Hi, we were talking quite a lot with @thepolynom about LES simulation, so I created a (probably temporary) channel focused on this

so if some other wish to participate, all good

I've found a very nice discussion yesterday about the Smagorinsky model in OpenFOAM : https://www.cfd-online.com/Forums/openfoam/72277-how-use-dynsmagorinsky-model-correctly.html

(Sorry for the long text, hope someone will read and comment 😄 ) I started reading about LES few days ago, and here is what I understood so far (to be corrected): Physically : Eddies are present in the flow, larger eddies are the size of the geometry (~characteristic diameter let's say) and take their kinetic energy from the mean flow. Large eddies then "decompose" into smaller eddies by the cascade principle. When the eddies are finally small enough, they dissipate into heat. Physically, the loss into heat happen at all eddies size, but it is assumed that >90% of the transfer into heat happens at very small eddies only (not modeled by CFD) Numerically : The goal is to resolve the grid-size eddies, and model the sub-grid scales ones. However in LES models (Smagorinsky atleast, the only one I read properly), small eddies are supposed to be isotropic, and that is the reason why your mesh should be fine. If your mesh is not fine enough, the calculated dissipation will assume your eddies are isotropic, when they might not be in real life, and so your dissipation will be badly calculated > wrong results. Mesh size : the mesh size should be: - normal to the wall : y+ < 1, with smooth grid expansion - streamwise : 50 < x+ < 150 - spanwise : 15 < z+ < 40 Values depends on the literature, but basically it's around those values, even though cases were shown to give proper results with x+ and z+ ~1500.

The first LES model to be created was the Smagorinsky one. The basic problem is the estimation of the turbulent viscosity : Viscosity_turb = [ Coef . Delta ]^2 * |S| |S| : magnitude of the strain-rate tensor Coef : Smagorinsky coef Delta : Filter width And here starts the problems : Coef is constant in this model, but in real life, it depends on the flow and is actually different everywhere in the domain, so a bad estimation of this value (it is bad anyway, since OF will assume it's the same in the whole domain) can lead to unproper results. Close to the wall, the viscosity should tend to 0, and it is not true with the given formula, so a damping function is applied (vanDriest), to reduce the value of the viscosity close to the wall. The damping is applied up to y+ = 500. Main disadvantages of the Smagorinsky model: - Too dissipative in laminar regions - Not good for laminar-turbulent transition (due to the constant Coef I guess) - Backstatter is not modeled (sometimes, smaller eddies are actually giving energy to bigger eddies)

So, to solve this problem, DynSmagorinsky is here, but wait for it: In the paper, the dynamic Smagorinsky is going to calculate the Coefficient depending on time and space. But it is not how it is implemented in OpenFOAM. In Openfoam it takes an average value for the whole domain, which changes yes, but it's still constant for the whole domain. So, for example, with the dynamic adaptation of the Coef, there shouldn't be a need for damping close to the wall, but here, it is maybe needed somehow, so it's weird and not a really useful model then.

". In Openfoam it takes an average value for the whole domain, which changes yes, but it's still constant for the whole domain." I used this model once some time ago and this is new too me, but if it is really that way I dont see a advantage in the dynamicSmagorinsky ^^

yes I saw this

Alberto changed it indeed and proposed to implement it in OF, it has not been done for some reasons idk why

That was the one I used ^^

you used alberto's one?

I think so 😄 I had it from a guy working in a similiar project but it should be this one

but I used it with openfoam 1812

or 1712

apart from this, do you agree with what is said? ^^

yeah I dont have the time to go in detail but I found nothing wrong ^^

new info:

oneEqEddy has the same principle as Smagorinsky, but the advantage will be in the transport equation of k, which gets rid of the assumption of Smagorinsky model saying "balance between the SGS energy production and dissipation". I would say this is true only if the eddies are "small enough". So oneEqEddy "is less wrong" than Smagorinsky for coarser meshes. Moreover, the locDynOneEqEddy is implemented, where the coefficients are locals, and not global. However I saw people complaining about negative values of nu_sgs using locDynOneEqEddy, and some people claim it's the backscatter, some say it's unphysical and should be limited to 0.

and as I get it, when the model is dynamic, the delta to use is the simple cubeRootVol, no need for damping

im not sure about the damping point

So if you use a dynamicSmagorinsky I think its still better to use vanDriest or am I wrong?

I guess so, but the approach is not really useful anyway (cause of the constant coef) But for the dynamicSmago from Alberto, you don't need to (that's what I learnt from the cfd-online topic I cited on the first post)

Basically : dynamicSmagorinsky will average the coeff over the whole domain, so, in average, the coef will be high next to walls (comparing to its 'real' value), where nuSgs should be small. In that case, then vanDriest is needed. It is maybe less needed somehow though

Hi, when looking at LES, you always see this chart :

it says that when the eddies are small enough, they are universal and independant of the problem

but why is that?

very nice question 😃 maybe because the Large Eddys depend on the geometry or the velocity but if you go in the direction of the dissipation area the geometry has no impact. But I never thought about that 😄

now I have to read a book 🤓 😒