A NN parameterization to test in the neverworld framework

[simone-silvestri]

This PR adds a NN closure to the neverworld representing the subgrid-scale fluxes of unrepresented mesoscale eddies in the large-scale circulation.

To be tested with a grid-size of 1/4 or 1/2 of a degree.
The neural network (provided by @Etienne-Meunier) uses values from https://github.com/chzhangudel/Forpy_CNN_GZ21/blob/smartsim/testNN.py

PR in collaboration with
@Etienne-Meunier and @vopikamm

[glwagner]

@xkykai, why don't you improve Oceananigans' RiBasedVerticalDiffusivity?

[simone-silvestri]

cc @xkykai

[xkykai]

@simone-silvestri should this be the nonlocal Ri-based parameterization that we have implemented? Let me retrain the model with varying Prandtl number in positive and negative Ri regimes then I can incorporate this.

[glwagner]

@simone-silvestri should this be the nonlocal Ri-based parameterization that we have implemented? Let me retrain the model with varying Prandtl number in positive and negative Ri regimes then I can incorporate this.

It looks like you have an improvement to the current RiBasedVerticalDiffusivity, right? We want the best model, whatever that is

[xkykai]

Yes I have, let me retrain with some slight modifications then I'll put it in!

[xkykai]

@simone-silvestri could you add me as a collaborator? I have updated the XinKaiVerticalDiffusivity with new parameter values as well as allowing 2 different turbulent prandtl numbers in convective and shear-driven regime, but I have no permission to push. It has also been calibrated against heating and precipitation scenarios. This scheme will preserve shear better than the current one.

[simone-silvestri]

@xkykai I added you as a collaborator, but I think the idea was to add this parameterization directly to Oceananigans (not here) and to remove the parameterization from here.

I think the idea is to modify the formulation and parameters RiBasedVerticalDiffusivity since this parameterization is basically the same in spirit with some minor changes in the functional form of the diffusivity and viscosity and in the parameters

[xkykai]

Okay, in practice that means renaming XinKaiVerticalDiffusivity into RiBasedVerticalDiffusivity right? Should I open a draft PR in Oceananigans to do that then?

Also do you think we should still keep the original RiBasedVerticalDiffusivity instead of modifying it? I feel like there might be merits in keeping it since it might have some advantages compared to the one I have right now.

[glwagner]

Okay, in practice that means renaming XinKaiVerticalDiffusivity into RiBasedVerticalDiffusivity right? Should I open a draft PR in Oceananigans to do that then?

Also do you think we should still keep the original RiBasedVerticalDiffusivity instead of modifying it? I feel like there might be merits in keeping it since it might have some advantages compared to the one I have right now.

What is the difference? They seem very similar, can they be implemented within a common framework using simple abstractions?