AMBER (`torchani-amber`)¶

Status: supported upstream with caveats.

AIMNet2 (and Nutmeg) are supported in AMBER through the torchani-amber integration from the Roitberg group. Despite the torchani- prefix the project explicitly handles AimNet2 and Nutmeg models in addition to ANI-family potentials. Reference paper: Bridging neural network potentials and classical biomolecular simulations.

Mechanism¶

torchani-amber is a compiled-in C++/Fortran integration, not a plugin. It hooks into AMBER's existing external-potential paths:

Mode	AMBER mechanism	Use case
Full ML	`iextpot = 1` + `&extpot` (`extprog = "TORCHANI"`)	All-atom ML simulation
ML/MM	`ifqnt = 1` + `&qmmm` (`qm_theory = 'EXTERN'`)	QM/MM with the QM region treated by the NNP

Both modes are pure mdin-file configuration -- no Python code at runtime.

Install¶

Not pip-installable. The integration must be built into AMBER from source:

Clone torchani-amber (git clone --recurse-submodules ...).
Build its bundled torchani_sandbox plus the C++ extensions (run-cmake).
Build AmberTools 25/26 from source with CMAKE_PREFIX_PATH pointing at the torchani-amber install. AMBER auto-links it.
For ML/MM where the QM region has a non-zero net charge (i.e. qm_charge != 0) with AimNet2 or Nutmeg, AmberTools 25 needs an additional patch (replace amber_src/AmberTools/sander/qm2_extern_torchani_module.F90). AmberTools 26 includes the patch.

A conda environment recipe (PyTorch, CUDA, GFortran, OpenMPI) is provided in the upstream repo.

User-facing input¶

Minimal AIMNet2 example (full ML):

&cntrl
    iextpot = 1,
/
&extpot
    extprog = "TORCHANI",
/
&ani
    model_type = "aimnet2",
    use_double_precision = .true.,
    use_cuda_device = .true.,
/

ML/MM example with electrostatic embedding via AIMNet2's predicted atomic charges:

&cntrl
    ifqnt = 1,
/
&qmmm
    qm_theory = "EXTERN",
/
&ani
    model_type = "aimnet2",
    mlmm_coupling = 1,
/

The mlmm_coupling keyword selects the QM/MM coupling scheme. The exact embedding details (which AIMNet2 charges are used, whether MM point charges polarise the QM region within a step, etc.) are documented in the upstream torchani-amber reference paper and source.

model_type may also be a full path to a TorchScript-jit-compiled .pt file. The v2 .pt assets shipped in aimnet/calculators/assets/ are torch.save state-dict archives, not TorchScript -- they cannot currently be passed as model_type directly. The TorchScript-export work in this repo would unblock shipping a self-contained AIMNet2 jit asset for torchani-amber rather than relying on the upstream-bundled model.

Caveats¶

Hard rebuild of AmberTools required. Once linked, the AMBER binaries depend on the torchani libraries even for non-ML runs.
Supported AMBER versions: AmberTools 25 (with patch) and 26.
The integration is tied to the torchani_sandbox build; the AIMNet2 model selection is constrained to what torchani-amber exposes upstream until a custom jit .pt path is provided.

Model coverage¶

What torchani-amber exposes upstream as model_type = "aimnet2". The NSE (open-shell) and rxn (reactive) AIMNet2 families are not currently exposed by the upstream integration.

Alternatives¶

If a full AmberTools rebuild is not acceptable, classical "AMBER force field plus ML region" workflows can sometimes be done via OpenMM + openmm-ml on AMBER-format topologies (loaded via ParmEd / OpenMM's AmberPrmtopFile). That route is pip-installable and needs no AMBER source build.

AMBER (torchani-amber)¶