Implementation of other interatomic potentials

[jmarbues]

Hi @brucefan1983 . You currently have a few interatomic potentials built in the GPUMD software (such as the Tersoff and LJ). I understand that the main focus of your software is the NEP implementation, however, I am writing you to inquire if you intend to add other forms of potentials to the software in the near future such as those implemented in the software LAMMPS (e.g. the Buckingham potential «https://docs.lammps.org/pair_buck.html» or the Morse potential «https://docs.lammps.org/pair_morse.html»)

Thank you for your time and for creating this amazing software.
Kind regards, José

[brucefan1983]

It is not hard to implement these two-body potentials. We actually have Buckingham+Coulomb before GPUMD-v3.4 and I removed it later.

The question is what would be the use cases for them nowadays. Could you elaborate?

[ejmeitz]

My unsolicited 2 cents....I uses LAMMPS because it has ALL the potentials you could ever want implemented. The lack of potentials in GPUMD makes it hard for me to use GPUMD as my go to MD library as I do a lot of phenomenological experiments and methodology development which often does not need full accuracy. Its also a lower barrier to entry to start with empirical potentials than training or finding a NEP potential for your material. For me at least, supporting a few more common empirical potentials would make me more likely to fully switch to GPUMD from LAMMPS. (despite this GPUMD is still super cool and I will continue to use it!)

[jmarbues]

Perfect answer to @brucefan1983's question. For me in particular, I am starting to get into MD simulations of glasses and before going into ab-initio MD simulations and training a NEP potential, I wanted to test some empirical potentials. For example, you have the Tersoff potential for silicon dioxide, which I was able to use. I specifically mentioned the Buckingham potential as I know that it is a common potential (or at least some extensions of the Buckingham) for some glass materials such as boron oxide or aluminium oxide.

But regarding your last message @brucefan1983, that would be really amazing. I know you already did that for 16 metals. I never thought that an entire periodic table potential would be possible given the huge differences in all the different combinations of materials one can obtain from combinations of elements. Could you perhaps give an estimation of when it will be implemented?

[brucefan1983]

Buckingham potential is usually used in combination of Coulomb potential. A proper evaluation of the Coulomb potential is highly nontrivial. I estimate that a professional, fast implementation of the Coulomb potential will take more time than completing the foundational NEP mode for the whole periodic table.

Let's see, when I have time, and if you still believe a Buckingham+Coulomb potential is highly desirable, I can try to do it. But now I am busy with developing the NEP-XXX extensions, such as NEP-Charge #880

[jmarbues]

Yes of course. It makes sense that you focus on the NEP model first. In that case I will try it for the system I wish to simulate when you release it. If the potential is accurate enough then there is no need to implement other empirical potentials.

Thank you for your help and good luck with the training of the NEP potential!

[brucefan1983]

How about if I tell you there will be a NEP model for the whole periodic table coming out soon?

[ejmeitz]

Yeah that would help lol. Any idea how soon is soon?

[brucefan1983]

I estimate it will be available within 3 months.

[brucefan1983]

Could you tell me how would you like to deal with the Coulomb potential? Currently I am only able to implement either the textbook ewald summation or the DSF method.

[jmarbues]

Hi @brucefan1983, sorry for the late reply.

I am not very familiar with the algorithms themselves used to efficiently calculate the different force terms. Between those two methods, I only know about the ewald summation, which I think is the most well known to deal with the coulomb potential in periodic systems. But in theory any method that calculates the same thing should yield the same result right? Maybe not for the absolute energies, but that should be the case for the relative energies and gradients/forces (at least within the error that comes with the cutoff parameters one imposes in the calculation).

So I would say that if you do have the time to implement the potential, you could just use the method that is the easiest for you (or that takes the least effort).

[ejmeitz]

I assume you already have a GPU kernel implemented for non-bonded pair potentials since you have LJ. You could implement Smooth Particle Mesh Ewald (SPME) without too much trouble. The real space part of the SPME is just the non-bonded kernel with a modified Coulomb potential and the reciprocal space is just an FFT which you can use an external library for. There is the matter of interpolating the system charges onto a grid so this is far from free, but I think you have most of the pieces in place to get SPME if that is something you deem worth developing.

[brucefan1983]

Hello, the NEP model for essentially the whole periodic table is published: https://github.com/brucefan1983/GPUMD/tree/master/potentials/nep/nep89_20250409