[Yeru]

Hi all,

I'm stuck with a should-be-easy simpleFoam case.
Long story short:

• mesh generated by SnappyHexMesh based on .STL file
  Y-junction followed by meandering pools, straight outlet sufficiently far away
• mesh_cropped.png
• BC inlet:
  U: surfaceNormalFixedValue p: zeroGradient
• BC outlet:
  U: zeroGradient p: fixedMeanValue 0;

simpleFoam should easily converge to a stationary solution.
The problem:
The first two time steps look fine.
But then, large fluctuations on the bordering walls appear, that don't dampen down anymore.
So: no convergence.

meshFault_01.png

meshFault_02.png

meshFault_03.png

You can find a GIF of the evolution here

What did I check so far?
checkMesh -> OK.
other BC -> still blows up
faulty assignment of patches in the STL file -> nope.

Honestly - what didn't I check?
The fvSolutions entries - since they're copied from an earlier case, that worked just fine.
The levels of refinement of SnappyHexMesh

What is my question?
What can I change to prevent this blow up and get nice convergence?

The STL file is 3.6MB large, so no nice upload here.
You can finde a zipped Minimal Working Example here

[alexeym]

Hi,

Just basics (I did not check any further after spotting these "features"):

- Your mesh is non-orthogonal (max 55).
- You have 0 non-orthogonal correctors (I would suggest 1 or 2).
- You use "Gauss linear" for gradient, "leastSquares" would be more suitable.
- You use "corrected" for laplacian and snGrad (try limited correction)
- Start with upwind, converge, and then proceed with second order schemes.

[Yeru]

Alexey, I owe you.
long story short: with some tweaks according to your advice, the error doesn't occur. Instead of 16h, my sim ran 53 min.

Originally Posted by alexeym

- Your mesh is non-orthogonal (max 55).

I adapted

maxNonOrtho 25;

in the snappyHexMeshDict. Earlier, the value was 75 and checkMesh yielded a max of non-orthogonality of 55.
Now, with a value of 25, checkMesh yields -can you guess?- 45. Don't know why.

Originally Posted by alexeym

- You have 0 non-orthogonal correctors (I would suggest 1 or 2).

Did that. Works magic.

Originally Posted by alexeym

- You use "Gauss linear" for gradient, "leastSquares" would be more suitable.

Strange, but: The sim breaks down when using "leastSquares" since the residiual explodes with values >e+22. "Gauss cubic" runs fine, but very long.

Anyway. Thanks for your advice!

[alexeym]

Originally Posted by Yeru

I adapted

maxNonOrtho 25;

in the snappyHexMeshDict. Earlier, the value was 75 and checkMesh yielded a max of non-orthogonality of 55.
Now, with a value of 25, checkMesh yields -can you guess?- 45. Don't know why.

If you check snappyHexMesh output, you find, that it complains about unsuccessful snapping. maxNonOrtho just defines quality check criterion, quality of generated mesh is defined by other parameters.

Strange, but: The sim breaks down when using "leastSquares" since the residiual explodes with values >e+22. "Gauss cubic" runs fine, but very long.

It is a question of relaxation. With 0.9 and leastSquares there are linear solver convergence problems, with 0.7 or 0.8 pressure equation linear solvers converge quite OK.

[Yeru]

Dear Alexey,

I've been running a few sims with this little tweak and that little tweak - all bearing in mind your recommendations.
Problem is: none of the "leastSquare" ones converge.

So I want to ask some things, just to make sure, I don't misunderstood:

1. You mentioned "pressure equation linear solvers". So do I have to switch the solver from simpleFoam to (e.g.) pimpleFoam?
2. You mentioned the relaxation factors. All the relaxation factors in the simpleFoam tutorials feature
   

   relaxationFactors
   {
       equations
       {
           U               0.9; // 0.9 is more stable but 0.95 more convergent
           ".*"            0.9; // 0.9 is more stable but 0.95 more convergent
       }
   }

   Should I put 0.7-0.8 as relaxation factor on U or should I simply define one for p?

Thanks.

[piu58]

I recommend to build up the case step by step:

- start with a single meander
- if possible (and its should be) make it with blockmesh, and make a mesh which is well balanced
- experiment a bit with different boundary conditions and mesh densities.
- then - with this experience - add a second meandering or the junction
- with that in heart you may built the case all in all

[alexeym]

Dear Dominik,

1. No, I did not mean that. In both cases you have "pressure equation linear solvers" (i.e. in simpleFoam you have SIMPLE, in pimpleFoam you have PISO with outer iterations).

2. Yes, relaxation factors influence your convergence.

Now real question is: do you need leastSquares? The reason for bad convergence could be:
- Oscillations due to second order. So basically, we have second order but no convergence.
- Since there are no plots of residuals in your post, I assume 'none of the "leastSquare" ones converge.' are in fact "I did not put enough iterations till convergence".
- For "consistent yes" relaxation factors of 0.9 are OK, yet in general it is 0.7/0.3 for U/p (0.8/0.2 in original papers, sum should be 1).