Training with virials

[Sokseiha]

Hi,

I've been training NequIP with energies and forces from DFT and so far, everything works very nicely. However, recently, I started to add virial in the training and things become, what should I say, strange.
As you can see from the learning cures, the the validation error for the virial plateaus while the training error keeps decreasing. I suspect that it has to do with the coefficients in the loss function so I've tried a few different values but that doesn't help. I'm wondering if any of you have seen this behavior before and if so, I'd really appreciate if you can suggest some ideas I can try to fix this problem.

Cheer.

[Linux-cpp-lisp]

Hi @Sokseiha ,

This just looks like the virial overfits sooner than the energy and forces. Which model gets picked will be determined by the best model on validation_loss, so yes, the balance will affect that. You can try weighting virial lower (virial is an extensive property, so PerAtom loss should also be used).

But I have seen this before, it doesn't look insane.