ML training in NpH ensemble leads to a soft stop based on dEPOT

[chamara_somarathna]

Hello everyone,

I am trying to construct a machine learning potential (using VASP ML_MODE = train) for a given phase-space region. I am reading a CONTCAR (only the atomic coordinates and velocities) from a previous equilibration run in NVT ensemble. Since I have the velocities in the CONTCAR, I started with an INCAR without specifying TEBEG for the NpH run. I am asking for 4000 MD steps (NSW = 4000). But VASP comes to a soft stop without any errors before completing 4000 steps. For reference, one of my simulations stopped at step 3794. This is not a time-constrained stop as OUTCAR prints all the timing info at the end of the file. I have never seen this and wanted to know the thoughts of VASP developers and advanced users.

As this is something new I experienced, I tried to understand why the simulation stops at such a random point. When I looked into the log file, I observed that there was a quantity called dEPOT printed at the end of every ionic step and there is some kind of threshold value printed right next to this dEPOT value (see the attached logfile_END.png). This is not written in OUTCAR. What I observe is If this dEPOT value comes below that threshold value, VASP comes to a soft stop regardless of the NSW specified in INCAR. For the case I mentioned above, the threshold value was 1E-3, and at step 3794, the dEPOT value dropped below 1E-3, and thereby VASP went to a soft stop. I could not find this explained in the VASP manual and wanted to know the perspective of VASP developers and advanced users.

The next thing I observe is if I specify a TEBEG, this strange stopping criteria does not occur and VASP completes the specified NO of NSWs in INCAR. Even though I specified the TEBEG, the initial velocities were read from the CONTCAR as I could see from the initial temperature being not equal to the specified TEBEG but to the final temperature of the previous run. So, there is no effect from specifying TEBEG to the simulation but it turns the simulation into a different path at least where it defines this stopping criteria.

I would like to know the meaning of the dEPOT value and why there is such a soft-stopping criteria.

Thank you very much for your time and advice.

Best,
Chamara Somarathna

[marie-therese.huebsch]

Hello,

The stopping criterion is active for TEBEG < 0.1, because in that case VASP assumes you are using that subroutine for relaxation and not an MD run.

dEPOT is the absolute value of the difference of the potential energy of the ionic degree of freedom for this MD step minus the previous MD step.

The potential energy of the ionic degree of freedom for each MD step is reported in the REPORT and the OUTCAR file.
In the REPORT file it is the second value in the line that contains "e_b>". Note the line above labels the value with "E_pot".

In fact, from the view point of MD, this is the potential energy of the ionic degree of freedom. However, from the viewpoint of the electronic calculation, this is the total free energy of the DFT calculation, considering the artificial electronic temperature introduced by the smearing factor SIGMA. Hence, it is called TOTEN in the OUTCAR file.

Now the actual stopping criterion is

Code: Select all

ABS(EPOT-EPOT_old)<DYN%EDIFFG

where EDIFFG can be set in the INCAR file. The default is 1E-3.

As you would like to run an MD calculation, I would recommend setting TEBEG to some value greater than 0.1. Does this clarify all your questions?

Best regards,
Marie-Therese

[chamara_somarathna]

As you would like to run an MD calculation, I would recommend setting TEBEG to some value greater than 0.1. Does this clarify all your questions?

Hi Marie,

Thank you very much for your very descriptive answer. It clarified all the questions I had. Thank you very much. Appriciate it.

Best,
Chamara

[marie-therese.huebsch]

Glad I could help!