I used the DeePMD kit to generate a neural network potential frozen in the .pb format for the H2O molecules. I have created this and I wanted to study these potentials using with and without electric field. I have used
pair_style deepmd graph.pb # DeepMD-kit model file for water
pair_coeff * * # Apply the DeepMD-kit model to all atoms
It ran without any errors or warnings. But the results I got from these are the same. No change in the energy with respect to the temperature and pressure. I want to know if there is an effect of Efield using deepmd potentials in LAMMPS or if we need to use some alternative method to do it. Because I want to study the dynamics of some systems with deep neural network potentials in the presence of an electric field.
Like many other ML MD potentials, DeepMD clearly doesn’t have any concept of partial charges. I’m old enough to remember when the first big splashy ML MD paper was published describing large-scale solid state silicon modelling – an application for which partial charges are clearly not very relevant or good descriptors for atomic configurations. That still seems to be the trend.
Without partial charges there is nothing for an electric field to act on, and so your results are perfectly sensible. You will either have to use conventional MD models, or find some ML potential with some notion of partial charges.
Recently, I came across this paper, which deals with the change of the bonds with respect to the electric field. They have used CP2K to generate data for the DeePMD kit and used that data in LAMMPS. They have studied the system with respect to different electric fields. Any commends on this?
Deepmd is not part of LAMMPS, so any deepmd related questions you have to direct to the deepmd developers in their forum or whatever way they use to interact with their users.
It is up to you to make sense of what was done in this paper. The fact of the matter remains: what fix efield does is to add a force to each atom that is the proportional to the product of the electric field vector and an the (partial) charge of an atom. Since machine learning potentials available in LAMMPS (and typical deepmd potentials apparently) do NOT assign (partial) charges to atoms, fix efield has no effect. This is plain and simple and no arguing will change that.