[Tobi]

Hi everybody,

I guess many people who are not specialists in the topic of turbulence models would like to know a bit more about the latest models available in science and FOAM. However, as I just know the theory behind the different approximations (RANS/LES), I would like to start a conversation about the turbulence models in FOAM. For example. If I am doing a industrial cfd simulation, the first model I am using is the old-well-known k-Epsilon model as it is stable. However, AFAIK, the following holds:

• k-Epsilon, well suited for turbulences far away from walls
• k-Omega, well suited for close wall turbulences
• k-Omega-SST, a combination of k-Epsilon and k-Omega, which summarizes the advantages of both (far-field and close wall areas)

Nevertheless, there are much more models available and the three mentioned above are already very old ones. As we are having two unknown in the RAS model (I am not so familiar with LES), my personal opinion is to use 2-equation models rather than 1-equation models. However, I also think, that this depends on the problematic one is investigating.


Assuming a compressible solver such as rhoSimpleFoam, FOAM offers 13 turbulence models:

RAS models
13
(
LRR
LaunderSharmaKE
RNGkEpsilon
SSG
SpalartAllmaras
buoyantKEpsilon
kEpsilon
kOmega
kOmegaSST
kOmegaSSTLM
kOmegaSSTSAS
realizableKE
v2f
)

Are there any information about the usage of the different models? Or are there any experimental papers for turbulence within simple geometries such that we could recalculate the data and compare the different turbulence models?

Thanks in advance and I am looking forward for some discussion and information.

[Bloerb]

Just to add to this. You mentioned you'd start with kEpsilon for industrial cases. I'd disagree...everyone i know uses kOmegaSST whenever they don't know shit about what is going on. Hence I use it all the time, I did many comparisons between the kOmega and kEpsilon based models and it has always been the one providing the results that match PIV measurements etc the closest. Which is probably your best bet if you are not a turbulence guy (at least from my limited perspective). Which model is used depends on the use case. First of all numerical effort increases with the number of equations solved. Hence let's order the ones you listed. For a car simulation e.g you could just use a 1 equation model to save some computation time.


1 equation models:
SpalartAllmaras

k-omega based models (most of them 2 equations)
kOmega
kOmegaSST (mixture of kEpsilon and kOmega)
kOmegaSSTLM (4 equation model to model transition)
kOmegaSSTSAS (hybrid LES / RAS model )

k-epsilon based models (most of them 2 equations)
kEpsilon
LaunderSharmaKE
RNGkEpsilon
realizableKE
v2f (4 equations to model anisotropic behaviour)
buoyantKEpsilon (well includes buoyancy effects)

Reynolds stress models (6 equations)
LRR
SSG

So more equations = better accuracy? Like everything in turbulence...maybe

Turbulence always depends on your Reynolds number. Different models perfom totally different at different Reynolds numbers. Lets take a look at an airfoil. A 1m long NACA0012 airfoil at Re=10^6 for example. The mach number is low hence we can assume this to be incompressible. There have been many posts in this forum for wrong results for this simple geometry. They always complain about the drag value. Which is about twice as high as measurements. This is partly due to y+ dependency, bad meshes, but also because of the transition from laminar to turbulent flow.
The boundary layers starts laminar and turns turbulent at some point. Most turbulence models, do not capture this. You can either use a transient simulation to increase accuracy or you could try the kOmegaSSTLM model. Because it models transition. If you do not like the effort of a full LES simulation you might also use something like kOmegaSSTSAS. v2f models the anisotropic behavior of turbulent flow near the wall. The problem is however that they all fail near the stall angle. Here about 15° as stall is a fully transient effect. You might want to use a 1eqn model as it might be more stable and allows a steady state simulation that actually still converges.....
Here is the case which lists the NASA sources and validation data. You can try out different models and see for yourself how they perform for this use case.
https://openfoam.com/documentation/g...irfoil-2d.html There are a few more to click through.

