Radial Turbine Simulation: Convergence Challenges with SST Model
 

Hey everyone,

I'm working on simulating a complete radial turbine geometry, including the volute, stator, rotor, and diffuser at a steady state. The interfaces between rotor-stator, volute-stator, and rotor-diffuser are considered stages. Initially, I conducted simulations without the volute and diffuser using the k-epsilon turbulence model. Then, I added these components, obtaining satisfactory results. However, rotor and stator until now had Y+ around 1 (not ideal for k-e, but done for further work).

To enhance accuracy and explore flow separation probability and boundary layer depth, I switched to the SST model. However, I'm encountering convergence issues. The RMS residuals remain above e-4, indicating potential numerical problems. The MAX residuals are significantly higher than RMS, hinting at numerical instability. The monitor points, evaluating the inlet-outlet mass flow difference, show fluctuations, indicating lack of convergence, especially at the volute (as seen in the screenshot).

While the k-e model struggles to capture stator flow separation, the SST model shows potential but lacks convergence. I tried adjusting the timescale for better convergence, but it doesn't resolve the flow separation issue.

I'm uncertain about the problem's source. Should I simulate the entire geometry? Could it be a mesh issue? Do I need to make adjustments? Or should I consider the problem converged? Currently, I'm simulating a single passage of both the rotor and stator.

It is quite likely you have a small separation which the k-e model suppresses but the SST model resolves. This means the k-e model converges nicely but will be wrong, but the SST model has problems converging.

Read the "Tips to obtain Convergence" in the CFX Modelling guide, but in short: try a larger time step, and if that does not work try a transient simulation.

Thank you for your response. I've previously gone through the section you referred to. I understand that it's not a transient flow (simulation) because, upon increasing the timestep, it converges well. However, despite this, the results fail to capture the separation accurately. It almost resembles the behavior seen when using the k-e model. This is the primary issue I'm encountering.

If you get good convergence with SST and a larger time step but the results are inaccurate then I would suggest a) Checking your boundary conditions match the experiment, including incoming turbulence levels and wall roughness, and b) doing a mesh sensitivity study to see if a finer mesh helps.

Did you check at which region your MAX Residuals are?
Maybe at the cutwater?

I simulated with different mesh sizes (excluding coarse meshes), yet the outcome remained unchanged. Despite elevating the wall roughness to 80, convergence was achieved; however, the persistent issue in capturing the separation region persisted.

It is evident that the flow in question is turbulent, as indicated by a Reynolds number of approximately 10^5. The maximum residuals are observed in the vicinity of both the rotor and stator blades.

Do you have any additional recommendations or suggestions to address this matter further?

The goal is to converge the residuals down to 0 using whatever timestep allows you do so. Verify the imbalances are also 0 (within reason)

Once converged, you analyze the results and compare them to existing data. If it does not match, you then review the model:
1 - Mesh quality, independence
2 - BC realism as level values, or wall settings
3 - Model validity

Dear Opaque,

I believe our focus shouldn't solely be on achieving converged residuals; rather, our primary aim should be to arrive at the correct solution. If the residuals were the sole concern, I would have opted for a k-e model with a coarser mesh.

Unfortunately, I lack comparative data. The available datasets pertain only to the stator, rotor, and diffuser with the k-e model. My attempts involved designing a volute and altering the diffuser's geometry for simulation, thereby eliminating the option for direct comparisons.

From my perspective, the model holds validity, and the boundary conditions are sound. This assertion stems from consistent results obtained using the k-e model. However, transitioning to the SS model inevitably leads to convergence issues.

Converged residuals are the primary requirement to evaluate the quality of any solution. If the residuals are not converged, your solution fields are suspect.

Anyone showing me results must 1st and foremost show me residual levels and convergence. Otherwise, I do not waste my time evaluating solutions for future decisions.

We must keep in mind there is a specific order in the modeling process

1 - Pick a model --> define which equations to solve
2 - Solve the equations such as the equations are satisfied --> residuals are zero (within reason), and imbalance are zero (within reason) as well
3 - Repeat (2) with a refinement of "dependency parameters" such as mesh, timestep and whatever the quality of the model may depend on until the solution no longer depends on such parameters
4 - Compare to data to see if the selected model is/was appropriate. If not, back to (1), improve the model
5 - Rinse and repeat until solutions make sense.

As an update and a question: When I modified the wall roughness of the volute, convergence was achieved. However, the converged SST model does not exhibit any flow separation, unlike the unconverged SST model.

1. Does it make sense to apply wall roughness only to the volute and not to other components, such as the diffuser?

2. The wall roughness in the volute influences the mass flow. Since I do not know the exact material of the volute, and the material has a range of wall roughness values, which value should I consider?

Wall roughness will generate more turbulence and reduce separations.

The default wall condition is smooth - which means perfectly smooth. In other words, a perfect mirror polish. Few real surfaces are this smooth, so you probably should use some wall roughness to reproduce this. For many simulations wall roughness makes no difference but in your case it might.

The best way forward is to measure the surface roughness on the part you are modelling. If you cannot measure it then you can infer it by adjusting it and finding which roughness gives the most accurate results, but this is risky and much less accurate so avoid it and do direct measurements if possible.

There is an add-on modification to the SST model, named Reattachment Modification. Some people have used it with success in certain cases. An alternative is to step back a bit from SST and use the BSL model.