[stimitkshah@gmail.com]

I am running a supersonic case with engine inflow/exhaust bcs (ENGINE_INFLOW_TYPE= FAN_FACE_PRESSURE), but the adjoint diverges. I have tried MARKER_INLET/OUTLET and MARKER_RIEMANN BCs as well. Other things tried include NAVIER_STOKES/EULER simulations, different meshes, and varied bunch of other parameters in configuration file. Any suggestions on what I can do to get the adjoint to converge with engine inflow/exhaust BCs will be appreciated.

Thanks

[pcg]

Which adjoint? Continuous or discrete?
I don't know much about continuous adjoint, but if you are running the discrete adjoint try running at the a high CFL and using RELAXATION_FACTOR_FLOW is it still diverges.

[stimitkshah@gmail.com]

Discrete. Tried your suggestions, but it still blows up.

[wallym]

From my experience the primal solution needs to be resolved fairly well to prevent the discrete adjiont doesnt blow up. What are the residuals you are reaching in the flow solution?

[stimitkshah@gmail.com]

Primal solution residual drop are 4-5 orders of magnitude.
How can I improve on it?

[wallym]

What order are the finalized residuals?



I do not have much experience with the BC's you have mentioned but here are somethings you cant try:
1.) Run your simulation longer (setting residual reduction higher) if you haven;t reached a plateau
2.) Use a numerical flux scheme for high-speed flows (AUSM family, HLLC?, maybe roe)

[stimitkshah@gmail.com]

>> What order are the finalized residuals?
-- Final residuals for density kinda plateaued at -4 (dropping from 1), momentum[0], [1], [2] are at -1 (dropping from 3) and energy is at 2 (dropping from max of 7). But why is the final value important?

>> I do not have much experience with the BC's you have mentioned but here are somethings you cant try:
1.) Run your simulation longer (setting residual reduction higher) if you haven't reached a plateau
2.) Use a numerical flux scheme for high-speed flows (AUSM family, HLLC?, maybe roe)
-- Still I tried running the simulation longer (20000 iterations instead of 10000) and usedCONV_NUM_METHOD_FLOW= AUSM numerical flux scheme for primal and CONV_NUM_METHOD_ADJFLOW= ROE for adjoint and that actually resulted in target CL not getting met (CL has significant oscillatory behavior with iterations) and residuals are flat at very high values.

[pcg]

That does not sound like good convergence, 6 orders would be preferable, and sometimes it is more important that you do not get limit-cycle behavior, which it seems you are getting.

CONV_NUM_METHOD_ADJFLOW is a continuous adjoint option I believe, the discrete uses whatever you have for the primal solver.

To rule out convergence issues I would start with the JST scheme which tends to be the most robust (there are recent but closed topics on github about how to get best performance out of JST).
If that works switch to 1st order Roe, then 2nd order Roe with the Venkat-Wang limiter. Roe still gives you control over numerical dissipation via the entropy correction coefficient, from my experience that makes it more robust than the other upwind schemes we have atm.

The convergence properties of the discrete adjoint are very similar to whatever you get in the direct simulation so you really need to take the time to make it converge properly.

[stimitkshah@gmail.com]

I have tried JST scheme before. Could you point me to the post on github that discusses how to get best performance out of JST? Maybe one or more settings that I have in my config file aren't right.
Thanks

[stimitkshah@gmail.com]

So I tried the recommendations for getting out best performance out of JST given in the post https://github.com/su2code/SU2/issues/793, but adjoint still diverges. Primal solution converges faster though

[pcg]

Can you make your case available? JST is quite robust so it might be a bug...
If it is too large or a proprietary 11 herbs and spices project there is not much we can do.

[stimitkshah@gmail.com]

Yes I can't upload the proprietary geometry, but I do have a made up oneraM6 + cylinder case that has same behavior. Note that the MARKER_RIEMANN (commented out in the config file) works fine but MARKER_ENGINE_INFLOW/EXHAUST does not.
To reduce the size of the attachment, I have uploaded .pw file without mesh in it.
.png file shows the setup and bcs
Let me know if you have questions about the setup.

[stimitkshah@gmail.com]

Following pictures show convergence of 2 cases
1) case with MARKER_ENGINE_INFLOW/EXHAUST BC (adjoint blows up)
2) with MARKER_RIEMANN BC (adjoint works fine)

Note that the setup with MARKER_RIEMANN does not work so well with the real proprietary case. Adjoint still has convergence issues with it.