Thermal Expansion using MLFF

[Fermi1976]

Dear VASP Forum,
I am interested in calculating the thermal expansion coefficient of graphite using MLFF, so I can utilize the temperature-dependent lattice constants later on to perform temperature-dependent phonon calculations. I have been trying using NPT, but I have some issues, I will be grateful if you could give some advices how to improve the calculations.

I am attaching the INCAR ML_LOGFILE files. I used 288 atoms, 3x3x3 k-mesh.
I am attaching also, the Bayesian error, RMSE, volume, Temp and c-lattice constant as a function of MD steps.

My questions are as follow:
1- There is a sharp drop in the volume (mainly due to the drop in c-lattice constant). This happened in the training stage (below 6000 MS steps). I appreciate your feedback for this issue
2- The number of the AB configurations is 94, how can increase this number to improve the quality of the FF?

Thank you

[ferenc_karsai]

Can you please send us all your files even if they are big (XDATCAR, POSCAR, POTCAR, ML_AB/N, ICONST, OUTCAR).
I would like to have a look at the trajectory, have you done so?

I already saw a type in your INCAR file. You wrote "Ecut = 600". I guess you meant "ENCUT = 600".

What you can try in the meantime:
-) Use ENCUT= 600 instead of the typo.
-) Try a different PMASS.
-) Try a different van der Waals (IVDW=10).

But these are only guesses, so I'm not sure any of these will help you.

[Fermi1976]

1- Thank you for finding the Ecut typo. I am repeating the calculations using ENCUT=600. I doubt this will make significant difference since ENMAX =400 in OTCAR and I am using PREC= Accurate (==> ENCUT ~520 eV).
I am wondering why VAPS did not show any warning of this typo.

2- I tried different PMASS values (1000, 2000, 3000), I did not get noticeable difference.

3- I tested different IVDW, IVDW=10 gave c = 6.39 A, which is far from the experimental value of 6.67 A. Whereas using IVDW=12 shows better prediction of c- value (6.66 A).

4- I will find away to send you the files you requested, they are so big.

Thank you again, I appreciate your willingness to help.

[Fermi1976]

Hi ferenc_karsai,
Could you please send me your email so I can share with you a link to access the files?

Regards,
Iyad

[Fermi1976]

Can you please send us all your files even if they are big (XDATCAR, POSCAR, POTCAR, ML_AB/N, ICONST, OUTCAR).
I would like to have a look at the trajectory, have you done so?

I have sent you the files, please let me know if you got them.

Thank you

[ferenc_karsai]

No I didn't get any files, but please upload them to something like dropbox and post the link here.

[Fermi1976]

No I didn't get any files, but please upload them to something like dropbox and post the link here.

Hi Fernec,

Please use this dropbox link to have access to the files. Please let me know if this works.
https://www.dropbox.com/scl/fo/3vhw4tp8 ... bd37m&dl=0

Thank you

[Fermi1976]

Hi Fernec,
Please let me know if you could access the files via the Dropbox link I posted.

Thank you

[ferenc_karsai]

We have looked at your calculation.

You had a quite large temperature range between 20 and 400 K in your calculations. Volume fluctuations should be expected in such a large range.
I don't know what your target temperature is, but if it's around room temperature it's good to learn up to 350 to 400 K.
So the upper bound is ok, but the lower is bound is too low.
You should equilibrate your cell at say 250 K (a few ps). Then take that structure as starting point. Then only ramp between 250 and 350-400 K.

Otherwise your calculational parameters look ok.