699999999999 -0.0029 
        H 0.1182 3.704 0.0818 
        H 3.1826 -2.6595 -0.0878 
        H -4.4532 -0.9316 0.003 
        H -3.3186 -2.1582 -0.6106 
        H -3.391999999999999 -1.8065 1.1328 
    End
End

Engine MLPotential
    Model AIMNet2-wB97MD3
EndEngine
eor

See also further examples.

ANI Models

ANI (Accurate NeurAl networK engINe for Molecular Energies) is a family of neural network potentials for efficient and accurate prediction of molecular geometries, vibrational frequencies, and reaction energies for gas-phase organic molecules.

Three pre-trained models are available in AMS: ANI-1x [4], ANI-1ccx [5], and ANI-2x [6]. ANI-1x and ANI-2x are trained to ωB97X/6-31G(d) DFT reference data, while ANI-1ccx is trained using transfer learning to target coupled-cluster quality reference energies CCSD(T)*/CBS.

Best for

• Very fast calculations of organic molecules (ANI-1x, ANI-1ccx), also including light halogens and sulfur (ANI-2x)

Limitations

• Restricted to a small set of light elements

• Does not support charged systems

Training data

Dataset	5.5 million conformers of small organic molecules (ANI-1x), supplemented with additional conformers containing light elements to make 8.9 million in total (ANI-2x)
Reference methods	ωB97X/6-31G(d) (ANI-1x, ANI-2x), DLPNO-CCSD(T)/CBS (ANI-1ccx)
Included elements	H, C, N, O (ANI-1x, ANI-1ccx), F, S, Cl (ANI-2x)

Notes

• Predictions from ANI models are calculated from committees (ensembles), meaning that the final prediction is an average over multiple independently trained neural networks.

Examples

An example AMS input file for the geometry optimization of methane with ANI-1ccx is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" --delete-old-results << eor
Task GeometryOptimization

System
    Atoms
        C 2.9166023165223268e-09 6.006194625417838e-09 7.204071579427031e-10
        H 0.5389120953335456 0.7623581289443943 -0.5992945752094279
        H 0.7312440919932264 -0.5966159238699776 0.5831823400796503
        H -0.5671285553273658 -0.6703024318238477 -0.6781076404295054
        H -0.7030276349160114 0.5045602207432393 0.6942198748388725
    End
End

Engine MLPotential
    Model ANI-1ccx
EndEngine
eor

eSEN Models

eSEN (Equivariant Smooth Energy Network) is a neural network potential developed for highly accurate computation of energies and forces for molecules with diverse chemistries. This includes the modeling of charged and open-shell systems.

In AMS, the pre-trained eSEN-S-Con-OMol [7] (con=conserving) model is provided, which is trained on the OMol25 dataset, designed to cover a very broad range of organic and bio-relevant molecules, including metal complexes and electrolytes.

Best for

• Highly accurate calculations of diverse organic and bio-relevant molecules

• Molecular systems containing main-group elements and many heavier elements up to Bi

• Inclusion of system charge and spin-multiplicity

Limitations

• Not intended for calculations on periodic inorganic bulk materials

Training data

Dataset	OMol25 dataset comprising over 100 million structures covering small molecules, biomolecules, metal complexes, and electrolytes
Reference methods	wB97M-V/def2-TZVPD, including non-local dispersion
Included elements	Elements from H .. Bi

Notes

• All training data is aperiodic, so any periodic systems should be treated with some caution

• eSEN models are made accessible for commercial and non-commercial use under the permissive FAIRChem license, which applies when using these models

Examples

An example AMS input file for the geometry optimization of methane with eSEN-S-Con-OMol is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" --delete-old-results << eor
Task GeometryOptimization

System
    Atoms
        C 2.9166023165223268e-09 6.006194625417838e-09 7.204071579427031e-10
        H 0.5389120953335456 0.7623581289443943 -0.5992945752094279
        H 0.7312440919932264 -0.5966159238699776 0.5831823400796503
        H -0.5671285553273658 -0.6703024318238477 -0.6781076404295054
        H -0.7030276349160114 0.5045602207432393 0.6942198748388725
    End
End

Engine MLPotential
    Model eSEN-S-Con-OMol
EndEngine
eor

M3GNet Models

M3GNet (Materials based on Graph Neural Networks with three-body interactions) is an interatomic potential designed for atomistic simulations of periodic materials.

