Magic numerics to get the new 'Standard' TET mesh for CFD to converge quicker (sometimes 1/11th real time and 1/6th core hours), to more stable CFD results

That is a big bold topic title but here is the project that seems to back it up.

EDIT 10-30-2019 : And if those numerics still do not converge, then simply Relax P to 0.75 and Relax U to 0.25 …
The default relax for those is P0.3 and U0.7. REALLY VERY SIMPLE, try relaxing first… (even for HEX meshes…)

EDIT: Yes they converge great to stable results but are giving me accuracy headaches in a NACA0012 verification project that I am working on. Stay tuned for updated numerics soon if I can more closely verify the NACA0012

PLEASE try these out and report back here on your difficult TETforCFD mesh, CFD convergence problems :wink:

EDIT: I just thought the topic was necessary to quickly put ‘out there’ as I do not want people getting discouraged when trying TETforCFD meshes with DEFAULT numerics. (I almost gave up and I have a lot of perseverance :wink: )

2 Likes

Hi Dale,

Just some advise, on these numerics.

Firstly for stability, I agree with you numerics, in terms of divergence schemes. Additionally, if any poor cells are present in the mesh, limited schemes to 0.33 or 0.5 will help for laplacian schemes. Also, for tet meshes, I add 2 non-orthogonal correctors. and that is my robust set of numerics for tet meshes.

However :slight_smile: Although this produces fantastic stability, it comes at a cost to accuracy, because the divergence schemes are first-order accurate (bounded gauss upwind is 1st order accurate) and for velocity and possibly the turbulence divergence schemes it’s better for accuracy if you used Gauss linear upwind (limited gradU or limited gradient) and are 2nd order accurate and in some industries where velocity gradients are critical to accuracy (such as aerospace) this might be the difference between invalid and valid results.

That’s just my penny for them, hope it helps.

Best,
Darren

1 Like

LOL, I am just coming to similar conclusions, not by knowledge but by iteration and perseverance…

I am trying to verify NACA0012 with TETforCFD mesh and my magic numerics are close on CL but I am not there yet on CD, but I know I will get there, even without knowledge :wink:

I will be updating this magic numerics topic soon to talk about the reduced result accuracy and perhaps show even better numerics if I can verify NACA0012…

EDIT: I just thought the topic was necessary to quickly put ‘out there’ as I did not want people getting discouraged when trying TETforCFD meshes with DEFAULT numerics. (I almost gave up and I have a lot of perseverance, as you know :wink: )

Thanks…

1 Like

Awesome work Dale, yes agreed.

The divergence sceheme for U using Gauss Linear Limited gradU was my ah ha moment for 0012 :slight_smile:

Best,
Darren

1 Like

COOL, I will have more comments in that area soon in your great NACA 0012 Verification project :wink:

1 Like

And recently, Darren has guided me to Relaxation values that are even more MAGICAL…

Simply Relax P to 0.75 and Relax U to 0.25 …

The default relax for those is P 0.3 and U 0.7

REALLY VERY SIMPLE MAGIC, try that first… (even for HEX meshes…)

The jury is still out on the new ‘Standard mesh for CFD’ accuracy but my conclusion is tending toward ‘not very good’…

1 Like