Unreliable Oceananigans extension. Need help keeping up with breaking changes.

[gaelforget]

Example for Oceananigans gives wildly different results depending on version being used.

@simone-silvestri , this is what I was telling you about yesterday.

Results are from running a version of this example, with diurnal forcing for a few days.

• Short term : would be great if we could review the v0.103.1 version together and correct it together if needed.
• Longer term : would be lovely if one of the dev could send a PR when breaking changes are made to Oceananigans. It's been really hard keeping up without any help.

[simone-silvestri]

There seems to be something wrong actually. Eddy viscosity it much much larger than before, convection is much faster and there is a lot of noise in the vertical velocity field. Is this that you are plotting the horizontal mean in time?
(I am upstairs in the lobby if you want to look at it together)

[simone-silvestri]

Can you try running the example using the previous AMD default? The default value of the Poincare coefficient was changed here CliMA/Oceananigans.jl#4494 from 1/12 (lower viscosity / diffusivity) to 1/3 so maybe that is the difference?

[glwagner]

cc @jagoosw who made the change and @tomchor who has a lot of experience with LES.

Results are from running a version of this example, with diurnal forcing for a few days.

the link above is broken. Can you provide another link?

In general I suggest using WENO only with closure = nothing (the default), which is simpler. If you would like to retain a turbulence closure, you might try using SmagorinskyLilly or DynamicSmagorinsky to see if they produce better results.

[gaelforget]

Can you try running the example using the previous AMD default? The default value of the Poincare coefficient was changed here CliMA/Oceananigans.jl#4494 from 1/12 (lower viscosity / diffusivity) to 1/3 so maybe that is the difference?

Thanks! Will run again

[glwagner]

I believe the example in question is:

https://github.com/gaelforget/ClimateModels.jl/blob/master/examples/Oceananigans.jl

@gaelforget please correct if wrong.

[glwagner]

PS the title of this issue gave me at first impression that there was a security risk associated with Oceananigans.

[gaelforget]

I believe the example in question is:

https://github.com/gaelforget/ClimateModels.jl/blob/master/examples/Oceananigans.jl

@gaelforget please correct if wrong.

correct

[gaelforget]

Can you try running the example using the previous AMD default? The default value of the Poincare coefficient was changed here CliMA/Oceananigans.jl#4494 from 1/12 (lower viscosity / diffusivity) to 1/3 so maybe that is the difference?

Thanks! Will run again

Fixed!

[glwagner]

It's unnecessarily provocative to refer to expected changes due to improvements in subgrid scale parameterizations as "unreliable".

[glwagner]

The real issue here is that the example is flawed rather than "unreliability". I've attempted to address them here: #144

It may need more work.