My Community

Hi,

I have some questions regarding the MLFF (i'm using vasp6.4.1v):

1) These are my INCAR file to generate MLFF, and I have been comparing both the geometry configuration and the single point calculation energies against the ab initio data. I think the fitted MLFF was qualitatively capable of capturing energy trends for my different geometries against the ab initio data. However, when compared against absolute energies the fitted MLFF energy predictions are overpredicting by ~0.8 eV lower compared to ab initio calculation energy. I would like some advice on possible any parameter that I can tune to close this discrepancy.
This is the INCAR file I'm using currently to fit the force field and I have used the same parameters(ex. ENCUT, EDIFFG, ISIF, etc.) excluding the MLFF training parameters as INCAR for the ab initio energy calculations.

ML_LMLFF = .TRUE.
ML_MODE = train

ENCUT = 400
PRECT = Normal
ALGO = Fast
LREAL = A
ISMEAR = 0
SIGMA = 0.1
EDIFF = 1E-05
EDIFFG = -0.05
NELM = 400

LCHARG = .FALSE.
LWAVE = .FALSE.
IVDW = 12
LDIPOL = .TRUE.
IDIPOL = 3
DIPOL = 0.5 0.5 0.5

MDALGO = 1
ISYM = 0
POTIM = 0.5
ML_CTIFOR = 0.02
ML_CX = -0.1
IBRION = 0
ISIF = 2
NSW = 6000
TEBEG = 500
ANDERSEN_PROB = 0.10

In addition to question 1 ), I even have attempted to validate the MLFF estimated energies using the training dataset as the validation dataset. Despite this redundancy for this validation process, MLFF was unable to estimate the energies with the same structure used in training (stills showing some discrepancy in energy although it could be able to predict the final geometry configuration even in geometry optimization).

Dear exk301,

To train a machine learning force field, the first step is always to collect local reference configurations.
After this we recommend to do a refitting with ML_MODE=REFIT. The discrepancy between your DFT data
and the MLFF is because your force field is not yet accurate enough. This has nothing to do with your
DFT settings in the INCAR file.
I would recommend you to read the Introductory and Best practice page about machine learning force fields which
can be found here
https://www.vasp.at/wiki/index.php/Mach ... ns:_Basics
and here
https://www.vasp.at/wiki/index.php/Best ... rce_fields

Maybe you also find the theory and tutorial page interesting:
https://www.vasp.at/wiki/index.php/Mach ... ld:_Theory
https://www.vasp.at/tutorials/latest/md/part2/

You should be aware that the machine learning force field and the DFT energy will never be spot on the same. The machine learning
force field will always show some deviation from the DFT calculation because of the final number of local reference configurations that can be picked up. Energy errors in the meV range per atom are tolerable when working with machine learning force fields.

I hope this helps to answer your question. If not please contact us again.

All the best Jonathan