Cut-offs, accuracies, and long-range forces

[NatashaMartinez]

Hi!

I am writing this hoping to get some advice on the best practice when simulating a solute-solvent system. I have run extensive tests with different descriptors, network sizes, cut-offs etc. I am a little bit confused about the results and how to use the trained networks.

1. The energies and forces are taken from CP2K AIMD which include PBC and long-range interactions are modelled with poisson solver. Does it mean that we do not need to include extra coul/long term on top of our deepMD potential when running simulation in LAMMPS? If we do include those, wouldn't that be double counting the interactions?

2. To maximally include the long-range interactions I tested higher cut-off values up to some very large distances. I have noticed that the accuracy of forces and energies drops with increasing cut-off. Is it normal? I can imagine the network becomes bigger and harder to fit but this happens very fast even for some reasonable cutoffs like 8 or 9A.

3. I have all the density cubes and trained atomic dipoles saved for my system. Is there any way to incorporate those to train DPLR model? The Wannier centroid approach is confusing for me. I can somewhat understand for a simple system like water but for my system with 6 different atom types and some heterogeneity, I have no idea how to approach it. Is it even worth trying?

[wanghan-iapcm]

1. yes the standard deepmd only considers the short-range part of the interaction. In most cases, it is a good approximation to the long-range interacting systems, however, in exceptions one needs to explicitly model long-range interaction on top of the short-range part by using e.g. dplr.
2. the long-range electrostatic interaction cannot be satisfactorily captured by a short-range model with large cut-off. Instead, one needs Ewald summation in pbc systems.
3. yes dplr works in a solute-solvent system. e.g. https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.131.076801