[Ashkan Kashani]

Hello all,
I've got a 2D simulation case in which the flow separates from the sharp leading edge of rectangular bluff body and reattaches to the wall some distance downstream. The main goal is an accurate prediction of pressure distribution along the body's face parallel to the flow.
I'm doing a transient simulation using SST model in conjunction with gamma-Re transition model. The time- cord-averaged y+ is less than 2~3 and the inflation layer around the face of interest contains 10 prism layers. The Re number based on the body's width (perpendicular to the flow) is 1.7e+4.
The problem is that my model overpredicts the reattachment length, which in turn leads to delayed pressure recovery.
I have a suspicion that longitudinal decay of turbulence values specified at the inlet might be to blame. Consulting the Ansys CFX-solver Modeling Guide, I learnt that one solution is to prescribe appropriate turbulence values at the inlet based on the desired values at the body. An alternative approach also suggests some additional source terms for k and w transport equations in order to preserve the inlet values up to some distance upstream the body, from where decay is allowed.
Here are my questions:
1- Is my suspicion valid in the case of my problem?
2- Is the decay of turbulence of physical basis or a numerical artifact?
3- which of the two methods works better? Are there any attempts in the literature?
I appreciate your comments.t

[ghorrocks]

Your question is a research question so you will need to investigate this in detail to get an answer - there will not be a simple answer, and the forum will not be able to answer it for you.

On your questions:
1) Maybe, do some experiments where you vary the initial turbulence values and dissipation and find out.
2) Turbulence does dissipate, of course; but whether your modelled dissipation is correct or not will require you to do a validation and verification study.
3) Why not try the both and find out?

I would like to put in an alternate hypothesis: The SST model cannot predict the reattachment length and you will require a model which goes beyond the RANS assumption, such as LES. I am not sure if that is correct, but you should not assume the SST model is capable of modelling your scenario exactly.

[Ashkan Kashani]

Dear Glenn,
Thank you so much for your comment. I agree with you about the tendency of SST (and all RANS models in general) to overpredict the reattachment due to the underlying assumptions. That's probably why CFX offers the so-called Reattachment Modification (RM) to promote reattachment by enhancing the turbulence production term. I activated RM and got a much better result indicating a faster reattachment, as shown in the figure attached. Based on this observation, it crossed my mind that further improvement to my results might be possible through the idea of adjusting the turbulence intensity around the body by preventing the immediate turbulence decay starting from the inlet. This is also in line with CFX's recommendation that the turbulence intensity around the body should be above 0.1%.
1- If the turbulence decay is to occur quickly from the inlet as the turbulence transport equation would predict, on what basis are we allowed to tamper with that with no concern about introducing any unphysical impact?
2- I want to test this idea anyway. However, I could not find a clear way of defining those source terms since I'm still new to CFX. What I understood from the CFX guide is that a subdomain has to be defined first. Are you aware of any tutorials that thoroughly go through the procedure?

Regards,
Armin

[ghorrocks]

1) You are assuming the turbulence dissipation is correctly modelled. I would not be so sure this assumption is correct. Decay of isotropic turbulence is one of the key cases turbulence models are tuned to capture, but if you can show that the default model is not correctly reproducing the physics then you have a good reason to change the model to improve it.

2) Before you try your suggestion I would try all the available options in CFX for the SST model (there are many), and the other turbulence model options as well. Using an existing model is preferable if possible.

But applying source terms is simple. Have a look in the CFX documentation. You can apply source terms to boundary faces, or if you want to apply a source term to a volume then use a subdomain. It is very simple, you should not need a tutorial. But working out what the source term should be is the challenging part (the CFX documentation shows how to set a variable to value in a source term).

[Opaque]

If you want to spend time (play/research) modifying a turbulence model, the state of the art is to use GeKO (Generalized K Omega) model where the most important features a two-equation model can capture have been summarized into a minimal set.

I would recommend reading about GeKO in the CFX/Fluent manuals and evaluating it before attempting modifications to SST or other existing models.

[Ashkan Kashani]

Dear Glenn and Opaque,

Thank you so much for the helpful tips and insights.
I am especially curious about GeKo as I had never heard of that before, probably because I do not follow the most recent improvements to Ansys products .
From my understanding, GeKo provides the flexibility to tune the model through a number of coefficients that have no/minimum overlapping impact on the ultimate turbulence behaviour, nor do they tamper with the model calibration to the most basic flow cases such as flow over a flat plate. Well, much as this looks very interesting, I have some concerns which I will explain in the following.
There is always a chance that fine-tuning a model makes it so closely tailored to a specific dataset as to reduce its power to "generalize", i.e. to predict more general cases beyond the dataset used for calibration; a concept that is known as "overfitting".
So here is my question:
(1) Is this a valid concern in the case of GeKo? and if yes, how to address that?
(2) Is there a systematic procedure for calibrating the Geko model?

I do appreciate your helpful comments.

[Opaque]

Originally Posted by Ashkan Kashani

Dear Glenn and Opaque,

Thank you so much for the helpful tips and insights.
I am especially curious about GeKo as I had never heard of that before, probably because I do not follow the most recent improvements to Ansys products .
From my understanding, GeKo provides the flexibility to tune the model through a number of coefficients that have no/minimum overlapping impact on the ultimate turbulence behaviour, nor do they tamper with the model calibration to the most basic flow cases such as flow over a flat plate. Well, much as this looks very interesting, I have some concerns which I will explain in the following.
There is always a chance that fine-tuning a model makes it so closely tailored to a specific dataset as to reduce its power to "generalize", i.e. to predict more general cases beyond the dataset used for calibration; a concept that is known as "overfitting".
So here is my question:
(1) Is this a valid concern in the case of GeKo? and if yes, how to address that?
(2) Is there a systematic procedure for calibrating the Geko model?

I do appreciate your helpful comments.

Suppose you review literature about developing the fundamental equations for turbulence modeling (Wilcox book for example). In that case, you will note that those coefficients (most people tinker with) are interrelated and CANNOT be changed at will.

Those coefficients are tuned to match a dataset comprised of "useful" flows and consistently modified to respect certain aspects of expected physical behavior. From what also understand, GeKO reduces all those "free" parameters to a few which represent important features a turbulence model must represent: separation, mixing, jet, near wall, etc.

Can you overdo the fitting to a specific dataset? Certainly, but that is the point. GeKO allows you to fit the dataset of the flow of interest to YOU, and remove compromises made in other models (different tuning). However, once you tuned to a certain dataset is YOUR responsibility to only use that tuning for the range of cases you have selected.

Hope that helps