Using MLFF from NPT in NVT ensemble

[xiliang_lian]

Hello,

I want to ask whether it is reasonable to use an MLFF trained with an NPT ensemble in an NVT simulation and whether the reverse is also possible. Do you have an idea of whether it works well? In some cases, we may only want to simulate with the NVT ensemble but we want to have better sampling with the NPT ensemble.

As far as I understand, if we train an MLFF with NPT, then the energy we get will consider the pressure contribution, which will make it inapplicable to NVT simulations and vice versa. But in an earlier publication by Georg Kresse(https://doi.org/10.1103/PhysRevLett.122.225701), for MAPbI3, it seems you only trained the model with NVT and apply it to NPT if I understand correctly.

Can you please comment on these questions? Thanks a lot in advance.

Best regards,
Xiliang

[ferenc_karsai]

I would recommend to always train in NpT with a temperature ramp. It should be perfectly usable for NVT and NVE calculations too.
Just don't forget to use higher ENCUTs in the learning. Please read the best practices site:
https://www.vasp.at/wiki/index.php/Best ... rce_fields

NVT can work for NpT, but I would test it always individually. If it won't work well, you will see it usually very fast in the stability of your results. You can also monitor your BEEF in the ML_LOGFILE and when the value goes up then you now the force field starts to extrapolate.

[xiliang_lian]

Thank you very much for the insightful comments. It is very helpful and clear to me.