[Ashkan Kashani]

I'm testing the performance of GeKo turbulence model of CFX on the problem of flow separation at the sharp leading edge of a rectangular body. My simulations are quasi-2D and transient.
The figure attached displays the streamlines obtained from the simulation (shown in black) and those measured experimentally (shown in red), along with the respective centers of the separation bubble denoted by '+'.
By adjusting the coefficient C_SEP of the GeKo model, my simulation is able to reproduce the experimental dimensions of the separation bubble with good accuaracy (i.e. the reattchment length and the thickness of the bubble). However, the prediction of the flow topology embedded in this region is way off, the streamwise location of the separation bubble core in particular.
My suspicion, that is also supported by findings of others, blames it on the inherent incapacity of the GeKo model to estimate the correct Reynolds stresses, perhaps due to ignoring locally significant flow features such as streamlines curvature and anisotropic turbulence.
I'm aware of CFX offering features such as curvature correction or even more sophisticated turbulence model such as EARSM that may perform better. However, before any attempt, i'm wondering if someone could provide good hands-on insights on the problem as well as the possible solutions. I would really appreciate it.

[ghorrocks]

You are in turbulence model development here - this is a very specialised and difficult topic. Tuning turbulence models to actually improve accuracy is not a trivial task.

I assume you have done convergence, mesh size and time step size independence studies. You are obviously not a beginner, so you would know you are wasting your time unless you have done these basic accuracy checks first.

After that, I would look at your assumption of quasi-2D. Do you get the same answer if you run full 3D?

The SST model in CFX has curvature correction options (and many others as well). If you want to stay with a 2 equation model this is definitely one to try. I would also try a few of the Reynolds Stress models as well.

If your flow Reynolds number is moderate or low then a LES model might be worth a try. These models introduce a whole new level of challenges in resolution, computational resources and long run times.

[Ashkan Kashani]

Hello

In my previous simulations using SST, I set the target RMS residual to 0.00001 based on the results of a sensitivity analysis, and I used the adaptive time-stepping method homing in on 3 ~ 5 loops to let the solver pick on the right timestep size automatically.

Originally Posted by ghorrocks

I would also try a few of the Reynolds Stress models as well.

Following your suggestion, and also based on the literature reporting a superior performance of Explicit Algebraic RSM models, I decided to give it a shot. I adopted the baseline version of this model (i.e. BSL EARSM). I initialized my simulation using the SST model results, with the mesh and all other settings (including the target RMS residual) being the same.

It turns out that the BSL EARSM model requires the adaptive time-stepping to pick on timestep sizes around 2 orders of magnitude smaller (compared to SST), making the transient simulation almost impracticable!
Another thing is that the EARSM seems to be very sensitive to the timestep size (); a slight decrease in leads to an early convergence after, let's say, two coefficient loop iterations. Then as the adaptive time-stepping slightly increases in the next timestep, no matter how many iterations the solver goes through, the RMS residual of the V-momentum equation stalls.
I also applied the identical model to another mesh that I deemed of higher quality to see if this would play a role. I couldn't detect any meaningful change in the overall convergence behavior though.

My questions:
1 - Are these observations normal in the context of EARSM?
2- Possible workarounds?

Any comment is appreciated.

[ghorrocks]

I do not have any experience with EARSM models, but I have some experience with full RSM models. I would give a full RSM model a try as well. My experience with them is that they are a bit harder to converge than a 2-eqn model but as long as your mesh quality is good and you keep the time step suitable they go OK.

Your report of going 100 times slow with EARSM sounds excessive. I would check what is causing that - output your residuals to the results file and see where the region of problematic convergence is. It probably lines up with poor mesh quality and/or difficult flow physics (eg a separation point).

You are likely to require a much smaller time step in any RSM model compared to a 2-eqn model. This, in addition to all the additional equations you need to solve will mean that the simulation runs slower.