[JulioPieri]

Originally Posted by LuckyTran

How are you deciding that the laminar case is better than RANS with no data? Because the results of the openfoam tutorial is more colorful?

I am distrusting that your laminar case is actually producing better results than the RANS because at the coarse mesh sizes we are talking about, the same spatial structures should be present in the urans as you see in the unsteady laminar case. I feel like there is more human interpretation error than actual CFD error.

LuckyTran:
That's exactly the point I'm bringing here... Why do the laminar results look (visually, judging only by human interpretation) better? I dont have a clue if they are better, and the discussion here is exactly to assess: is there a way to check if the unstructured DNS quantitative results are better than the uRANS? Assessing just trends, rather than precise numbers, in both cases are ok? Is my mesh so coarse that both results are distrustful?

From the pictures attached before, the results looks more what one would expect to see. They are much more convincing, though I know they are not trustful. Where is the flaw, and why is this situation counter-intuitive? Is it because de uRANS model is a closed model and, even though it does not resolve small structures, it models the influence of those structures on the flow, leading to better overall (averaged) results? Why the turbulence structures being partially resolve is worse than a model that does not resolve them at all?

Prof. Denaro, thank you for all your help. I'll work a little more on my geometry to increase the overwall resolution of the domain. Some other questions and comments:
1) I'm unsure whether to remove the tubes and replace them by a wall around them (like a chamfer on the geometry), or to create a porous region to capture some of its effect. See picture attached. I'm not very fond of porous regions as they are yet another model to pile-up, validate etc... What is your opinion about it?
2) I'll create this test case in my code (OpenFOAM), and see if the results from an equally unresolved DNS are anyhow comparable to an uRans and experiments. Do you think it would be possible to infer conclusions to my big case from this small one? Such as: if the results are ok from the small case, I can assume that with same Re/resolution the results would also be ok.
3) Could you point me books and literature to further study these foundations of CFD (mesh, numerics, etc)? To judge reliability of the results, etc. I do have (what I thought to be) an ok base knowledge from my time in Cornell, but as time passes, it seems I get lost in "practical" (and imprecise) CFD use.

[FMDenaro]

Some hints:


- unresolved DNS means that both the grid size and the discretizations act as a filter on the smaller spatial components of the flow. At this point, one has only the implicit action of the local truncation error of the scheme, not a physical closure model. But some kind of schemes produce a local truncation error that mimicks the dissipative effects others not.


- URANS: this formulation is largely debated, it is not clear its rigorous theoretical foundation. Theoretically it is based on ensemble averaging in case of an external time dependent driving force. That is in case a clear separation of physical unsteadiness is present. In your case, you have no external forcing, your flow appears statistically steady. Thus, what you would describe is the URANS based on a local time averaging, nothing but what is also called Temporal LES (TLES). But here, the action of the closure model is not limited to the small spatial scales. Actually, it is not clear at all what this kind of simulation means!
Have a look here:
https://www.researchgate.net/post/UR...5-PszMTgPWFclL


- For the LES literature you can read Sagaut, for URANS Wilcox. But many details can be found in published articles. Have alos a look to the reports at the CTR.

[LuckyTran]

The laminar results "look better" because it has the instantaneous structures you expect in an instantaneous snapshot of the flow (i.e. more colorful).


The variables in RANS (steady or unsteady) are the statistics of the flow. If you want to compare apples to apples, do a steady laminar calculation and compare it with the steady RANS. Or, take the time-average of the unsteady laminar results and compare it with the time-average of the unsteady RANS. And for both of these, you should compare it with a super-long exposure time photograph of the flow for example.


Consider rolling a die. The average roll is a 3.5. But a die will never land on 3.5. It can be considered counter-intuitive why a die that does not have a 3.5 on it will roll an average of 3.5. Does that mean all dice should be labeled with sides only 3.5 or is the mathematical formula for an average wrong? No and No. Apples vs Onions.


If you are actually seeing these structures in your coarse grid laminar case, the structures you see will be spatially and temporally coherent (i.e. not turbulence at all) and they will show up in both unsteady laminar calculations and unsteady RANS calculations. I think you are making too hasty generalizations about turbulence modeling without having actually done the simulations yet.


Emissions regulations are based on hourly averages (and even daily averages) because it is understood that instantaneous ppm of pollutants can temporarily exceed the average, because that's how averaging works. The EPA knows we are not doing spatially and temporally resolved measurements and that we are not capable of doing spatially and temporally resolved simulations.

[sbaffini]

Don't take me wrong but, why do you think you have the capability to visually assess your computational results?

Also, visually from what? A somehow arbitrary mapping between your instantaneous numerical solutions and some colors?

Everything stated by Filippo is correct but, it should really be put in the right perspective, otherwise there is the risk of accepting a sense of arbitrariness that really isn't there.

Concerning your unsteady solutions, assuming they are both 3D (otherwise, there is really nothing to compare here), and without knowing your grid and numerical schemes (but assuming they are the same in both simulations), we can surely say you are comparing an URANS with a sort of LES. Let us also assume that your URANS is correct. Your "sort of LES" could still be totally wrong and only apparently resemble something meaningful.

