MLFF Nose-Hoover thermostat and Langevin thermostat

[xiliang_lian]

Hello,

I read from the wiki that during the MLFF training, it is good to use the Langevin thermostat due to its stochastic nature. I was wondering whether you could comment on the use of the Nose-Hoover thermostat for the NVT ensemble during the training, does this leads to a worse model compared to the model trained with the Langevin thermostat?

I also want to ask a question that has not been very clear on the forum yet, how do we decide on the LANGEVIN_GAMMA parameter? It seems that with a large LANGEVIN_GAMMA value, the code will sample a lot of local structure during the training which is obviously not desired, but with a too small value, the model is undertrained and can't be used for future work. If you could also give some comments on this, it would be great.

Thanks a lot in advance.

Best regards,
Xiliang

[ferenc_karsai]

It is also fine to use NVT ensemble using the Nose-Hoover thermostat for many systems and you can get very accurate force fields with that too. But generally one can say that the simple Nose-Hoover thermostat, that we have implemented, is not a very ergodic thermostat. So even using an NVT ensemble I would suggest to use Langevin. Regarding the ensemble NpT gives you more robustness due to the volume fluctuations during the MD, which are not present for NVT.

In principle your results should be independent of the LANGEVIN parameters. So as long as you don't choose too extreme parameters they are free to choose. Keep LANGEVIN_GAMMA between 3-10.

[xiliang_lian]

Okay, I see, get it. Thank you very much for your answer. It is very helpful for me.