[panda60]

Dear foamers,

I have trouble with using dynSmagorinsky model. Firsr I use Smagorinsky model. then I change to dynSmagorinsky, some times later, it blows up. I don't know how to use it correctly.

This is my LESproperties file, is it right ?

FoamFile
{
version 2.0;
format ascii;
class dictionary;
location "constant";
object LESProperties;
}
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * //

LESModel dynSmagorinsky;

delta vanDriest;

printCoeffs on;

dynSmagorinskyCoeffs
{
filter simple;
ce 1.05;
}

cubeRootVolCoeffs
{
deltaCoeff 1;
}

PrandtlCoeffs
{
delta cubeRootVol;
cubeRootVolCoeffs
{
deltaCoeff 1;
}

smoothCoeffs
{
delta cubeRootVol;
cubeRootVolCoeffs
{
deltaCoeff 1;
}

maxDeltaRatio 1.1;
}

Cdelta 0.158;
}

vanDriestCoeffs
{
delta cubeRootVol;
cubeRootVolCoeffs
{
deltaCoeff 1;
}

smoothCoeffs
{
delta cubeRootVol;
cubeRootVolCoeffs
{
deltaCoeff 1;
}

maxDeltaRatio 1.1;
}

Aplus 26;
Cdelta 0.158;
}

smoothCoeffs
{
delta cubeRootVol;
cubeRootVolCoeffs
{
deltaCoeff 1;
}

maxDeltaRatio 1.1;
}


and I don't know why dynSmagorinsky model need k file. Because we don't need to solver k equation.

Could someone help me ?

[panda60]

My simylation always blow up. What I should pay attation to when using dynSmagorinsky model ?

Thanks.

[foamWang]

Co-ask. My case blows up every time. I want to try update the fvSchemes. Let's see what will happen.

[foamWang]

Hi panda,

I updated the fvSchemes with QUICK phi. So far, the simulation doesn't easily blow up. I will try other method to further improve.

Roro

[xiao]

As I learned from other threads of the forum, the wall damping is in
conflict with "dynamic models", that is, if you are using dynamic LES
models, you should _not_ use "delta vanDriest". The dynamic model is
capable of treating the wall boundary condition giving a laminar nuSgs
near the wall, at least theoretically.

However, I am not sure that caused your blow up. I had the same blow up
by using other delta (cubeRootVol), but when switched to "delta smooth",
it became stable! (I should add that my mesh is not overly stretached,
it is about 1.05 in the wall normal direction.) HOWEVER, the result
seems to be seriously wrong in that it under-predict the friction
velocity (utau) by 10% compared to static LES model and DNS results.
Both dynSmagorinsky and dynOneEqEddy model have the same problem.

Anyone has similar experience or any hints on this? Thank you in
advance.

[foamWang]

Hi Xiao,

I used delta smooth with the "upwind phi" in fvSchemes::divScheme.
Generally speaking, the simulation is stable, except max co goes up to <10 and drop back occasionally.

I would say smooth delta helps.

Btw, anyone knows more stable fvSchemes? Now I can do is to use upwind to increase the numerical viscosity.

Other suggetion?

Thanks.
roro

[foamWang]

Hi xiao,

Using smooth delta, I also get a very weired result.
I changed the delta back and improved my grid. The simulation doesn't blow up.
My case is axis-symmetric. So I use some wedges at the center before.
It turns out that wedges significantly destroy the grid quality.

It concludes that dynSmagorinsky is sensitive to grid quality.

Hope this help you.
Regards,

roro

[xiao]

hi roro,

Thanks for the communication.

I have two questions for you (and for other foamers on the forum as well)
1. You said you changed delta back. do you mean back to "vanDriest" or "cubicRootVol"? As I said in my post, I think vanDrist is not compatible with dynamic models theoretically. I wouldn't do that. Experts... please comment on this issue.

2. Is the "upwind" 1st order or 2nd order? This is a question I am currently looking into. This is important because LES required minimum 2nd order accurate numerical methods, otherwise the numerical error would dominate the SGS contribution.

