[bikalpa10]

Hello everyone,

I am using SU2 version 7.0.6. and trying to perform 3D optimization of delta wing Micro Air Vehicle at Re 10^5. with following settings.

Incompressible RANS model,
Turbulence Model: SA,
CFL: 20,
Euler implicit method for the flow equations.
FGMRES is used for solving the linear system.
Using a ILU(0) preconditioning.
Max number of linear iterations: 10.

Mesh quality metric
+--------------------------------------------------------------+
| Mesh Quality Metric| Minimum| Maximum|
+--------------------------------------------------------------+
| Orthogonality Angle (deg.)| 15.9908 | 89.9951|
| CV Face Area Aspect Ratio| 1.01323| 591.614|
| CV Sub-Volume Ratio| 1 | 732.846|
+--------------------------------------------------------------+

Direct simulation converged up to rms[P]= -10.4.

However, the discrete adjoint solver diverged from just 20 iterations.
I have also tried various other settings, all settings gave divergence.

• CFL= 5
• Max linear Solver Iteration = 50
• Frozen the turbulent viscosity in the discrete adjoint formulation (NO, YES)
  FROZEN_VISC_DISC= YES
• JACOBI precondition

I have attached the direct log file, adjoint log file, and adjoint history for reference

[pcg]

Hello,
You can try the option QUASI_NEWTON_NUM_SAMPLES= 20
It is a method to stabilize the discrete adjoint (not sure if it is in 7.0.6 or if you need to update).
Note that the relative residual reduction is only -2, so that might be an issue as well.

[bikalpa10]

Originally Posted by pcg

Hello,
You can try the option QUASI_NEWTON_NUM_SAMPLES= 20
It is a method to stabilize the discrete adjoint (not sure if it is in 7.0.6 or if you need to update).
Note that the relative residual reduction is only -2, so that might be an issue as well.

Hello Pedro,

Thank you for your advice,
I tried QUASI_NEWTON_NUM_SAMPLES=20, and it worked in 7.0.6
( adjoint residual did not converge beyond -3 but it did not diverge either).


And regarding the relative residual reduction, I ran a simulation with a much finer mesh as well, but could not reduce the relative rms below -2.4. Mesh quality of the fine mesh is as follows.
+----------------------------------------------------------------------+
| Mesh Quality Metric | Minimum | Maximum|
+---------------------------------------------------------------------+
| Orthogonality Angle (deg.) | 27.5705 | 89.986 |
| CV Face Area Aspect Ratio | 1.00468 | 218.596 |
| CV Sub-Volume Ratio | 1.00107 | 95.716 |
+--------------------------------------------------------------------+

Also, adaptive CFL for coarse mesh did not improve relative RMS.

Could you suggest numerical settings for better convergence or tips on mesh quality improvement to improve convergence?

Thank you

Regards
Bikalpa Bomjan Gurung

[pcg]

In my experience mesh quality metrics are not always directly related with convergence, yours look fine.

I don't have much experience with the incompressible solver, but you can try changing the limiter constant up/down to see what effect it has (sometimes if it is too low the solution flip-flops because it is too limited, but sometimes if it is too high the solution becomes unstable) similarly you can try the VAN_ALBADA_EDGE limiter.
Your mesh is completely unstructured so multi-grid will probably not help.

I find that the adjoint solver responds well to high CFL values, for my problems I usually use around 25 for the primal problem and 100-200 for the adjoint.

[bikalpa10]

Thank You, Pedro, for your suggestions.