[aakie]

Dear OpenFOAMers,

I have trying to achieve grid convergence for the mean drag coefficient and RMS of the lit force coefficient on a stationary, fixed cylinder at Re = 3.6*10^6. Yet, it seems that the drag and lift force coefficient keep on increasing with relatively big jumps with every mesh refinement. I have used the following meshes:

1. Coarse mesh: N = 69.069
2. Regular mesh: N = 98.730
3. Fine Mesh: N = 176.635

All the meshes were done in 100Dx100D domain. The mesh was divided into multiple blocks, where the cylinder was surrounded by the typical O-grid block. The turbulence model that I am using is k-w SST without wall models. Y+ was smaller than 1 and equal in all three meshes. I have seen in literature another numerical study (title: Numerical simulation of flow around a smooth circular cylinder at very high Reynolds numbers. M. Ong et al., 2009) where grid convergence was reached for a 2D stationary cylinder at Re =3.6*10^6 with approx. N = 50.000. Yet, the paper I am referring to used the k-epsilon model with wall functions and they used a smaller domain.

I have also worked with smaller computational domains, however, this did not necessarily lead to different results compared to the larger domains (esp. because I kept the topology of the mesh the same).

Personally, I think the meshes I have created are fine and should produce GC. I have uploaded the mesh topology of the fine mesh as a screenshot in the attachment. The regular and coarse mesh have the same topology, yet the amount of cells allocated to the all the edges are decreased with a factor of 1.25. Btw; I created the mesh in ANSYS ICEM and converted it with the utility fluentMeshToFoam (which seemed to work).

A summary of my Mesh properties/CFD settings:

- Flow is resolved in 2D
- Re = 3.6*10^6
- Turb. model: k-w SST without wall function
- yPlus < 1
- I use standard BC for a cylinder (Uniform inlet velocity, zero ref. pressure at the outlet, fixedValue of U = (0 0 0) around cylinder wall.
- I use a fixed timestep. Al the disc. schemes used are 2nd order. The final residuals have been put on a tolerance of 1e-08. For the p-U coupling I use the PIMPLE family, with 20 outer and 3 inner Correctors and some PIMPLE tolerances.
- The case has been run in parallel on four processors.
- I know the Courant number in all three cases was well below Co < 0.7
- The case files have been uploaded in the attachment.

Things I have tried so far:
- I have tried a different mesh topology
- I Used less tight PIMPLE settings
- I have tried to run with a smaller domain size
- I have played a bit with the disc. schemes (tried some 1st order as well)
- I tried to use the adjustable timestep scheme as well (Co < 0.7)

The problem is that especially the lift coefficient keeps on increasing with big numbers, which makes the grid convergence very hard. I have tried to already run with N = 400.000 cells, but still the lift seemed to go to even higher numbers. It looks like the finer the mesh, the higher the lift force (without reaching a ceiling for grid convergence). I dont believe N>400.000 would be required to achieve grid convergence. Personally, I would think that the meshes are fine (esp. since y+ < 1 and the O-grid is very fine) but there might be a problem with the CFD settings.

I am out of ideas at this moment and I am wondering if some of you has faced a similar problem before with achieving grid convergence for a turbulent flow. Please let me know if you have any advice/tips/tricks on this matter. Thanks in advance! And feel free to ask for more information

[aakie]

Bump! I am very eager to hear your opinion!

[JNSN]

Hi,


have you tried kLowReWallFunction for k and nutLowReWallFunction for nut?


No idea, if this helps, but maybe worth to check.


Best,
Jan

[aakie]

Originally Posted by JNSN

Hi,


have you tried kLowReWallFunction for k and nutLowReWallFunction for nut?


No idea, if this helps, but maybe worth to check.


Best,
Jan

Hi Jan,

Thanks for your reply. Is maybe not a bad idea, since I am not using a typical wall function for BL prediction but I am resolving up to the viscous sublayer. This could indeed be seen as a 'Low-Re' approach. Thanks!