The M3GNet-UP-2022 [8] model available in AMS is intended to be “universal”, i.e., applicable to a broad range of materials containing elements from across the periodic table, although the training data is primarily made up of crystal data from inorganic materials from the Materials Project [12].

Best for

• Fast calculations of inorganic crystalline materials

• Periodic systems (bulk solids, surfaces, interfaces)

Limitations

• Not designed for accurately modeling small organic molecules or biomolecules

Training data

Dataset	187k structures from 63k materials from the Materials Project
Reference methods	PBE, PBE+U
Included elements	Elements from H .. Pu (except Po .. Ra)

Notes

• M3GNet-UP-2022 can be fine-tuned with ParAMS

Examples

An example AMS input file for the geometry optimization of NaCl with M3GNet-UP-2022 is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" --delete-old-results << eor
Task GeometryOptimization

System
    Atoms
        Na 0.0 0.0 0.0 
        Cl 2.815 2.815 2.815 
    End
    Lattice
        0.0 2.815 2.815
        2.815 0.0 2.815
        2.815 2.815 0.0
    End
End

Engine MLPotential
    Model M3GNet-UP-2022
EndEngine
eor

See also further examples.

MACE Models

MACE (Message Passing Atomic Cluster Expansion) [13] is a family of equivariant neural network interatomic potentials designed for accurate prediction of energies and forces in atomistic simulations.

MACE-MP-0 and MACE-MPA-0 models [14] available in AMS are pre-trained foundation potentials targeting inorganic materials chemistry and are intended for periodic systems. For MACE-MP-0, multiple model sizes are provided (Small, Medium, Large), offering a trade-off between computational cost and accuracy. MACE-MPA-0 is equivalent in size to “Medium”, but is trained on a larger dataset with additional crystal structures for improved accuracy.

Best for

• Accurate periodic calculations of inorganic materials

• Tuning speed/accuracy via Small/Medium/Large variants

Limitations

• Not designed for accurately modeling small organic molecules or biomolecules

Training data

Dataset	Materials Project MPtraj dataset comprising 1.58 million structures from 146k materials (MACE-MP-0), supplemented with the sAlex dataset comprising a further 10.4 million structures from 3.23 million materials (MACE-MPA)
Reference methods	PBE+U
Included elements	Elements from H .. Pu (except Po .. Ra)

Notes

• MACE-MP-0 and MACE-MPA-0 can be fine-tuned with ParAMS

Examples

An example AMS input file for the geometry optimization of methane with MACE-MPA-0 is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" --delete-old-results << eor
Task GeometryOptimization

System
    Atoms
        Na 0.0 0.0 0.0 
        Cl 2.815 2.815 2.815 
    End
    Lattice
        0.0 2.815 2.815
        2.815 0.0 2.815
        2.815 2.815 0.0
    End
End

Engine MLPotential
    Model MACE-MPA-0
EndEngine
eor

See also further examples.

NEP89

NEP89 is a general-purpose Neuroevolution Potential (NEP) for inorganic and organic materials across 89 elements [18]. It combines a compact neural network with local atom-centered descriptors to enable fast, large-scale atomistic simulations.

Best for

• Fast molecular dynamics of large, 3D-periodic inorganic and organic systems

• Chemically diverse systems such as alloys, solid-state electrolytes, water, solvated biomolecular systems, and reactive hydrocarbons

• Simulations spanning many elements or compositions with a single model

Limitations

• Only overall neutral systems are supported (no net system charge).

• The supported atomic numbers are 1–83 (H through Bi) and 89–94 (Ac through Pu).

• As for any general-purpose potential, validate the model for the target system and properties. The publication recommends fine-tuning when the out-of-the-box accuracy is insufficient.

