Restarting MLFF retraining

[sonaguluzade]

Hello,

I have trained a force field using default cutoff value (5 A). Now I am trying to retrain this FF with larger cutoff (6 A) using following settings

ML_LMLFF = .TRUE.
ML_ISTART = 3
ML_CTIFOR = 0.008
ML_MCONF_NEW = 16
ML_CDOUB = 4
ML_ICRITERIA = 0
ML_RCUT1 = 6.0
ML_RCUT2 = 6.0
ML_MB = 3500
ML_LBASIS_DISCARD = .TRUE.

and ML_AB file edited with all the 1330 structures and all the basis sets as following

The numbers of basis sets per atom type
--------------------------------------------------
1 1 1
**************************************************
Basis set for Li
--------------------------------------------------
1 1
**************************************************
Basis set for Si
--------------------------------------------------
1 1
**************************************************
Basis set for S
--------------------------------------------------
1 1

The training is pretty long, so it does not fit to single walltime. At the end of training I am getting ML_ABN file only with structures that passed through the retraining process. Is there a simple way of restrarting the retraining from where it left off, other than doing another retraining with reamaining structures and merging the local configurations in a single file? Morever, would you recommend other tags to speed up the retraining?

Thank you very much for helping out!

[jonathan_lahnsteiner2]

Dear sonaguluzade,

If I understand correctly you already have a ML_AB file and want to retrain the force field with larger cutoff.
What you are doing is actually not a retrain of your force field. What you are selecting with ML_ISTART=3
is a reselect of local reference configurations. I would recommend you to take a look
at the vasp wiki. There is a page about machine learning best practices.
wiki/index.php/Best_practices_for_machi ... rce_fields
In principle it is not necessary anymore to use ML_ISTART as mentioned on its wiki page. It is recommended to use the
ML_MODE tag instead wiki/index.php/ML_MODE
So if you want to refit the selected ab-initio data with a different cutoff radius you should
use ML_MODE = refit. This will be much faster compared to ML_ISTART=3/ML_MODE=select.
I hope this helps to answer your question.

All the best Jonathan