Obtain AceFF NNP models#

You will learn: how to download an AceFF neural-network checkpoint manually, for offline use or version pinning.

Prerequisites:

  • ACEMD installed with NNP support.

AceFF is the family of neural-network potentials developed by Acellera. The current release is AceFF 2.0, distributed under the Apache 2.0 license: no HuggingFace account, no access request, and no token needed.

Note

For most users, the easiest path is to let ACEMD download and cache the model automatically. Set nnp.name: AceFF-2.0 in your input.yaml and skip this page. See Run an NNP/MM simulation or Run a pure NNP simulation.

This page covers the manual flow for offline machines, version pinning, or sharing a checkpoint across a cluster.

Download#

Pull the checkpoint directly:

curl -L -o aceff_v2.0.ckpt https://huggingface.co/Acellera/AceFF-2.0/resolve/main/aceff_v2.0.ckpt

Reference the resulting file from your input.yaml with nnp.name: TorchMD-Net and an explicit nnp.file:

input.yaml#
nnp:
  file: aceff_v2.0.ckpt
  name: TorchMD-Net
  type: torch

What AceFF 2.0 supports#

  • Elements: H, B, C, N, O, F, Si, P, S, Cl, Br, I.

  • Total charge: -2, -1, 0, +1, +2.

  • Recommended timestep: 2 fs (with hydrogen-mass repartitioning).

  • Architecture: TensorNet v2 inside the TorchMD-Net runtime.

Gotchas#

Warning

Small molecules only. AceFF 2.0 is trained on a curated PubChem dataset of small molecules. Proteins, water, and other biomolecular polymers are not in the training set, and applying AceFF 2.0 to those species (in either pure NNP or as the NNP region of an NNP/MM run) will give wrong forces. For NNP/MM, restrict nnp.sel to a single small-molecule ligand or cofactor.

Warning

Total charges outside the -2…+2 range are not supported. Charged species at the edge of this range can still extrapolate poorly; if you see crashes, drop the timestep to 1 fs or fall back to an NNP/MM run with a smaller NNP region.

See also#