CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   OpenFOAM Running, Solving & CFD (http://www.cfd-online.com/Forums/openfoam-solving/)
-   -   strange behaviour of the one-equation SGS model (http://www.cfd-online.com/Forums/openfoam-solving/66703-strange-behaviour-one-equation-sgs-model.html)

cfdmarkus July 22, 2009 04:42

strange behaviour of the one-equation SGS model
 
1 Attachment(s)
Hi all,

I am simulating the flow around two tandem cylinders using the one-equation SGS model with wall-functions. I set the the wall boundary conditions for nuSgs to nuSgsWallFunction and k to zeroGradient.
The first grid spacing is between y+=30-60 and second order schemes are used.

In my simulations the separation form the cylinders is extremely delayed which results in a suppression of any kind of vortex shedding (see attached picture). Using the dynamic oneEqn. SGS model seems to works fine and flow separates at the expected location (see picture).

I am running out of ideas why the oneEqnEddy-SGS model fails in my case.
Has anyone seen such a behaviour before or does anyone know what could have gone wrong here?

Markus

alberto July 23, 2009 02:26

What Re? Also in the second case the vorteces seem very diffuse, I'd say.

There is a long discussion on flows around a cylinder (Frank Bos did a lot of work on it), and some problems with the setup of the outlet BC were observed in that case. I'd recommend you search for it.

Best,

cfdmarkus July 23, 2009 03:45

Hi

The Reynolds number is ReD=1.6*10^5.
In the plot you can see the velocity magnitude, it would have probably been better to plot somethings like vorticity. I guess thats why it looks a bit diffuse.

I will search the forum for the flow around a cylinder.

Thanks a lot

Markus

alberto July 24, 2009 02:54

If possible, let me know if you find the answer ;-)

Thanks in advance,
Alberto

lakeat July 24, 2009 05:10

Hi Markus!

Have you tried a single cylinder 1st, I remember I did single cylinder flow b4, but using S and DS model, it seems to me both are okay. I guess some ic or bc setting wrong with your 1st case. See your first picture, is it you velocity field? I just can't imagine it.
Frank did a lot of good jobs, but it seems he focuses only on low Re. :rolleyes:

Cheers,

cfdmarkus July 24, 2009 07:58

The setting for both cases are exactly the same since I just copied the entire folder. Also, if I start the simulations from an solution of the dynOneeqnEddy with a resolved flow field, the oneEqnEddy does not sustain the turbulence and I end up with the flow in picture 1.

Initially, I blamed the wall functions for this behaviour, but if it would be a problem with the wall functions the dynOneEqEddy probably wouldn't work either.

I would still like to know how the wall functions compare to other hybrid RANS/LES methos but at the moment I am stuck here.

Markus

nikos_fb16 July 24, 2009 08:57

Hello Markus,

just an idea: looks like there is too much damping in your oneEqEddy.
Maybe decreasing the ck constant in the model could improve the behaviour.

cfdmarkus July 24, 2009 09:14

You are right, it seems that the turbulent viscosity near the wall is too high, which in turn is responsible for a significantly delayed separation point and hence the suppression of vortex shedding. But, somehow, I dont like messing around with the calibrated constants.

At the moment I am using the cubeRootVol filter length scale.
Maybe I should try using the vanDriest damping in conjunction with the wall-functions.

Markus

alberto July 24, 2009 10:12

To know if it's due to the wall functions or not, simply check at what y+ you have high viscosity values.

I don't think it's right to combine wall-functions and damping functions, which should be in principle mutually exclusive and have different requirements. In particular, if you use wall-functions you do not need any other wall treatment because you are simply not resolving the zone near the wall, but replacing the solution in that zone with what the wall-function says. Van-driest functions require exactly the opposite (you need to resolve the wall zone).

Best,

nikos_fb16 July 24, 2009 10:57

I would not mind too much to change the constant.
At equilibrium conditions the oneEqEddy reduces to the Smagorinsky model.
Everyone plays with the smago constant depending on the case being investigated, although its calibrated too.
So why not playing with ck in oneEqEddy.

in point of fact that your damping is too high, its very likely that the oneEqEddy has difficulties in getting the separation region.

At least you should see, if nothing changes when decreasing ck, probably the dynamic version of oneEqEddy is the better choice for this case.

alberto July 24, 2009 11:38

Quote:

Originally Posted by nikos_fb16 (Post 224036)
I would not mind too much to change the constant.
At equilibrium conditions the oneEqEddy reduces to the Smagorinsky model.
Everyone plays with the smago constant depending on the case being investigated, although its calibrated too.
So why not playing with ck in oneEqEddy.

Because it is sort of cheating ;)

nikos_fb16 July 25, 2009 09:14

Having a look at Yoshizawas paper about the oneEqEddy shows that the recomended constants are far from universal and are optimized for channelflow.

