OrbMol

OrbMol: Extending Orb to Molecular Systems

Built on the Orb-v3 architecture, OrbMol is a universal interatomic potential trained on the Open Molecules 2025 (OMol25) dataset—over 100M high-accuracy DFT calculations (ωB97M-V/def2-TZVPD).

Unlike its predecessors, OrbMol accepts total charge and spin multiplicity, enabling it to accurately model open-shell, ionic electrolytes, metal complexes and biomolecules.

---

Model checkpoints

There are two model checkpoints available:

• OrbMol: conservative default model
• OrbMol-direct: direct variant (i.e. forces are not the exact gradient of the energy)

Both models use a 6 Å radius cutoff with "infinite" (120) max neighbours in the neighborlist computation.

---

🚀 Quick Start

Head to the orb-models Github repository for complete instructions. In a nutshell:

pip install orb-models
pip install --extra-index-url=https://pypi.nvidia.com "cuml-cu11==25.2.*"  # For cuda versions >=11.4, <11.8
pip install --extra-index-url=https://pypi.nvidia.com "cuml-cu12==25.2.*"  # For cuda versions >=12.0, <13.0

import ase
from ase.build import molecule
from orb_models.forcefield import atomic_system, pretrained
from orb_models.forcefield.base import batch_graphs

device = "cpu"  # or device="cuda"
orbff = pretrained.orb_v3_conservative_omol(
  device=device,
  precision="float32-high",   # or "float32-highest" / "float64
)
atoms = molecule("C6H6")

atoms.info["charge"] = 0  # total charge
atoms.info["spin"] = 1  #  spin multiplicity
graph = atomic_system.ase_atoms_to_atom_graphs(atoms, orbff.system_config, device=device)

result = orbff.predict(graph, split=False)

---

Support

If you run into any issues feel free to post your questions or comments on our Github Issues page.

License

ORB models are licensed under the Apache License, Version 2.0. Please see the LICENSE file for details.

---

Citation

If you use this work, please cite:

@misc{rhodes2025orbv3atomisticsimulationscale,
      title={Orb-v3: atomistic simulation at scale}, 
      author={Benjamin Rhodes and Sander Vandenhaute and Vaidotas Šimkus and James Gin and Jonathan Godwin and Tim Duignan and Mark Neumann},
      year={2025},
      eprint={2504.06231},
      archivePrefix={arXiv},
      primaryClass={cond-mat.mtrl-sci},
      url={https://arxiv.org/abs/2504.06231}, 
}

@misc{neumann2024orbfastscalableneural,
      title={Orb: A Fast, Scalable Neural Network Potential}, 
      author={Mark Neumann and James Gin and Benjamin Rhodes and Steven Bennett and Zhiyi Li and Hitarth Choubisa and Arthur Hussey and Jonathan Godwin},
      year={2024},
      eprint={2410.22570},
      archivePrefix={arXiv},
      primaryClass={cond-mat.mtrl-sci},
      url={https://arxiv.org/abs/2410.22570}, 
}