Training data

Dataset	Descriptor-space-selected and iteratively refined subsets of OMat24, MPtrj, SPICE, ANI-1xnr, SSE-ABACUS, SSE-VASP, Protein, UNEP-v1, CH, and water datasets, supplemented with a reactive CHONPS dataset
Reference methods	Mixed quantum-mechanical levels, primarily DFT. D3 dispersion was added to datasets that did not already include it; the water data uses MB-pol, which has near-CCSD(T) accuracy.
Training targets	Energies, forces, and stresses. MPtrj energies were excluded because of an incompatible treatment of DFT+U.
Included elements	Atomic numbers 1–83 and 89–94 (H through Bi and Ac through Pu)

Notes

• NEP89 provides energies, gradients, and stress tensors. Dispersion is represented implicitly by the model, so no separate D3 correction is needed.

• No Python machine-learning environment or additional installation is required, and the native worker has minimal startup overhead.

• By default, AMS uses the first compatible visible CUDA device and falls back to CPU/OpenMP when CUDA is unavailable.

• The included parameters are read from $AMSRESOURCES/MLPotential/NEP/2026_Ting_89/nep89_20250409.txt. $AMSRESOURCES is typically $AMSHOME/atomicdata.

• You can also use Backend NEP with Model Custom and a ParameterFile (which should end in .txt for NEP models).

• Download other NEP models for particular systems from https://gitlab.com/brucefan1983/nep-data

To check the installed NEP worker and the automatically selected device, use the ML backend diagnostic tool.

Examples

An example AMS input file for a single-point calculation on an organic molecule with NEP89 is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" <<EOF
Task SinglePoint

Properties
    Gradients True
    StressTensor Yes
End

System
    Atoms
        C 1.1705 0.2287 -0.3792
        C 0.0000 -0.6414 0.0001
        C -1.1706 0.2288 0.3792
        O -2.2331 0.0875 -0.1770
        O 2.2330 0.0876 0.1771
        H 1.0604 0.9753 -1.1517
        H -0.2743 -1.2708 -0.8464
        H 0.2743 -1.2705 0.8469
        H -1.0592 0.9787 1.1484
    End
    Lattice
       7.0000 0.0000 0.0000
       0.0000 7.0000 0.0000
       0.0000 0.0000 7.0000
    End
End

Engine MLPotential
    Model NEP89
EndEngine

EOF

QET Models

The QET-PBE-2025 and QET-r2SCAN-2025 models are pre-trained foundation potentials based on the charge-equilibrated TensorNet (QET) [11] architecture, which extends TensorNet with an analytically solvable charge-equilibration scheme. In addition to energies, forces, and stresses, they predict atomic partial charges, making them well suited to systems where electrostatics and charge transfer are important.

They are trained on MatPES, a dataset developed by the Materialyze Lab and the Materials Project to provide broad elemental coverage together with more extensive sampling of non-equilibrium structures than datasets derived mainly from relaxed DFT calculations. MatPES contains single-point energies, forces, and stresses for structures sampled from molecular dynamics using stringent DFT convergence settings.

Two model variants are provided, differing in the reference functional used for the training data. The QET-PBE-2025 model is trained on PBE data (without PBE+U). The QET-r2SCAN-2025 model is trained on r2SCAN data.

Best for

• Systems where electrostatics and charge transfer are important, such as ionic materials, electrolytes, and interfaces

• Molecular dynamics and geometry optimization of inorganic materials

• Simulations where atomic partial charges are of interest

Limitations

• Not designed for accurately modeling small organic molecules or biomolecules

• Coverage is limited for noble gases, most rare earth elements, and some radioactive elements

Training data

Dataset	MatPES, with single-point energies, forces, and stresses computed with stringent DFT convergence settings for structures sampled from molecular dynamics
Reference methods	PBE, r2SCAN
Included elements	Elements from H .. Pu (except Po .. Ra)

Notes

• The QET models compute atomic charges, which are available in the AMS output.

Examples

An example AMS input file for the geometry optimization of NaCl with QET-PBE-2025 is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" --delete-old-results << eor
Task GeometryOptimization

Properties
    Charges Yes
End

System
    Atoms
        Na 0.0 0.0 0.0 
        Cl 2.815 2.815 2.815 
    End
    Lattice
        0.0 2.815 2.815
        2.815 0.0 2.815
        2.815 2.815 0.0
    End
End

Engine MLPotential
    Model QET-PBE-2025
EndEngine
eor

TensorNet Models

The TensorNet-PBE-M-2025 and TensorNet-r2SCAN-M-2025 models [9] are pre-trained foundation potentials based on the TensorNet [10] architecture, an O(3)-equivariant graph neural network using Cartesian tensor representations; the M suffix denotes the medium-sized variant of the architecture.

