Pure NNP simulations

ACEMD supports running molecular dynamics simulations with neural networks in so-called pure NNP simulations. In these simulations the whole system is described with a neural network potential without any classical force field.

You can download here the example input file for a pure NNP simulation of benzamidine in vacuum.

The input file of ACEMD can then be configured as follows to run a pure NNP simulation (see Neural Network Potentials for more details). Since the whole system is described with a neural network potential, we can skip the sel option of the NNP configuration.

thermostat: true
thermostattemperature: 300
velocities: 300
structure: BEN_pH7_4.cif
coordinates: BEN_pH7_4.cif
minimize: 500
timestep: 2
run: 2ns
nnp:
    file: aceforce_v1.0.ckpt
    name: TorchMD-Net
    type: torch

Compared to NNP/MM simulations, pure NNP simulations don’t require a classically built system. So you can directly use PDB, CIF, binary CIF, SDF or MOL2 files as structure and coordinate files.

A timestep of 2 fs is recommended for NNP simulations as NNP simulations are not very stable at larger timesteps.

To obtain a NNP model you can follow the instructions in Obtaining AceForce NNP models.

Running the simulation

You can execute the simulation directly with a simple acemd call:

acemd