Poor cylinder wake results on unstructured meshes in Code_Saturne 9.1

[wayne_cai]

Hello,

We are testing the flow around cylinder case with Code_Saturne 9.1.

With structured meshes, the case works correctly in Code_Saturne 9.1. However, with unstructured meshes, we could not obtain reasonable results.

We tested:

- 2D / quasi-2D unstructured mesh
- 3D unstructured mesh
- Re = 200
- Re ≈ 3900

The observed problems in Code_Saturne are:

- No clear vortex shedding / Karman vortex street
- Poor wake field
- Abnormal velocity or pressure distribution near the outlet
- Outlet backflow messages such as `Incoming flow detained ... outlet faces`

For comparison, we ran the same meshes and the same Reynolds numbers in OpenFOAM, and the results were normal, with the expected wake and vortex shedding behavior.

So the issue seems related to Code_Saturne when using unstructured meshes for the cylinder wake case. Structured meshes do not show this problem.

Could you please advise whether this is a known limitation, a setup issue, or a possible bug? Are there recommended settings for unstructured cylinder wake simulations, especially regarding:

- outlet boundary condition
- gradient reconstruction
- extended neighborhood settings
- turbulence model setup for Re ≈ 3900
- mesh requirements near the outlet

Thank you.

[Yvan Fournier]

Hello,

Note that v9.1 is not a fully validated version, so may have some bugs.

In any case, the first thing to try is to add at least 1 layer of extruded cells at the outlet (you can do this directly in code_saturne, see the preprocessing section of the GUI).

For outlet conditions, trying both standard outlet and imposed pressure outlet is interesting. Imposed pressure is often more robust, but can cause issues (especially with buoyant cases, but it is safe to check.

We have observed at least one case where the extended gradient neighborhood causes issues at the outlet (even with extrusion layers), and are investigating this. This may also depend on the outlet type.

So for gradients, I recommend standard or reduced (heuristics or opposite face center) neighborhood, default (LSQ/Gauss hybrid) gradient.

If this is not sufficient, try adding some RHS reconstruction sweeps (instead of 1 for velocity and 2 for pressure, try 2/4, or even 10/5, which is the default for LES).

FYI, our reference cylinder test case uses a block-hexahedral mesh, but recent tests have also been done by the colleagues working on CDO schemes, and from discussions with them, their feedback is quite similar. I'll point them to thi post to see if they have any specific suggestions (or questions).

Best regards,

Yvan