Instabilites of dynamic core causes problems with AD

[pat-schmitt]

Instabilites of dynamic core causes problems with AD

[pat-schmitt]

A short update:

• I tried out different smoothing/filter strategies in this branch https://github.com/OGGM/COMBINE/tree/smoothing. Unfortunately, no one was successful...

• But in the meantime, @dngoldberg developed an implicit solver for the numeric core, together with an adjoint model. I did some optimisation of the code and the runtime looks quite promising. Also, the first tests show that the implicit solver is more stable and no instabilities occur, with faster overall runtime compared to the explicit solver. Therefore my plan is now to implement the implicit solver in COMBINE and compare the whole minimisation with the current implementation.

I will update my progress here!

[pat-schmitt]

In PR #13 I finally added the implicit solver to COMBINE.
The first test results look quite promising:

• The CPU time is comparable (a little faster) than the FluxBasedModel and less RAM is used for experiments where no instabilities occur with the FluxBasedModel.
• If instabilities occur in the FluxBasedModel the implicit solver is clearly better (faster, less memory, works...)

I am quite optimistic that this solved the problems with instabilities and AD (at least for the few test cases I have looked at so far) and will close the issue for now.

Thanks, @jrmaddison and @dngoldberg for your input!

[dngoldberg]

that is great @pat-schmitt! really promising! i can't recall if the tests are on realistic geometries? it looks like you define them here...

[pat-schmitt]

Yes, that is right, these are the five glacier geometries I use for my first idealized experiments.