Exlucing stres learning in MLFF

[burakgurlek]

Hi,

I have an ab inition data that I would like to use MLFF. However, I do not have the stress of the structre which seems to be important in learning force filed (e.g. wiki/index.php/Machine_learning_force_field:_Theory). Do you know is there a way to turn-off learning stress and just keep energy and forces?

Regards,

[ferenc_karsai]

Yes, there is a way.

Set ML_WTSIF to a small value like ML_WTSIF=0.0001 in the INCAR file. This tag sets the relative weight of the stress compared to forces and energy.
Caution: You still need to have some dummy values for the stress in the ML_AB file.
Here is the documentation of the flag:
wiki/index.php/ML_WTSIF

[burakgurlek]

Dear ferenc_karsai,

thanks for your reply. I did your suggestion, but ML_REG file ?shows values like 1e2 for the fitted stress values, although my dummy variable in the ML_AB file is zero for all structures. Does it normal or I should lower ML_WTSIF further?

[ferenc_karsai]

The fitted values for stress are garbage, but you weight them with a very small factor when solving the normal equations so it doesn't matter.