Tabulated non-bonded potentials

User discussions
1

GROMACS version: 2022
GROMACS modification: No

Hi,

Is there any (short/mid-term) plans to re-implement tables in GROMACS?

We are thinking of doing some development with the Martini FF using tabulated NB potentials with earlier GROMACS versions. However, that only makes sense if tables will be back to GROMACS in foreseeable future versions.

Thanks!

2

Hi,
As far as I understood, there is not short term plan to re-implement tabulated potential. @pbauer may know more on mid/long term plan.
Anyway to use tabulated non-bonded potentials, you can use version 2018 (but maybe you know already)
Best regards
Alessandra

3

For what it’s worth, LAMMPS has tabulated non-bonded potentials. It might be slower than an equivalent GROMACS computation (especially on CPUs), but at least the feature is portable across hardware platforms (i.e. GPUs).

MARTINI was designed to be 1:1 compatible with biomolecular atomistic force fields, with portability in mind. Other CG force fields have tried different approaches, e.g. sacrifice portability for accuracy by using different functions than those of atomistic force fields. You would probably find that many of those force fields have a LAMMPS implementation using e.g. dedicated potential functions or tabulated ones.

With some exceptions, it has been standard practice for people developing those force fields to also invest effort developing the code needed to run them.

I don’t know how much work is required by the GROMACS core team, but certainly it’s not trivial or it’d have been done already. If there is a broad demand to use GROMACS with force fields other than atomistic and MARTINI, they may consider adding it back in the future.

Giacomo

4

Thanks for the info Alessandra! Indeed, I am aware that we can use version 2018. But it would be good to know that our development would later be supported modern GROMACS versions in the future. @pbauer , any more insight on the topic?

@giacomo.fiorin , thanks. I am indeed aware that we can use other software for this purpose, but as you mentioned, compatibility is important for any Martini development, and for that reason, we like to stick to GROMACS (as long as it is possible).

5

Hello @selimsami,

sorry for the long silence here. The issue with the tabulated potentials is that we can’t just write a direct replacement for the verlet kernels, but need some major work on the internals for having a proper integration in GROMACS. This has been in the work for a while, but it still work in progress. I hope we can share a version with them integrated when the GROMACS 2023 beta comes out in autumn this year, but I’ll make sure to keep the community informed here anywhere.

I know that this has been a major issue with GROMACS since we removed this with the 2019 release, and that it is something we need to urgently address.

Cheers

Paul

6

Hi @pbauer,

Thanks for the update, this is very useful!
We’ll keep our eyes open for the 2023 beta then.

Cheers
Selim

7

hi everyone,
any update on reintegrating tabulated potential in gromacs 2024?

8

Our hope is to reimplement tabulated nonbonded potentials during 2024, i.e., for GROMACS 2025. But since we’ve been hoping to have time and resources to do that for many releases now, I will definitely not promise anything.

9

Any updates on this? It would seem that with the new implementation of the neural network potentials, there may be a new way to incorporate non-standard potentials.

10

Yes, this is a way to incorporate non-standard potentials. But this is in practice only useful for neural network type potentials. These are orders of magnitude slower than standard or tabulated pair potentials.

11

Hi everyone,

I wanted to share a concern from my current workflow in case it helps inform future development priorities.

I am currently using GROMACS 2018.8 because I rely on non-standard 12-6-4 tabulated potentials, and I couple this with PLUMED for metadynamics. In practice, I am running into serious performance and scalability limitations. The older GROMACS version does not benefit from the GPU improvements available in newer releases, and the lack of Verlet-scheme support for tabulated potentials makes these simulations especially slow.

I would really appreciate continued support for users who need flexible non-standard interaction forms. This could be through robust tabulated-potential support in modern kernels, or more general support for flexible 1/r^n-type terms that work with Verlet and free-energy workflows.

At the moment, it feels like users with these kinds of systems are effectively locked into older versions, which is a difficult place to be.

Thanks for considering this.

Hossam

12

Tabulated potentials will be back in the 2027 release. But that might not be sufficient for your 12-6-4 potentials. There are only two parameters, not three. If you are lucky that you can live with two, it is now quite easy to modify standard LJ to 12-6-4, as the interaction form only appears on two lines in the SIMD and GPU kernels.

13

Hello Hess,

That is really great news. I am very happy to hear that tabulated potentials will be back in the 2027 release. Thank you for sharing this.

For my current 12-6-4 workaround, since GROMACS supports only two parameters, I combine the 12 and 6 terms into one tabulated column and place the interaction parameters directly into the tabulated potential and force values. In the topology, I then set the non-bonded parameters to 1 so the scaling is fully carried by the table.

This has been a practical solution for my case, so I really hope a similar workaround will still work in the new release.

Thanks again for the encouraging update.

Best,

Hossam