3. Did any one try "QUICK" or "MUSCL" or "limited Cubic" with LES simulations?

Thanks!

Best,
heng

Originally Posted by foamWang

Hi xiao,

Using smooth delta, I also get a very weired result.
I changed the delta back and improved my grid. The simulation doesn't blow up.
My case is axis-symmetric. So I use some wedges at the center before.
It turns out that wedges significantly destroy the grid quality.

It concludes that dynSmagorinsky is sensitive to grid quality.

Hope this help you.
Regards,

roro

[xiao]

I think "Gauss upwind" for divScheme is not appropriate for LES simulation, becausue this is only first order. See User Guide pp 116, Table 4.10.

[foamWang]

Hi, Xiao,

Thanks for prompt reply.

1. "Change delta back" means back to default cubicRootVol. Now I am just not comfortable with the smooth delta. Hope others can comment.

2. I agree that "upwind" is the first order scheme. I made this choice coz dynSmagorinsky has too strong turbulent dispersion for my grids. So I need some artificial viscosity. I also found in XiFoam tutorials (or somewhere else) the fvSchemes::divScheme was all set to upwind for LES. So I just follow. I am not sure whether 2nd order accuraccy is necessary. Please comment.

3. I did try QUICK. It has less numerical viscosity and requires a lot of memory. Once I change all the divScheme to QUICK, the memory error is appeared. Maybe I can try later in the parallel machine and see any improvement.

Anyway, I am happy, coz I feel we are approaching a sound solution.

Cheers.

roro

Originally Posted by xiao

hi roro,

Thanks for the communication.

I have two questions for you (and for other foamers on the forum as well)
1. You said you changed delta back. do you mean back to "vanDriest" or "cubicRootVol"? As I said in my post, I think vanDrist is not compatible with dynamic models theoretically. I wouldn't do that. Experts... please comment on this issue.

2. Is the "upwind" 1st order or 2nd order? This is a question I am currently looking into. This is important because LES required minimum 2nd order accurate numerical methods, otherwise the numerical error would dominate the SGS contribution.

3. Did any one try "QUICK" or "MUSCL" or "limited Cubic" with LES simulations?

Thanks!

Best,
heng

[panda60]

Dear Roro Wang and heng Xiao,
You are right, I have wrong here.
VanDriest should not be use in dynamic model.
But even I use cubeRootVol or delta smooth, dynSmagorinsky model alwayse diverge after several thousands steps later. But there is no problem with Smagorinsky model using the same mesh.
So I wonder if dynamic model is more sensitive to mesh than Smagorinky model ? or if I should use more mesh near wall ? or Corrunt number should less enough ?

I think upwind scheme shouldn't be use in LES.

[xiao]

Hi Jiayi,

I don't know exactly why your simulations blow up. As I said, when I use dynamic model in a channel flow, it just give low shear stress. For me, the blow up occurred for one of the meshes when I used cubeRootVol. I am still struggling to figure out why the shear stress is seriously under-precicted, since the dynamic model is supposed to be better than its static counterapart! Those who has experience on this problem, please give some hints, thanks!

For your problem, I would recommend that you check your mesh quality, and more importantly, control your Co number, i.e. reduce your time step. For Crank Nicholson and backward timestepping scheme, Co should be below 0.5. I would say, try to bring Co under 0.2 to start with.

Also, are you using pisoFoam or channelFoam? How complex is your problem? Complex flows (backward facing step, flow around cylinder, etc) are certainly more difficult to compute. Also, what spatial discretization schemes are you using?

