Facundo Sapienza

Completely agree! I was postponing this for some time, but worth creating the issue for it!

@ChrisRackauckas @wsmoses @avik-pal we have fix/improve the following features in the ODINN architecture 1. As @JordiBolibar mentioned, all code is now pure Julia, so hopefully all problem that potentially could...

@JordiBolibar I started doing this in Sleipnir (see [here](https://github.com/ODINN-SciML/Sleipnir.jl/blob/b76862b9852bd03dc62cd15a47cc0f210cd0ea72/src/glaciers/glacier/Glacier2D.jl#L112-L125)), it is very simple but it does require playing a bit with print functions.

We should. Why you don't like the @info flags? They look great.

Jaja @albangossard I think this is not your first opened issue jaja you had been opening a long list of (very good!) issues so far.

@albangossard @JordiBolibar I think this PR is ready for some minor small changes and a merge. Please let me know if you have some final thoughts on this, or if...

@albangossard maybe we can discuss this points next week. I am a bit confused about how to interact with the cache here.

@albangossard I have addressed all your points. The main thing that is left is to make the target D to work with the velocity loss. I tried to do this...

Yes! Let's discuss all this tomorrow. I would like to keep the custom NN test (at least one) since I noticed in many cases the simple NN tests work, but...

Notice that the performance of these in the Halfar tests is actually very bad: ***Continuous adjoint, discrete VJP*** ```bash l_enzyme=[1869.7360497324557] ∂A_enzyme=[9.91829420032986e20] ratio = -3.23e-01 angle = 2.22e-16 relerr = 4.77e-01...