[hejiandong]

Dear developers:

I tried to run the test case of unsteady rans optimization of pitching NACA64A010 but finally failed.

because the sensitivity of many points are too large and finally collapsed at the begin of second optimization loop..

i want to ask if the test case has been tested?

[economon]

Hi,

Yes, this test case has worked in the past. However, this is a very active research area at the moment. If you haven't already, you might try the files that are available in the test cases folder under TestCases/optimization_rans/pitching_naca64a010/.

All the best,
Tom

[EMolina]

Hi Tom.

I succesfully performed the pitching naca64a10 euler optmization. And I noticed that the objective function to minimize, for example DRAG, is actually the mean value of this objective function over time. I would like to change this objective function to be the standard deviation not the mean value.That is, the code is doing right, the average of the objective function is decreasing, but after each iteration of the design space, the value of the standard deviation increases.

I searched the part of the code where the mean of the objective function is calculated for unsteady adjoint optimizations and I have not found. If it is possible, I would like your help on where I could perform this modification in the code.

Thank you in advance.

Eduardo

[hejiandong]

Originally Posted by economon

Hi,

Yes, this test case has worked in the past. However, this is a very active research area at the moment. If you haven't already, you might try the files that are available in the test cases folder under TestCases/optimization_rans/pitching_naca64a010/.

All the best,
Tom

Dear Tom

I have performed TestCases/optimization_rans/pitching_naca64a010/ with the newest version of SU2, but the averaged drag coefficient increased up to 10 times of original configuration in the second optimization loop. and the thickness increased almost two times of the original configuration. is there any problem of the code?

[economon]

Hi,

I do not expect that anything significant has changed related to the unsteady adjoint recently. However, sometimes small changes can affect the performance of the shape design process, since there are so many different components at work.

To help figure out what is going on, can you check that the gradient computed by the continuous adjoint method is approximately on the order of a small deformation relative to the airfoil chord length (1-2% chord, for instance)? If the values are large, you can scale the objective & gradient in the opt. portion of the config file so that the first step of the optimizer remains feasible, i.e., the airfoil doesn't invert or cause the mesh deformation code to fail.

Cheers,
Tom

[hejiandong]

Originally Posted by economon

Hi,

I do not expect that anything significant has changed related to the unsteady adjoint recently. However, sometimes small changes can affect the performance of the shape design process, since there are so many different components at work.

To help figure out what is going on, can you check that the gradient computed by the continuous adjoint method is approximately on the order of a small deformation relative to the airfoil chord length (1-2% chord, for instance)? If the values are large, you can scale the objective & gradient in the opt. portion of the config file so that the first step of the optimizer remains feasible, i.e., the airfoil doesn't invert or cause the mesh deformation code to fail.

Cheers,
Tom

Dear Tom

I have attached my cfg file and history project file, I,m sure that the mesh deformation is successfully running, and after the second loop, the time averaged drag coefficient began to decrease. however, after 10 optmization loops, drag coefficient is still larger than baseline configuration.

I will also give a try to further scale the opt_obj.

[hejiandong]

Originally Posted by hejiandong

Dear Tom

I have attached my cfg file and history project file, I,m sure that the mesh deformation is successfully running, and after the second loop, the time averaged drag coefficient began to decrease. however, after 10 optmization loops, drag coefficient is still larger than baseline configuration.

I will also give a try to further scale the opt_obj.

I found that maybe the problem is sensitivity of geometry can not converged in unsteady adjoint, this figure become larger and large even after lots of steps. however, in steady cases, everything seems OK , anyone can give me some suggestions? thanks a lot!

Another things to mention is that this figure of unsteady adjoint became negative while the according steady case of NACA64- A010 is positive numbers.

[hejiandong]

Originally Posted by economon

Hi,

I do not expect that anything significant has changed related to the unsteady adjoint recently. However, sometimes small changes can affect the performance of the shape design process, since there are so many different components at work.

To help figure out what is going on, can you check that the gradient computed by the continuous adjoint method is approximately on the order of a small deformation relative to the airfoil chord length (1-2% chord, for instance)? If the values are large, you can scale the objective & gradient in the opt. portion of the config file so that the first step of the optimizer remains feasible, i.e., the airfoil doesn't invert or cause the mesh deformation code to fail.

Cheers,
Tom

Dear Tom

I have scaled the opt.portion. But it seems not the problem.
Like what i posted above, time averaged drag coefficient increased in first loop and decreased in following loops. but after 12 loops, optimization terminated, the final configuration and drag coefficient is almost the same as baseline configuration...so it is more like that i was carrying a inverse design problem rather than drag reduce optimization..

Another thing to mention is that only in first loop, the code run the adjoint calculation! that is only in DSN_001, I can find the adjoint information.

Hopes your response!!

[economon]

Hi,

Just wanted to say that I plan to run this case again soon for a paper. Hopefully, I will be able to get back to you relatively soon with more details on this case...

Cheers,
Tom

[hejiandong]

Originally Posted by economon

Hi,

Just wanted to say that I plan to run this case again soon for a paper. Hopefully, I will be able to get back to you relatively soon with more details on this case...

Cheers,
Tom

Thanks a lot!!

I also found that unsteady rans optimization can work on euler grid.

Now i have another question, i want to run SU2 in our computer clusters, however the code can only run on the control cluster, because the linux system only installed on control cluster. maybe i need some script or other method to apply the code on computation clusters. can u give me some advice?