Documenting experience with the Kolmogorov-to-grid scale ratio required for numerical stability

[glwagner]

@SandreOuza has found that dedaLES will run stably only if the ratio between the grid scale Δ and the Kolmogorov scale, defined as

ℓ = ∜(ν³ / ε) ,

where ν is viscosity and ε is turbulent dissipation, is below a certain number (perhaps around Δ/ℓ = 5). We should document this empirical rule. (Related to #19).

[sandreza]

Definitely one of the many interesting things about 3D turbulence is how quickly the small scales show up as compared to the 2D one! I am still working on getting a good intuition for exactly why the numerics breaks (or doesn't) for a given situation

[wenegrat]

Do you have a sense of why this is the case? It seems that, aside from the question of accuracy, these subgrid models should be able to scale to coarser grids and still remove sufficient grid-scale energy. If you are finding stability only in regimes where the grid scale is only ~5 times Kolmogorov, it might be possible that the model is running essentially as a DNS. That is, do the eddy viscosity terms matter to the stability in those runs, or would the model run stably with no closure? I'm thinking of a paper by Bill Smyth where he discusses DNS runs as often having 3-6 x Kolmogorov scale for grid spacing (I can dig this up if needed). This is of interest to me, as I am still failing to get reasonable geophysical type setups to run, regardless of closure type. It appears to me that in these runs the eddy viscosity often takes on negative values (despite the zero_max function). I wonder if this may be due to under-resolving sharp features in the calculated eddy viscosity/diffusivity. I've tried a test run increasing dealiasing, but that doesn't seem to solve the problem. A spectral filter could be added in addition to the eddy viscosity, but it would be nice to avoid this sort of ad hoc choice. Cheers, Jacob

…

On Thu, Mar 7, 2019 at 6:53 PM SandreOuza ***@***.***> wrote: Definitely one of the many interesting things about 3D turbulence is how quickly the small scales show up as compared to the 2D one! I am still working on getting a good intuition for exactly why the numerics breaks (or doesn't) for a given situation — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#20 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGIDJ1lFpMPHgyQzBHCgTtC8h0Eg-Byqks5vUdCWgaJpZM4bkeDC> .

[kburns]

It looks like the zero_max function in AMD is implementing a hard max rather than a soft max. This may be part of the problem -- computing the hard max on a certain grid may result in oscillations leading to negative values at different points.

[glwagner]

@kburns this wouldn't affect ConstantSmagorinsky though, correct?

Can we use a spectral filter on the eddy viscosity/diffusivity itself, rather than the solution?

[kburns]

Well I'm not sure which closure @wenegrat was referring to, but it doesn't look like we currently have any regularization on the constant smagorinsky eddy viscosity, right?

[glwagner]

Right @kburns there is no regularization on Smagorinsky; it is "guaranteed" > 0. So perhaps negative viscosity is only possible with regularization?

Still puzzled how LES is ordinarily done with spectral methods.

[kburns]

I think it's only guaranteed to be positive on the dealiased (scales=3/2) grid points, where it is calculated, but the spectral interpolant may end up being negative at other points, perhaps including the output grid (scales=1) if there are sharp features in the field.

[wenegrat]

Yes, I have noticed similar behavior (negative eddy viscosity) with the Smagorinsky closure as well, which as Greg says should be positive definite. Am I the only one seeing this behavior? Keaton, I understand your comments to imply that if I ran with no dealiasing that the negative values should disappear, is this correct? I can try to test this, although it might take me a few weeks to get to it.

…

On Tue, Mar 12, 2019 at 5:40 PM Keaton J. Burns ***@***.***> wrote: I think it's only guaranteed to be positive on the dealiased (scales=3/2) grid points, where it is calculated, but the spectral interpolant may end up being negative at other points, perhaps including the output grid (scales=1) if there are sharp features in the field. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#20 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGIDJ-C2liiP1zp-nHg1--xe9F4kKkDWks5vWEkSgaJpZM4bkeDC> .

[kburns]

It's a little subtle -- during the simulation, the subgrid viscosity will be calculated on the grid, where it should be non-negative, and immediately multiplied by the other terms to produce the subgrid stress. However, if you designate that you want to save the subgrid viscosity as an output, there may be a round-trip transform to coefficient space and back to grid space before it is saved to disk. Even without dealiasing, this may slightly change the grid values since the Fourier Nyquist mode is dropped during these transforms, so the FFT is not quite exactly 1-to-1 (as an aside, this is a bit of a pain, but when you think about applying operators in coefficient space, it doesn't make much sense to keep and work with the Nyquist mode). So this may result in slightly negative values, on the order of the Nyquist mode amplitude. This may seem bad, but it's also the order-of-magnitude of the truncated terms, so its probably a decent estimate of how negative the truncated spectral interpolant may be between the grid points, anyways.