They are trained on MatPES, a dataset developed by the Materialyze Lab and the Materials Project to provide broad elemental coverage together with more extensive sampling of non-equilibrium structures than datasets derived mainly from relaxed DFT calculations. MatPES contains single-point energies, forces, and stresses for structures sampled from molecular dynamics using stringent DFT convergence settings.

Two model variants are provided, differing in the reference functional used for the training data. The TensorNet-PBE-M-2025 model is trained on PBE data (without PBE+U). The TensorNet-r2SCAN-M-2025 model is trained on r2SCAN data.

Best for

• Molecular dynamics and geometry optimization of inorganic materials

• Periodic systems such as bulk solids, surfaces, and interfaces across a broad range of chemistries

• Systems with mixed ionic and van der Waals bonding, where the r2SCAN variant can be advantageous

Limitations

• Not designed for accurately modeling small organic molecules or biomolecules

• Coverage is limited for noble gases, most rare earth elements, and some radioactive elements

Training data

Dataset	MatPES, with single-point energies, forces, and stresses computed with stringent DFT convergence settings for structures sampled from molecular dynamics
Reference methods	PBE, r2SCAN
Included elements	Elements from H .. Pu (except Po .. Ra)

Examples

An example AMS input file for the geometry optimization of NaCl with TensorNet-PBE-M-2025 is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" --delete-old-results << eor
Task GeometryOptimization

System
    Atoms
        Na 0.0 0.0 0.0 
        Cl 2.815 2.815 2.815 
    End
    Lattice
        0.0 2.815 2.815
        2.815 0.0 2.815
        2.815 2.815 0.0
    End
End

Engine MLPotential
    Model TensorNet-PBE-M-2025
EndEngine
eor

UMA Models

UMA (Universal Model for Atoms) [15] is a foundation neural network potential with models trained on large-scale atomistic datasets spanning molecules, materials, surfaces, adsorption systems, and molecular crystals, which are intended to provide broad transferability across diverse chemistry.

Several pre-trained UMA variants are available in AMS. Each variant is specialized towards a particular training domain while retaining a shared underlying model architecture.

These include:

• UMA-S-1.2-OC20: adsorbates on catalytic surfaces (solid-gas interfaces)

• UMA-S-1.2-OC22: adsorbates on catalytic oxide surfaces (solid-gas interfaces)

• UMA-S-1.2-OC25: solvent layers, ions and electrolytes with catalytic surfaces (solid-liquid interfaces)

• UMA-S-1.2-ODAC: CO₂/H₂O adsorbed in MOFs

• UMA-S-1.2-OMat: inorganic materials

• UMA-S-1.2-OMC: organic molecular crystals

• UMA-S-1.2-OMol: organic molecules, biomolecules, metal complexes, electrolytes

Best for

• High accuracy calculations on a broad range of systems including molecules (OMol) and inorganic materials (OMat) with diverse chemistry

• Charged / open-shell molecules including radicals (OMol)

• Adsorption and surface chemistry with solid-gas interfaces (OC20), oxide materials (OC22) and electrocatalysts with solid-liquid interfaces (OC25)

• Porous framework adsorption (ODAC)

• Organic molecular crystals (OMC)

Limitations

• Predictions are best within the dominant chemistry represented in the chosen UMA variant

• Relatively computationally expensive compared to other, more targeted models

Training data

Model	UMA-S-1.2-OC20
Dataset	OC20 dataset comprising >100 million calculations of small molecules adsorbed on catalyst surfaces formed from materials in the Materials Project
Reference methods	RPBE, no dispersion
Included elements	Elements H .. Cl .. Bi (except other grp. 7, grp. 8 or Mg, Ba, Ln)

Model	UMA-S-1.2-OC22
Dataset	OC22 dataset comprising ~10 million calculations of small molecules adsorbed on catalyst surfaces formed from unary and binary oxide materials in the Materials Project
Reference methods	PBE/PBE+U, no dispersion
Included elements	Elements Li, Be, O, Na, Mg, Al, Si, K .. Ce, Lu .. Bi (except grp. 7, grp. 8, As, Tc, Te)