Another quick test would be a simple pipe flow. As measurements for the velocity profile are readily available. At modest reynolds numbers like e.g 40000 most of the kEpsilon models perform poorly. Especially if there is a bend. Since this is just a 1D simulation you could try that to calculate how close you can match the friction coefficient or how closely you can match the velocity profile in that case. I have done that at some point in the past and might still have a case for that to recreate the moody chart with different models. Here kOmegaSST was the best allrounder from my tests.

Apart from that I think that turbulence models are like religions...But maybe some expert turbulence guy can limit where to use which of these models more closely. For me it is always testing with some measurement data etc, or fall back to kOmegaSST.

[Santiago]

Originally Posted by Tobi

Hi everybody,

I guess many people who are not specialists in the topic of turbulence models would like to know a bit more about the latest models available in science and FOAM. However, as I just know the theory behind the different approximations (RANS/LES), I would like to start a conversation about the turbulence models in FOAM. For example. If I am doing a industrial cfd simulation, the first model I am using is the old-well-known k-Epsilon model as it is stable. However, AFAIK, the following holds:

• k-Epsilon, well suited for turbulences far away from walls
• k-Omega, well suited for close wall turbulences
• k-Omega-SST, a combination of k-Epsilon and k-Omega, which summarizes the advantages of both (far-field and close wall areas)

Nevertheless, there are much more models available and the three mentioned above are already very old ones. As we are having two unknown in the RAS model (I am not so familiar with LES), my personal opinion is to use 2-equation models rather than 1-equation models. However, I also think, that this depends on the problematic one is investigating.


Assuming a compressible solver such as rhoSimpleFoam, FOAM offers 13 turbulence models:

RAS models
13
(
LRR
LaunderSharmaKE
RNGkEpsilon
SSG
SpalartAllmaras
buoyantKEpsilon
kEpsilon
kOmega
kOmegaSST
kOmegaSSTLM
kOmegaSSTSAS
realizableKE
v2f
)

Are there any information about the usage of the different models? Or are there any experimental papers for turbulence within simple geometries such that we could recalculate the data and compare the different turbulence models?

Thanks in advance and I am looking forward for some discussion and information.

Smagorinsky models for explicit LES perform poorly, unless the anisotropicFilter is used. The implementation of the dynamic lagrangian model is not as the one of Meneveau & Cabot.

And then there is the over-dissipative character of PISO. Strangely enough, pisoFoam works really badly when doing unresolved DNS. Usually 2nd order FD codes run using LES model or uDNS you see little difference in the results.

[cryabroad]

Hi Santiago,

You mentioned something I'm really interested in. The Germano's dynamic Smagorinsky is not officially included in OpenFOAM, but the Meneveau's one is in it, do you have any ideas why? The reason I'm asking about this is because in my field (turbulent combustion), we have a lot of swirling flows, and people tend to (and it has become kind of standard) use Germano's dynamic model. The results are usually pretty good. I've seen people from CERFACS using WALE for swirling flows as well, but not that popular (not trying to offend anybody but that's what I found). WALE is very popular for wall-bounded flows though. Before Germano, people uses the traditional Smagorinsky, but you always have to tune the coefficient.

I actually did my own survey for LES SGS models some time ago, and came to an conclusion that as the mesh refines, the different models seem to give similar results. What do you think about it? Do you have any experience in other models, e.g., kEqn?

Thanks in advance.

Ruiyan

[cryabroad]

I'm not an expert in RANS, but based on my limited experience, the RNG and realizable kEpsilon work better than kEpsilon. The Reynolds Stress Model works the best but it's kind of hard to use it. Hope others can share more experience!

For LES, I really recommend Fumiya's blog (https://caefn.com/openfoam/turbulence), it is very clear and thorough. By the way I think he is on this forum as well.

[Santiago]

Originally Posted by cryabroad

Hi Santiago,

You mentioned something I'm really interested in. The Germano's dynamic Smagorinsky is not officially included in OpenFOAM, but the Meneveau's one is in it, do you have any ideas why? The reason I'm asking about this is because in my field (turbulent combustion), we have a lot of swirling flows, and people tend to (and it has become kind of standard) use Germano's dynamic model. The results are usually pretty good. I've seen people from CERFACS using WALE for swirling flows as well, but not that popular (not trying to offend anybody but that's what I found). WALE is very popular for wall-bounded flows though. Before Germano, people uses the traditional Smagorinsky, but you always have to tune the coefficient.

