Learning parameters to train with stress, forces and energy

[mesonepigreco]

Dear developers,
I am trying to train a nequip potential with data with different cell shape and a different number of atoms.
I followed the example files provided, however I am getting not very good values of errors in energies and forces and I am worried I'm missing some of the meaning of the hyper parameters that could maybe be optimized better:

For example, I'm using a loss_coeffs of:

What is the meaning of these parameters? how can they be optimized?

Also, this is how the validation plots look (taken form wandb.ai, validation energy/N_mae and validation_forces_mae on the two atomic species; units are eV and A). I see a lot of sharp peaks toward high values. it seems there s an overall trend to go down, but seems there are this spikes that are hampering getting to very good precision. Is it normal or it is a signature of something wrong happening with my dataset/parameters?

Thanks for the attention,
Best regards,
Lorenzo Monacelli

[Linux-cpp-lisp]

HI @mesonepigreco ,

For example, I'm using a loss_coeffs of:

Was this problem introduced by adding stress to your training, or this is just an issue you are facing in general?

Without knowing more, I would suggest trying a lower learning rate... spikes like this can be a sign of the training "blowing up" when it takes too large a step in weight space.

[mesonepigreco]

Hi Alby,
Thanks for the response,
This was a general issue I was facing even before introducing the stress. I am now trying to lower the learning rate (from 0.005 to 0.001) and keep you updated if the issue is solved.
Cheers,
Lorenzo