What LES should do for you?
 

I often read people writing about using "LES turbulence model", therefore I open this post since I am curious about your ideas of what LES is and what LES should do for you.
You are also probably aware of the paper http://iopscience.iop.org/1367-2630/6/1/035

Working on it since 2008... still have to find out :rolleyes:

Jokes apart, i don't trust Unsteady RANS (or RANS, for that matter)

I'd say provide better results than ranse (not necessarily the right ones, just better would be enough) in all the areas where ranse fail.

what "better" means for you? does LES should be just "different" from RANS?

Good question! :D

I expect from LES just what its definition foresees, i.e., an under-resolved, poorly modeled direct numerical simulation of Navier-Stokes equations. LES should provide, with a reasonable computational cost, the correct underlying physical processes (within modeling limits), should automatically distinguish between laminar and turbulent zones, catch unsteady phenomena, give accurate macroscopic results… in a few words, provide something which is as close as possible to reality.

Considering the endless modeling, numerical, filtering, mathematical and even philosophical issues inspired by LES during the last decades (see ten questions by Pope, for instance), I think this is the minimum we must expect from it!

what an Irony: In RANS we first average the equations and filter all fluctuations and solve for the mean flow quantities only. Then we add the effect of fluctuations through turbulence model. Then in URANS we add the time factor through single time term which is not able to resolve all time scales.

Can any body tell me, what is the URANS actually do?

Why do you want to resolve all time scales? For uRANS you do the filtering before the calculation and when doing an LES you have to apply "the filtering" after the calculation to be able to interpret the results.

So whats the cheaper way of getting an engineering problem solved?

I know it is viable engineering approach. I actually cant understand the URANS philosophy. This approach may be good for the laminar transient flow, but not for the highly turbulent transient flows.

Just think of these two philosophies:

LES -> filter G(x-x'; Delta_x)
URANS -> filter G(t-t'; Delta_t)

Actually, an example would be useful, the cycle of a piston in engine. RANS will set a (hypotethically) statistical steady-state for all the cycles (1 cycle = 360 degree). URANS will give you the unsteady (statistical) behaviour during the 360 degree. Each time is representative of the average of N events at that time.

While in LES we get the all times and all N events at each time?

no.... in LES you do not have the statistical meaning of the solution at time t as the average of N sample... Indeed, in LES you have to simulate the N sample and explicitly do the average. LES require that you solve in time for all the scales, as happens in DNS

Does LES is useful only if RANS/URANS fails or you really need as much details as LES can provide?

LES for me is one thing above all: an intellectual challenge, and a good meeting place for fluid physicists, mathematicians, numerical analysis, high performance computing people....

plus it is damn fascinating and also frustrating :)


seriously: For me, it is a great research tool to do research with and to do research on!

good !

just to stimulate the discussion, I do a simple example... consider the "academic" test-case of the lid-drive cavity at Re=1000. It is a laminar case.
Then you run your code as it is on a uniform grid of 10 cells for side. You have a Re cell number =100 which means you are running the code in unresolved condition.
For you, is that an LES? In other words, any time you use a CFD code in unresolved conditions, you are filtering something and implicitly doing an LES?

well, my answer to this question is twofold:

1) yes, everytime you are using any type of discretization, you are doing an "unresolved DNS".... one might call that an implicit LES.... As soon as you are introducing a filter through approximation(by projection onto your approximation basis, e.g. phi=1 (the mean) in FV, dirac in FD, polynomials in FE) , you are no longer solving for u (the DNS), but for u(bar) (the filtered u). Plus you are adding approximation errors through your discretization of continuous term (like difference equations for derivatives), so, you what you are in reality dealing with is something like u(hat)_h.

2) Should every underresolved DNS be called an LES? No, I don't think so, because for me there's more to a LES than just picking a low resolution for a high frequency problem. But that's open to discussion, implicit filter implicit SGS LES-people would disagree maybe :)

Very interesting discussion! What are your thoughts about this?

Cheers, all the best!

I agree :)

I think that a cfd user must be aware of the fact the running any numerical simulation will allways produce an intrinsic filtering. The "label" of the simulation can be DNS or LES depending on the ratio between the Kolmogorov and the Nyquist frequency scale. I think that if this ratio is no longer than 10 we can still talk of real DNS. When this value exceeds 10 I use to call "unresolved DNS" or, better, "no-model LES".
In my experience, despite we know that is wrong, no-model LES often produces better results that LES modelled with an eddy viscosity SGS model.

