LES vs RANS
 

Please help me to understand the difference between LES and RANS.
I'd read some article in this forum, but I still confused about them.

The point is this,

1) What is the exact difference between LES that uses box filter and RANS.
2) Is it true that RANS does not use time averaging?
Then why it's named after Reynolds-Averaging?


Thanks, YE

Dear Ye;
Consider a turbulent flow doamin. RANS can only give a time averaged mean value for velocity field. since it is based on time averaging. in fact velocity field in this method is averaged over a time period of "t" which is considerably higher than time constant of velocity flucctuations. therefore within the period of "t" we have only a constant mean velocity and could not monitor its time-dependent variations. for example, suppose that you can take successive pictures of velocity vector in a specific point of a turbulent flow during "t". clearly you can see that this velocity vector varies with the time. but if you use RANS you can only see a constant velocity vector which is an average of pictures you have already taken!.
On the other hand, LES is based on filtering rather than averaging. in this method you need to choose a filter size first. all flow scales larger than the filter size specified will be exactly calculated and the scales smaller than filter size will be modeled. now consider the picture taking again. if you use LES you can clearly see the variation of velocity vector at that specific point. smaller the filter size, more exact the time variation resolution of velocity vector.
as filter size reaches to zero, LES results turns to DNS ones.

regards

I'm also confused by RANS and LES.

If the grid size is the same, then what's the difference?

It seems that everybody says that LES need higher resolution grid than RANS, but I did not get a criterion.

Well, i think different turbulence model has different result. because its equation different (energy generation, diffusion, dissipation).. LES is better than RANS but less than DNS.. LES is good to capture small length scale (eddy scale).. therefore LESs grid must be better than RANs. because, if the grid is same.. LES cannot work well.. For the size of your grid, you must choose your length scale model (Macro scale, integral scale, taylor scale or kologorov micros scale)..
hope it helps
teguh

I think you have to read a course on turbulence modeling like (Moin, Sagaut, Poinsot ...) because it is more complex than what is exposed in the answers.

The main point is than RANS methods model all the turbulence spectrum and on the other hand, LES model only high frequency while computing the smaller ones (high frequency = small scales, which are supposed to be independent to the configuration of the flow).

The choice of the mesh depends directly of the geometry, the Reynolds number etc ... and will drive the quality of the LES (with a given model of course).

To conclude my post, the best advise I can give is to read a book !!

Thanks for your explanation.

I guess the main problem for me is that for all the books I have read, "LESs grid must be better than RANs".

But for some journal articles I have read, say, somebody use RANS with grid size A, while some others do LES with grid size B, and B >> A (order of magnitudes). I'm confused by this fact.


So my question is that, is there any universal criterion about the grid size between LES and RANS, if the grid size is larger than a specific value, one should use RANS, if the grid size is smaller than a specific value, one should use LES?

or the grid size is dependent on specific phenomenon, which allow the grid size for some LES simulations larger than (orders of magnitude) that for some RANS simulation?

To answer your question, their is no universal criterion ! That's why I advise you to read interesting papers and books (I have forgotten Pope work also).

What you have to keep in mind is that for a theoretical view point, using LES require to use a mesh and a model that provide a cut-off in the inertial range of the turbulence spectrum. And most of the time (for practical applications), you just don't known these scales before doing the computation. That's why, you have to check the quality of your computation after doing the computation and check the sensitivity of the results to the resolution.

From a practical point of view, most of the time when people do RANS computations and want to go further, the first idea is to use the same mesh (in order to keep a low cost computation and compare the results given with the same mesh). This is only due to practical reasons. Then, when you have a certain experience with LES, you can feel before computation the mesh resolution needed by the physics. But never forget to check if the resolution is sufficient.

To conclude:
- their is not any universal criterion, most of the time LES meshes are more fine that RANS meshes,
- one have to check the quality of his mesh after the computation (Pope criterion for example),
- finally, the quality of the mesh largely depends on the price you can pay for you computation: CPU !!!! Normally, a good LES model goes to 0 when the resolution is increase ...

Thanks Florent for your explanation.

LES is the field I have never tried, but on my todo list in the near future. Surely I have to do more reading. Your post helps me a lot.

From the limited reading I have done, I just have the impression that in journal articles, the authors just write they use LES because it's better than RANS, while they do not give explanation for the reason why the specific grid resolution they choose is fit for LES, while not RANS. From my limited knowledge now , I could not figure it out. Perhaps I have igored something.


Thanks for kind post again.

Thanks for everyone.



Actually, I've read an article in here.

http://www.cfd-online.com/Forums/mai...-dns-rans.html

In this article, one said that

For RANS, there is no time averaging (this is a common misunderstanding). The average is done - virtually - on a large number of the same experiment. Hence, as turbulence is chaotic, all the experiments are slightly different and RANS captures the mean. You are thus able to capture - if grid size is fine enough - some structures, for instance Von-Karman street eddies.


It makes me confused.
If he is right, RANS do ensemble by itself?
I mean, if I operate RANS one time, it done several time and output the ensembled one?
If he is incorrect, please notice me which is correct.