Share those info with the forum, and maybe some experts would be able to give you some hints. Looking forward to it.

best,
heng

[fs82]

Did you check where your simulation blows up? My experience is, that the scheme influences your solution but do not blow up the whole simulation. May be your boundary condition isn't set correctly or your mesh has some inconsistencies.
After some testing with a circular cylinder I have choosen "Gauss limitedLinear 1" for the convective term because all other schemes (QUICK, linear ...) leads to a unphysical behaviour of my SGS Model (turbulence was produced far away of the boundary layer of my cylinder). And up to now I noticed no problems in using dynamic Smagorinsky and vanDriest wall damping. But to be honest I just tried this one time. But as far as I know one needs a wall damping even for a dynamic SGS model.

kind regards,
Fabian

[panda60]

Dear Heng Xiao,
I think you are right. I change the near wall mesh again, use smooth ratio 1.25. Then use "delta smooth", and change Co number to 0.3 , now simulation has going on for 5000 steps, the dynSmagorinsky model seems OK.
I am very interenting your dynamic model result in channel flow. You said"it just give low shear stress", That means if shear stress is low, the velocity will be larger.The following is my Smagorinsky+VanDriest result.
I use buoyantBoussinesqPisoFoam to calculate heat transfer problem. I compare a horizontal line AB in front of building, which is 4mm above the ground. from result, RANS model velocity transgfer faster near ground, but Smagorinsky model velocity is more lower than experiment and RANS, that means shear stress got from LES is larger. So I think three reasons may cause this problem. First, no inflow turbulence. Second, SGS model. Third, near wall mesh is not fine enough.
The following is my yPlus in ground, most place are nearly 1.5 , only small place are 6. So I think my near wall mesh is OK. So differdent SGS model and inflow turbulence may cause my problem. I am very glad to hear dynamic model will give low shear stress. I will try dynamic model. I want to know if you have used inflow turbulence ? not random, but recyle method.

Dear Fabian Schlegel,
I don't agree with you. wall damping function only can be used for standard Smaorinksy model and standard one equation model. dynamic model doesn't need this , That is why people develop denamic model.

[fs82]

Hello Jiang Guoyi,

just a short comment to:

dynamic model doesn't need this , That is why people develop denamic model.

The idea of a dynamic Smagorinsky model is not to get rid of a wall damping, but to get rid of the Smagorinsky constant. Because this constant changes for every flow problem. The dynamic Smagorinsky introduces some ideas to adopt this constant dynamically to your flow problem or short: the model calculates the constant itself.
The wall damping is due to the different behavior of your viscous sublayer which has a anisotropic turbulence (the idea of Smagorinsky is indeed isotropic turbulence) where the viscous forces dominates the flow. And this leads to the idea of wall damping. You want your model to work outside the viscous sublayer and not inside as far as I understand the idea of LES. But please correct me if I am wrong.

kind regards,
Fabian

[xiao]

Hi Fabian,

To figure out this issue, I read again the original paper of dynamic SGS model in detail, and insist that wall functions are not necessary for dynamic SGS models. In the following discussion, I will refer to this paper by Germano et al. If you need a copy, please let me know.

In the second last paragraph of Section I (introduction), page 1761, it says:

"The model presented here requires a single input parameter and exhibits the proper asymptotic behavior near solid boundaries or in laminar flow without requiring damping or intermittency functions. The model is also capable of accounting for backscatter. "

Also, at the end of the first paragraph of the Conclusion section (IV), page 1765, it reads:

"Among the useful properties of the model is its proper asymptotic behavior near the wall without the use of ad hoc damping functions."

So, to summarize what you said, what is presented in the original paper, and what I discussed with my colleague Florian K., the dynamic model serves three purposes (or equivalently, has 3 advantaged over static ones), at least theoretically:

1. Model the near-wall region by dynamically determining the C constant (or equivalently, eddy viscosity).
2. Model the laminar to turbulent transition process, also by dynamically adjust the C constant.
3. Account for the energy backscatter.

Item (1) is a remedy for the spacial variation of C, and (2) the remedy for the temporal variation of C.

Just to step off the topic a bit, my own experience with channel flow also confirmed the second item. I wanted to obtain a fully turbulent flow initial condition by starting from a perturbed laminar flow field, the static One Eq. Eddy model just damped out the perturbation, and re-laminarized the flow, while the dynamic model achieved the transition to turbulent flow.

Back to the topic, I think one should not need wall damping for dynamic LES models, according to the original paper. However, there may be later development in the community, which I missed (i.e. another researcher might have showed evidence that wall damping is needed ...). I will certainly dig further in the literature, and report in this thread if I found any new information.

