[Xikii]

Nice to me you!

I'm totally beginer using SU2 for shape optimization, when I run the testcase 'disc_adj_turbomachinery/transonic_stator_2D returns: Exception: unknown function name, FLOW_ANGLE_OUT. Please check config_template.cfg for updated list of function names.

I look it up in github issue and discussion, and changed the OPT_OBJECTIVE to ENTROPY_GENERATION, then it returns: KeyError: 'ENTROPY_GENERATION'. The solution in github seems to be using FADO, but the tutorial is very hard for me, I wonder is there any more specific tutorial or guidence about FADO or is there any way that can run the testcases without using FADO.

I'm using SU2 8.0.1 and python 3.9.13 running in Windows Powershell.

It would be very greatful if you can offer some help! Thank you!

[bigfootedrockmidget]

What is the exact command that you are using? I could use SU2_CFD as well as SU2_CFD_AD with the testcase and they both run fine. So I guess you are using maybe the shape_optimization.py scripts?


Remember that all these scripts are just automating the optimization for you, they are still doing basically
SU2_DEF -> deforming the mesh, for instance with an FFD box

SU2_CFD -> computing CFD solution
SU2_CFD_AD -> computing adjoint solution
SU2_DOT_AD -> computing dot product
Now you have a bunch of sensitivities at e.g. the nodes of your FFD box. You can feed them to SU2_DEF to deform your mesh again.

[Xikii]

Thanks very much for your reply！

Affirmative！I‘m using shape_optimization.py script, and I found that using FADO may solve this but its too hard for me, so I still have no progress by now.

In your means that I can manually run the SU2_DEF>SU2_CFD>SU2_CFD_AD>SU2_DOT_AD repeatly and evey time means finish a iteration right? And I should run SU2_CFD>SU2_CFD_AD>SU2_DOT_AD before first SU2_DEF right?

I had switch my work enviroment into linux which may help my work.

Apologize for my ignorance
THANKS AGAIN for your help!

[bigfootedrockmidget]

yes, in principle you can do a 'manual' optimization. But if you can get FADO to work that would be better. The FADO tutorial on the website is a good starting point.

[Xikii]

Thanks for your reply!

I think I'd like to do it manually first to check the result and try the FADO later for futher learning.

[Tongtong1919]

Thank you Xikii and bigfootedrockmidget,

I have carefully read your conversation, and the content discussed is exactly what I am interested in. Following the process SU2_DEF > SU2_CFD > SU2_CFD_AD > SU2_DOT_AD, I finally obtained some results.



I offset (deformed) five control points (the deformation is quite exaggerated just to verify each step of the process), and the control points and grid are shown in the figure below.



Through the SU2_DOT_AD process, I finally obtained the SURFACE PRESSURE DROP gradient at the five control points. Literally, it seems to be a gradient (like df/dx, can anyone confirm if this is indeed the case?). Judging by the numerical values, the SURFACE PRESSURE DROP gradient seems more like the pressure drop itself (i.e., df), which confuses me a bit.


Assuming I have obtained the sensitivity of the objective function to the design variables, according to the common optimization framework, the next step should be to transmit these sensitivities to the optimization algorithm to obtain the next set of shape offsets (i.e., mesh updates), which is what bigfootedrockmidget mentioned as “Now you have a bunch of sensitivities at e.g. the nodes of your FFD box. You can feed them to SU2_DEF to deform your mesh again.” I don’t quite understand “feed them to SU2_DEF.” How is this step specifically implemented? Do you have any good suggestions?

[bigfootedrockmidget]

Hi Tongtong, SU2_DEF uses the keyword DV_VALUE to deform the mesh. So you copy the projected sensitivities coming from SU2_DOT_AD and give them as values to SU2_DEF.

[Tongtong1919]

Hello, bigfootedrockmidget,
Thank you for your response. I have been delayed in discussing this question due to some issues over the past few days. As you mentioned, running SU2_DOT_AD will yield the projected sensitivities result file. I believe that this result cannot be used directly as the input for SU2_DEF DV_VALUE. Instead, it should first go through the optimization algorithm, which will determine the geometric design variables. These variables will then be used as the input for DV_VALUE. Does this explanation sound more reasonable?