Barostats for NPT

[jcwang587]

Dear Development Team,

I noticed that Torch‑Sim provides an npt_langevin integrator and found this example script, which looks very useful:
https://github.com/Radical-AI/torch-sim/blob/main/examples/scripts/3_Dynamics/3.12_MACE_NPT_Langevin.py

Since ASE currently offers only two NPT algorithms (the Berendsen barostat and the Nosé-Hoover-Parrinello-Rahman barostat), which appear to have limitations in stability. Just want to get some idea if the NPT‑Langevin integrator here is now the recommended choice for running NPT simulations with the MACE MLIP, compared to the existing options in ASE? Thank you!

[abhijeetgangan]

There is a Nosé-Hoover NPT but currently doesn't support batching. From my testing the Langevin was more stable but you can take a look. You can track the invariant and see how it performs. NPT Berendsen is not implemented because it breaks the equipartition.

torch-sim/torch_sim/unbatched/unbatched_integrators.py

Line 1652 in 22aa785

def npt_nose_hoover( # noqa: C901, PLR0915

[jcwang587]

Thanks for the information!