Noise below the mixed layer when using AnisotropicMinimumDissipation

[jhdong2016]

Hi all, I am simulating an ocean surface mixed layer. However, I always get lots of noise in the ocean interior below the mixed layer. I have applied a Gaussian-mask sponge layer near the bottom. But the noise is still there. Below are two figures of the vertical velocity and buoyancy perturbations. Any suggestions?

Here are the codes for the sponge.

ps. the whole depth is Lz = 150 m and the mixed layer is about 80m.

######========================================
damping_rate = 1/1800 # relax fields on a 100 second time-scale
bottom_mask = GaussianMask{:z}(center=-grid.Lz, width=grid.Lz/3)
uvw_sponge = Relaxation(rate=damping_rate, mask=bottom_mask)

model = NonhydrostaticModel(; grid, coriolis,
advection = WENO(),
timestepper = :RungeKutta3,
tracers = :b,
buoyancy = BuoyancyTracer(),
closure = AnisotropicMinimumDissipation(),
stokes_drift = UniformStokesDrift(∂z_vˢ=∂z_vˢ),
boundary_conditions = (v=v_boundary_conditions, b=b_boundary_conditions),
forcing=(u=uvw_sponge, v=uvw_sponge, w=uvw_sponge))
#####============================================

[navidcy]

Hi @jhdong2016, thanks for your interest in Oceananigans!

Personally I find it very difficult to help out without something that I can run and reproduce your results. Could you post your script somewhere also for making the plots. Something I can copy/paste in the REPL or just write include("jhdong_script.jl")?

It's often useful to post the version of Oceananigans you are using (and the rest of dependencies). E.g., post the output of using Pkg; Pkg.status().

Do I understand correctly or is it my impression that you have depth in the horizontal axis?

[jhdong2016]

THanks for the reply. Here are the files that may help.

si_langmuir_case0.8m.jl.txt

[navidcy]

Thanks!

You seem to be using a very very old version of Oceananigans; v0.79.0 (from 2022).
Could you try to update to the latest version and see if the issue(s) remain?

[jhdong2016]

Thanks for the suggest! Let me update it and see if the issue is fixed.

[glwagner]

This is common to observe with AnisotropicMinimumDissipation. If you don't need the dissipation rate or eddy diffusivities explicitly, you could try using closure=nothing and advection=WENO(order=9) instead. This might solve your noise problem and also give you a higher effective resolution.

Can we convert this to a discussion? I don't think this is a bug in the code.

[glwagner]

@tomchor might have more to say

[glwagner]

And yes more generally it's best if you include your whole script so that the simulation can be reproduced. In particular here we need the grid, initial condition, and time step to fully grasp what's doing on.

[jhdong2016]

This is common to observe with AnisotropicMinimumDissipation. If you don't need the dissipation rate or eddy diffusivities explicitly, you could try using closure=nothing and advection=WENO(order=9) instead. This might solve your noise problem and also give you a higher effective resolution.

Can we convert this to a discussion? I don't think this is a bug in the code.

Please convert it into a discussion. And I am updating the version and will try your suggestion! Thanks!

[glwagner]

I'd like to revisit this. @jhdong2016 can you provide your entire script? Alternatively, @tomchor do you have a script that illustrates the noise issue? I would like to see if there is a simple fix.

[glwagner]

I'd suggest changing one thing at a time from my working example.

They are already identical except for parameter choices and adaptive time stepping.

[glwagner]

Also just as a PS --- powers of 2 are fastest for NonhydrostaticModel due to the FFT.

Multiples of 10 are ok, because cufft supports prime factorizations of 2 and 5, for sure. However, power of 2 are still faster. I think (but am not sure) that powers of 3 are faster than powers of 5. So it is nice to follow a convention that uses just powers of 2 and 3.

From an aesthetic point of view, powers of 2 are actually nice. It's the least arbitrary basis. (It's believed we use base 10 because we have 10 fingers, which is also kind of cool but still, a bit more arbitrary.)

Personally I think communication is enhanced and mental noise is reduced if we always use powers of 2 and 3 for resolution, plus grid spacings that are simple (ie do not combine "Lx = 100" with "Nx = 64" --- use "Lx = 128" instead).

You can read more here: https://docs.nvidia.com/cuda/cufft/

[glwagner]

The other parts of the "MWE" that I found to be non-minimal were the interpolation of output to cell centers. This is extraneous to the AMD issue (hopefully)... but complicates the script since I spent some time reading it. I also think MWEs should use JLD2 at the moment, since NetCDF still (and unfortunately) does not support FieldTimeSeries which complicates visualization. I think the "minimal" part of MWE should really be taken seriously.

[tomchor]

I agree with you. If you read my post I called it a "small working example", not a minimal one ;)

I interpolated the outputs to make plotting easier with NetCDF (locations are the same) and I was using NetCDF because I can explore the solutions faster with xarray. I of course could have spent the time to trim down those things before posting, but I was running out of time and I wanted to make sure I posted it here, rather than have the example forgotten in my laptop (which is what happened the last time I investigated this issue).

[glwagner]

If FieldTimeSeries is missing useful features from xarray then we should have an issue open about it I think