[MaxGains]

Hello,


I am currently trying to compare gradients obtained from the continuous and
discrete adjoint solvers.
First, I am trying to replicate the comparison of surface sensitivities from the
continuous and discrete adjoint solvers, using drag coefficient as the objective
function, as demonstrated at the end of the QuickStart tutorial:
https://su2code.github.io/docs_v7/Quick-Start/
However, I obtain quite different results to those illustrated in the figure at
the end of the tutorial.
Notably, when using the discrete adjoint solver, the surface sensitivities are
all zero.
This result seems incorrect; although we are using an inviscid solver, the flow
is supersonic in some regions, which should lead to nonzero drag.
Moreover, the sensitivity profiles I obtain from the continuous adjoint solver
look quite different to those in the tutorial figure.
The angle-of-attack and Mach number sensitivities are also different, though
this could be due to the different adjoint methods.

As a next step, I am trying to compare results on the Turbulent ONERA M6 case:
https://su2code.github.io/tutorials/Turbulent_ONERAM6/
Here, I obtain nonzero sensitivities with the discrete adjoint solver, though
these differ considerably from the results given by the continuous adjoint.

I have replicated these results on three machines now, which hopefully rules out
any local install issues.
I am particularly concerned about the QuickStart case, as I'm unsure where I
could have made an error.

Thanks in advance for any help.

Viv

[MaxGains]

With SU2_DOT and SU2_DOT_AD I can get the correct surface sensitivities in the surface_sens.csv file.
Is this missing from the tutorial?

[TKatt]

Hi MaxGains,

Can you post some pictures please of the two results please. I.e. the results after just the adjoint step SU2_CFD(_AD) and then SU2_DOT(_AD).

I just ran the thing with v7.3.1 and the most recent v8.0.1 and both times I was able to recreate the thing as shown below.

The SU2_DOT step is not necessary for re-creation as it projects the sensitivies onto the surface mesh first and then on the DV's (here Hicks henne bump functions). In this regard the one with the discrete adjoint (i.e. using SU2_DOT_AD) worked fine for me but the continuos adjoint seems to already use the surface_adjoint from the Continuos adjoint run which I think is not correct. But again, the DOT step is not part of the tutorial that is for sure. In doubt by the way, the discrete adjoint stuff is used much more regularly compared to the continuous adjoint, so I would naturally trust the correct working condition of the discrete adjoint.

Please make sure the set the MATH_PROBLEM correctly as stated in the tutorial and you take a converged primal solution file.

I added SURFACE_CSV to OUTPUT_FILES to look at the plot in paraview

I do not know if that helps, but maybe, let me know.

best, Tobi

[TKatt]

Here an attempt to add a screenshot i made (dont know, couldnt attach it before, forum said the post was too long *shrug*)

black - continuous ad
blue - discrete ad


fits pretty well with the tutorial I'd say
So I am positive we can get it to run for you as well