Originally Posted by Tobi (Post 805611)

Honestly, since I am using FOAM, I never used tetrahedral meshes. However, there is a talk on YouTube from Hrv. Jasak. He is talking about tetrahedron meshes and numerics. So in principal I would say, we can work with tet meshes.

Originally Posted by olesen (Post 805617)

Even in the old days with STARCD and/or STARCCM+ always stayed far away from tet meshes. If you have a tet mesh, can try using its poly dual - should be better. However, the polyDualMesh in OpenFOAM will look a bit dodgy at concave geometry features. To be really useful it would also need some cell splitting there (easy to conceptualize in 2D, not so easy to implement in 3D).

Originally Posted by Tobi (Post 805624)

Obvious, I forgot the polyDualMesh application. Nevertheless, I never worked with tet-meshes and hence, no experience. Is it still tricky to work with tets and FOAM?

Originally Posted by Tobi (Post 805532)

the schemes do look okay. However, the gradient scheme could be modified by using the cell or edge limiter. Gradients should be calculated corrected in order to get correct results. Furthermore, nCorrectors could help stabilizing the solution within one SIMPLE iteration (not sure right now which solver you are using but for simpleFoam we don´t have that option).

Originally Posted by kishpishar (Post 805647)

Hi Tobi,



As much as I remember, I did a few tests with all the 3 least square schemes - the standard one, edgeCell and the pointCell versions and observed really no substantial improvement with any of these. I've only attempted steady cases with simpleFoam with relatively simpler geometries, but have some observations.

In my experience, it is not that tetrahedral meshes are useless with OF, rather it is often the case that convergence is quite bumpy and the residuals don't fall to the same levels as with Hex meshes. It may still be possible to obtain reasonable looking solutions.

Here are a few suggestions from my limited experience:

Mesh:

As far as possible, generate tet meshes with nearly equilateral triangles with smooth size variation. This would make the average non-orthogonality higher compared to Hex meshes, but the maximum non-ortho will be around 60. The solutions might still oscillate but the solver crashes etc. can be avoided.

Turbulent model:

kEpsilon seems to work better with this type of meshes than the other variants realizable, RNG etc.

fvSchemes:

• Gradient: use leastSquares with or without a cell Limiter. If using cell limiter, use it only for U, k and epsilon (not for p)
• divSchemes: bounded Gauss upwind (use always first order schemes)
• snGrad: limited corrected 0.333 (or in worst case, uncorrected)
• laplacian: Gauss linear limited corrected 0.333 (or in worst case, Gauss linear uncorrected)

fvSolution:

• relaxationFactors: use 0.5 for U, k, epsilon

If the above combination doesn't work, there could be mesh problems in the regions close to boundary layers. Better remesh without boundary layers and see if the problem persists.

Originally Posted by hbulus (Post 805649)

Thanks for information. Do you have any experiment with tetrahedral meshes with this settings? I mean can you verify what you are saying ?
If you can enlighten us, i'll be very grateful.

Originally Posted by kishpishar (Post 805652)

All that is written above comes from my experience on fully tet meshes.

Originally Posted by kishpishar (Post 805647)

Hi Tobi,

Turbulent model:

kEpsilon seems to work better with this type of meshes than the other variants realizable, RNG etc.

fvSchemes:

• Gradient: use leastSquares with or without a cell Limiter. If using cell limiter, use it only for U, k and epsilon (not for p)
• divSchemes: bounded Gauss upwind (use always first order schemes)
• snGrad: limited corrected 0.333 (or in worst case, uncorrected)
• laplacian: Gauss linear limited corrected 0.333 (or in worst case, Gauss linear uncorrected)

Originally Posted by Tobi (Post 805689)

Based on your output it seems that the more diffusion you add, the better the results get (as it is a common point in fluid dynamics).

Originally Posted by Tobi (Post 805689)

Originally Posted by Tobi (Post 805689)

The only thing that I am not happy about is that for bad things you set an uncorrection term to the laplacian and snGrad scheme? Hence, you don't take into account for non-orthogonal correction. I am not sure if you introduce (again) diffusion here? But commonly, the error introduced for non-orthogonality should be taken into account.

Originally Posted by Tobi (Post 805689)

[/LIST]Based on your output it seems that the more diffusion you add, the better the results get (as it is a common point in fluid dynamics).

• kEpsilon should have more diffusion compared to others such as RNG
• divergence scheme upwind will introduce a lot diffusion error which smooths out the gradients and support stability

The only thing that I am not happy about is that for bad things you set an uncorrection term to the laplacian and snGrad scheme? Hence, you don't take into account for non-orthogonal correction. I am not sure if you introduce (again) diffusion here? But commonly, the error introduced for non-orthogonality should be taken into account.