MODELS.md

May 26, 2026 · View on GitHub

Pretrained models

We provide several pretrained models that can be used to calculate energies, forces & stresses of atomic systems. All models are provided in the orb_models.forcefield.pretrained module.

OrbMol Models

These models are a continuation of the orb-v3 series trained on the Open Molecules 2025 (OMol25) dataset—over 100M high-accuracy DFT calculations (ωB97M-V/def2-TZVPD) on diverse molecular systems including metal complexes, biomolecules, and electrolytes. Note: The training data does not contain periodic systems and these models have not been carefully tested on periodic systems.

There are two options:

  • orbmol-v1-conservative
  • orbmol-v1-direct

See below for more explanation of this naming convention. Both models have inf neighbors, ensuring a continuous PES.

OrbMol-v2 (learnable electrostatics)

  • orbmol-v2

OrbMol-v2 extends the OrbMol architecture with learnable per-atom electrostatics: a LatentChargeHead predicts per-atom latent charges constrained to sum to the system total charge, and a CoulombModule adds a long-range Coulomb energy on top of the GNN, direct bare-1/r Coulomb sum for non-periodic systems, Particle Mesh Ewald via nvalchemiops for periodic systems. The energy head (ChargeConditionedEnergyHead) is conditioned on the per-atom charges. Similar to orbmol-v1, system-level total charge and spin are required.

Trained on OMol25 and OPoly26 (ωB97M-V/def2-TZVPD); supports both periodic and non-periodic systems. Stress is enabled via model.enable_stress() if needed.

from orb_models.forcefield.pretrained import orbmol_v2
model, atoms_adapter = orbmol_v2(device="cuda")
# atoms.info["charge"] and atoms.info["spin"] (multiplicity, = 2S+1) must be set.

Caution: While the model does predict per-atom charge values as a latent feature in the charge head, the model has not seen any per-atom charge values during training; these are emergent from optimisation against energies and forces alone. They should therefore be treated with caution: while in at least some cases they appear to correspond to the correct physical values, the reliability and generality of this correspondence is unclear and is the subject of ongoing investigations.

V3 Models

V3 models use the following naming convention: orb-v3-X-Y-Z where:

  • X: Model type - direct or conservative. Conservative models compute forces and stress via backpropagation, which is a physically motivated choice that appears necessary for certain types of simulation such as NVE Molecular dynamics. Conservative models are significantly slower and use more memory than their direct counterparts.

  • Y: Maximum neighbors per atom: 20 or inf. A finite cutoff of 20 induces discontinuities in the PES, which can lead to significant inaccuracies for certain types of highly sensitive calculations (e.g. calculations involving Hessians). However, finite cutoffs reduce the amount of edge processing in the network, reducing latency and memory use.

  • Z: Training dataset - omat or mpa. Both of these datasets consist of small bulk crystal structures. We find that models trained on such data can generalise reasonably well to non-periodic systems (organic molecules) or partially periodic systems (slabs), but caution is advised in these scenarios.

Features:

  • Model compilation using PyTorch 2.6.0+, enabling faster inference while maintaining support for dynamic graph sizes
  • Wider architecture (1024 vs 512) with fewer layers (5 vs 15) compared to v2, resulting in 2-3x faster performance with similar parameter count
  • Two variants available: direct models and conservative models (forces/stress computed via backpropagation)
  • Trained on the larger, more diverse OMat24 dataset
  • Improved edge embeddings using Bessel-Spherical Harmonic outer products (8 Bessel bases, Lmax=3)
  • Enhanced stability through Huber loss and a ZBL pair repulsion term added to forces
  • Models available with both unlimited neighbors and 20-neighbor maximum configurations
  • New confidence head providing intrinsic uncertainty estimates for predictions

Advice / Caveats

  • Consider using orb-v3-conservative-120-omat for initial testing, specifying precision='float32-highest' when loading the model. This is the most computational expensive but accurate configuration. If this level of accuracy meets your needs, then other models and precisions can be investigated to improve speed and system-size scalability.
  • We do not advise using the -mpa models unless they are required for compatibility with benchmarks (for example, Matbench Discovery). They are generally less performant.
  • Orb-v3 models are compiled by default and use Pytorch's dynamic batching, which means that they do not need to recompile as graph sizes change. However, the first call to the model will be slower, as the graph is compiled by torch.

V2 Models

  • orb-v2 - trained on MPTraj + Alexandria.
  • orb-mptraj-only-v2 - trained on the MPTraj dataset only to reproduce our second Matbench Discovery result. We do not recommend using this model for general use.
  • orb-d3-v2 - trained on MPTraj + Alexandria with integrated D3 corrections. In general, we recommend using this model, particularly for systems where dispersion interactions are important. This model was trained to predict D3-corrected targets and hence is the same speed as orb-v2. Incorporating D3 into the model like this is substantially faster than using analytical D3 corrections.
  • orb-d3-{sm,xs}-v2 - Smaller versions of orb-d3-v2. The sm model has 10 layers, whilst the xs model has 5 layers.

Features

  • v2 models use a smoothed cosine distance cutoff for the attention mechanism, ensuring a continuous PES.
  • The force predictions now have net zero forces, meaning they are much more stable for MD simulations.
  • The models are generally more accurate (Increase in 2-3% on the matbench discovery dataset).

V1 Models

Our initial release. These models were state of the art performance on the matbench discovery dataset at time of release, but have since been superseded and removed.