Model	UMA-S-1.2-OC25
Dataset	OC25 dataset comprising ~8 million calculations of solvents, ions and electrolytes with catalyst surfaces formed from materials in the Materials Project
Reference methods	RPBE+D3
Included elements	Elements H .. Pm, Hf .. Bi

Model	UMA-S-1.2-ODAC
Dataset	ODAC23 dataset comprising >10 million calculations of CO2/H2O molecules adsorbed in Metal Organic Frameworks sampled from various open databases like CoreMOF
Reference methods	PBE+D3
Included elements	Elements H, Li .. Np (except grp. 8, K, Rb, Tc, In, Pm, Yb, Ta, Os, Ir, Tl, Pb, Po .. Ac, Pa)

Model	UMA-S-1.2-OMat
Dataset	OMat24 dataset comprising >100 million calculations of inorganic materials collected from many open databases like Materials Project and Alexandria, and randomly sampled far from equilibria
Reference methods	PBE/PBE+U, no dispersion
Included elements	Elements H .. Pu (except Po .. Ra)

Model	UMA-S-1.2-OMC
Dataset	OMC25 dataset comprising ~25 million calculations of organic molecular crystals from random packing of OE62 structures into various 3D unit cells
Reference methods	PBE+D3
Included elements	H, B, C, N, O, F, Si, P, S, Cl, Br, I

Model	UMA-S-1.2-OMol
Dataset	OMol25 dataset comprising over 100 million structures covering small molecules, biomolecules, metal complexes, and electrolytes. Supplemented with OPoly26 dataset comprising ~6.35 million structures of various polymeric systems including traditional polymers, fluoropolymers, optical polymers, peptoids, lipids and polymer electrolytes.
Reference methods	wB97M-V/def2-TZVPD, including non-local dispersion
Included elements	Elements from H .. Bi

Notes

• A UMA variant should be selected based on the dominant system type

• Only UMA-S-1.2-OMol supports charged systems/spin multiplicity

• All UMA-S-1.2-OMol training data is aperiodic, so any periodic systems should be treated with some caution

• UMA models are made accessible for commercial and non-commercial use under the permissive FAIRChem license, which applies when using these models

• UMA models are much faster to run on a GPU than on the CPU.

Examples

An example AMS input file for the geometry optimization of methane with UMA-S-1.2-OMol is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" --delete-old-results << eor
Task GeometryOptimization

System
    Atoms
        C 2.9166023165223268e-09 6.006194625417838e-09 7.204071579427031e-10
        H 0.5389120953335456 0.7623581289443943 -0.5992945752094279
        H 0.7312440919932264 -0.5966159238699776 0.5831823400796503
        H -0.5671285553273658 -0.6703024318238477 -0.6781076404295054
        H -0.7030276349160114 0.5045602207432393 0.6942198748388725
    End
End

Engine MLPotential
    Model UMA-S-1.2-OMol
EndEngine
eor

An example AMS input file for the geometry optimization of sodium chloride with UMA-S-1.2-OMat is as follows:

#!/bin/sh

export NSCM=1

"$AMSBIN/ams" --delete-old-results << eor
Task GeometryOptimization

System
    Atoms
        Na 0.0 0.0 0.0 
        Cl 2.815 2.815 2.815 
    End
    Lattice
        0.0 2.815 2.815
        2.815 0.0 2.815
        2.815 2.815 0.0
    End
End

Engine MLPotential
    Model UMA-S-1.2-OMat
EndEngine
eor

See also further examples.

