[Ryan.]

Hi everyone,

I am using the discrete adjoint solver and I am interested in surface sensitivities in the normal direction. However, I found that the "Surface_Sensitivity" inside "surface_adjoint.vtk" is not the same as the "Sensitivity" in "surface_sens.vtk". The Surface_Sensitivity is actually by 5 orders of magnitude larger than the Sensitivity. I thought they're both computed by taking the inner product of the surface mesh sensitivities and the cell normal vector, but apparently I was wrong. I would be glad if someone could explain how they are calculated.

Many thanks in advance

[TKatt]

Hi Ryan,

Upfront: I use a recent develop version of SU2 in which 'Surface_Sensitivity' is used for the variable in surface_sens.* (* is vtk for your paraview output) and surface_adjoint.*, I try to be clear with what I mean.

I'll try to explain briefly what I found in the code and what is happening here in general:

1. The surface_adjoint.* is written by SU2_CFD_AD (just as adjoint.*) and surface_sens.* by SU2_DOT_AD.

2. The sensitivity information in both files "mean different things". In adjoint.* and surface_adjoint.* you'll find the "Volume Sensitivities", i.e. the sensitivity with respect to every(!) mesh node coordinate alone. In surface_sens.* the information of the mesh deformation algorithm is incorporated and they are only available on the surface nodes (which control the geometry/shape).

3. The computation of Surface_Sensitivity is, as you said, the dot product between normal and sensitivity vector. But I am not sure whether that computation is correct for surface_adjoint.* and therefore also adjoint.*. I have to take a closer look to be sure though, and one has to be careful whether one has a unit normal or Area normal. For now I think a factor of -1/Area is the difference to the computation in surface_sens (based on a given sensitivity vector).

Summing up: The sensitivities in both files are different in what they represent and shouldn't be the same (in general). However, the computation of Surface_Sensitivity looks inconsistent to me for both files which might be an error.

I hope what I wrote was somewhat understandable. Let me know if an explanation was bad. I'll take a closer look at the computation in the next days and report here. If you want to look in the code yourself and try out things I can give you the lines of code where to computation takes place.

Cheers, Tobi

[Ryan.]

Hi Tobi,

thanks a bunch for the thorough explanation. I believe you're right, just by comparing the result in Paraview it seems the difference between them comes from the fact that one of them uses the area normal vector and the other uses the unit normal vector.

Cheers and happy new year,
Ryan

[angupraveen]

Dear all,

I'm trying to do a 3D SE2A Aircraft wing optimization using SU2. Unfortunately, I am unable to get the following things working.

1. The solution doesn't converge even after 5000 iterations both in ADJOINT and DIRECT solvers. I mean it keeps computing until the max iter value.

2. After the iterations are done, I tried to compute the sensitivity using SU2_DOT_AD and it gives of_grad.dat file with '0' as the value with the following message before

"Design variable (FFD_SETTING) number 0."

Any help would be highly appreciated

Thanks in advance

Angu

[Haoge Li]

Hi TKatt,
Thanks for your thorough explanation, as I computed the naca0012 airfoil with the continuous adjoint solver, and i'm confused about the Surface_sensitivity in the surface_adjoint.dat. Does the surface area normal vector point to the flow field, or the interior of the body?

Any help and explanation will be appreciated!

Thank you a lot

Haoge Li

Originally Posted by TKatt

Hi Ryan,

Upfront: I use a recent develop version of SU2 in which 'Surface_Sensitivity' is used for the variable in surface_sens.* (* is vtk for your paraview output) and surface_adjoint.*, I try to be clear with what I mean.

I'll try to explain briefly what I found in the code and what is happening here in general:

1. The surface_adjoint.* is written by SU2_CFD_AD (just as adjoint.*) and surface_sens.* by SU2_DOT_AD.

2. The sensitivity information in both files "mean different things". In adjoint.* and surface_adjoint.* you'll find the "Volume Sensitivities", i.e. the sensitivity with respect to every(!) mesh node coordinate alone. In surface_sens.* the information of the mesh deformation algorithm is incorporated and they are only available on the surface nodes (which control the geometry/shape).

3. The computation of Surface_Sensitivity is, as you said, the dot product between normal and sensitivity vector. But I am not sure whether that computation is correct for surface_adjoint.* and therefore also adjoint.*. I have to take a closer look to be sure though, and one has to be careful whether one has a unit normal or Area normal. For now I think a factor of -1/Area is the difference to the computation in surface_sens (based on a given sensitivity vector).

Summing up: The sensitivities in both files are different in what they represent and shouldn't be the same (in general). However, the computation of Surface_Sensitivity looks inconsistent to me for both files which might be an error.

I hope what I wrote was somewhat understandable. Let me know if an explanation was bad. I'll take a closer look at the computation in the next days and report here. If you want to look in the code yourself and try out things I can give you the lines of code where to computation takes place.

Cheers, Tobi

[lalatuobu]

Hi Ryan,

I want to compute the surface sensitivity for my configuration as well. But it seems to me that SU2 surface sensitivity requires some deformation by FFD or by user defined deformation function. What I want is small perturbance on surface and I want to calculate the current surface sensitivity. I wonder how to do that?

Originally Posted by Ryan.

Hi everyone,

I am using the discrete adjoint solver and I am interested in surface sensitivities in the normal direction. However, I found that the "Surface_Sensitivity" inside "surface_adjoint.vtk" is not the same as the "Sensitivity" in "surface_sens.vtk". The Surface_Sensitivity is actually by 5 orders of magnitude larger than the Sensitivity. I thought they're both computed by taking the inner product of the surface mesh sensitivities and the cell normal vector, but apparently I was wrong. I would be glad if someone could explain how they are calculated.

Many thanks in advance

[Daaman]

Originally Posted by lalatuobu

Hi Ryan,

I want to compute the surface sensitivity for my configuration as well. But it seems to me that SU2 surface sensitivity requires some deformation by FFD or by user defined deformation function. What I want is small perturbance on surface and I want to calculate the current surface sensitivity. I wonder how to do that?

Hi, if I understand you correctly- you want to deform the surface and then perform the sensitivity analysis of that deformed geometry? If that's the case, then you can deform the surface by displacing the FFD control point using SU2_DEF. Then use the resulting .su2 file with deformed geometry for performing the sensitivity analysis.