AIMNet2

Fast neural-network interatomic potentials for organic, elemental-organic, open-shell, reactive, and palladium chemistry.

• Documentation: https://isayevlab.github.io/aimnetcentral/
• Repository: https://github.com/isayevlab/aimnetcentral
• PyPI: https://pypi.org/project/aimnet/
• Demo: https://huggingface.co/spaces/isayevlab/aimnet2-demo

What It Does

AIMNet2 predicts energies, forces, atomic charges, stress tensors, and Hessians for molecular and periodic atomistic simulations. It combines AIMNet2 neural potentials with adaptive dense/sparse neighbor handling, optional long-range electrostatics, DFT-D3 dispersion, and integrations for simulation workflows.

Highlights

• Pretrained model families for general organic chemistry, B97-3c, open-shell chemistry, palladium catalysis, and reactive paths
• Core inference through AIMNet2Calculator
• ASE, PySisyphus, Sella, TorchSim, Hugging Face Hub, and CLI workflows
• Periodic systems with DSF, Ewald, and PME Coulomb methods
• DFT-D3(BJ) dispersion through GPU-accelerated nvalchemiops kernels
• Python 3.11, 3.12, and 3.13 support; TorchSim extra is available on Python 3.12+

Install

AIMNet2 requires Python 3.11+ and PyTorch 2.8+.

# CPU or PyTorch-default install
pip install aimnet

# CUDA example: install PyTorch first, then AIMNet2
pip install torch --index-url https://download.pytorch.org/whl/cu126
pip install aimnet

With uv:

uv pip install torch --index-url https://download.pytorch.org/whl/cu126
uv pip install aimnet

Optional integrations:

pip install "aimnet[ase]"              # ASE calculator
pip install "aimnet[pysis]"            # PySisyphus reaction paths
pip install "aimnet[sella]"            # Sella TS optimization, includes ASE
pip install "aimnet[hf]"               # Hugging Face Hub model loading
pip install "aimnet[torchsim,ase]"     # TorchSim, Python 3.12+
pip install "aimnet[train]"            # training and export commands

For a full development environment:

git clone https://github.com/isayevlab/aimnetcentral.git
cd aimnetcentral
make install
source .venv/bin/activate

Quick Start

from aimnet.calculators import AIMNet2Calculator

calc = AIMNet2Calculator("aimnet2")

result = calc(
    {"coord": coordinates, "numbers": atomic_numbers, "charge": 0.0},
    forces=True,
)

print(result["energy"])   # eV
print(result["forces"])   # eV/Angstrom
print(result["charges"])  # e

ASE

from ase.io import read
from aimnet.calculators import AIMNet2ASE

atoms = read("molecule.xyz")
atoms.calc = AIMNet2ASE("aimnet2", charge=0)

energy = atoms.get_potential_energy()
forces = atoms.get_forces()

TorchSim

import ase.io
import torch_sim as ts

from aimnet.calculators import AIMNet2Calculator, AIMNet2TorchSim

atoms = ase.io.read("molecule.xyz")
model = AIMNet2TorchSim(AIMNet2Calculator("aimnet2"))

results = ts.static(system=atoms, model=model)
print(results[0]["potential_energy"], results[0]["forces"])

Periodic Systems

calc = AIMNet2Calculator("aimnet2")
calc.set_lrcoulomb_method("ewald", ewald_accuracy=1e-6)

result = calc(
    {
        "coord": coordinates,
        "numbers": atomic_numbers,
        "charge": 0.0,
        "cell": cell_vectors,
        "pbc": True,
    },
    forces=True,
    stress=True,
)

Model Families

Model	Elements	Best for
aimnet2	H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I	General organic and elemental-organic chemistry, wB97M-D3
aimnet2-2025	H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I	B97-3c with improved intermolecular interactions
aimnet2-b973c	H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I	Legacy B97-3c reproducibility
aimnet2-nse	H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I	Open-shell systems and radicals
aimnet2-pd	H, B, C, N, O, F, Si, P, S, Cl, Se, Br, Pd, I	Pd catalysis with B97-3c/CPCM(THF) reference data
aimnet2-rxn	H, C, N, O	Reactive chemistry, transition states, NEB, IRC; net-neutral systems only

Each family has four ensemble members indexed 0 through 3. Short aliases load member 0 by default. Previously published aliases such as aimnet2_2025, aimnet2nse, and aimnet2pd continue to resolve.

