[MaxGains]

Hello,

I am looking for some assistance tuning SU2 settings to accelerate my
simulations.
I am currently performing RANS simulations on the ONERA M6 wing for a variety of
boundary conditions (i.e., variations in Mach, alpha, and geometry changes using
an FFD box).

My simulations converge, however, I think they could be accelerated with
appropriate tuning of the CFL adapatation parameters and multigrid settings.
I'm considering transsonic Mach numbers and am prioritizing simulation
robustness, so I've chosen the JST scheme.
I'm using implicit Euler with FGMRES and am currently obtaining best performance
with CFL fixed to 100 (seems high, but it works).
I've also found that three levels W_CYCLE multigrid gives good performance, though
I'm unsure if this is the optimal setting.
I imagined that I would be able to get better performance with CFL adaptation,
but so far I've not had good results.
Can anyone help me with tuning some of these parameters?

Given that I'm performing RANS simulations, it seems as though CFL=100 might be
a reasonable max value (given linear solver considerations); is this correct?

I've attached a sample .cfg file (renamed as txt for upload).
I cannot share my mesh files here due to their size (fine mesh is approx 5mil
cells).
I generated the (tetrahedral) meshes using Nishikawa's code from the following
page:
https://turbmodels.larc.nasa.gov/one...ics_grids.html

Regards,
Viv

[bigfootedrockmidget]

Hi,
So what is the current performance? How many iterations does it take you to get a converged solution? How does it compare to the existing Onera cases that we have in the Testcases repository?
You are saying that your cases converge with CFL=100. So you could simply start with that CFL and see if the adaptive CFL gives you a higher value:

CFL_NUMBER= 100.0
% (factor-down, factor-up, CFLmin, CFLmax, linear solver convergence)

CFL_ADAPT_PARAM= ( 0.99, 1.01, 10.0, 200, 0.001 )

[MaxGains]

Sorry - I should have given more detailed information.
On my coarser meshes, I'm obtaining very similar convergence behavior as seen in the Onera M6 examples in the Tutorials repository.
Specifically, I'm using the convergence behaviour of the example in
/Tutorials/compressible_flow/Turbulent_ONERAM6
with the mesh file switched to (the more refined mesh)
mesh_ONERAM6_100k.su2
as a benchmark.

However, on my fine mesh, I'm having convergence problems.
See the following image for convergence behaviour with bigfootedrockmidget's suggested CFL parameters, but with the lower limit set to 1:
https://imgur.com/a/0wo7Ugv

The solver reports a min CFL of 1 after ~150 iterations, which is making me worry about mesh quality, given I'm using implicit Euler.
However, for my fine mesh, I get the following mesh quality readouts


+--------------------------------------------------------------+
| Mesh Quality Metric | Minimum | Maximum|
+--------------------------------------------------------------+
| Orthogonality Angle (deg.) | 1.92128 | 85.5444|
| CV Face Area Aspect Ratio | 1.89781 | 45457|
| CV Sub-Volume Ratio | 1.00013 | 3310.34|
+--------------------------------------------------------------+


whereas for a coarse mesh, with good convergence behaviour, I get similar numbers,

+--------------------------------------------------------------+
| Mesh Quality Metric | Minimum | Maximum|
+--------------------------------------------------------------+
| Orthogonality Angle (deg.) | 2.71967 | 85.7994|
| CV Face Area Aspect Ratio | 1.84488 | 36105.1|
| CV Sub-Volume Ratio | 1.00026 | 2916.91|
+--------------------------------------------------------------+


Thanks

[MaxGains]

I've addressed some issues with mesh quality and am getting better, but still not perfect, convergence behaviour.
I'm now using the following settings for CFL adaptation, which are resulting in the min CFL being driven to 1.0 at the end of the simulation:
CFL_ADAPT_PARAM= ( 0.95, 1.01, 1.0, 200, 0.001 )



I'm now observing convergence behaviour as shown in the following two images (residuals and CL, CD):
https://imgur.com/a/pvFTHLq
https://imgur.com/a/aGmapb7


I still think there is something wrong with my simulation though, as the CFL should become large at the end of the simulation?


Viv

[bigfootedrockmidget]

Well, the CFL should increase when the linear solver converges sufficiently.

You could try either to increase the number of linear solver iterations or reduce the linear solver criterion in the adaptive CFL:

LINEAR_SOLVER_ERROR= 1E-6
LINEAR_SOLVER_ITER= 5