Devasmit/AD_gsoc

[DevasmitDutta]

No description provided.

[oriolcg]

@DevasmitDutta I added a minor comment. The priority and intensity need to be filled in as well.

I think it would be good to involve someone that has been involved in the latest devs of GridaDistributed. @amartinhuertas @JordiManyer , are you interested?

[JordiManyer]

Hi @oriolcg , @DevasmitDutta . Sorry if I may be ruining the project, but this is done already. It's just not part of the main repo. I plan on bringing it to GridapDistributed at some point, when time allows. If it's something you need, I can speed it up.

The main idea is that automatic differentiation is a cell-wise thing, which means we want it to happen locally. However, when we have a functional defined by an anonymous function, like we generally do, the Measure attached to the weakform is a DistributedMeasure. This means we cannot evaluate that functional locally.
Because of how anonymous functions work in Julia, we cannot retrieve the DistributedMeasure from the function, so it is not possible (without rewriting most of the AD code for distributed) to keep using anonymous functions. This is even more tricky when we have multiple triangulations/measures involved.

All in all, the solution to this problem is quite elegant, and involves creating a new structure

struct IntegrandWithMeasure{A,B<:Tuple}
  F  :: A       # F(u,v,dΩ...)
  dΩ :: B    # Measures that F needs
end

which holds the same info as the anonymous function, but in a way that can be accessed. The evaluation and gradient functions are then quite trivial.

Moreover, this also has benefits when considering functionals that depend on more variables than just u and v. For instance, in Topology Optimization scenarios you have functionals that depend on the level set functions. We want to be able to take directional derivatives wrt those variables as well. This structure allows us to do so.

[DevasmitDutta]

@oriolcg looks like there might be a new release to GridapDistributed where this is probably already done by @JordiManyer. So, now having said that, would you like to propose some new topic for the GSoC ? (@amartinhuertas @JordiManyer probably you may also like to chime in for proposing any new topic that might be good to work on with GridapDistributed for this year's GSoC as well. )

[JordiManyer]

Just as another comment, in the PR that is currently been considered to overhaul the ODE module, Alex and I have fixed a couple edge cases where AD in serial failed. We might also try to solve the AD for multifield bug.

[JordiManyer]

See changes in gridap/Gridap.jl#978 and gridap/GridapDistributed.jl#144

[oriolcg]

@DevasmitDutta, given that these developments are already on the way, I would suggest to look into another topic. I will propose the same topic as last year on adjoint-based PDE constrained optimization.

You can close this PR and add this new one on the list.