Best,
Heng

Reference:
Germano et al. A dynamic subgrid‐scale eddy viscosity model. Physics of Fluids A: Fluid Dynamics (1991)

Originally Posted by fs82

Hello Jiang Guoyi,

just a short comment to:

The idea of a dynamic Smagorinsky model is not to get rid of a wall damping, but to get rid of the Smagorinsky constant. Because this constant changes for every flow problem. The dynamic Smagorinsky introduces some ideas to adopt this constant dynamically to your flow problem or short: the model calculates the constant itself.
The wall damping is due to the different behavior of your viscous sublayer which has a anisotropic turbulence (the idea of Smagorinsky is indeed isotropic turbulence) where the viscous forces dominates the flow. And this leads to the idea of wall damping. You want your model to work outside the viscous sublayer and not inside as far as I understand the idea of LES. But please correct me if I am wrong.

kind regards,
Fabian

[fs82]

Allright, I do not have the time at the moment to deal with this in detail. But the article refer to a dynamic Smagorinsky? Smarorinsky uses the assumption of isotropic turbulence which isn't fullfiled in the viscous sublayer. But I could imagine that other dynamic models do not use this assumption and so far do not need a wall damping if they adjust the constant for the viscous sublayer in a appropriate way. I would suggest a test if the solution changes if you enable vanDriest damping in combination with a dynamic smagorinsky or not.

[fs82]

Allright, I had a deeper look into my literature and I agree with you. A dynamic model allways needs no wall damping. It adjust the constant in a way that the influence of the model is reduced near a wall. But this should not blow up the whole simulation if you use a dynamic model with a wall damping, because there is nothing to damp at the walls. So the cause of this error should be somewhere else.

kind regards,
Fabian

[xiao]

Hi Fabian,

Yes, I also agree. The set up of wall damping (with smooth delta) should actually increase the stability performance, compared to cubeRootVol delta. The blow-up is caused by something else.

I looked further in the literature, and it is likely that the advantage of "dynamic SGS modeling not needing wall function" is only theoretical. Using a wall model (like wall damping, or wall stress model) may sometimes be desired. The paper below seems to suggest that wall treatment is necessary when you use a "very coarse grid". Here is a quote from the paper. See page 277-278. (Hope this would not form a copyright infringement. I will remove it if it does):

"In plane-parallel flows, filtering and averaging are performed only on horizontal planes parallel to the walls. When the simulation uses a refined wall-normal grid, this may be justified, since the filtering over very small wall-normal distances has little effect. This approach is not at all justified when the mesh is so coarse in the wall-normal directions that large flow variations are being averaged over. "

"We therefore recognize from the outset that the SGS model may be highly inaccurate in the near-wall region, and that improvements are clearly needed in SGS/wall modelling in this area."

Note that here the authors are referring to the dynamic SGS models.
In addition, the "wall modeling" in their paper is a much more general concept than the "wall damping" in openFoam. This paper also talk about the potential instability issue caused by dynamic models. But I will need more time to digest it and to really draw a conclusion from it.

As of now, I think this is still a more or less unanswered questions for me. Would any gurus/experts (like Eugene de Villiers, HrV) shed some light this problem?

Fabian, Guoyi, and Roro, thank you all for the discussions, which have been very helpful to me. I will research further and keep you updated on what I found related to this issue.

Best,
Heng

Reference:
Cabot and MOIN. Approximate wall boundary conditions in the large-eddy simulation of high reynolds number flow. Flow Turbul Combust (2000) vol. 63 (1-4) pp. 269-291

Originally Posted by fs82

Allright, I had a deeper look into my literature and I agree with you. A dynamic model allways needs no wall damping. It adjust the constant in a way that the influence of the model is reduced near a wall. But this should not blow up the whole simulation if you use a dynamic model with a wall damping, because there is nothing to damp at the walls. So the cause of this error should be somewhere else.

kind regards,
Fabian

[fs82]

Yeah nice discussion and I have learned something :-D