How to make sure the learned physics doesn’t contain the dynamical part during traing?

[weatherforecasterwhai]

If in train, the learned part can not divide from dynamical part,in inference, adding them together will double count the mixed part?

[shoyer]

I'm not sure exactly what you're trying to do, but you will probably need to modify the code.

[weatherforecasterwhai]

I just try to understand how could neuralGCM be much better than other pure data-driven models in my running case.
And how to make use of it.