[pgdelaserna]

Hi all,

I am using the discrete adjoint solver based on the RANS equations (with SA as turbulence model) to obtain the surface sensitivities embedded in SU2. My issue is that, even though I achieve a reasonable convergence of the direct problem (where I reduce the residuals about 5 orders of magnitude and the residual plots are smooth, with no oscillations towards the end), the discrete adjoint is oscillating.

For the direct problem, I have used JST as the convective numerical method, with MUSCL flow activated and VAN_ALBADA_EDGE as slope limiter. Now, my questions are:

1. For the discrete adjoint problem, do I also have to activate MUSCL_ADJFLOW, as I have used MUSCL scheme for the flow solution?
2. Does necessarily the convective numerical method for the adjoint problem need to be the same as the one used for the flow? (e.g. if I've used JST for the flow, do I need to use JST as well for the discrete adjoint).

Any help related to how to avoid oscillations in the discrete adjoint residuals is much appreciated

Thanks in advance!

[wallym]

Hiya,

Thanks for your question. My suggestion for you would be to try and converge your direct problem even more to get a better discrete adjoint solution.

To answer your questions, these options are only required for the continuous adjoint formulation (both in 1 and 2). So messing with these wont affect your solution.

Best,
Wally

[pgdelaserna]

Hi Wally,

Thanks for your help! I did not know that those options were only for the continuous adjoint. But indeed I noticed that did not make any effect on my results if they were or not activated.

I will take into account your advise and try and converge better the direct problem.

Regards,
Pablo

[pcg]

Hi Pablo,

JST is a central scheme, it does not use MUSCL reconstruction.

For best convergence of the discrete adjoint everything should be consistent, otherwise you are linearizing around an unconverged solution.

From a mathematical stand point the discrete adjoint should converge better if you increase the CFL, and if relaxation becomes necessary use RELAXATION_FACTOR_FLOW in the range 0.5 - 0.8.
JST results in relatively hard linear systems so BCGSTAB+ILU are your best bet, the former to avoid using too much memory.

Cheers,
Pedro