So, further question arise, what an SGS model should do for you?

Ok, I guess we agree on the essence of the issue, although I tend to be a little bit more strict with what I really would call a DNS... in fact, (I have never tested it), I guess one would have to resolve at least half of the Kolmogorov scale to be fully fully fully DNS, since the non-linear terms will produce (for incompressible eqn) a frequency that's two times Kolmogorov... so to capture that (without any need for dealising), one you have to resolved the non-linear terms computed from the smallest scale...

but I have never tested that, and is doesn't really matter, I guess, plus that's then often so close to the numerical double precision barrier of 10E-16.... so it is just a theoretical point...

I think that's due to two issues:
a) interaction of the numerical errors and SGS model... I have the feeling that this often makes the overall results worse than without a model.
b) here's another idea, although it just a feeling, not a fact:
eddy viscosity models (at least the Smagorinsks) rely on two things: 1) the equilibrium of dissipation and production, i.e. isotropic turbulence and 2) the alignment of resolved strain and stress. Both assumptions are invalid in many cases, BUT (here comes my feeling) if you use your discretization error as the model (or the Riemann solver /flux function in FV), then your discretization has the chance of introducing the anisotropy naturally... I'm not sure if I can express it in a good way, but I have the feeling that eddy viscosity models are more "isotropic" than your discretization dissipation.

the textbook answer would be to model the effect of the unresolved on the resolved scales, I guess :)
For me, it should
a) turn off when laminar
b) allow backscatter
c) remove the energy, of course, at the correct rate
d) (maybe) even be clever enough to remove any aliasing errors I'm introducing through inexact nonlinear terms...
e) not assume isotropy of strain and stress
f) not destroy the time step
g) not interfere with parallelization

wow..... can think of more things, but have to get dinner now!

until then!

Very interesting discussion! :)

In my opinion the major drawback of many "classical" SGS models is that they have been derived within a formal mathematical framework which starts from the analytically filtered Navier-Stokes equations. As a consequence, such models imply a filter to the simulation rather than adapting to the real effective filter imposed by discretization and numerical scheme. One should rather start from the discretized form of the governing equations, but unfortunately the discretization operator is often unknown... however one interesting contribution in this vein was the one by Carati & al:

http://journals.cambridge.org/action...line&aid=82969

Another interesting point is the one made by cfdnewbie about the alignment between resolved strain and SGS stress models, and the inherent anisotropy introduced by numerical schemes. In fact, in the book on ILES by Grinstein et al. the authors show via Modified-Equation Analysis how some flux reconstruction schemes are equivalent to scale-similarity SGS models, which are able to capture anisotropic effects. I also agree on what a SGS model should do, and would also add: catch correct near-wall dependence.

I agree that the eddy viscosity assumption suffers of many limitations. However, non-isotropic eddy viscosity SGS provides somehow disappointing results, as well ...

In my experience the two-parameter dynamic mixed model is the more accurate, it provides good zero-order statistics as in the no-model LES but also quite good high order statistics.
Deconvolution-based modelling are also advisable for smooth filters.

Now, my next question is: what do you intend for an LES?
- a numerical approximation of a mathematical system of PDE obtained by filtering the NS in continuous form
- a way to model all is unresolved by the discretized NS equations

In the first case, in order to accurate represent the filtered solution, the ideal numerical method should minimize the local truncation error, that is exactly in opposite sense of the ILES approach.

In what way would that be possible? by selecting your number of DOF, you restrict the dimensionality (and thus exactness) or your solution space... so there is no real way to recover all that is lost....you might get first and second order relations correct, but beyond that, I find that hard to imagine....

or am I missing something?

No, I wanted to say "model", not "recover"... recovering outside the Nyquist freqeuncy is impossible as you correctly said.
Therefore, modelling all means modelling physical unresolved scale as well as taking into account for numerical errors. A report on CTR was based on a dynamic procedure taking into account also for numerical error

What should LES do for me?

Simple, fix the base flow issue with RANS!! And, I assume URANS has the same problem.

http://www.hegedusaero.com/examples/...celerator.html

This can be a HUGE problem with RANS.

So, for base flows, is URANS or LES better?

