[Tommy Chen]

Hi Guys:

I want to know some details about the computation of the gradient of Volume grid node movement respect to the surface grid node in SU2.

Could anyone point out that part of code related to this functionality ?

Thanks !

[fpalacios]

Originally Posted by Tommy Chen

Hi Guys:

I want to know some details about the computation of the gradient of Volume grid node movement respect to the surface grid node in SU2.

Could anyone point out that part of code related to this functionality ?

Thanks !

Mmm, maybe I have misunderstood the question. But, if you are asking for the volumetric sensitivity (how changes the volumetric mesh when we perturb the surface geometry), the continuous adjoint formulation does not require that step (everything is computed only using surface information.

Best,
Francisco

[Tommy Chen]

Originally Posted by fpalacios

Mmm, maybe I have misunderstood the question. But, if you are asking for the volumetric sensitivity (how changes the volumetric mesh when we perturb the surface geometry), the continuous adjoint formulation does not require that step (everything is computed only using surface information.

Best,
Francisco

Hi Francisco:

Is this partially due to the assumption that the perturbation of the surface grid node is normal to the local wall surface when you formulate the sensitivity equation?

Since I am much more familiar with discrete adjoint, in which the chain rule has to be used , including the derivatives of volume mesh node displacements respect to every surface mesh node displacement, in order to compute the final gradient of dv. I do not understand why on surface information is necessary.

So , the outputed surface adjoint file after running the SU2_CFD adjoint solver is the only file needed to compute the projected gradient, is it right ?

many thanks

[economon]

Hi Tommy,

Indeed, this is the case for the type of continuous adjoint based on shape calculus that we have implemented in SU2. The resulting variation of the objective function is simply an integral over the design surface (no volume mesh sensitivities), and the integrand is what we call the surface sensitivity, which is computed at each mesh node. The surface sensitivity is a measure of the change in the objective function given a small perturbation in the local normal direction for each surface mesh node.

Hope this helps clear things up,
T

[Tommy Chen]

Originally Posted by economon

Hi Tommy,

Indeed, this is the case for the type of continuous adjoint based on shape calculus that we have implemented in SU2. The resulting variation of the objective function is simply an integral over the design surface (no volume mesh sensitivities), and the integrand is what we call the surface sensitivity, which is computed at each mesh node. The surface sensitivity is a measure of the change in the objective function given a small perturbation in the local normal direction for each surface mesh node.

Hope this helps clear things up,
T

Hi Tom

Could you recommend some related material to me about the adjoint method used in SU2 ? Paper, books or whatever that can interpret the theory

Thank you