Requests for technical support from the VASP group should be posted in the VASP-forum.
ML FF ISAMPLE
ML_FF_ISAMPLE = [integer]
Default: ML_FF_ISAMPLE = 3
Description: This tag controls the sampling in the machine learning force field method.
- ML_FF_ISAMPLE=1: If the estimated error is larger than the pre-determined threshold (ML_FF_CTIFOR for the Bayesian error, or ML_FF_CSF for the spilling factor), a first principles calculation is performed and the structure is added to the first principles dataset. If the number of structures in the data set reaches ML_FF_MCONF_NEW the force field (FF) is updated. Collecting a set of structures allows for efficient blocking strategies in the FF update and makes the calculations significantly more efficient.
- ML_FF_ISAMPLE=2: If the estimated error is ML_FF_CDOUB times larger than one of the thresholds (ML_FF_CTIFOR or ML_FF_CSF), a first principles calculation is performed and the force field is immediately updated. If the estimated error is larger than one of the thresholds (ML_FF_CTIFOR or ML_FF_CSF), the structure is added to the first principles dataset. As before, the FF is only updated after collecting ML_FF_MCONF_NEW datasets.
- ML_FF_ISAMPLE=3: This is similar to ML_FF_ISAMPLE=2. The only difference is that as long as the upper threshold is not exceeded (e.g. ML_FF_CDOUB times ML_FF_CTIFOR), at least ML_FF_NMDINT MD steps are preformed using the MLFF (i.e. no first principles calculation is performed). This is the default method, and avoids that many nearly identical structures are added.
We do not recommend to change this from the default value.