Chemistry & ADMET · isayevlab
AIMNet2-rxn
A neural-network interatomic potential for predicting energies, forces, and pathways in organic chemical reactions.
What AIMNet2-rxn is
AIMNet2-rxn is a neural-network interatomic potential developed by the Isayev Lab, designed for generalised organic chemical reaction modelling. It delivers energies, atomic forces, partial charges, and electrostatic moments across reaction coordinates without requiring system-specific retraining.
Conventional quantum-chemical pathway searches with density functional theory (DFT) are computationally demanding, limiting their use in high-throughput mechanism exploration. AIMNet2-rxn handles bond forming and breaking, transition states, and reaction profiles at near-ωB97M-V/def2-TZVPP accuracy while drastically accelerating minimum-energy-pathway searches.
The model is supplied as an ensemble of four models for uncertainty estimation, enabling tasks such as transition-state optimization and large-scale nudged-elastic-band (NEB) reaction pathway exploration.
How it works
AIMNet2-rxn relies on the AIMNet2 architecture, which combines Atomic Environment Vectors with message-passing multilayer perceptrons, complemented by prediction heads for molecular dipoles and quadrupoles. It employs a 5.0 Å cutoff radius and handles electrostatic interactions via an embedded short-range Coulomb treatment (≤ 4.6 Å) combined with an external long-range Coulomb module (≥ 4.6 Å). Dispersion is captured implicitly through the VV10 non-local correlation kernel present in the underlying DFT reference functional, omitting any separate empirical dispersion correction.
The model was trained on 4,685,165 geometries derived from the AIMNet2 base set and the RGD-1 reaction database, all computed at the ωB97M-V/def2-TZVPP level. Outputs are expressed on a learned, per-element shifted electronic-energy scale in electronvolts (eV), intended strictly for evaluating energy differences between configurations of identical atomic composition.
Key facts
- Architecture
- AIMNet2 (Atomic Environment Vectors + message-passing MLPs) with dipole and quadrupole heads
- Parameters
- 4 ensemble members
- Reference DFT level
- ωB97M-V/def2-TZVPP
- Training data
- 4,685,165 geometries
- Cutoff radius
- 5.0 Å
- Supported elements
- H, C, N, O
What you can use it for
- Minimum-energy pathway searches
- Execute minimum-energy pathway calculations using nudged elastic band (NEB) or batched NEB workflows in tools such as ASE and pysisyphus.
- Transition-state optimization
- Locate and refine organic reaction transition states using energy and force evaluations or Hessian-based algorithms.
- Intrinsic reaction coordinate (IRC) profiling
- Map continuous reaction coordinates and energy profiles to obtain accurate barrier heights and reaction energies.
- High-throughput reactivity screening
- Screen vast reaction spaces and enumerate reaction mechanisms across millions of pathways on closed-shell organic compounds.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
AIMNet2-rxn
Strengths and limits
Good at
- Attains near-ωB97M-V/def2-TZVPP accuracy, with ~1 kcal mol⁻¹ error across IRC profiles and ~1.6 kcal mol⁻¹ RMSD on reaction energies.
- Facilitates high-throughput pathway searches at approximately 10⁶ reactions per day via batched NEB on a single GPU.
- Supplies four ensemble members to enable automated uncertainty quantification across predictions.
- Includes built-in safeguards to detect unsupported elements, non-neutral systems, and incompatible compilation configurations.
Watch out for
- Restricted strictly to closed-shell species composed of hydrogen, carbon, nitrogen, and oxygen (H, C, N, O).
- Does not support net-charged species like isolated cations or anions, although net-neutral zwitterions are supported.
- Outputs energies on a learned, shifted scale, meaning single-point energies are not absolute and cannot be compared across different AIMNet2 model families.
- Hessian and transition-state workflows hang under torch.compile due to double-backward issues through GELU activations, requiring model compilation to be disabled.
Inputs and outputs
You provide
Molecule (SMILES)
We generate and relax a 3D conformer before the model sees it.
You get back
Energy, maximum force and per-atom partial charges.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use AIMNet2-rxn commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run AIMNet2-rxn without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on CPU only, typically in 12s, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
Which elements are supported by AIMNet2-rxn?
The model supports only hydrogen, carbon, nitrogen, and oxygen (atomic numbers 1, 6, 7, and 8). Unsupported elements are not evaluated and will cause input validation errors.
Can AIMNet2-rxn model radicals or charged molecules?
No. The model is strictly trained on closed-shell, net-neutral organic molecules. Open-shell species like radicals and net-charged ions are unsupported, though net-neutral zwitterions with localized charge separation are permitted.
How should energy outputs from this model be interpreted?
Energies are given on a learned, per-element shifted electronic-energy scale in electronvolts. They must always be evaluated as relative energy differences (such as barrier heights, reaction energies, or conformer energies) between geometries of identical atomic composition, rather than as absolute single-point energies.
Why is an external dispersion correction like D3 not needed?
The ωB97M-V reference functional incorporates the VV10 non-local correlation kernel directly into all training-set energies and forces. Adding an external dispersion correction at inference would double-count dispersion interactions.
What setting is required for computing Hessians or running transition-state optimizations?
You must set compile_model=False when initializing the calculator. Autograd double-backward operations through GELU activations during Hessian calculation cause torch.compile to hang.
How to cite it
Anstine, Dylan M. and Zhao, Qiyuan and Zubatiuk, Roman and Zhang, Shuhao and Singla, Veerupaksh and Nikitin, Filipp and Savoie, Brett M. and Isayev, Olexandr. AIMNet2-rxn: A Machine Learned Potential for Generalized Reaction Modeling on a Millions-of-Pathways Scale. 2025. doi:10.26434/chemrxiv-2025-hpdmg.
Written from AIMNet2-rxn's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.