Thanks!

When you use Reynold's Averaged Navier-Stokes, what you get, is the mean of the velocity at one point in an interval, but there are also fluctuations which you are trying to model in some way (clossure problem). You realize only one simulation, and what you are getting is the huge part of the velocity, the variance (or fluctuation) of the velocity is deduced from other equations that aren't Naver-Stokes.

I recommend you to read the mathematical deduction of the RANS, it is simpler than you imagine and it can help you to understand it. It is much difficult LES's one.

Hope it helps

Thank you, Davahue!

You should notice that one of differences between RANS models and LES models is grid. for RANS models results are independence on grid sizes. i mean if your mesh size be decreased from specific size your results be same. it is desirable mesh size for RANS models and you don't need to fine it more.
But in LES models, results have grid dependency. less sizes more accuracy. in very fine meshes your results can be reach to DNS results. in this situation you solve all of your flow instead of modelling.

Thanks for your explanation. Personally I accept the idea that DNS need finer grid than LES, LES needs finer grids than RANS.

The main problem for me now is , for a specific simulation, how to determine if the grid size I choose is suitable for RANS, LES, or DNS? The limited papers I have read just present the the grid and the method they choose, without giving the reasons why the grid they choose is suitable for LES or DNS. I need more reading to solve this issue. Surely I appreciate your help if you could give more idea.

In fact, the mathematic equations don't know if they are RANS or LES.
Actually, the equations are same to some extend!

Imagine you use a same uniform mesh to do two simulations, one URANS with mixing length model and with the Lm chosen to be Cs*dx, and the other LES, with Smagorinsky model. You should get exact same results.

Often this aspect leads to confusion... URANS and LES equations can be quite similar in the appearence. The key is in the way in which you model the unresolved terms: it is different. In URANS the unresolved part is related to a time averaging field while in LES is related to a spatial filtering.
Consequently, is the model that creates the difference between LES and URANS, and the resolved fields have different meaning.

Although the physics of RANS and LES are different (as explained by Florent), the equations are basically the same. The Reynolds stresses are simply replaced by the SGS stresses. Frohlich makes the following distinction between RANS and LES: If the turbulence model depends on the grid size, you have a LES model. If it does not, then you have a RANS model. You can see how people turned classic RANS turbulence models such as Spalart-Allamaras or kw-SST into DES models by simply introducing a mesh dependency. It does not make a lot of sense physically (again, because the physics of RANS is not the same as LES), but that is pretty elegant... :D

Fröhlich, Jochen, and Dominic von Terzi. "Hybrid LES/RANS methods for the simulation of turbulent flows." Progress in Aerospace Sciences 44.5 (2008): 349-377.

In my opinion LES and RANS equations can not be confused, simply because the RANS equations have no time derivative by definition.
Conversely URANS and LES can have a similar formulation, as a matter of fact URANS can be seen as a form of LES with time-filtering

This is a very useful document in simple language provided by Prof. Lars Davidson. You can find several fundamental points related to turbulence and fluid dynamics in it, and ansewer of most of these questions were asked in the threat.

"Fluid mechanics, turbulent flow and turbulence modeling"; you can download it from:
http://www.tfd.chalmers.se/˜lada/MoF/lecture notes.html
and
http://www.tfd.chalmers.se/˜lada/comp turb model/lecture notes.html

if you have problem in downloading it, send me an email. I 'll willingly send it to you.

Hi sorry to bring this up as I have some queries regarding RANS and LES.

I just want to sum up the above dicsussion from what I have learnt. please correct me if i am wrong.

RANS
- The equations to be solved are obtained by applying a time-averaging and reynolds decomposition technique on the NS equations.

- Time averaging is to obtain the mean quantity over a certain time interval (i.e. timestep?).

- I also read that RANS equations is also ensemble averaging, which is averaging over a large number of the same experiment? Why would the result be different if they are all numerical experiments (shouldn't the result be always the same)?

- The unsteady RANS equations has a time dependent term (URANS and LES equations look almost identical actually). Does that mean that the resolved time dependent term give unsteady behaviour of the mean quantities? Does that also mean if my time interval (i.e. timestep?) goes very small, solving RANS would give a accurate time series result?

LES

I also have a quesition on the grid size on LES.

So the advantage of LES over DNS is that the grid size of LES can be larger, thus reducing computation time.

LES equations are obtained by a spatial filter of the NS equations. The filter width should be larger than Kolmogorov scale (smallest length scale in the problem). LES is then performed on the filtered NS equations. The scales which are smaller than the filter width are modelled by a Sub-Grid Scale (SGS) model, and the effects are therefore accounted in the solution.

If I understand LES correctly, filter width = grid size?

