Discrete adjoint returns almost zero sensitivities with inverse design objective

Zen · March 12, 2018, 15:35

Hello,

I downloaded the test case 'steady_inverse_adjoint' from the GitHub page of SU2 and I tried to run a shape optimization with a discrete adjoint method.

The optimization terminates after one single function and adjoint evaluation since the sensitivities are O(1e-11).

On the other hand, when running the optimization with a continuous adjoint technique, everything works fine and a minimum is reached after 28 function and 11 gradient evaluations.

Any idea why this happens?

Thank for the help,

Z

talbring · March 17, 2018, 07:28

Hi Zen,

the answer is simple: this objective function cannot be used for the discrete adjoint at the moment. It is not a big deal to implement it ...actually is very simple. Do you want to give it a try ? Let me know, I can point you to the correct places.

Tim

Zen · March 17, 2018, 10:10

Hi Tim,

I solved the problem by inserting the following lines after DirectRun(); in CDiscAdjFluidDriver::SetRecording(unsigned short kind_recording):

Code:

 if (config_container[ZONE_0]->GetInvDesign_Cp() == YES){
output->SetCp_InverseDesign(solver_container[ZONE_0][MESH_0][FLOW_SOL],
geometry_container[ZONE_0][MESH_0], config_container[ZONE_0], ExtIter);
}

This actually solves a small bug concerning the evaluation of Total_CpDiff and makes the optimization work correctly.

Z

P.S. Probably, this should also be done for INVERSE_DESIGN_HEATFLUX.

fpalacios · March 18, 2018, 21:43

Hi Zen,

Glad you are working on this. As you know, Cp inverse design is a very useful capability in aero design. It seems that you are also familiar with the software as developer, if you have a while the SU2 community would appreciate if you could add the changes in GitHub.

By the way, we are creating the SU2 International Developers Society, it maybe be interesting for you (www.su2devsociety.org).

Best,
Francisco

March 12, 2018, 15:35	Discrete adjoint returns almost zero sensitivities with inverse design objective	#1
Zen Member Zeno Join Date: Sep 2013 Location: Delft, The Netherlands Posts: 63 Rep Power: 13	Hello, I downloaded the test case 'steady_inverse_adjoint' from the GitHub page of SU2 and I tried to run a shape optimization with a discrete adjoint method. The optimization terminates after one single function and adjoint evaluation since the sensitivities are O(1e-11). On the other hand, when running the optimization with a continuous adjoint technique, everything works fine and a minimum is reached after 28 function and 11 gradient evaluations. Any idea why this happens? Thank for the help, Z

March 17, 2018, 07:28		#2
talbring Super Moderator Tim Albring Join Date: Sep 2015 Posts: 195 Rep Power: 11	Hi Zen, the answer is simple: this objective function cannot be used for the discrete adjoint at the moment. It is not a big deal to implement it ...actually is very simple. Do you want to give it a try ? Let me know, I can point you to the correct places. Tim __________________ Developer Director @ SU2 Foundation Get involved: Facebook https://www.facebook.com/su2code Github https://github.com/su2code Twitter https://twitter.com/su2code

March 17, 2018, 10:10		#3
Zen Member Zeno Join Date: Sep 2013 Location: Delft, The Netherlands Posts: 63 Rep Power: 13	Hi Tim, I solved the problem by inserting the following lines after DirectRun(); in CDiscAdjFluidDriver::SetRecording(unsigned short kind_recording): Code: if (config_container[ZONE_0]->GetInvDesign_Cp() == YES){ output->SetCp_InverseDesign(solver_container[ZONE_0][MESH_0][FLOW_SOL], geometry_container[ZONE_0][MESH_0], config_container[ZONE_0], ExtIter); } This actually solves a small bug concerning the evaluation of Total_CpDiff and makes the optimization work correctly. Z P.S. Probably, this should also be done for INVERSE_DESIGN_HEATFLUX.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Discrete vs Continuous Adjoint solution variable: Scaling/normalization issue?	asitav	SU2	2	November 21, 2018 20:11
Shape optimization doesn't converge - Adjoint log file	JPBLourenco	SU2 Shape Design	0	December 11, 2017 09:41
Difficulties running 3D inverse design optimisations	dominico	SU2 Shape Design	1	September 19, 2017 09:56
Is SU2 able to handle adjoint for pressure distribution when doing inverse design?	Xiaosong	SU2	2	December 4, 2016 01:25
Info: Short Course On Thermal Design of Electronic Equipment	Arnold Free	Main CFD Forum	0	August 10, 1999 11:18

March 18, 2018, 21:43		#4
fpalacios Super Moderator Francisco Palacios Join Date: Jan 2013 Location: Long Beach, CA Posts: 404 Rep Power: 15	Hi Zen, Glad you are working on this. As you know, Cp inverse design is a very useful capability in aero design. It seems that you are also familiar with the software as developer, if you have a while the SU2 community would appreciate if you could add the changes in GitHub. By the way, we are creating the SU2 International Developers Society, it maybe be interesting for you (www.su2devsociety.org). Best, Francisco