[Eduardo Carvalho]

Hi all,

I'm simulating a two-dimensional flow past a bluff body. I computed this case using v6 and v7 and found a concerning problem in the surface file generated. After examining the surface_flow file, I noticed some surface cells lacking in the one from v7.

The following images show details of the geometry where this problem happens. The red line is from the surface_flow v7, and the white one is from v6.

https://imgur.com/gWrcUIY

https://imgur.com/Uw4CpI9

Although these cells do not appear in the surface_flow, they do exist in the flow file. The flow variables near them seem to behave normally, considering the affected surface as having a wall boundary condition.

Since I'm trying to optimize this shape, I'm concerned with the possibility of this problem disturbing the gradient calculation.

Is there a way of avoiding it? If not, should I be worried about its impact on the optimization? I found just this error in just these two cells, and they are very small when compared to the full geometry.

The two versions are 6.2.0 and 7.0.4.

Thanks for your attention.

Eduardo

[pcg]

Hi,

What export format are you using?

The information for gradients always comes from the native volume file, it does not have anything to do with the surface outputs (which are just for post processing).

In any case I imagine this is an MPI-related issue, so you can try running the code on different number of processors and see if there is a significant impact in the gradients.

Regards,
Pedro

[Eduardo Carvalho]

Hi, Pedro

Thanks a lot for your quick answer.

I haven't defined an OUTPUT_FILES command, so I'm considering that SU2 is using the default values, or "(RESTART, PARAVIEW, SURFACE_PARAVIEW)". I also checked the files and all restart_flow are binary with a ".dat" extension.

You're correct it does look like a parallelism problem. I tested your guess, executed the simulation without parallelism, and the problem didn't happen. I think this is due to a problem with the differences in installation between the mpi (SU2 v6) and the mpich (SU2 v7).

Do you think this could indicate a parallelism problem that could potentially spoil the optimization?

Eduardo

[pcg]

I don't think it should affect gradients (they are based on volumetric values).
In any case if the gaps in the output are a single element, then all nodal values are still there (silver linings...).
But the way to test would be the same, run it serial and run it in parallel.
The problem should not be due to different MPI distributions, most likely it's a bug I'll bring it up with the other developers.
Regards,
Pedro

[Eduardo Carvalho]

Thanks a lot for your answer, Pedro, it really helped me.

I know that this question is off topic, but do you know if it is possible to edit the project.pkl file? Is there a way of restarting the optimization cycle from a specific step (e.g. after the adjoint simulation of the 3rd design)?

Best regards

Eduardo

[pcg]

If there is a way, I do not know it.
Regards,
Pedro