Trouble With Adjoint Convergence Using SST K Omega

[PeterWalsh]

When using SA turbulence models I see rapid convergence of the adjoint:

`
Solver Type: gmres
GMRES Restart: 2000
ASM Overlap: 1
Global PC Iters: 0
Local PC Iters: 1
Mat ReOrdering: natural
ILU PC Fill Level: 1
GMRES Max Iterations: 2000
GMRES Relative Tolerance: 1e-06
GMRES Absolute Tolerance: 1e-14
Solving Linear Equation... 938.55 s
Main iteration 0 KSP Residual norm 1.316119039528e+02 950.44 s.

Main iteration 100 KSP Residual norm 2.062007312202e+01 1002.15 s.
Main iteration 200 KSP Residual norm 9.167684159407e-01 1060.79 s.
Main iteration 300 KSP Residual norm 2.907447219647e-02 1127.95 s.
Main iteration 400 KSP Residual norm 1.046934348066e-03 1203.50 s.
Main iteration 461 KSP Residual norm 1.310959627761e-04 1254.09 s
Completed! Total iterations: 461. PetscConvergedReason: 2. 1254.09 s
`

However when using SST K Omega turbulence model I see this:
Solver Type: gmres
GMRES Restart: 2000
ASM Overlap: 1
Global PC Iters: 0
Local PC Iters: 1
Mat ReOrdering: natural
ILU PC Fill Level: 2
GMRES Max Iterations: 2000
GMRES Relative Tolerance: 1e-06
GMRES Absolute Tolerance: 1e-14
Solving Linear Equation... 1361.5 s
Main iteration 0 KSP Residual norm 1.331066571411e+02 1641.91 s.
Main iteration 100 KSP Residual norm 1.074773093342e+02 1722.17 s.
Main iteration 200 KSP Residual norm 8.285818651111e+01 1810.37 s.
Main iteration 300 KSP Residual norm 6.738514638871e+01 1907.94 s.
Main iteration 400 KSP Residual norm 5.952249148796e+01 2015.06 s.
Main iteration 500 KSP Residual norm 5.058505901916e+01 2131.76 s.
Main iteration 600 KSP Residual norm 4.270603602417e+01 2257.65 s.
Main iteration 700 KSP Residual norm 3.865112264224e+01 2395.99 s.
Main iteration 800 KSP Residual norm 3.382575491736e+01 2543.04 s.
Main iteration 900 KSP Residual norm 3.066078405984e+01 2697.49 s.
Main iteration 1000 KSP Residual norm 2.733987011118e+01 2860.95 s.
Main iteration 1100 KSP Residual norm 2.491887903865e+01 3034.22 s.
Main iteration 1200 KSP Residual norm 2.093743925734e+01 3216.92 s.
Main iteration 1300 KSP Residual norm 1.764545050563e+01 3409.54 s.
Main iteration 1400 KSP Residual norm 1.669450067479e+01 3614.63 s.
`

I have followed the recommendations on the FAQ changing ILU PC Fill to 2 and div(phi, U) to bounded Gauss upwind.
I am curious if there is anything else I can try?

I also changed the adqEqnOption from:

"adjEqnOption": {
"gmresRelTol": 1.0e-6,
"pcFillLevel": 1,
"jacMatReOrdering": "rcm",
"gmresMaxIters": 1000,
"gmresRestart": 1000,
},

To this:

"adjStateOrdering": "cell",
"adjEqnOption": {
"gmresRelTol": 1e-6,
"pcFillLevel": 2,
"jacMatReOrdering": "natural",
"gmresMaxIters": 2000,
"gmresRestart": 2000,
},

I'd really appreciate any assistance or additional information / things I can try to improve the adjoint convergence.

Thanks for the help!

[PeterWalsh]

I see I have read further discussion on Github here specifically this one: #74 (comment)

You don't recommend resolving down to Y+ < 1 as a general rule?

You mention in the conversation: "We managed to run optimization with many y+=1 cases, the adjoint just converged slower; it just needed more iterations." Could you let me know how many iterations? Was it many thousands of iterations for the adjoint to converge? Does my rate of convergence appear similar to that case?

Thanks again for any help!

[friedenhe]

Yeah.. SST's adjoint converges more slowly than SA's, especially for y+=1 cases. One more thing you can try it so use the kahip decomposition: check here