My Community

Dear all,

I would like to apply machine learning force fields (MLFF) to the study of strongly correlated electron systems.
So, I'm considering settings for the first principles (FP) calculation to be performed during the MLFF(ML_ISTART=0).

My question is whether the DFT+U method can be used for the FP calculation when learning force fields.
Also, some tags such as GGA or ISPIN were not included in the tutorial INCAR, can they be appled to the FP calculation?

Would you tell me what DFT calculation settings can be applied?

Best regards,
Kai

I confirmed that those settings were applied to my MLFF job with " Startparameter for this run:" in OUTCAR.

Hello Kai,

in principle all ab initio features which provide energies and forces (and stresses) can be combined with the machine learning force fields. However, mind that from the perspective of the ML force field there is only structural information from atomic positions. No information about electronic (electronic densities, bands,...) or spin degrees of freedom is passed on to the ML force field. For example, consider a system with multiple spin configurations corresponding to different energies and forces but identical atom positions and lattice. In such case the ML force field cannot distinguish the spin configurations and will be unable to learn them correctly. Does this answer your question?

All the best,

Andreas Singraber and Ferenc Karsai

Dear Andreas Singraber and Ferenc Karsai,

Thank you very much for kind reply.

I'd like to confirm about your reply, just in case.

Do you mean that detailed settings of DFT can be applied to the FP calculations when creating the Force Field, but the "trained" Force Field (ML_FF) does not include information such as electronic state or spin?

In other words, the settings of DFT can be used when ML_ISTART=0 or 1, but they do not apply when ML_ISTART=2?

Is my understanding correct?

Best regards,
Kai

Hello Kai,

indeed, ab initio settings in the INCAR file are ignored when VASP is started with ML_ISTART = 2 because this executes the ML force field in "prediction-only" mode. Then, energies and forces are always used directly from the ML prediction to drive the MD simulation and no ab initio calculation is ever performed.

On the other hand, with ML_ISTART = 0 or 1, whenever an ab initio calculation is triggered (check column 5 "doabin" of STATUS line in ML_LOGFILE) the corresponding INCAR settings are applied to carry out first principles calculations.

Best,
Andreas Singraber

Dear Andreas Singraber,

My problems were solved!

Thank you for all your help!!!

Best,
Kai