I actually did my own survey for LES SGS models some time ago, and came to an conclusion that as the mesh refines, the different models seem to give similar results. What do you think about it? Do you have any experience in other models, e.g., kEqn?

Thanks in advance.

Ruiyan

I guess by Germano's model you refer to Lilly's approach for the Dynamic Smagorinsky model, the latter proposed by Germano...

Well, Meneveau's Model is an improvement from the localized dynamic model of ghosal, and the plane averaged version of Lilly, as is more stable and it averages cs in a "generalized" way. So moving towards this model is the correct step.

I dont find any particular reason why the dynamic model should be preferred over the lagrangian dynamic, except perhaps if one wants to save computational time.

[Lookid]

Dear Tobi,

from my small experience, I would agree with Bloerb, and say kOmegaSST is the best first try, however, it can lead to bad stuff also.

Example with a low specific speed pump calculation with kOmegaSST in CFX. By comparing a WF mesh (y+ ~100) and a Low Reynolds mesh (y+ ~1), the difference in the main characteristics can be up to 50% at high flow rates (see attachment, excel curve is my case, the second graphs are from this thesis : http://rosdok.uni-rostock.de/resolve...hab_0000001868)

The reason is due to that, at high flow rates, some detachment causing massive losses is simply not captured by the WF mesh. But I guess at this y+, it behaves like kEpsilon, but I did not check.

Here is a very nice course describing different RANS, URANS and LES models:
http://www.tfd.chalmers.se/~lada/pos...trgjmrDCKskNCE


I would like to extend the conversation to LES, since I just start it now with foam-extend, and it has been brought to the conversation already.

foam-extend will propose also a large number of models:

DeardorffDiffStress
LRRDiffStress
Smagorinsky
SpalartAllmaras
SpalartAllmarasDDES
SpalartAllmarasIDDES
dynMixedSmagorinsky
dynOneEqEddy
dynSmagorinsky
kOmegaSSTSAS
laminar
locDynOneEqEddy
mixedSmagorinsky
oneEqEddy
spectEddyVisc

How to choose between them?
I am aware of the problems of Smagorinsky (constant coef), but dynSmagorinsky is just not implemend as in the theory as you said, the coefficient does not vary "by cell" but just changed over the whole domain, making dynSmagorinsky pretty useless. AlbertoPa did implement a proper dynSmagorinsky though (https://github.com/AlbertoPa/dynamicSmagorinsky).

How to evaluate if an LES simulation is good?
I saw 2 ways so far : Evaluating res_turb (How to use dynSmagorinsky model correctly ?)
And plot somehow the -5/3 decay thingy.
Does it mean there are some ways to see if the LES simulation is good without experimental comparison?


EDIT:
@cryabroad, I went through the blog and indeed it is full of information, thanks!
I guess Smagorinsky problems are solved using OneEqEddy (basically is "less wrong" than Smagorinsky when using relatively coarse meshes, as I understood). dynOneEqEddy have the same problems as dynSmagorinsky but they actually also implemented a locDynOneEqEddy for local refinement. But I saw people arguing about that they obtain negative nuSgs values. Some say it is backscatter, some say it is unphysical and should be limited to 0.

[Tobi]

Hi all,


a very nice discussion already started. I just want to clarify my statement:

You mentioned you'd start with kEpsilon for industrial cases. I'd disagree...everyone i know uses kOmegaSST

Generally, I start with kEpsilon as it is known to be very diffusive. Thus, the solution will be more stable. Surely, I will switch to kOmegaSST later on as it combines kEpsilon and kOmega and therefore, we get "better" more "realistic" values. Personally, I would use the kEpsilon model just for free-stream flames as I did a lot of validation with the laminar flamelet model which gave extreme good results.

[akidess]

Originally Posted by Tobi

Generally, I start with kEpsilon as it is known to be very diffusive. Thus, the solution will be more stable.

In my experience, the convergence behaviour between kEpsilon and SST is quite similar (this is also confirmed in the CFX Theory Manual), and there is no need to start with kEpsilon if you plan to use SST anyway.

