Adjoin Output Options

Eduardo Carvalho · August 10, 2020, 18:38

Hi all,

What are the options of the adjoint solver's output?

I failed in executing a dryrun to verify the options of output available for the adjoint solver. I also didn't find this in the user's guide or in the template.

Thanks

pcg · August 10, 2020, 20:16

Hello,

There are no options, you get the coordinates, the solution (the adjoints of the primal solution variables), and the sensitivities (volume and surface).
In the case of the continuous adjoint only the surface sensitivities matter, for discrete adjoints only the volume sensitivities matter (the surface values are just for post processing and they do not reflect what would happen if the mesh was deformed).

There is not much else we could output for adjoint solvers.

Cheers.

Eduardo Carvalho · August 10, 2020, 20:42

Thank you for your answer, Pedro.

I'm trying to reproduce the plot from Figure 2 of the unsteady optimization's tutorial (https://su2code.github.io/tutorials/..._Opt_NACA0012/).

I'm trying to find the "drag sensitivity" output, but the original tutorial's adjoin history file doesn't have any variable with this name. I've thought that maybe was some variable not listed in the history output command (maybe the SENS_GEO for instance).

Do you know how to get this output?

pcg · August 11, 2020, 05:35

Ah ok, that's different (I thought you were talking about volume outputs).
History outputs are specified in groups, those sensitivities are in group SENSITIVITY (and they should be part of the defaults).

Eduardo Carvalho · August 11, 2020, 15:03

Ok, so just to clarify: which output variable is being plotted as the drag sensitivity in the Figure 2 of the Unstady Shape Optimization Tutorial???
(https://su2code.github.io/tutorials/..._Opt_NACA0012/)

pcg · August 11, 2020, 18:42

I'm afraid that is yet a different thing, the author is looking at drag sensitivity w.r.t. the design variable setup in the config file, this is not a direct output of SU2 as some post processing is required.
Derivatives w.r.t. shape parameters are output to a file named according to GRAD_OBJFUNC_FILENAME (default is of_grad.dat) when running SU2_DOT_AD.
When you run the tutorial do you see that file?
The author might have post processed all time steps manually to build that plot, in practice (i.e. for optimization) you do not need the history of the sensitivity, just the averaged value.

Eduardo Carvalho · August 12, 2020, 14:57

Yes, I've found the file. It cointains a value of "GRADIENT" for each "VARIABLE", as you mentioned.

So, to summarize:
1 - The SENS_GEO described in the history file is a vector containing a projection of the volumetric sensitivity into the surface for each time step.
2 - The gradient described in the of_grad_cd.csv file is calculated once per design and is a "windowed mean" of the volumetric sensitivities multiplied by the sensitivities w.r.t. your deformation method.
Is that correct?

Thanks again for your help.

pcg · August 13, 2020, 06:44

SENS_GEO is the sum of the sensitivities at the surface, it does not include the projection step required to compute the values that go into of_grad_cd.csv.
The purpose of SENS_GEO is just to monitor convergence.
The projection step is, like you say, the sensitivities w.r.t. the deformation method, or more formally the partial derivatives of the mesh coordinates w.r.t. the design parameters.

August 10, 2020, 18:38	Adjoin Output Options	#1
Eduardo Carvalho New Member Eduardo de Oliveira Carvalho Join Date: Mar 2018 Location: São José dos Campos - SP. Brazil Posts: 15 Rep Power: 8	Hi all, What are the options of the adjoint solver's output? I failed in executing a dryrun to verify the options of output available for the adjoint solver. I also didn't find this in the user's guide or in the template. Thanks

August 10, 2020, 20:16		#2
pcg Senior Member Pedro Gomes Join Date: Dec 2017 Posts: 466 Rep Power: 14	Hello, There are no options, you get the coordinates, the solution (the adjoints of the primal solution variables), and the sensitivities (volume and surface). In the case of the continuous adjoint only the surface sensitivities matter, for discrete adjoints only the volume sensitivities matter (the surface values are just for post processing and they do not reflect what would happen if the mesh was deformed). There is not much else we could output for adjoint solvers. Cheers. Eduardo Carvalho likes this.

