MLFF Mode SELECT & ALGO = All

[rwindiks]

Dear Ferenc, dear Andreas,

please perceive this ticket as a follow-up of our discussion last week, which I enjoyed very much.

Regarding, the SCF algorithm preconditioned conjugate gradient aka ALGO = All: Please reduce the default time step to 0.2. With this, the SCF procedure has a much better performance in terms of reliability and amount of steps than with the current default time step of 0.4.

Regarding the ML_MODE = SELECT: I would appreciate very much if you outline how this mode can be used to reduce the MLFF training set from, e.g. 6000 configurations per element to, e.g 3000 or even more with maintaining the accuracy. Already you provided good advices at wiki/index.php/ML_MODE However, I understood there are a few more MLFF flags to consider.

Definitely, I'll take your advice very seriously to examine generated MLFFs with a validation test set which can be easily done within the MedeA software environment.

In case you want to continue the discussion outside the VASP forum then feel free to contact me via rwindiks@materialsdesign.com

Thanks.

Rene

[rwindiks]

Dear forum administrator,

I am glad that we can discuss the topics in the subject in this forum, as they can be interesting for a plenty of VASP users, too.

Also, I want to apologize that I forgot to address Jonathan in my first message, as he was also involved in the fruitful face-to-face discussion last week.

Thanks.

Rene