Using ML_LMLFF too large RMSE

[dominika_melicherova1]

Dear vasp developers,

I would like to create ML potential for polymeric liquid nitrogen under pressure.
I got liquid nitrogen by heating crystal starting with temperature 500 K and I was increasing
temperature with step 500 K until it melted at 4500 K. I used ML_LMLFF = .TRUE. for acceleration
of the calculation. After melting, I cooled the liquid down to 2000 K and I started a new training
from scratch to obtain ML FF for liquid. However, ML model has too large RMSEs and RMSEs
are increasing with growing number of MD steps.
Can I expect RMSEs decline at some point? Or do I have to change some parameters to model
improvement?

Thank you for reply

[andreas.singraber]

Hello!

Indeed the RMSEs you observe are very high. To analyze further it would be helpful if you could send us the OUTCAR file and also the OUTCAR, INCAR and POSCAR file of the heating run. One thing we noticed is that you are simulating a liquid in the NpT ensemble (MDALGO=3, ISIF=3). Are you using an ICONST file to fix the box shape, as shown in (3) and (4) here? If not, the box can "collapse" under these conditions, i.e. it gets elongated in some directions and squeezed in others, to the point were the box is only a single atom layer thick.

All the best,
Andreas Singraber

[dominika_melicherova1]

Thank you for your reply,

I don't use ICONST file so the box shape isn't fixed and yes, I see that the box is skewed after melting
but there isn't box "collapse" as you describe that there could be a problem. In addition, I have some
calculations of liquid at given condition obtained by purely DFT calculation. The liquid obtained by DFT
with ML has RDF similar to the liquid obtained by purely DFT calculation.

OUTCARs from the simulations are too large and I am not able to attach it.

Thank you for your help

Dominika