Suggestions for interfacing with RMC

[mikejwaters]

Hi, My experimental colleagues and I were discussing using my MLIPs to add physical guidance to their XPDF reverse Monte Carlo fitting procedure. They currently use powerful GPU nodes to perform ~ 10**6 structure trials with 10,000 to 30,000 atoms. Can I use the tensorpotential backend to quickly evaluate my MLIPs on a GPU? Or should I focus on using LAMMPS for large-scale CPU parallelism?

This might seem like a silly question from someone that's been using PACEMAKER for almost a year, but the ASE calculator has been fast enough for my purposes until now.

[yury-lysogorskiy]

Hi! PACEMAKER is based on TensorPotential package which is based on TensorFlow. It can be used for evaluation of potential, but I would suggest to try LAMMPS + ML-PACE + KOKKOS on GPU, which is faster for evaluation purposes. Large-scale CPU parallelism with LAMMPS would also work

[mikejwaters]

Thanks for the quick answer! It looks like I need to talk to the RMC developers about how they normally interface with LAMMPS.

[danielhollas]

Hi @yury-lysogorskiy. I have a related question going in an opposite direction. I would love to use (M)ACE potentials for small gas phase molecules (without GPU) in our MD program written in Fortran. I looked a bit over the codebase and on the LAMMPS integration (lammps-user-pace), and it seems that the C++ code here is not fully standalone, i.e. it relies on some LAMMPS internals (e.g. the neighborlist).

I am still very very new to ACE so I am not sure how much work would it be to create a new LAMMPS-like interface. If you have any thoughts about how best to proceed, let me know. Thank you!
(btw: I have really enjoyed your workshop on the (M)ACE meeting last month, it was very useful!)

[mikejwaters]

@yury-lysogorskiy I talked to the RMC developers, the good news is that they have already been testing with pacemaker potentials! The bad news is that current configuration only supports interfacing with OpenMP parallelized Lammps. I'm not sure if this is a lammps as a library issue or something they are working on. I'll find out more next week.