Number of basis functions in ML_FF capped below ML_MB

[andreas_mlkner]

Hello all,

during the learning process of a ternary liquid metal system the number of basis sets in the ML_ABN file caps at 2825 per element despite of the given ML_MB = 3500.
This behavior occurred at other systems with different metal combinations.

Attached are the initial Input-files and the header of the ML_ABN file (whole file too large to upload ~170MB) resulting from the calculation after ~800000 steps.
Due to walltime restrictions the calculation was restarted multiple times to continue the learning process.

Best regards,
Andreas

[ferenc_karsai]

ML_MB is just the maximum number of local reference configurations allowed to store.
Depending on your structure (liquid-solid, temperature range, etc.) and force field settings (ML_CTIFOR, ML_SCLC_CTIFOR, etc.) the required accuracy may be reached with much fewer number of local reference configurations.

As long as the accuracy of the force-field is satisfying for your problem, it is totally normal to see this kind of behaviour and the computational performance of the force-field will benefit from a lower number of local reference configurations.

[andreas_mlkner]

Different Systems at different stages of the learning process cap at the same number of 2825 basis functions per element, while the number of configurations in the ML_ABN file is growing.

Furthermore we observed multiple times the dissociation of single atoms into the gas phase during the running, which could be related to the insufficient number of basis functions.

[ferenc_karsai]

Which version of VASP do you use, can you please upload the ML_LOGFILE. The behaviour sounds like it's not the latest version.

[andreas_mlkner]

The Version is 6.4.0.

The ML_LOGFILE of a part of the learning process is attached. At this point the cap 2825 was already reached.

[ferenc_karsai]

Oh yeah 6.4.0 is a problem. Please use the latest version, we changed the behaviour of ML_MB in the latest version.

Previously ML_MB was the maximum size you were allowed to have including the buffer for new candidates etc.
Now ML_MB is the maximum target value for the number of local reference configurations for a single atom type but the buffer is usually added to this value .

The actual value used for allocation (including all buffers) is written out in the ML_LOGFILE as:
Maximum number of local reference configurations in memory (max. buffer size before sparsification)