LAMMPS simulation not starting when vacuum in structure

[VictorBrouwers]

I'm modelling perovskite structures, trained with FLARE++ and then mapped to LAMMPS. This generally works wonderfully well, so a big thank you to making this code openly available in that sense :).

My process

1. I train a bulk perovskite structure in FLARE++.
2. I map the model to LAMMPS.
3. I run LAMMPS simulations for longer timescales.
   [4. I run SGP simulations for uncertainty checks and uncertainty visualization.]

My problem
The process outlined above usually works fine. However, recently, I've started to simulate surfaces with the bulk trained model (to check transferability). Here, I run into issues.

• My LAMMPS simulation with the surface structure (so including vacuum) does not start, and it does not produce any errors. This is the case for both NpT and NVT simulations.
• The exact same model IS able to run an SGP simulation (in both NpT and NVT).
• I've tried to train a model (in NVT) including vacuum and map this, but this does not solve my problem.

Screenshots

This is an image of my surface material (called "cub_I_LARGE") in the attached ZIP-folder. Again, this does not work in LAMMPS, while it does work with just the SGP model in Python.

This is an image of my bulk material (in a different, less symmetric phase). Simulating this works perfectly fine.

Desired outcome
I'd like to use a mapped model to run large-scale LAMMPS simulations with surfaces. If I understand correctly, you are also able to do this in your FLARE++ paper for your hydrogen-platinum catalysis.

Additional context

• LAMMPS has been installed according to https://github.com/mir-group/flare_pp/blob/master/lammps_plugins/README.md.
• FLARE++ version 0.0.24 is installed.
• I've uploaded a ZIP-file to my Drive with my files for potential reproduction. It contains:

1. LAMMPS_bulk: the working LAMMPS simulation
2. LAMMPS_surface: the not-working LAMMPS simulation with both the bulk-trained and surface-trained systems
3. SGP_I_surface: the SGP simulations that dó work with the bulk-trained system and the surface-training system.

The ZIP file can be downloaded from here: https://drive.google.com/file/d/1e5g_ysL5Uu_VKz62Rd4_lO5E4EVZypeY/view?usp=sharing

I hope this somewhat clearly illustrates the problem I'm facing.

Kind regards,
Victor Brouwers

[anjohan]

Hi Victor,

I've tried running your Surface_issue/LAMMPS_surface/Surface_Trained_NVT/lammps.in, and the atoms are moving for me - with and without Kokkos[*]. This is with the master branch of flare_pp.[**] My commands are

$HOME/lammps/build/lmp -in lammps.in 
LAMMPS (29 Sep 2021)
OMP_NUM_THREADS environment is not set. Defaulting to 1 thread. (src/comm.cpp:98)
  using 1 OpenMP thread(s) per MPI task
...
Step Time Temp PotEng KinEng TotEng Press Volume 
       0            0          600   -1328.6804    34.667571   -1294.0129    1824.9892    39847.344 
       1        0.001    599.90273   -1328.6751    34.661951   -1294.0132     1823.651    39847.344 
       2        0.002    599.70034   -1328.6644    34.650257   -1294.0141    1823.0908    39847.344 
       3        0.003    599.39306   -1328.6482    34.632503   -1294.0157    1823.3072    39847.344 
       4        0.004    598.98119   -1328.6266    34.608705   -1294.0179    1824.2977    39847.344 
       5        0.005    598.46509   -1328.5996    34.578885   -1294.0207    1826.0589    39847.344 
       6        0.006    597.84518   -1328.5672    34.543067   -1294.0241    1828.5866    39847.344 
       7        0.007    597.12195   -1328.5294     34.50128   -1294.0281    1831.8755    39847.344 
       8        0.008    596.29598   -1328.4863    34.453555   -1294.0327    1835.9193    39847.344 
       9        0.009    595.36788   -1328.4378    34.399931   -1294.0379    1840.7108    39847.344 
      10         0.01    594.33837   -1328.3841    34.340446   -1294.0436    1846.2419    39847.344 
...

$  $HOME/lammps/build/lmp -sf kk -k on t 12 -pk kokkos newton on neigh full -in lammps.in 
LAMMPS (29 Sep 2021)
KOKKOS mode is enabled (src/KOKKOS/kokkos.cpp:105)
  will use up to 0 GPU(s) per node
  using 12 OpenMP thread(s) per MPI task
...
Step Time Temp PotEng KinEng TotEng Press Volume 
       0            0          600   -1328.6804    34.667571   -1294.0129    1824.9892    39847.344 
       1        0.001    599.90273   -1328.6751    34.661951   -1294.0132     1823.651    39847.344 
       2        0.002    599.70034   -1328.6644    34.650257   -1294.0141    1823.0908    39847.344 
       3        0.003    599.39306   -1328.6482    34.632503   -1294.0157    1823.3072    39847.344 
       4        0.004    598.98119   -1328.6266    34.608705   -1294.0179    1824.2977    39847.344 
       5        0.005    598.46509   -1328.5996    34.578885   -1294.0207    1826.0589    39847.344 
       6        0.006    597.84518   -1328.5672    34.543067   -1294.0241    1828.5866    39847.344 
       7        0.007    597.12195   -1328.5294     34.50128   -1294.0281    1831.8755    39847.344 
       8        0.008    596.29598   -1328.4863    34.453555   -1294.0327    1835.9193    39847.344 
       9        0.009    595.36788   -1328.4378    34.399931   -1294.0379    1840.7108    39847.344 
      10         0.01    594.33837   -1328.3841    34.340446   -1294.0436    1846.2419    39847.344 

Do your results differ? Are the forces on your atoms zero?

[*] Note that unless you are running with GPUs/CUDA, the non-Kokkos version will mostlikely be faster, since the Kokkos version has been optimized for GPUs in terms of memory layouts etc.
[**] There's been a format change in the coefficient file, so you need to add the power (hopefully 2) on the second line of beta.txt

DATE: Thu Feb 10 15:53:24 2022 CONTRIBUTOR: YX
2
chebyshev
...

Best wishes,
Anders
log.lammps.txt
traj.lammps.txt

[VictorBrouwers]

Hi Anders,

Thanks for your quick and clear reply! Running both a Kokkos and non-Kokkos version on just 1 core works for me. Extending to multiple cores on the non-Kokkos version works just fine, while this is not the case for the Kokkos version (and this is not a problem with my other systems for some reason).

I was not aware that the non-Kokkos version is (probably) faster when not running on GPUs, so thank you for notifying me! The non-Kokkos version also runs just fine on multiple cores, so you can consider my issue resolved!

Kind regards,
Victor

[anjohan]

Hm, this is strange. I'll have a look later today. It could be that if you use a lot of MPI ranks, some of them end up without atoms. This should, of course, not be an issue, but it's not something I have ever tested. You could try the balance command (which will generally improve performance if there is vacuum in your system).

Otherwise, let me know if other issues show up!