August 11, 2020, 18:42		#6
pcg Senior Member Pedro Gomes Join Date: Dec 2017 Posts: 466 Rep Power: 14	I'm afraid that is yet a different thing, the author is looking at drag sensitivity w.r.t. the design variable setup in the config file, this is not a direct output of SU2 as some post processing is required. Derivatives w.r.t. shape parameters are output to a file named according to GRAD_OBJFUNC_FILENAME (default is of_grad.dat) when running SU2_DOT_AD. When you run the tutorial do you see that file? The author might have post processed all time steps manually to build that plot, in practice (i.e. for optimization) you do not need the history of the sensitivity, just the averaged value. Eduardo Carvalho likes this.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
OpenFOAM - Control and Editing of Lagrangian Output Variables	gmag	OpenFOAM Programming & Development	1	June 25, 2019 13:29
[ICEM] Appropriate bocos for POPINDA output	FlorisvdBeek	ANSYS Meshing & Geometry	1	November 19, 2017 05:22
Projected area to output parameters	Daniil85	Fluent UDF and Scheme Programming	1	March 16, 2016 16:44
Generating output data on AQWA Naut	afitriadhy	ANSYS	0	July 2, 2015 03:37
lift and drag on ship superstructures	vaina74	OpenFOAM Running, Solving & CFD	3	June 8, 2010 13:30

August 10, 2020, 20:42		#3
Eduardo Carvalho New Member Eduardo de Oliveira Carvalho Join Date: Mar 2018 Location: São José dos Campos - SP. Brazil Posts: 15 Rep Power: 8	Thank you for your answer, Pedro. I'm trying to reproduce the plot from Figure 2 of the unsteady optimization's tutorial (https://su2code.github.io/tutorials/..._Opt_NACA0012/). I'm trying to find the "drag sensitivity" output, but the original tutorial's adjoin history file doesn't have any variable with this name. I've thought that maybe was some variable not listed in the history output command (maybe the SENS_GEO for instance). Do you know how to get this output?

August 11, 2020, 05:35		#4
pcg Senior Member Pedro Gomes Join Date: Dec 2017 Posts: 466 Rep Power: 14	Ah ok, that's different (I thought you were talking about volume outputs). History outputs are specified in groups, those sensitivities are in group SENSITIVITY (and they should be part of the defaults).

August 11, 2020, 15:03		#5
Eduardo Carvalho New Member Eduardo de Oliveira Carvalho Join Date: Mar 2018 Location: São José dos Campos - SP. Brazil Posts: 15 Rep Power: 8	Ok, so just to clarify: which output variable is being plotted as the drag sensitivity in the Figure 2 of the Unstady Shape Optimization Tutorial??? (https://su2code.github.io/tutorials/..._Opt_NACA0012/)

August 12, 2020, 14:57		#7
Eduardo Carvalho New Member Eduardo de Oliveira Carvalho Join Date: Mar 2018 Location: São José dos Campos - SP. Brazil Posts: 15 Rep Power: 8	Yes, I've found the file. It cointains a value of "GRADIENT" for each "VARIABLE", as you mentioned. So, to summarize: 1 - The SENS_GEO described in the history file is a vector containing a projection of the volumetric sensitivity into the surface for each time step. 2 - The gradient described in the of_grad_cd.csv file is calculated once per design and is a "windowed mean" of the volumetric sensitivities multiplied by the sensitivities w.r.t. your deformation method. Is that correct? Thanks again for your help.

August 13, 2020, 06:44		#8
pcg Senior Member Pedro Gomes Join Date: Dec 2017 Posts: 466 Rep Power: 14	SENS_GEO is the sum of the sensitivities at the surface, it does not include the projection step required to compute the values that go into of_grad_cd.csv. The purpose of SENS_GEO is just to monitor convergence. The projection step is, like you say, the sensitivities w.r.t. the deformation method, or more formally the partial derivatives of the mesh coordinates w.r.t. the design parameters.