See the model selection guide for limitations, aliases, citations, and download links.

Hugging Face Models

Install the HF extra and pass a repo ID directly:

pip install "aimnet[hf]"

from aimnet.calculators import AIMNet2Calculator

calc = AIMNet2Calculator("isayevlab/aimnet2-wb97m-d3")
calc = AIMNet2Calculator("isayevlab/aimnet2-wb97m-d3", ensemble_member=2)
calc = AIMNet2Calculator("isayevlab/aimnet2-wb97m-d3", revision="v1.0")

Published repositories:

HF repo	Alias	Use
isayevlab/aimnet2-wb97m-d3	aimnet2	General organic chemistry
isayevlab/aimnet2-2025	aimnet2-2025	Improved intermolecular interactions
isayevlab/aimnet2-nse	aimnet2-nse	Open-shell chemistry
isayevlab/aimnet2-pd	aimnet2-pd	Palladium chemistry
isayevlab/aimnet2-rxn	aimnet2-rxn	Reactive paths and transition states

Private repos can be loaded with token= or the HF_TOKEN environment variable. Local HF-style directories with config.json and ensemble_N.safetensors are also supported.

Outputs

Key	Shape	Units
energy	scalar or (B,)	eV
forces	(N, 3) or (B, N, 3)	eV/Angstrom
charges	(N,) or (B, N)	electron charge
stress	(3, 3)	eV/Angstrom^3
hessian	(N, 3, N, 3)	eV/Angstrom^2

Hessians are single-molecule only and are incompatible with compile_model=True. Long-range backends differ in derivative support; see the calculator and long-range docs for the exact contracts.

Training and CLI

pip install "aimnet[train]"
aimnet train --config my_config.yaml --model aimnet2.yaml

The aimnet entry point is installed with the core package. Training, export, and self-atomic-energy commands require the train extra.

Development

make check       # formatting, linting, dependency checks
make test        # pytest suite
make docs        # serve docs locally
make docs-test   # strict docs build

CI runs the core test suite across Python 3.11-3.13 and separate optional-extra lanes for ASE, PySisyphus, Sella, Hugging Face, TorchSim, docs, and security checks.

Citation

If AIMNet2 is useful in your work, please cite the relevant model papers:

AIMNet2

@article{aimnet2,
  title={AIMNet2: A Neural Network Potential to Meet Your Neutral, Charged, Organic, and Elemental-Organic Needs},
  author={Anstine, Dylan M and Zubatyuk, Roman and Isayev, Olexandr},
  journal={Chemical Science},
  volume={16},
  pages={10228--10244},
  year={2025},
  doi={10.1039/D4SC08572H}
}

AIMNet2-NSE: Kalita, B.; Zubatyuk, R.; Anstine, D. M.; Bergeler, M.; Settels, V.; Stork, C.; Spicher, S.; Isayev, O. AIMNet2-NSE: A Transferable Reactive Neural Network Potential for Open-Shell Chemistry. Angew. Chem. Int. Ed. 2026. DOI: 10.1002/anie.202516763

AIMNet2-Pd: Anstine, D. M.; Zubatyuk, R.; Gallegos, L.; Paton, R.; Wiest, O.; Nebgen, B.; Jones, T.; Gomes, G.; Tretiak, S.; Isayev, O. Transferable Machine Learning Interatomic Potential for Pd-Catalyzed Cross-Coupling Reactions. ChemRxiv 2025. DOI: 10.26434/chemrxiv-2025-n36r6

License

See LICENSE.