I have started off with steady RANS and moved to unsteady simulation. Now, I am confused whether LES or URANS is giving me a correct representation.

if the filter width Delta is implicitly related to the grid size h, then yes, the computational grid measure the smallest resolved scale..
However, it is possible to use an explicit filtering in LES. In such a way, the filter width can be as larger as you want, for example 4 times the computational grid size or greater

Let me pic up this topic.
Recently, I experimented with a Rayleigh Taylor Instability using Fluents VOF method just for fun (see the picture on the left).
As it turned out, even with an undisturbed initial solution (a perfectly flat interface) I was able to trigger the instability. It is the roundoff errors that accumulate and produce the initial disturbance.
What is even more surprising in the first place is that for example reordering the grid and running the simulation again produced an entirely different solution.
So the roundoff errors that are random to some extent are usually enough to produce different solutions for two otherwise identical DNS.

well, this is a typical effect, even simply porting the code on a different computer o using different compliers you see different numerical transient...
However, it is really important assessing that the statistics are all equal, whatever the numerical transient is

thanks for the reply. To be explicit in my understanding, is there any recommended criterion on how to set the width of the filter? Must the size be for e.g., be in the inertial subrange?

for implicit-based filtering, the size is determined by the grid and numerical method, you can only set exactly the grid size, not the filter width...

were you solving for a steady solution?

Im rather confused. if i understood correctly, the production of different results has nothing to do with the equations to be solved (e.g. RANS, LES), but rather how the grid is structured and the precision handling of the computers on which the simulation is run.

in LES you compute statistics after a long-time integration is solved and several flow fields are sampled...
Difference in the statistics can be due to the grid, to the numerical model, to the SGS model, etc.

to compute transient (unsteady) solutions, how should one choose between URANS or LES? kindly explain in simple terms.

I have the same question. For one specific problem, I can use the same grid sizes for URANS and LES. Which one to choose, URANS or LES?

URANS and LES provides solutions that have different meaning! the URANS is statistically averaged, the LES is filtered in space but not in time (if you do not use specific time-filtering apporach).

Using URANS o LES depends on the problem and on what you want to compute...but LES requires more refined grids than URANS

Thanks for your reply.

From the definition, they indeed have the different meaning. But the final form of the averaged equations (URANS) and filtered equation (LES) are nearly identical, except that the form of the sub-grid parameterization. To me it just like one method with two different parameterization methods.

I've read this text many times, but the answer I want to know is how to decide to choose RANS or LES for one specific problem, but not "it depends". I'm also curious to know, for one specific LES study, if I do a URANS study with the same grid size, what' wrong?

Thank you.

the mathematical equations looks similar but they differ for the expression of the unresolved terms. They take into account different unresolved tensors.

URANS can be used when an external unsteady force is present, for example the compression/expansion flow in cylinder due to the piston motion.

comparing LES and URANS on the same computational grid can give very different results for many reasons...
just as example, if you use a very very refined grid, LES solution will automatically tend to DNS, URANS does not.

Without regarding computation time, can i say, one should always trust LES to give a better result?

Can you help me understand how fine the grid should be for LES cases?

the answer is...it depends!
you can find very bad LES simulations... (too coarse grid, to dissipative model, etc)

the guideline for a LES grid is to solve wall boundary by using at least 3-4 cells within y+<=1.
Unbounded flow regions can be meshed by grid sizes dx+,dy+,dz+ = O(10)-O(20)

thanks. im new to CFD and appreciate the reply.

why would LES be too dissipative? i thought LES resolves all large scales, and only small scales are modelled - unless you mean that the SGS model are also too dissipative in the first place.

refining the mesh to achieve a 3-4 cells representation of y+ = 1 would require a very fine mesh.

if im interested only in wake structures, which are far away from walls, do I have to resolve my grid to such a resolution?

Some SGS models are purely dissipative in nature (e.g., the Smagorinsky) and could strongly affect the energy transfer of large resolved scales.

Furthermore, if you have vortical structures in wakes but that interacts with a wall (that is are generated by stress or are dissipated within the B.L.) you need to have a refined grid at wall. Some wall model exists in LES that allows to apply special B.C. at y+>1 but the full success is debatable.

i have been rereading this statement many times.

since the size of the filter is implicity implied by the grid size, does it always mean that the mesh density of a LES case is always higher than URANS?

yes, generally LES grid is more computational extensive even for the fact that URANS could be used (depending on the problem) in 2D.

in a few posts back, u mentioned that

can you explain how you judged that URANS can also be suitable (i assume LES is just as suitable in the quoted case ; please correct me if wrong).

in my current problem, im only interested in the wake (and my domain is set so large that it will not interact with the walls). For LES, must I still be bothered with refining the grid then? Will URANS suffice?

your question is not so well-posed ...
let me explain better, you have a flow problem to solve and you have to choose the suitable tool.
Now you said to be interested in wake..that does not say nothing useful to me...what parameter you need from your solution? It is the type of information you need in the wake that can drive to the correct tool.
Therefore, if the wake is generated my a mixing of two or more corrent, you have a sort of evolving mixing layer that requires fine grid and correct time accuracy. You need to analyze special frequency analysis? that drives to LES as the time scales are solved directly.

Consider my suggestion, ask yourself what you really need in your problem and step-by-step decide the correct tool.

Some reading of a book such http://www.cambridge.org/us/academic...urbulent-flows

cna be useful