Operator prediction becomes constant

[dezenn]

I train a DeepONet model to approximate a ODE system of my own.
As a loss I get the following

and the results looks like this

where the red curve are the predictions of my learned operator on some of the test set curves and the blue curves are some of the target curves from the test set.

I do not think that I have an error in my plotting code because I do get different prediction curves when I only train for like 10 iterations/epochs, so I really wonder why I get this strange behaviour.
PS: the loss is so high compared to the test metric as one sees the non-standardized loss but the test metric is a relative error.