Model

Type:

Multiple Choice

Default value:

ANI-2x

Options:

[Custom, AIMNet2-B973c, AIMNet2-NSE, AIMNet2-Pd, AIMNet2-Rxn, AIMNet2-wB97MD3, ANI-1ccx, ANI-1x, ANI-2x, eSEN-S-Con-OMol, M3GNet-UP-2022, MACE-MP-0-Large, MACE-MP-0-Medium, MACE-MP-0-Small, MACE-MPA-0, NEP89, QET-PBE-2025, QET-r2SCAN-2025, TensorNet-PBE-M-2025, TensorNet-r2SCAN-M-2025, UMA-S-1.2-OC20, UMA-S-1.2-OC22, UMA-S-1.2-OC25, UMA-S-1.2-ODAC, UMA-S-1.2-OMat, UMA-S-1.2-OMC, UMA-S-1.2-OMol]

Description:

Select a pre-parameterized or custom model. AIMNet2 variants: best for fast calculations of small, drug-like molecules; limited to systems of 14 elements (H, B, C, N, O, F, Si, P, S, Cl, As/Pd, Se, Br, I). ANI-(1x/1ccx/2x): best for very fast calculations of organic molecules; limited to elements H, C, N, O (ANI-1x/1ccx), F, S, Cl (ANI-2x). eSEN-S-Con-OMol: best for highly accurate calculations of diverse organic and bio-relevant molecules; not intended for calculations on periodic inorganic bulk materials. M3GNet-UP-2022: best for fast calculations of inorganic crystalline materials; not designed for accurately modeling small organic molecules or biomolecules. MACE-MP-0-(Small/Medium/Large): best for accurate periodic calculations of inorganic materials; size trades speed/accuracy; not designed for accurately modeling small organic molecules or biomolecules. MACE-MPA-0 has improved accuracy vs MP-0. NEP89: Neuroevolution potential for neutral systems. UMA-S-1.2 variants: best for high accuracy calculations on a broad range of systems; choose from OC20 (adsorption and surface chemistry), OC22 (oxide catalysis), OC25 (electrocatalysis), ODAC (adsorption in porous frameworks), OMat (inorganic materials), OMC (organic molecular crystals), OMol (molecules, biomolecules, metal complexes, electrolytes); can be computationally expensive compared to other, more targeted models. Set Custom to choose a backend and provide your own parameters.

Custom models (custom parameters)

Note

You can use Engine ASE to use any ASE calculator as the engine.

Note

You can use ParAMS to train your own ML potential parameters.

Set Model to Custom and specify which backend to use with the Backend option. In a typical case, you would have used that backend to train your own machine learning potential.

The backend reads the parameters, and any other necessary information (for example neural network architecture), from either a file or a directory. Specify the ParameterFile or ParameterDir option accordingly, with a path to the file or directory. Read the backend’s documentation to find out which option is appropriate.

Example:

Engine MLPotential
    Backend MACE
    Model Custom
    ParameterFile mace-custom.model
EndEngine

Backend

Type:

Multiple Choice

Options:

[AIMNet2, FAIRChem, M3GNet, MACE, MatGL, NEP, NequIP, TorchANI]

Description:

The machine learning potential backend.

ParameterDir

Type:

String

Default value:

GUI name:

Parameter directory

Description:

Path to a set of parameters for the backend, if it expects to read from a directory.

ParameterFile

Type:

String

Default value:

Description:

Path to a set of parameters for the backend, if it expects to read from a file.

Backends

Table 1 Backends supported by the MLPotential engine.

	AIMNet2	FAIRChem	M3GNet	MACE	MatGL	NEP	NequIP	TorchANI
Reference	[1]	[15]	[8]	[13]	[9]	[17]	[16]	[19]
Parameters from	ParameterFile	ParameterFile	ParameterDir	ParameterFile	ParameterDir	ParameterFile	ParameterFile	ParameterFile
Included models	AIMNet2-B973c, AIMNet2-NSE, AIMNet2-Pd, AIMNet2-Rxn, AIMNet2-wB97MD3	eSEN-S-Con-OMol, UMA-S-1.2-OC20, UMA-S-1.2-OC22, UMA-S-1.2-OC25, UMA-S-1.2-ODAC, UMA-S-1.2-OMat, UMA-S-1.2-OMC, UMA-S-1.2-OMol	M3GNet-UP-2022	MACE-MP-0, MACE-MPA-0	TensorNet-PBE-M-2025, TensorNet-r2SCAN-M-2025, QET-PBE-2025, QET-r2SCAN-2025	NEP89	none	ANI-1x, ANI-2x, ANI-1ccx
ML framework	PyTorch 2.8.0	PyTorch 2.8.0	TensorFlow 2.9.1	PyTorch 2.8.0	PyTorch 2.8.0	Native CPU/OpenMP	PyTorch 2.8.0	PyTorch 2.8.0

Note

Starting with AMS2023, PiNN [21] is only supported as a custom Calculator through Engine ASE [20].

