[cfd-jg]

Hi all,

I am doing my thesis on aerodynamic shape optimisation of winglets. However, I am struggling to make this work on SU2.

My aim is to minimize drag maintaining lift fixed. The approach I was following is setting an FFD box around the winglet (taking part of the wing as well). Then setting control points, which are allowed to move vertically (so it can adjust the cant angle, radius of rotation of the winglet, AoA and twist). And once I had this working, my plan is to add more complexity letting the control points to move both in x and z directions (to optimise sweep angle as well).

However, I am having issues with implementing this. The problem I am finding is that the objective function (drag) is reduced minimally (only 4th decimal place). I first thought that the problem might be that I had to tune the parameters OPT_GRAD_FACTOR and OPT_RELAX_FACTOR, so I tried it with multiple values (see fig Cd-gradFactor-4.jpg), but none of them reduced drag significantly, while a couple of them diverged.

To give an example of the output: (fig surface_deformed-relaxFactor_600.jpg) this is what the solver generates (first few iterations). This mainly modifies the AoA to reduce drag but not the cant angles much. Also, despite having the 2nd derivative continuity constraint on the configuration file, the output produced is not physically viable. This is because the gradients on the control points next to the boundary between wing-winglet are quite high compared to the rest of control points (seen in fig gradients.jpg).

I was expecting the deformation of the geometry to be larger (at least in cant angle). Interestingly, when I set a larger OPT_GRAD_FACTOR (1E3) the cant angle was deformed significantly (see fig surface_deformed-relaxFactor_1E3.jpg) but maintaining the discontinuity in the boundary between wing and winglet. As a result, the adjoint solution diverges after this point.

I have also tried with different scales in the objective function, but similar results are obtained.

To check if there was something wrong in my configuration file, I checked its validity with the ONERA wing mesh tutorial, and it worked fine (for that mesh and FFD box configuration). Therefore I suspect that my approach may not be the correct one for my case.

I would highly appreciate if anyone more experienced could help or indicate how to approach this optimisation or any suggestion that could help me.


Thank you very much for your help in advance,


The configuration file I am using is: EULER-winglet-shapeOptimisation.txt

[bigfootedrockmidget]

The first step is to check the CFD setup.



Do you have a fully converged CFD simulation (independent of the adjoint setup) for a specified alpha that shows the expected CL (I guess there is measurement data)? This to check that the CL that you choose is a viable option, the the lengthscales were chosen correctly to get to this CL and that the solver settings that you use realize a converged solution.

[cfd-jg]

Originally Posted by bigfootedrockmidget

The first step is to check the CFD setup.



Do you have a fully converged CFD simulation (independent of the adjoint setup) for a specified alpha that shows the expected CL (I guess there is measurement data)? This to check that the CL that you choose is a viable option, the the lengthscales were chosen correctly to get to this CL and that the solver settings that you use realize a converged solution.

Thanks for your reply.


Yes, I first ran the simulation with MATH_PROBLEM= DIRECT and got a converged result of CL= 0.853490 and CD= 0.011720 cfd-residuals.jpg. So then I used CL = 0.85 as the target lift. When running the shape optimisation, the results also converged log_DIRECT.jpg. These are the residuals of the direct solution of a random design iteration.


Even though the convergence criteria are not reached in the pictures, the residuals are quite low (around -9/-10, and both coefficients reach a steady value).

[bigfootedrockmidget]

You need to have lower residuals. The adjoint can be quite sensitive to left-over errors. I would go to at least 1e-13, preferably 1e-14. You can also see CL still changing so that's an indication that you need to go lower.
Then, when you go to the adjoint with this converged solution, please check that the adjoint has converged as well. You can also load the history.csv file directly in paraview to visually inspect all the saved residual values.