Hi Martin,
first I am curious to know what do you expect from RANS in your case... Any computed variable is statistical, that means you have a function like <f>(x) therefore, in your opinion, in what your RANS plots are wrong?

Second, have you tried to do a simple test, running your code without any turbulence model? This should give you the feeling of what the models really do...

Third, try LES ....;)

Sorry, I may have stated my post incorrectly. With the eddy viscosity model and with how flow features set themselves up, I don't expect RANS to be different. The results were not a surprise. However, I don't believe that in reality RANS is truly statistical. Yes that's what the math says for small local areas, but the flow in small local areas sets up the flow for the entire problem and this then feeds into other small areas. In other words, the flow state trajectory for solution convergence leads to a fictitious (to a degree) answer. For example, base flows. I believe, in general, the time averaged true pressure values along many base flows are (more or less) constant along the base. The deaccelerator is an example of it. However the RANS code gives a different statistical time averaged answer. In the end, a RANS result for base flow is an answer for a flow with a lot of local (created locally and/or transported in) viscosity. So, IMO, RANS flow does not necessarily represent flow values which have been averaged over an infinite amount of time. That's a problem with RANS. And that is the nature of the beast.

So I would like LES, URANS, DES, or anything else, to get me a better physical model.

Unfortunately Reynolds number is too high, so going without a turbulence model is questionable. It definitely will not get skin friction correctly. And, that feeds into some types of base flow.

I haven't tried LES/DES for this problem. Too expensive. The intent for the deaccelerator example was just to point out the issues of RANS to people. But, given what I've heard, I'm not confident I'm aware of all the ins and outs with LES. LES/DES "should" give me better base flow results, but I am also not sure where it falls apart.

For base flows, will URANS and DES results be similar? I've never tried URANS for it.

I believe that RANS solution are statistical in global sense, often they fail in providing correct (statistical) results only because the model is not perfect to take into account the fluctuations over the entire spectrum. ;)

For your problem I suggest to run an unresolved DNS just beacause it is wrong.... see the solution and then compare to a LES solution to assess the relevance of the model.

LES and DES are different formulations.... in DES you get mixed solutions type

Since there were a lot of discussion about RANS vs. URANS, how do people view the acceptability of using URANS on base flow?

Just to address an issue ... if someone sees the continuous URANS equations and compare to the LES equations (for implicit filter) he would for sure wonder what about the difference ... The answer is that there are apparently no differences ... is the type of closure that distinguish URANS from LES.
But is a model really capable to do this distinction?

LES is the process in which instead of modeling everything like RANS we apply filtering process and separate large eddies form the smaller one.

Based on Kalmogorov principle the smaller scales of motion are universal (isotropic) hence can be modeled similar to RANS. As the larger scales depend upon the boundary and flow conditions, hence solved like DNS.

So we can say LES is compromise between DNS and RANS both in terms of accuracy and computational cost.

ok, but this is just the theory...

- the filter process is never applied in practical LES (apart the explicit filtering approach). Filtering is only an implicit process due to discretization of equations and domain.
- the universailty of the SGS model based on isotropy of the unresolved scales is almost never obtained in practical LES since you can not produce a filter width so small in all the regions of a wall-bounded flow. As a consequence, often you have unresolved scales that are far from being isotropic.

Thus, what do you expect when analyse your LES solution?

what to expect from LES
 

...modelling mountain torrent flows, I can define what I expect from LES and how far it works:
I need to get the influence of the complex geometry on the turbulence structure and that of the presence of the free surface. I accept huge errors, because there is nothing better available.
Based on the thesis of I. Wendling (2007) "Dynamische Large-Eddy Simulationen turbulenter Strömungen in komplexen Geometrien" I started to consider the dynamic mixed SGS model as that one that accounts at least for some influence of the complex geometry, and with Keylock et. al (2005) "The theoretical foundations and potential for large-eddy simulations (LES) in fluvial geomorphic and sedimentological research" in Earth-Science Reviews 71 I started to see that the backscatter is important to catch the influence of the free surface, and that the dynamic mixed scale similarity model is ok.

So in the current version of OpenFOAM, the dynLagrangian SGS model is promising, but since the authors of the corresponding paper ('A lagrangian dynamic subgrid-scale model of turbulence' by Charles Meneveau 1996) themselves point out that the scale similarity should be included, I am looking forward for the day when there will be a lagrangian dynamic mixed SGS model in OpenFOAM.

