Problem regarding discrete adjoints

[SumanVajjala]

Hi, I have a problem computing the surface sensitivities for drag (CD) using discrete adjoints in SU2. I have computed the direct solution with density residuals converged to E-12. I am using this to compute the surface sensitivities using SU2_CFD_AD (SU2 version 8.0). I am facing 3 issues.

1. The intermediate surface_adjoint files (VTU to be specific) show surface sensitivities as 1e-38 when plotted. Adjoints for density and other variables are non zero except for sensitivities.

2. When the run gets completed, I get non zero values of surface sensitivities. However, I get speckled like contours and also it is only displayed for a portion of the surface.

3. My adjoint solution's residual (density) starts decreasing and reaches a level of -8. Then it starts increasing and it looks like it will diverge.

Can anyone help me regarding these issues? I am using SU2 v 8.0 for these computations.

Suman

[SumanVajjala]

Also forgot to add that I can't open the surface_adjoint.szplt/ surface_adjoint.vtu in tecplot. I can open the vtu file in paraview though!

[joshkellyjak]

Converted to discussion as it appears to be a setup issue rather than an issue with the code. In future I would reccomend CFD online forum for these sorts of issues.

By the looks of things something is wrong in your config file. I would reccomend running one of the example tutorial cases and comparing your setup at each step. As for the convergence it appears that either the primal solver is not converged sufficiently or the adjoint solver cannot solve your system, try decreasing CFL first then changing some of the solver settings.

[SumanVajjala]

Sorry for the late reply. I have followed the configuration file as given in the standard test cases. The primal solver has converged very well reaching -12 level in residue. The sensitivity computations are not working. The residuals for the computations reach -8.3 and then increase.

Moreover when visualizing the sensitivity values only a part of the contours are shown.

Do you have any comments/suggestions as to why this is happening? I can share the config file if some one wants to have a look at it and point out any mistakes.

[joshkellyjak]

The solution looks like it hasn't converged, this can be remedied by playing with your solver settings, I would start with lowering the CFL. This looks like quite a large test case with a reasonable complex geometry, have you run on something smaller? You are not visualising the sensitivity in this image you attach, this is the adjoint density residual.

[SumanVajjala]

@joshkellyjak Thanks! Lowering the CFL produces the same convergence behavior as shown in the previous post. I have not shown the sensitivity here but I've started another discussion: #2553.

I ran the airfoil case given in the Testcases suite and then migrated to a wing-body configuration. I have also setup an ONERA M6 test case as well and will share my experience after I get the results.