High error in energy

[baoanh13]

Hello,

I am trying to train a model of a system that contains > 4000 training data with 3 type of elements: Fe, C, H, by Nequip and Allegro. The error in training session by nequip seems good but the training time is rather slow, so I tried Allegro to accelerate, but it seems the error in training session by Allegro is not good, especially in energy.

This is the latest results after 2 days for Nequip:

And this is by Allegro after 3 days:

In the train.yaml file, in both case I used the loss coeffs as the default
loss_coeffs:
forces: 1.
total_energy:
- 1.
- PerAtomMSELoss

I don't know what is the reason causes the bad performance in energy training by Allegro. Are there any better loss function values or another hyperparameters to improve the accuracy?

I attach the yaml file for training in both cases I use as below.

train_allegro.txt
train_nequip.txt

Thank you for your concern.