Training MLFF to existing ab initio data

[payam_kaghazchi3]

Dear all,

we are currently trying to fit a MLFF using existing ab initio data in VASP 6.4.3 (CPU, MPI, 1 node, 128 cores). In accordance to the documentation we created our own ML_AB file from the data and are using ML_MODE=select.

However, there are two issues:
1. When following the documentation of ML_AB and introducing a dummy basis set section of 1 basis set per atom and 1 1 for every local reference, no ML_* files are written after the calculation.
2. We overcame the first issue by incorporating a longer dummy section, e.g. 3 basis set per atom of 1 1, 2 2, and 3 3 local reference structures. Then all output files are written, however, the ML_ABN file has the same contents as the initial ML_AB, e.g., still a dummy selection rather than proper local references. The same applies to the ML_FFN file. We also tested to manually increase the ML_MCONF and ML_MB tags by setting their values in the INCAR file explicitly as suggested in the documentation for ML_AB but this had no effect.

The calculations are always terminating normally and there are no (obvious) warnings or errors in the corresponding out-files (stdout, stderrt, OUTCAR, ML_LOGFILE).

We would be grateful for any advice on how to proceed in fitting a MLFF with VASP to our existing ab initio data and attach the input and output files for a minimum working example with 5 reference configurations of a simple system with 3 ions.

Many thanks in advance and best,
Konstantin Köster

[max_liebetreu]

Hello Konstantin,

Thank you for reaching out to us, and welcome to the VASP Forum!
I took a look at your input files and noticed a few things.

• For one, your ML_MODE=refit, not select. You're probably aware of this, and likely want to deliberately use this setting (see ML_MODE).
• For ML_MODE=refit, as per the wiki entry:

  Sparsification is performed on the local reference configurations, so the resulting ML_ABN file will contain the same number or fewer local reference configurations than the ML_AB file.

  This is in contrast to ML_MODE=select, which determines new selections.

In short, I believe there is some confusion regarding ML_MODE=select and refit, which each have their own unique behavior. For ML_MODE=refit, you can reference ML_AB or, perhaps more relevant, our MLFF tutorial. You can also download an example ML_AB file from the tutorial page to see how to set up such a file for ML_MODE=refit.

Please let me know if that helps, or if you require further clarification!

Best regards,
Max

[konstantin_kster]

Dear Max,

thanks for the prompt reply. Sorry for providing the wrong test input file. I am aware of the difference between select and refit, the described behaviour was actually correct for refit and the code runs smoothly with select.

In my real calculation I had the select mode but another issue with my self-generated ML_AB file. Once I fixed my ML_AB file, the calculation also starts selecting. I apologize for the confusion.

Again, many thanks and best,
Konstantin

[max_liebetreu]

Dear Konstantin,

Glad I could help, and no worries. I'm happy to hear your calculation is now running as expected!
Good luck with your training!

Best regards,
Max