Default MLFF hyperparameter values discrepancy between vasp 6.4 and 6.4.2
Posted: Thu Jan 04, 2024 2:54 pm
Hello all,
I was hoping to get clarification on what the default values for MLFF hyperparameters should be.
I noticed when running an MLFF training using vasp 6.4.0, and without explicitly setting any values for the various hyperparameters, some of the values for the hyperparameters as written in the ML_LOGFILE are different than the default values stated on the vasp wiki. For example, ML_RCUT1 default should be 8 angstrom according to the vasp wiki, but in the ML_LOGFILE it is written as 5 angstrom whenever I do not explicitly define the value in the INCAR file for the calculation.
I also tested this similar calculation - doing an MLFF training without explicitly setting hyperparameter values in the INCAR file - using vasp version 6.4.2 instead, and found agreement between the hyperparameter values written to the ML_LOGFILE and the default values stated on the vasp wiki. Maybe this was an issue in the earlier version then?
I attached a zip containing input and standard output and ML_LOGFILE files for each calculation (6.4.0 and 6.4.2).
Thank you,
Dylan
I was hoping to get clarification on what the default values for MLFF hyperparameters should be.
I noticed when running an MLFF training using vasp 6.4.0, and without explicitly setting any values for the various hyperparameters, some of the values for the hyperparameters as written in the ML_LOGFILE are different than the default values stated on the vasp wiki. For example, ML_RCUT1 default should be 8 angstrom according to the vasp wiki, but in the ML_LOGFILE it is written as 5 angstrom whenever I do not explicitly define the value in the INCAR file for the calculation.
I also tested this similar calculation - doing an MLFF training without explicitly setting hyperparameter values in the INCAR file - using vasp version 6.4.2 instead, and found agreement between the hyperparameter values written to the ML_LOGFILE and the default values stated on the vasp wiki. Maybe this was an issue in the earlier version then?
I attached a zip containing input and standard output and ML_LOGFILE files for each calculation (6.4.0 and 6.4.2).
Thank you,
Dylan