Facundo Sapienza

Hi @albangossard ! Thanks for the comments, I will address them all soon and merge this PR. I completely agree with your comment on having a small glacier (I actually...

@albangossard I fixed the original problem with the docs (which actually forced me to change the API, which I don't know if it is a desirable behaviour of the docs)...

I agree with you @albangossard ! There should not be averaging with respect to the initial condition. However, notice that in general the loss function should include a weight (a.k.a...

One source of the error here is probably how we impose the boundary condition and how we do the clipping. This definitively add some artifacts on the border. Another source...

Also, for the Halfar solution we can probably design a loss function for which the derivative is analytical: no need to test agains Enzyme or something else, since H(x,y,t) is...

I am continuing this discussion here, which is attached to the [PR in Huginn](https://github.com/ODINN-SciML/Huginn.jl/pull/87). As mentioned in the paper Bueler et al 2005, there is a source of error coming...

Also important for this discussion, it is important to see that the effect of the border condition does not increase as we move forward in time. Here a longer simulation...

Hi @albangossard @mleseach @JordiBolibar Thank you for putting this together. I think I will have to get a more careful explanation on our next meeting to understand a bit the...

Following my comment in Slack, I think it makes more sence to move things to ODINN (and not Huginn) since the type of inversion will determine some internals of the...

Thank you for the comments @albangossard ! I partially disagree with the first argument, but I fully agree with your second argument. **However, the VJPs will depend of the neural...