[syavash]

Originally Posted by Lookid

Dear Tobi,

from my small experience, I would agree with Bloerb, and say kOmegaSST is the best first try, however, it can lead to bad stuff also.

Example with a low specific speed pump calculation with kOmegaSST in CFX. By comparing a WF mesh (y+ ~100) and a Low Reynolds mesh (y+ ~1), the difference in the main characteristics can be up to 50% at high flow rates (see attachment, excel curve is my case, the second graphs are from this thesis : http://rosdok.uni-rostock.de/resolve...hab_0000001868)

The reason is due to that, at high flow rates, some detachment causing massive losses is simply not captured by the WF mesh. But I guess at this y+, it behaves like kEpsilon, but I did not check.

Here is a very nice course describing different RANS, URANS and LES models:
http://www.tfd.chalmers.se/~lada/pos...trgjmrDCKskNCE


I would like to extend the conversation to LES, since I just start it now with foam-extend, and it has been brought to the conversation already.

foam-extend will propose also a large number of models:

DeardorffDiffStress
LRRDiffStress
Smagorinsky
SpalartAllmaras
SpalartAllmarasDDES
SpalartAllmarasIDDES
dynMixedSmagorinsky
dynOneEqEddy
dynSmagorinsky
kOmegaSSTSAS
laminar
locDynOneEqEddy
mixedSmagorinsky
oneEqEddy
spectEddyVisc

How to choose between them?
I am aware of the problems of Smagorinsky (constant coef), but dynSmagorinsky is just not implemend as in the theory as you said, the coefficient does not vary "by cell" but just changed over the whole domain, making dynSmagorinsky pretty useless. AlbertoPa did implement a proper dynSmagorinsky though (https://github.com/AlbertoPa/dynamicSmagorinsky).

How to evaluate if an LES simulation is good?
I saw 2 ways so far : Evaluating res_turb (How to use dynSmagorinsky model correctly ?)
And plot somehow the -5/3 decay thingy.
Does it mean there are some ways to see if the LES simulation is good without experimental comparison?


EDIT:
@cryabroad, I went through the blog and indeed it is full of information, thanks!
I guess Smagorinsky problems are solved using OneEqEddy (basically is "less wrong" than Smagorinsky when using relatively coarse meshes, as I understood). dynOneEqEddy have the same problems as dynSmagorinsky but they actually also implemented a locDynOneEqEddy for local refinement. But I saw people arguing about that they obtain negative nuSgs values. Some say it is backscatter, some say it is unphysical and should be limited to 0.

Regarding the LES models, I have got decent results using the local dynamicSmagorinsky. However, it should be noted that OpenFOAM mesh requirements for LES is quite rigorous and x+ should be below 30. You can also find the local dynamicSmagorinsky, implemented for higher versions of OpenFOAM (3.0, 4.1, ...) on the forum.

I have also tested WALE model (the one I implemented my self, the one native to OF in higher versions is not correctly implemented IMO) and it correctly predicted the near-wall asymptotic behavior of O(y^3) for nuSgs.

I am currently implementing more recent functional and structural SGS models in foam. I hope to make them available soon.

Best Regards,
Syavash

[Santiago]

Originally Posted by Lookid

Dear Tobi,

from my small experience, I would agree with Bloerb, and say kOmegaSST is the best first try, however, it can lead to bad stuff also.

Example with a low specific speed pump calculation with kOmegaSST in CFX. By comparing a WF mesh (y+ ~100) and a Low Reynolds mesh (y+ ~1), the difference in the main characteristics can be up to 50% at high flow rates (see attachment, excel curve is my case, the second graphs are from this thesis : http://rosdok.uni-rostock.de/resolve...hab_0000001868)

The reason is due to that, at high flow rates, some detachment causing massive losses is simply not captured by the WF mesh. But I guess at this y+, it behaves like kEpsilon, but I did not check.

