Hi all,

I was wondering if anyone has already validated the implementation of IDDES for planar channel flow.

I am currently validating IDDES on a channel flow at REtau=395 and I am using the hybrid convection scheme, which reduces to almost a full CDS. The flow is well resolved with 96x96x96 grid points.

After fixing the bugs in IDDESDelta.C and in the function alpha() I still cannot reproduce the results presented in the IDDES paper.

According to the paper I should be able to match the log-layer nicely. In my case, I significantly over-prdict the log-layer constant, see figure.

Has anyone managed to reproduce the channel flow results presented in the IDDES paper?

Any ideas or comments are much appreciated.

Markus.

Hi Markus,

It's nice to hear that you are looking into all this - if you look at Ulf's minutes of the Berlin BoF session, I have an action item to test OF's IDDES in the channel!

First up, I can't find the picture that you refer to...?

Nonetheless I have some comments that might be helpful:

Re=395 and such high resolution should result in a fully-resolved LES, and is hence not the most ideal test of IDDES running as a wall-modelled LES (WMLES). I'd agree that IDDES should, in the limit of a fully-resolved LES reduce to pure LES mode and should give good results, however that would first be worth testing once you've validated the implementation with a WMLES case.

I would also recommend switching off the hybrid scheme and running with pure CDS ("linear") - for this case, the grid is fine enough everywhere for full CDS and channel flow is extremely sensitive to just a little bit of dissipation. As above, a later test case with the hybrid scheme would be an interesting idea, but pure CDS gives you the least setup uncertainty to start with.

It was all a long time ago, but I think I remember that we found problems both with IDDES for fully-resolved LES (low-Re) and for channel flow in combination with the hybrid scheme... all in the direction of an excessively "high" log law.

For a basic verificaton I'd suggest the following case, for which I could provide ASCII data of various blending functions, eddy viscosity, velocity profiles etc. from our in-house code and (I'll ask them for permission first) the results of Shur et al. An example plot is attached.

SA-IDDES
Re_tau = 18000
(fix pressure gradient dp/dx = -1 and set rho = 1 and viscosity = 1/18000)
Domain size L_x = 6.4 \delta, L_y = 2 \delta, L_z = 3.2 \delta.
Tangential grid resolution: Delta_x = 0.1, Delta_z = 0.05. (i.e. 64 cells in the x and z directions).
Wall-normal grid with y^+ of first cell vertex = 1., stretching ratio around 1.12.
"linear" convection scheme



Would you be interested in running such a simulation for a direct code-to-code verification? If I forgot any details or if you have any questions don't hesitate to ask.

Best regards,

Charlie.

Thanks for the detailed answer.

Here is another attempt to attach the image:


In the original IDDES paper they seem to get the correct log-law even for low Reynolds number. However, the mesh resolution might be very different.

Also, I think that I will run another case with REtau=18000. If you are interested in a code-to-code comparison, it would be useful to use identical meshes. In case you still have access to the mesh that you have been using, I would be happy to run IDDES in OF on that mesh.
Otherwise, I will just generate a new mesh.

Thanks,
Markus

Hi there,

You're right, they also computed Re_tau = 400, so it "should work" and the test is worthwhile: If the implementation is otherwise correct then perhaps the problem is the little bits of upwind introduced by the hybrid scheme.

Great that you want to run the 18000 case! It would be easier for you to generate your own mesh using blockmesh than to try to convert mine (and get the cyclic patches right!). Using the information in my last post for the tangential grid should be easy enough. This file contains the wall-normal distribution of the cell vertices from my grid:



The precise stretching ratio was 1.13654. Try to get roughly this distribution using blockmesh, but don't be too precise, it isn't so important.

To attach an image, you select the file to attach and then you have to press the button to the right of "Browse..." - this inserts a tag in the text. It is limited to 800 pixels wide.

Good luck, I look forward to comparing results!

Best regards,

Charlie.

Hi Charlie,

I will run the REtau=395 case again but this time without the hybrid convection scheme.

I think my image was too large, so here is another attempt:

I will keep you posted whether the results change with a full CDS.

Markus.

Hi Markus,

Thanks for the picture - reading off some values, it looks as if you have about 9% under-prediction of Cf. Let's see how it looks with pure CDS.

You'll see from the picture I posted that we have a slightly higher LES log law than the Shur et al. results, which we don't fully understand. One theory is that it is because our code is 2nd order whereas the Shur et al. solver is 4th order. Adding evidence to this theory, I have seen a similar upward shift for IDDES implemented in another 2nd order solver. It'll be interesting to see what OpenFOAM® does.

The same effect might be responsible for a higher log law region in your low-Re calculation.

Best regards,

Charlie.