One more thing I would like to see is a discussion about how to determine if the grid resolution is fine enough for LES. I use two-point correlations to check if the bigger structures are resolved well, since Lars Davidson shows in "Quality and Reliability of Large-Eddy Simulations II" Vol. 16, pp 269-286, Springer (2010) that the energy spectra and ratios of SGS viscosity to physical viscosity are not reliable quantities to estimate the grid resolution.

from my experience, I agree that dynamic mixed model is superior, in order to use on non regular grids (e.g. unstructured) the LES procedure must be properly written.

As far the grid resolution is concerned, this issue is somehow misleading for LES based on implicit filtering. The computational grid is also a measure of the filter grid, therefore the question is "what do you want that a filtered field resolve for you?"
I know the work of Davidson as I was present at the QLES conference and some conclusions he draw were criticized

I would like to add that, still in my opinion, LES is also the most natural and (in a certain sense) trivial way to approach a fluid flow computation when the computational resources are not enough for a DNS.

The main problem here is that we still don't know what is the most appropriate large scale discrete system of equations to be solved in order to remain consistent with the original Navier-Stokes equations... that is, a universal framework taking also the full numerical implementation into account seems far to come at the moment.

Of course, in this sense, the task is also far from trivial

...concerning the dynamic mixed SGS models and the lagrangian dynamic mixed model, I am happy they have been realized for OF 2.1. Here is the link, I post it here because I did not find it by searching cfd-online:
http://www.lemos.uni-rostock.de/en/cfd-software/

Hi All! this thread is very useful, there are lots of information about LES. But the fact is that most of them are advance level information and hard to get into it for the beginner. Moreover, most of the people in CFD are using different black box (fluent, cfx, openfoam etc)and they actually don't know how LES works and what are the criteria to meet the LES requirements and other relevant things. For the new comer in this challenging field you all could do a great help!

If some of you (it seems most of you are expert in LES, I will definitely mention the name of FMDenaro and Sbafinni, they are always helping lots of people and students like us regarding LES all the time in CFD online -respect!) post some lectures on YouTube including.......how LES equations are derived and filterd, .....how filtering operation accomplished, .......why FFT and other mathematical stuffs are required for the LES solution, .....how discretization method and modelling parameters impacts the solution etc. in sequential manner with pictorial presentation and required diagrams, it will be a great help for all new comers to get on the wheels and a great contribution to the knowledge in the field of LES as well.

I can definitely say that most of the new comers often go to the YOUTUBE (best open source e-learning resource) to have some visual idea about LES and actually how it works? And most of them are disappointed after wasting their plenty of time for nothing. Addition, there is no information about LES in Youtube.

Its my personal suggestion, and would like to place in front of you all for kind consideration.

Regards

Hi, what are you asking for is something that is only possible in the academic site...attending CFD and turbulence courses is necessary and have a practice with other students is very helpful.
I don't think that forums, youtube or other can go further in details.
Of course, don't forget that many books about LES are now available

:eek::eek::eek:
Youtube is definitely not the place to learn something about CFD.
Many of the tutorials there are crap and for a beginner in CFD it is impossible to tell which ones.

Sorry for picking up this specific aspect, I know it has nothing to do with the subject of the thread.
I didnt mean to distract from the topic, I just could not leave this statement uncommented.
Because apparently, many of the misconceptions that pop up in the subforums on commercial cfd software after long discussions about what appears to be a simulation results could be traced back to the errors and omissions made in the tutorials on said platform.
I am not saying that the "bad practice" was necessarily acquired there, but it might be.
And as Paolo has just shown, there is no lack of good material to learn from, it just has to be digested in a less convenient form.

Now back to topic.

If we can forget youtube for a while (sorry, but FMDenaro, flotus1 and Martin Hegedus are completely right and, definitely, that's not how we learned the few things we know), as you can imagine from my commitment, i would like to share everything i know to the best i can.

However, the required effort, especially in LES, is really far from trivial. Mostly because you cannot have a single target for your explanations.

While, probably, some time, i will try to put some effort on the work you suggest, this is not certainly the way a forum is intended to work and not the way it can be best used.