Here is a very nice course describing different RANS, URANS and LES models:
http://www.tfd.chalmers.se/~lada/pos...trgjmrDCKskNCE


I would like to extend the conversation to LES, since I just start it now with foam-extend, and it has been brought to the conversation already.

foam-extend will propose also a large number of models:

DeardorffDiffStress
LRRDiffStress
Smagorinsky
SpalartAllmaras
SpalartAllmarasDDES
SpalartAllmarasIDDES
dynMixedSmagorinsky
dynOneEqEddy
dynSmagorinsky
kOmegaSSTSAS
laminar
locDynOneEqEddy
mixedSmagorinsky
oneEqEddy
spectEddyVisc

How to choose between them?
I am aware of the problems of Smagorinsky (constant coef), but dynSmagorinsky is just not implemend as in the theory as you said, the coefficient does not vary "by cell" but just changed over the whole domain, making dynSmagorinsky pretty useless. AlbertoPa did implement a proper dynSmagorinsky though (https://github.com/AlbertoPa/dynamicSmagorinsky).

How to evaluate if an LES simulation is good?
I saw 2 ways so far : Evaluating res_turb (How to use dynSmagorinsky model correctly ?)
And plot somehow the -5/3 decay thingy.
Does it mean there are some ways to see if the LES simulation is good without experimental comparison?


EDIT:
@cryabroad, I went through the blog and indeed it is full of information, thanks!
I guess Smagorinsky problems are solved using OneEqEddy (basically is "less wrong" than Smagorinsky when using relatively coarse meshes, as I understood). dynOneEqEddy have the same problems as dynSmagorinsky but they actually also implemented a locDynOneEqEddy for local refinement. But I saw people arguing about that they obtain negative nuSgs values. Some say it is backscatter, some say it is unphysical and should be limited to 0.

oneEqEddy is a different model, a semi-empirical approach, compared to Deardorf's, Lilly's, Ghosal's, which have a purely mathematical foundation.

w.r.t negative nuSgs, is an error! antidissipation is NOT backscattering, physically it means that you're moving back in time (!!!). Backscatering can only be modelled by an appropriate deconvolution of the residual reynolds stress (i.e. Bardina, Stoltz, mixed models).

[Santiago]

Originally Posted by syavash

Regarding the LES models, I have got decent results using the local dynamicSmagorinsky. However, it should be noted that OpenFOAM mesh requirements for LES is quite rigorous and x+ should be below 30. You can also find the local dynamicSmagorinsky, implemented for higher versions of OpenFOAM (3.0, 4.1, ...) on the forum.

I have also tested WALE model (the one I implemented my self, the one native to OF in higher versions is not correctly implemented IMO) and it correctly predicted the near-wall asymptotic behavior of O(y^3) for nuSgs.

I am currently implementing more recent functional and structural SGS models in foam. I hope to make them available soon.

Best Regards,
Syavash

some comments:
(1) I guess the tests you've conducted are for resolved LES. An X+ of 30 is a bit low considering a streamwise lenght of 2 pi. Do you have a reference for that?
(2) You say that you get decent results using the local dynamic smagorisky model. does this mean that you're able to get the y+3 growth for nuSgs? Did you implement a filter?
(3) the y+3 behavior when using the WALE model, seems to me, is obtained partially because sqrt(ksgs) changes with the 3/2 power of sqr(S) (so third power). I'm not too familiar with WALE in any case.


The near wall behavior of nuSgs depends on the filter used. Literature (Piomelli, Meneveau, others) reporting on y+3 for the growth of nuSgs is because they use tophat filters in the horizontal (uniform) directions only. I find it difficult to get such a behavior with the default filters in OF, unless you implement one that only filters horizontal homogeneous planes.

The local dynamic needs in any case some sort of local averaging in order to avoid negative nusgs. Is that done in OpenFOAM? last I checked, it did not. but i checked some years ago.

[syavash]

Originally Posted by Santiago

some comments:
(1) I guess the tests you've conducted are for resolved LES. An X+ of 30 is a bit low considering a streamwise lenght of 2 pi. Do you have a reference for that?
(2) You say that you get decent results using the local dynamic smagorisky model. does this mean that you're able to get the y+3 growth for nuSgs? Did you implement a filter?
(3) the y+3 behavior when using the WALE model, seems to me, is obtained partially because sqrt(ksgs) changes with the 3/2 power of sqr(S) (so third power). I'm not too familiar with WALE in any case.


