Computing surface sensitivity using discrete adjoints

[SumanVajjala]

Hi, I have a problem computing the surface sensitivities for drag (CD) using discrete adjoints in SU2.

The primal solution's density residual converged to E-12. The adjoint solution's residual (density) decreases to a level of -8 and starts increasing. Reducing the initial CFL number for the adjoint computations makes the convergence even worse.

I have used the same settings for computing primal and adjoint soutions.

SOLVER= RANS
KIND_TURB_MODEL= SA
SA_OPTIONS= NEGATIVE, EXPERIMENTAL
MATH_PROBLEM= DISCRETE_ADJOINT
RESTART_SOL= NO

REF_DIMENSIONALIZATION= FREESTREAM_PRESS_EQ_ONE

NUM_METHOD_GRAD= WEIGHTED_LEAST_SQUARES
CFL_NUMBER= 4.0
MAX_DELTA_TIME= 1E10
CFL_ADAPT= YES
CFL_ADAPT_PARAM= ( 0.995, 1.005, 3.0, 8.0, 0.00001 )
ITER= 100000

LINEAR_SOLVER= FGMRES
LINEAR_SOLVER_PREC= ILU
LINEAR_SOLVER_ERROR= 1E-5
LINEAR_SOLVER_ITER= 20

CONV_NUM_METHOD_FLOW= JST
TIME_DISCRE_FLOW= EULER_IMPLICIT
OBJECTIVE_FUNCTION= DRAG

CONV_NUM_METHOD_ADJFLOW= JST
SLOPE_LIMITER_ADJFLOW= VENKATAKRISHNAN
ADJ_SHARP_LIMITER_COEFF= 3.0
ADJ_JST_SENSOR_COEFF= ( 0.0, 0.02 )
TIME_DISCRE_ADJFLOW= EULER_IMPLICIT
CFL_REDUCTION_ADJFLOW= 0.8
SENS_REMOVE_SHARP= YES
SENS_SMOOTHING= NONE
FROZEN_VISC_DISC= YES

CONV_NUM_METHOD_TURB= SCALAR_UPWIND
MUSCL_TURB= YES
SLOPE_LIMITER_TURB= NONE
TIME_DISCRE_TURB= EULER_IMPLICIT
CFL_REDUCTION_TURB= 1.0

I have tried with SU2 V8 and V8.2 without any success. Can anyone help me regarding these issues?

Suman

[SumanVajjala]

Update: When I let the adjoint solver converge to -8 then I get the following output for surface sensitivity. The visualization does not look right to me.

[joshkellyjak]

Is this the full config? You have no boundary conditions.

[SumanVajjala]

@joshkellyjak Hi, this is for the full configuration. There are boundary conditions in the file. I did not share them here.

[joshkellyjak]

OK without all the information it can be difficult to determine if there is a mistake in the config somewhere. Please attach the full config in future.

Your primal is converged to 1E-12, where does it start? How many orders of magnitude reduction do you get?

[bigfooted]

my suggestion is to take an existing configuration that works, like this one:https://su2code.github.io/tutorials/Inc_Turbulent_Bend_Opt/
and replace the mesh with your own mesh and see if that works.

[SumanVajjala]

@joshkellyjak I've shared the cfg files and primal computation's history file. Can you please tell me if I have made any mistakes in the setup?
Config_Log.zip

@bigfooted Thanks! I will try that.

[SumanVajjala]

@joshkellyjak @bigfooted Will the adjoint solver's convergence worsen when the CFL number is reduced? This is happening for my simulation. I feel that should not happen.

[SumanVajjala]

@pcarruscag if the direct solution has separated zones then will there be an impact(adverse) on the adjoint solution's convergence?

[joshkellyjak]

Sorry, just to be clear, when I said lower the CFL I meant in the direct solution, apologies if this was unclear in my initial reply. The convergence in direct looks good but one comment I would make is this looks to be a fairly complex geometry, have you run something simpler first to test the workflow? The convergence appears to be relatively quick, how big is your mesh? This could be beneficial in this case. Did you try running the adjoint with an increased CFL as suggested?

[SumanVajjala]

@joshkellyjak I did run the Onera M6 case provided by SU2 and the adjoint solver converges. However, when I use a finely resolved mesh, the adjoint solver diverges. Direct solver converges excellently though.

[joshkellyjak]

Could be worth simplifying your geometry somewhat to see if you can achieve adjoint convergence on something specific to your problem, then build back up by each component. This could reveal where any issues are arising from.

[SumanVajjala]

@joshkellyjak I don't know where to look anymore, honestly speaking. I ran the Onera transonic viscous flow case and the adjoints do not converge for a fine mesh. Increasing CFL does not help. Direct solution converges very well.

I have an observation though. Starting the adjoint solver from the direct solution, the initial residue is computed as -7.27! Isn't that strange? I thought the residue should be larger given the initial solution is bound to be a bad guess.

[SumanVajjala]

@joshkellyjak I have another query. Shouldn't the intermediate files atleast show some meaningful contours? I do not see them in my simulations.

[joshkellyjak]

I mean simplify your case, only run the wing, then wing and engine, etc., you are starting of with quite a complex geometry, simplify the case and work up and try and identify issues in this manner.

[SumanVajjala]

@joshkellyjak Ok. I did use adjoints in an older version of SU2 and the results looked meaningful. I am just not able to run the adjoints in the latest version. I will try to simplify the problem and figure out issues one by one.