[CellZone]

Hello,

where does it come from, that I have an oscillatory behavior when I am doing a grid convergence study?

So when I pick for example the heat transfer coefficient, temperature and pressure in my cfd simulation, I can observe that it varies:

Number of Cells (Mio) htc (W/m²K)
8.5: 50.25
9 : 49.7
10: 49.5
12.5: 49.95
16.5 : 50.25


Number of Cells (Mio) Temperature in °C
8.5: 60.85
9: 60.62
10: 60.7
12.5: 60.9
16.5: 61.05


Number of Cells (Mio) Pressure in Pa
8.5: 9813
9: 9835
10: 9814
12.5: 9810
16.5: 9812

For my grid study I kept the nearest wall distance and the prism layers constant. I only changed the target mesh size (core mesh). So at the walls the cells have the same height, but change in their width.
In another study I looked for the effect of changing the prism layers (numbers and y+) , this did not show any big effect.


Why do the results vary? Is it, because when I change the mesh size, there might be the effect, that suddenly there are bigger cells which meet smaller cells what can lead to changes in the results?

Thank you for your hints and explenations!

[FMDenaro]

Are you using a fully unstructured grid everywhere?

[CellZone]

Yes I do. Automated Mesh with StarCCM+

[FMDenaro]

Well, it is quite important to define how do you perform the grid convergence study. Your aim is to reduce the magnitude of the local truncation error by refining the gird, therefore you should consider that each tethraedral mus be reduced on the more refined grid.
Finally, I don't know if you are also using some turbulence model, it can have effects on the solutions.

[CellZone]

Yes, I am using k-omega sst.

So you mean, instead of keeping the y+ at the walls constant, I should refine ALL cells, so not only the one in the core mesh?

So my idea was:
-- > keeping the y+ constant (under 1) in areas, where I expect high gradients and where it is important to have the correct htc and temperature profile

+
change the core mesh (away from the walls)



I could now also coarsen my mesh, so that I reach for example y+ = 5 to 30 for the nearest wall cell. But then I can not clearly say: different results for temperature etc. arise from a coarse base mesh or a coarse boundary layer?


Thank you for your help Mr. FMDenaro

[FMDenaro]

It is not a straightforward issue to define a grid refinement techique on unstructured grid. But in principle you should work by reducing the local truncation error everywhere (see for example an illustration in the book of Peric& Ferziger).
Be aware that some turbulence RANS models do to allow to decrease the y+ value at the wall.

[CellZone]

Okay.

But do you know where it comes from, that the results vary (temperature goes up and down for finer grid).

Where do this phenomenon come from? What happens here numerically?

Thanks!

[FMDenaro]

Originally Posted by CellZone

Okay.

But do you know where it comes from, that the results vary (temperature goes up and down for finer grid).

Where do this phenomenon come from? What happens here numerically?

Thanks!

Difficult to say one specifi reason...it could be due to the turbulence model that works indipendently from the grid size or you need to decrease the convergence threshold on the finer grid...
Do not forget that also for simple linear test, the convergence is monotone only for sufficiently fine grids.

[JBeilke]

How do you obtain your numbers? Are they averaged over a number of iterations or is it the report value from the last iteration?