[rgbatista]

Dear someone who might be able to help,


For my master's thesis, I am working on the optimization of a benchmark case from the ADODG group. The objective is to minimize drag (Cd) in a transonic flow over the NACA0012 airfoil. The only restriction is y_optimized_airfoil > y_baseline at each point on the airfoil's surface. However, I have encountered challenges as I am unable to optimize this case, and Cd does not decrease.


I have attempted various approaches, including FFD Set Box with different numbers of design variables (DVs), adjusting upper and lower bounds, imposing a scale of 0.001 in the restrictions and optimal_objective, and applying the Hick-Hennes geometric parameterization method. Unfortunately, in most cases, Cd fails to decrease and the imposed restriction is not met.


Interestingly, when working on another case from the ADODG group—drag minimization on a viscous transonic flow over RAE 2822—I achieved improvement, reducing drag counts from 190 to 126 using the FFD-setting box option and meeting all associated restrictions (Cl=0.824, Cm >=-0.092, and A_optimized >= Area_baseline).


I am perplexed about why the NACA 0012 case is not showing improvement. Is there something crucial that I might be overlooking? I will provide a link with the documents for both cases, including .cfg files, .su2 meshes, and PDFs with the benchmark case restrictions.


The link: https://unlpt-my.sharepoint.com/:f:/...UtGrT4DchRJjHQ



https://unlpt-my.sharepoint.com/:f:/...UtGrT4DchRJjHQ


If someone could offer insights or assistance, I would be extremely grateful!


Best Regards,
Rafael Batista

[bigfootedrockmidget]

Here are some things to check:
1. How well does the primal converge?
2. How well does the adjoint converge?
3. What does the sensitvity vector field on the surface look like?
4. How large are deformations on the mesh?

5. If you do not use constraints, do you get a drag reduction?

[rgbatista]

Originally Posted by bigfootedrockmidget

Here are some things to check:
1. How well does the primal converge?
2. How well does the adjoint converge?
3. What does the sensitvity vector field on the surface look like?
4. How large are deformations on the mesh?

5. If you do not use constraints, do you get a drag reduction?

Hello bigfootedrockmidget,

Thank you for the quick answer.

1. I think it doesn't converge perfectly, but it converges normally.
2. Here, are you referring to the "history_adjoint" .csv file, am I correct? There might be an issue here because my adjoint just ends when it reaches the maximum number of iterations set for the direct/primal simulations, "ITER=NUMBER." How do I set the convergence criteria for the Adjoint simulations? Is it by setting the minimum residuals value, "CONV_RESIDUAL_MINVAL"? If yes, I've set the residuals to a value of 10^-4 just to test, and the adjoint residuals surpass the -4 value, continuing to run until it reaches the "ITER=". I will attach my history adjoint in the link I've given before.
3. When you mention the sensitivity vector, do you mean it's supposed to be like the ones in the images - in the link I've sent before?
4. They are very small.
5. I haven't tried, but I used OPT_OBJECTIVE=EFFICIENCY without any restrictions, and it almost didn't decrease the Cd, even that the main objective function was efficiency. It has decreased from 459 to 432 drag counts. Anyway, I will try it, and if I have updates, I will post them here.

Thank you for your assistance and attention.

Best Regards,
Rafael Batista

[bigfootedrockmidget]

Hi, Can you post pictures of the residuals in the history file (just plot them in paraview,matlab or python)?
For the adjoint to converge, it is important that the primal has also converged. If the adjoint reached the max. nr of iterations there might be a problem, but this could aready be with the convergence of the primal.

[rgbatista]

Originally Posted by bigfootedrockmidget

Hi, Can you post pictures of the residuals in the history file (just plot them in paraview,matlab or python)?
For the adjoint to converge, it is important that the primal has also converged. If the adjoint reached the max. nr of iterations there might be a problem, but this could aready be with the convergence of the primal.

Good afternoon, bigfootedrockmidget,

I will post pictures of plot of residuals, in the link below and also in zip file, from the 1 design iteration and last design iteration from both primal and adjoint simulations.


By the way, I tried again with FFD_THICKNESS_2D and it's improving the Cd mainting the y>ybaseline. I will also attatch this residuals from design iteration 34 (random iteration) and the last design iteration from both primal and adjoint simulations. Maybe this case is better.


Link: https://unlpt-my.sharepoint.com/:f:/...UtGrT4DchRJjHQ

Thank you so much for all your help and quick replies.

Best Regards,
Rafael Batista

[bigfootedrockmidget]

Hi,
OK, so you have some optimized solutions now?

I had a look at the residuals, to look mostly OK, but some simulations stop before reaching very low residual values, it is best to try and converge everything to a residual value of -14 for the primal and the adjoint. At least it looks like your cases are converging well. You might even be able to increase the CFL to make it converge faster.

[rgbatista]

Good morning, bigfootedrockmidget,


I would like to thank you once more for all the help, advice, and enlightenment about certain aspects in this case.


Yes, it's optimizing and fulfilling the restriction of y_optimized > y_baseline.
Okay, not bad news. So why do the adjoint simulations go until the limit of iterations "ITER="? -14 isn't a very low residual, maybe too low? Isn't there a possibility of all residuals stabilizing at another order of magnitude before reaching all residuals -14?


Regarding the CFL number, I've set CFL=15 because a PhD student, from his experience, told me I should be careful with the CFL number in the design/optimizations overall simulations. What CFL would you suggest? Maybe going to 25, 30? Or much higher?


Best Regards,
Rafael Batista