The near wall behavior of nuSgs depends on the filter used. Literature (Piomelli, Meneveau, others) reporting on y+3 for the growth of nuSgs is because they use tophat filters in the horizontal (uniform) directions only. I find it difficult to get such a behavior with the default filters in OF, unless you implement one that only filters horizontal homogeneous planes.

The local dynamic needs in any case some sort of local averaging in order to avoid negative nusgs. Is that done in OpenFOAM? last I checked, it did not. but i checked some years ago.

In response to your comments:

1) Yep. It was a resolved LES. You can find the LES mesh comparisons in the following paper:

https://www.sciencedirect.com/scienc...21999117304059

2) When I stated decent, I meant in terms of first- and second-order statistics. Check the following paper:

https://www.sciencedirect.com/scienc...45793019301872

3)Honestly, I did not go in deep. I just checked the local growth of nuSgs for WALE model against local dynamicSmagorinsky and I found it closer to y^3.
BTW, the original WALE model developed by Nicoud et al. does not establish a relation between kSgs and nuSgs. It obtains the nuSgs by establishing an "operator" which is composed of symmetric and anti-symmetric parts of the resolved velocity gradient.

A simpleFilter (box filter) was used in all my calculations with dynamic models (well, the dynamic procedure requires a filter).

Regarding your statement about the behavior of nuSgs, I guess you were addressing the "dynamic" models as only they require a "test" filter in OpenFOAM. In case of WALE model, no such filter is required.

The local averaging has been implemented already as mentioned in the previous posts, by Alberto Passalacqua and then upgraded for use in higher versions, see:

Improved implementation of dynamic Smagorinsky

However, the averaging does not eliminate the negative values of nuSgs. Another modification was done to clip the values of nuSgs such that (nuSgs + nu) >= 0.

Regards,
Syavash

[mehrabadi]

Hi,

In OpenFoam7, I could get access to the turbulent thermal diffusion as:

const compressible::turbulenceModel& turbulence(*turbulencePtr);
const tmp<volScalarField>& alphat = turbulence.alphat();

NOTE: turbulencePtr is a pointer to a compressible::turbulenceModel object I have instantiated in my code.

In OpenFOAM8, it seems that fluidThermophysicalTransportModel has been introduced and alphat is somehow constructed by the diffusivity class through the fluidThermophysicalTransportModel. I have not managed to access alphat in OpenFOAM8, though. Does anyone know how to get access to it in OpenFOAM8?

[Tobermory]

I agree that the two front-runner RANS closure models for "industrial use" are indeed the k-omega SST and realizable k-E; standard k-epsilon generally has too many flaws (turbulence over-production for bluff body separation, round jet anomoly etc.).

The choice is probably heavily influenced by the nature of the flow being solved (is the boundary layer the purpose of the simulation, or is the real "action" elsewhere), and therefore industry it is being applied to.

For atmospheric dispersion studies I always used realizable KE with commercial solvers, but I have found that this is not very stable in OpenFOAM's implementation. I.e. I almost always have to start with kEpsilon before moving to realizableKE ... but I never had to do that with commercial solvers. Does anyone have an idea why that might be? Is it just that it has cleaner numerics, and therefore less numerical (artificial) stabilising?

[syavash]

Originally Posted by mehrabadi

Hi,

In OpenFoam7, I could get access to the turbulent thermal diffusion as:

const compressible::turbulenceModel& turbulence(*turbulencePtr);
const tmp<volScalarField>& alphat = turbulence.alphat();

NOTE: turbulencePtr is a pointer to a compressible::turbulenceModel object I have instantiated in my code.

In OpenFOAM8, it seems that fluidThermophysicalTransportModel has been introduced and alphat is somehow constructed by the diffusivity class through the fluidThermophysicalTransportModel. I have not managed to access alphat in OpenFOAM8, though. Does anyone know how to get access to it in OpenFOAM8?

I noticed it in the release note of ver. 8. I wonder if there is an appropriate way for calling the turbulent diffusion.