On one side, you could be doing an actual DNS, with mesh and time step sufficiently fine to actually resolve everything.

But, you could have actually been doing what is termed an ILES, or coarse DNS. Your mesh, time step and schemes are not really there but, "you know what, it works".

Finally, there is the chance that you are doing none of the above and your grid, time step and scheme are having a strong influence on the flow dynamics.

The crucial point here is that you don't typically catch these differences by eye. Actually, experienced eyes can rapidly catch signatures of BS (linkedin is flooded with such videos of self-proclaimed LES which are polluted by numerical artifacts), but you can't do the opposite, validate an LES simulation by eye. Actually, it is also not completely clear what an LES solution should be, so you typically look for certain standard facts you know about LES (mostly spectral content and correlations, but eventually also mean fields).

For my phd thesis, I've perfomed hundreds of LES on different geometries, different grids, sgs models, numerical schemes, etc. Except for pathological cases, all the rest had pretty much similar instantaneous snapshots of the fields, according to my eyes. Yet, numerical results varied greatly when actually analyzed. For example, certain time discretization schemes, when used with too large time steps, strongly affects the dynamics of intermediate scales (not nyquist ones, not mean ones, not directly at least). How can you possibly catch that by eye? LES channel flow results are strongly affected by the spanwise grid resolution but in a way that is typically evident in the mean quantities, not much in spectral content. Again, how can you perceive a wrong mean flow if you don't already know the final result (as is the case for the channel)?

So, trust me, that's not how this is done (despite most people/companies doing exactly this to market themselves).

Stressing instead what Luckytran has evidenced, I would also point to the fact that when you do some simulation you actually have a definition, implicit or explicit, of the meaning of the variables you are simulating for. If you are doing a DNS you should expect a solution resembling the actual flow field. If you're doing an LES, you should expect a spatially/temporally filtered counterpart of the actual flow field (plus dynamical artefacts due to the limited resolution). If you're doing an URANS there is actually a debate about what you should expect, but let's neglect that and practically assume that you should expect some form of ensemble average of the flow field over a possibly infinite number of experiments. These are very different variables and you should not really compare them.

In partial justification, I can tell you that I also had that "WTF, you gotta be kidding me" moment when I saw my first RANS results, but you will learn the difference.

[FMDenaro]

Originally Posted by JulioPieri

I see... So I might be too optimistic when seeing the results from unresolved DNS. I'm mostly using limitedLinear discretization scheme.

Do you think I can extract useful results from my grid as is? Even from a "simple" turb model like KE?

I'm using wall functions at the BCs for the turbulence quantities.

One option is that I might be able to simplify the geometry by removing the tubes inside the domain (either completely, or replacing it by a porous zone); that would give me a bunch of elements to redistribute throughout the domain. Maybe that will improve the accuracy.

About the lack of data, it is indeed a problem. I'm not expecting to get quantitatively accurate results. I intend to study "trends". I proceed tweeking the parameters to get an abstract "base case" and assess changes from that case. Then, I intend to observe trends of improvement, rather than the _actual_ (quantitative) gain. So, I can say "changing this operational parameter proved to have a good effect on the results, about 10-15% improvement", for instance.

You can have some idea about the LES no-model results for several codes here:

https://www.researchgate.net/publica...extFileContent

[JulioPieri]

Dear all,

Thank you for your time to answer my questions. In fact, I have a long way to go to precisely understand the mathematics behind turbulence.

The thing firstly got me confused because I was able to see structures (big vortices) with laminar, but not with RANS (supposedly unsteady). But now I understand what was going on. Doing the time average of both unsteady cases (laminar vs RANS), I'd probably get "ok" results (under some margin), but the laminar would underestimate turbulent effects, thus being somewhat (even) less correct than RANS. I've also realized that OpenFOAM doesn't actually solve uRANS (correct me if I'm wrong), or at least is not something fully agreed upon.

Now, I'm focusing more time on simplifying the geometry to save computer capacity to solve a more refined mesh. Also, I saved some references you suggested to study more.

[LuckyTran]

OpenFOAM and all unsteady RANS solvers are (numerically) doing the correct URANS. What is not agreed upon is whether the same k-epsilon model or k-omega model, which is based on statistically stationary variables, can be blatantly applied to unsteady cases or if these models need to be modified for URANS scenarios. Maybe the fundamental approach of the models is okay but the model coefficients need to be tuned, maybe a certain term is missing entirely, etc. These are tough questions to ask considering that the models in question aren't even correct when used "as intended." For example, there are a lot of scenarios in turbulent backscatter where a negative turbulent viscosity is the correct desired result, but a k-epsilon and k-omega will never yield a negative turbulent viscosity. So now we go and use these models that are known to be wrong under many scenarios and apply them to a situation where we have very little clue what they are even doing. Suffice it to say, we are fairly confident they're not always doing the right thing. Some people lose their minds and can never sleep because this bothers them. And then there are those that understand "All models are wrong, but some are useful" -Box