Starting with AMS2026, SchNetPack [22] and sGDML [23] are also only supported as a custom Calculator through Engine ASE [20].

Note

If you use a custom parameter file with TorchANI, the model specified via ParameterFile filename.pt is loaded with torch.load('filename.pt')['model'], such that a forward call should be accessible via torch.load('filename.pt')['model']((species, coordinates)). The energy shifter is not read from custom parameter files, so the absolute predicted energies will be shifted with respect to the reference data, but this does not affect relative energies (e.g., reaction energies).

References

[1] (1,2)

D. M. Anstine et al., Chem Sci. 16 (2025) 10228–10244. https://doi.org/10.1039/D4SC08572H

[2]

B. Kalita et al., Angew. Chem. Int. Ed. 65, e16763 (2026). https://doi.org/10.1002/anie.202516763

[3]

D. M. Anstine et al., ChemRxiv (2025). https://doi.org/10.26434/chemrxiv-2025-n36r6

[4]

J. S. Smith et al., J. Chem. Phys. 148 (2018) 241733. https://doi.org/10.1063/1.5023802

[5]

J. S. Smith et al., Nat. Commun. 10 (2019) 2903. https://doi.org/10.1038/s41467-019-10827-4

[6]

C. Devereux et al., J. Chem. Theory Comput. 16 (2020) 4192-4202. https://doi.org/10.1021/acs.jctc.0c00121

[7]

D. S. Levine et al., arXiv:2505.08762 (2025). https://arxiv.org/abs/2505.08762

[8] (1,2)

C. Chen, S. P. Ong. Nature Computational Science 2, 718–728 (2022). https://doi.org/10.48550/arXiv.2202.02450

[9] (1,2)

T. W. Ko, B. Deng, M. Nassar, L. Barroso-Luque, R. Liu, J. Qi, A. C. Thakur, A. R. Mishra, E. Liu, G. Ceder, S. Miret, and S. P. Ong. npj Computational Materials 11, 253 (2025). https://doi.org/10.1038/s41524-025-01742-y

[10]

G. Simeon, G. de Fabritiis. Advances in Neural Information Processing Systems 36 (2024). https://doi.org/10.48550/arXiv.2306.06482

[11]

T. W. Ko, R. Liu, A. R. Mishra, Z. Yu, J. Qi, S. P. Ong. arXiv:2511.07249 (2025). https://doi.org/10.48550/arXiv.2511.07249

[12]

A. Jain et al., APL Materials 1 (2013) 011002. https://doi.org/10.1063/1.4812323

[13] (1,2)

I. Batatia et al., Advances in Neural Information Processing Systems (2022). https://openreview.net/forum?id=YPpSngE-ZU

[14]

I. Batatia et al., arXiv:2401.00096 (2023). https://arxiv.org/abs/2401.00096

[15] (1,2)

B. M. Wood et al., arXiv:2506.23971 (2025). https://arxiv.org/abs/2506.23971

[16]

S. Batzner et al., Nat. Commun. 13 (2022) 2453. https://doi.org/10.1038/s41467-022-29939-5

[17]

Z. Fan et al., Phys. Rev. B 104 (2021) 104309. https://doi.org/10.1103/PhysRevB.104.104309

[18]

T. Liang et al., arXiv:2504.21286 (2025). https://arxiv.org/abs/2504.21286

[19]

X. Gao et al. J. Chem. Inf. Model (2020). https://doi.org/10.1021/acs.jcim.0c00451

[20] (1,2)

https://wiki.fysik.dtu.dk/ase/index.html

[21]

Y. Shao et al., J. Chem. Inf. Model. 60 (2020) 1184-1193. https://doi.org/10.1021/acs.jcim.9b00994

[22]

K. T. Schütt et al., J. Chem. Theory Comput. 15 (2019) 448-455. https://doi.org/10.1021/acs.jctc.8b00908

[23]

S. Chmiela et al. Comp. Phys. Commun. 240 (2019) 38-45. https://doi.org/10.1016/j.cpc.2019.02.007