My Community

Dear VASP users,

When restarting ML MD (ML_ISTART=1) from a previous ML_AB(N) file, the following two things happen which I want to optimize. Version=6.3.2

(1) ML_MCONF=1500 and ML_MB=1500 are set to a new value (as quoted below) which is similar to the restarted ML_AB file + a small overhead (see wiki quote below). Often this increase is too small and leads to premature crashes. This seems to assume that the first calculation will cover a lot of the configuration space for learning. Often the values of ML_MCONF and ML_AB are below the original 1500. Trying to specify these in the INCAR on restart is overwritten (seen in the ML_LOGFILE) to values based on the ML_AB restart file. Can I fix this?

(2) Restarted runs still do the first 10 steps with ab initio calculations. I think this should not be the default for ML_ISTART=1, and only for fresh calculations (i.e. ML_ISTART=0). Perhaps this can be fixed manually by changing "ML_MHIS" if this is defining the 10 steps that have to be done ab initio? Or is there another way.

Thanks, Victor

Caution: number of structures and basis functions

The maximum number of structure datasets ML_MCONF and basis functions (local reference configurations) ML_MB potentially constitutes a memory bottleneck, because the required arrays are allocated statically at the beginning of the calculation. It is advised not to use too large numbers initially. For ML_ISTART=0, the defaults are ML_MCONF=1500 and ML_MB=1500. For ML_ISTART=1 and 3, the values are set to the number of entries read from the ML_AB file plus a small overhead. If at any point during the training either the number of structure datasets or the size of the basis set exceeds its respective maximum number, the calculation stops with an error message. Since the ML_ABN is continuously written during on-the-fly learning, not all is lost though. Simply copy the ML_ABN to ML_AB and CONTCAR to POSCAR, increase ML_MCONF or ML_MB, and continue training (see restart the calculation).

1) Will be fixed in the newest update 6.4.1. The new default for continuation runs is:
MB_AB + min(1500, max(NSW , 2*ML_MCONF_NEW * MAXAT_SP)

where
MB_AB = greatest number of local reference configurations within all species in the ML_AB file.
MAXAT_SP = greatest number of atoms within all species among the current structures and the structures in the ML_AB file (if present).

This information will be soon online on the ML_MB wiki page.

2) The 10 new steps are because the default value for ML_CTIFOR is and the new value is only obtained at the 10th step. As you mentioned one could try to set ML_MHIS to a smaller value but that doesn't avoid the problem rather just makes less steps. The simplest solution is to set ML_CTIFOR to the last value from the previous learning. THen the code will take this avalue as the starting value for the autmatic adaptation of ML_CTIFOR (if ML_ICRITERIA=1 is used).

Hi Ferenc,

Thanks for that. Glad to hear that (1) is changing on restart mode. I also wondered if the MD step counter in the new OUTCAR could even reflect this (i.e. starting at step 4212 rather than 1).

For (2) this is very useful, I thought it would be ML_CTIFOR. I wonder if this could also update automatically on restart mode, like the changes for mentioned (1).

(3) I also noticed that "ML_ISTART=1 and NSW=0" still did 1 ab initio step before fitting the potential from ML_AB. While the wiki wiki/index.php/ML_ISTART says there should be no ab initio step.

I also noticed using ML_ISTART=4 to make an FF from an ML_AB file gave me segfaults.