My suggestion to markus is not to cheat. Its trying to find the reason for the behaviour of the oneEqEddy in that case.

Maybe you would call it: learning by cheating :)

Nikos

lakeat July 25, 2009 10:42

Quote:

Originally Posted by lakeat (Post 223980)
Hi Markus!

Have you tried a single cylinder 1st, I remember I did single cylinder flow b4, but using S and DS model, it seems to me both are okay. I guess some ic or bc setting wrong with your 1st case. See your first picture, is it you velocity field? I just can't imagine it.
Frank did a lot of good jobs, but it seems he focuses only on low Re. :rolleyes:

I want to add that I have never tried a wall-function modeling b4.

So, may I know your grid number, is it possible that you take a try of wall resolved (without wall function) method, and see if it is the COMBINATION of "c_k problem" and "wall-function" responsible for that.

So to dear Nikos, my question is:
Which one should be blamed for the strange behaviour, "c_k problem" and "wall-function" together or just "c_k problem"?

And Markus, I'm looking forward to your results. :D

alberto July 25, 2009 14:05

Quote:

Originally Posted by nikos_fb16 (Post 224116)
Having a look at Yoshizawas paper about the oneEqEddy shows that the recomended constants are far from universal and are optimized for channelflow.

My suggestion to markus is not to cheat. Its trying to find the reason for the behaviour of the oneEqEddy in that case.

Maybe you would call it: learning by cheating :)

Nikos

Hehe I didn't mean to say anything personal. Just my two cents ;-)

I think playing with constants is wasting time, in general. I know parameters are adapted to the specific case quite often. I guess we all did that at least once, and I agree you can calibrate them once for a given class of problems (confined flows, external flows, ...).
However, if this becomes a habit, so that you adapt it to each specific case (geometry configuration, for example) just to reproduce experiments, it becomes a funny game of fitting data, and I just do not see the advantage, because at that point the model is not predictive anymore.
If you need to do that, you need another, more complete model, because you are not including the physics necessary to describe your model. So I would suggest to run with the same constants you use in the dynamic version of the model, and just say that it did not predict the properties of the flow under those conditions.

To understand what is going on in term of physics you don't need to play with the constants, but to compare quantities predicted by model with experiments (assuming they are right) or with reference results (DNS, but Re is high here!). In this way you figure out where the model fails.

In addition, you can study the effect of the wall treatment by comparing a run with the wall functions and one without them (van Driest). As a side note, damping functions should not be used with the dynamic models, if implemented correctly, with a local value of the constant (not true in the dynSmagorinsky implemented in OpenFOAM: the constant is averaged on the domain), because the dynamic procedure makes the model correctly degenerate to the laminar limit.

Best

cfdmarkus July 27, 2009 09:42

I am aware that people adapt the Ck and also Cdes in DES to get the desired behaviour of the model. However, I also agree with Alberto's opinion and therefore try to avoid changing the constants if possible.

I have done a simulation with a "wall-resolved" oneEqnEddy, i.e. y+=0.8 and 2.2 mio grid points. The attached figure shows a snapshot of this simulation.
The flow now separates at the correct location, however the development of the dettached shear layer of the first cylinder seems wrong (but thats another issue which I need to resolve).

I also had a look at the wall-function formulation in OF. It seems that the wall functions are based on y+ which probably will cause troubles in situations where the wall-shear-stress vanishes.

Could it be possible that the presence of stagnation/separation points with vanishing wall-shear upsets the wall-function formulation and as a consequence contaminates the results?

Markus

cfdmarkus July 27, 2009 09:46

1 Attachment(s)
ok here is the figure

lakeat July 27, 2009 10:57

Quote:

Originally Posted by cfdmarkus (Post 224275)
I have done a simulation with a "wall-resolved" oneEqnEddy, i.e. y+=0.8 and 2.2 mio grid points. The attached figure shows a snapshot of this simulation.
The flow now separates at the correct location, however the development of the dettached shear layer of the first cylinder seems wrong (but thats another issue which I need to resolve).

Why it is wrong? Can you explain it in detail?

So it is not a matter of Ck value, is it? It is bcoz of wall-function, right?

cfdmarkus July 27, 2009 12:31

yes, it seems to me that the failure is due to wall functions.
In the figure, the shear layer is very stable and does not break down to turbulence quick enough. I suppose that the the contribution from the SGS model is too much which damps the natural instabilities and hence the delays the breakdown.

Markus

waynezw0618 August 14, 2009 05:33

how about the developing of the turbulence? i mean the time effect.i face the problem is that the flow is going smooth after some time. i guess that may be the problem of the inflow condition.i don`t know if it is righ.


All times are GMT -4. The time now is 03:05.