When i first used CFD-ONLINE, back in 2005, i was unable to use few lines in a UDF for Fluent. They were taken straightly from the manual and they didn't work. Stacked. Someone just answered with the correct way to do it, something i could have never achieved by myself in the time i had... even because the Fluent ufficial support could not help neither!!!

The second time i really needed the forum, i had a problem with the principal value component in vortex panel methods and the way these behave for closed surfaces with lift. In this case i got the answer after few weeks.

What i'm trying to say is that if you have a problem with something, you can come here and ask for an answer. Soon or later someone is gonna answer.

If you need a general explanation of a method, that could still work (but notice that, ususally, in this case, you get very short answers, possibly even incorrect).

If you need a full knowledge, that is not gonna work. More specifically, in LES you would need to first give details of numerical methods, how they work and how they are implemented. Then you should give details of CFD and RANS/URANS. How they work, how they are effectively implemented, how the Whole thing is actually set up to get some answers. After that you can introduce LES. But even in that case, trust me, you're not gonna get it in any case.

The first time i read the book of Sagaut was in 2007 and i made i full translation, from english to italian (my native language), of almost all the first 7 chapters, including all the derivations in the models. 2 Years later i was working on my master thesis and i read it again few additional times. While i would have not changed a word of what i wrote 2 years before, i definitely had a different view on the topic, and my 80 pages thesis chapter on LES was clearly different. Today, i'm writing my Ph.D. thesis (which is on LES of course). While i'm possibly using some material from my old master thesis, my view is still completely diffferent and in 80 pages i wrote completely different things, none of which i had any chance to get 4 years before.

To give you an example, consider that i just got, few weeks ago, that, possibly, the basis of LES was already put down by Reynolds in 1895,
including the Favre filtering for compressible flows and some very advanced concepts which would only emerge again in 1974!!!
That is certainly something you don't read in books (not all at least).

So, to better clarify what i'm saying: if you need an answer, you can get it here. If you want to understand something, you can still hope. If you want a full course on a topic and hope to understand it... well, forget it. That's simply not possible, no matter how effort someone could put on it.

Besides this, there is already some material on internet which is not bad at all.

http://www.eng.utah.edu/~rstoll/LES/Lectures.html

Still, thank you very much for the consideration ;)

Regards


Edit

Dear MMS, i hope you get this whole thing right.

P.S.
Just for curiosity, back in 2005, what i needed in my UDF was just the following line:

for (n=0; n <= cell_type_nnodes[(int)C_TYPE(c,t)]; n++)
{
...stuff...
}

which does a loop over the nodes of a finite volume cell. Now, consider the version reported in the manual:

c_node_loop(c,t,n)
{
...stuff...
}

Today i completely understand where the answer comes from (but i still don't get why the manual version should have been wrong). But, 9 years ago i was practically a child, i would have never managed it by myself. This is the best place for this kind of answers and, despite all, will always be.

1) In Pope's book, he has explained the difference between LES, DNS & RANS based on the spectrum. I highly doubt any "better" explanation is possible atm

2) In one of Dr Moin's interviews, he has pointed out the effective use of LES in acoustics which was one of the main reasons for which the funding was possible to arrange to continue the research on LES.

3)I think it has been explicitly mentioned by other members but still let me point out:-
DNS > LES > RANS

Conclusion:- LES can be used to:-
1)show-off (which is the most common case now-a-days)
2)acoustics
3)turbulence and related work

Hello,
I don't agree with your statement at all. In my experience, there are usually two scenarios:
a) Industrial applications of LES, where people either check the "Do LES" Box in their solver and don't know much about what happens behind the scenes, and it just like "better URANS" to them or they realize that they cannot do LES due to the parallel licences of their commercial codes or the CPU time required- because while LES is of course a lot cheaper than RANS, it is usually the temporal averaging that requires a lot of CPU time. So they consider LES unnecessary and stay away from it
b) researchers in the field, who usually are a humble crowd and are aware of the tons of open issues in LES and the complexity involved.

Using LES to show off is a very stupid idea, because it is so easy to point out the open issues / unresolved problems and shortcomings :)

From my experience, people working in the industrial field consider LES as a more accurate turbulence modelling than RANS, but then are frustrated while discovering the computational cost for problems with complex geometries and quite large lenght scales. RANS is still largely used.
Conversely, during the last years I observed a reduced interest in LES researches at CTR, see for example http://ctr.stanford.edu/ResBriefs/2012/index.html.

What do you think about?