Run a pure NNP simulation#
You will learn: how to simulate a system in which every atom is described by a neural-network potential, with no classical force field involved.
Prerequisites:
ACEMD installed with NNP support.
A structure file in any of: PDB, CIF, BCIF, SDF, MOL2. No classical topology is needed.
Setup#
Worked example: benzamidine in vacuum, all-AceFF. Download benzamidine_nnp.zip for the input file.
thermostat: true
thermostattemperature: 300
velocities: 300
structure: BEN_pH7_4.cif
coordinates: BEN_pH7_4.cif
minimize: 500
timestep: 2
run: 2ns
nnp:
name: AceFF-2.0
type: torch
The nnp block has no sel — that’s what makes it a pure NNP simulation. ACEMD applies the NNP to every atom. ACEMD auto-downloads AceFF-2.0 from HuggingFace on first use and caches it under ~/.cache/acemd/nnp/. See Obtain AceFF NNP models if you need to manage the checkpoint manually.
Run#
acemd
Why no classical topology?#
In NNP/MM mode you need a PSF / PRMTOP because the rest of the system has to be parameterised classically. In pure-NNP mode there is no “rest of the system” — the NNP supplies all forces — so any file that gives atom positions and elements is enough. Use PDB / CIF / SDF / MOL2 directly.
Parameters that matter#
timestep— 2 fs is the recommended maximum. NNPs are less numerically stable than classical force fields.nnp.name— model identifier. Use"AceFF-2.0"to auto-download the latest AceFF, or"TorchMD-Net"if you’re providing a checkpoint path viannp.file.nnp.file— optional. Only set when supplying a TorchMD-Net checkpoint manually.
See NNP options for the full reference.
Gotchas#
AceFF 2.0 is small-molecule only. Don’t pure-NNP a protein, nucleic acid, or solvent box — those species are outside the training distribution. For mixed systems, use NNP/MM and keep
nnp.selon the small molecule.AceFF 2.0 supports total charges of -2 to +2. Other charges produce wrong forces.
If the simulation crashes, drop
timestepto 1 fs. Either that, or wrap the species in classical solvent and switch to NNP/MM.