NPT simulations with LAMMPS+Allegro

[EnricoTrizio]

Morning guys,

I've used NequIP before and now I'm testing Allegro for the study of some phase transition processes in the future, thus I need to run NPT simulations on my systems.
The patch with LAMMPS is super convenient as it can run very fast but it's not clear to me if the pair_allegro patch can be safely used with barostats to run NPT. So far, I tried kind of blindly and always got strange results for systems in which NVT was fine and stable, so maybe I've been missing something even if it doesn't raise any error.
I've also seen that for NequIP you have a stress/develop branch on Github even if only compatible with ASE and still under testing, so I was wondering about the possibilities with Allegro.

In case the answer is no, is this something you are planning to add in the near future and/or can you give me eventually some indications about where these things should be handled within the code?

Thanks in advance for the help!

[simonbatzner]

Answered in email, summary: stress fully working on the PyTorch side, see develop branch of nequi repo. Stress in LAMMPS is implemented in pair_nequip (for NequIP and Allegro, albeit Allegro less well-tested). This let's you do single-GPU NPT runs. We're working on a parallel pair_allegro implementation of stress.

[sigbjobo]

Hi, what is the status of the NPT implementation?

[Linux-cpp-lisp]

Hi @sigbjobo

Everything should now be working on the stress branch of pair_allegro combined with the latest develop of nequip. That said, it's still pretty new so please let us know if you encounter any problems.

[Linux-cpp-lisp]

See also mir-group/pair_allegro#12

[weilaigit]

To follow up on this, in order to use stress/LAMMPS/Allegro, we need to use the develop branch of nequip and the stress branch of pair_allegro. What about allegro, main or develop branch? Thanks.