Enzyme and structures and Oceananigans

[swilliamson7]

A potentially vague question/issue regarding how Enzyme works. In order to run the Oceananigans barotropic gyre model, we need to initialize this structure

model = HydrostaticFreeSurfaceModel(; grid, closure,
                                   free_surface = ExplicitFreeSurface(),
                                    momentum_advection = VectorInvariant(),
                                    coriolis = HydrostaticSphericalCoriolis(),
                                    boundary_conditions = (u=u_bcs, v=v_bcs),
                                    tracers = nothing,
                                    buoyancy = nothing)

which contains all the needed information about the model that we're running, and so needs to be given to time_step! to update the velocity fields and so on. The function that I call autodiff on currently takes this as input, and so we need to wrap it in Duplicated, but the vast majority of the entries in model are not something we want to take the derivative with respect to. Is there a way to tell Enzyme to take this into account?

Another weird thing is that the choices (such as ExplicitFreeSurface()) will impact the derivative, so we do still want Enzyme to go through those functions, meaning they shouldn't be ignored in their entirety (if any of that makes sense). I guess I'm a little lost about how best to set this up to try and make the MWE of Enzyme and Oceananigans not getting along, but I think understanding how to deal with this structure is at least a start.

[swilliamson7]

Maybe @vchuravy or @wsmoses could help a bit?