Problem with Discrete adjoint based on RANS

SumanVajjala · June 18, 2018, 05:30

Hi. I am using SU v 6.1 to extract adjoints for a simple onera M6 wing (mesh provided in the SU2 Testcases suite) using discrete adjoint for RANS and the simulations are blowing-up when the solver starts the AD part after I get the direct solution. However, the same simulation does not blow when running Euler simulation instead of RANS. I have tried many options- reducing CFL no, changing the flux calculation (Least squares and Green-Gauss) but none worked. Can you please suggest what to do in this regard.

Best
Suman

Zen · July 4, 2018, 04:36

Hi Suman,

Can you attach your config file?

Zen

SumanVajjala · July 12, 2018, 02:37

Hi Zen!

Quote:

Originally Posted by Zen

Hi Suman,

Can you attach your config file?

Zen

Sorry for the late reply. I have attached the config file. Please let me know if there are any issues with it.

Thanks
Suman

pcg · July 31, 2018, 06:36

Hi Suman,

The universal advise with discrete adjoint is that you should converge the direct solver very well. Do note that the turbulence residuals usually take longer to drop, I usually use a much larger CFL for the turbulence equations (by setting CFL_REDUCTION_TURB > 1).

With regards to solver settings you can try:
MGLEVEL= 2 or 3

LINEAR_SOLVER= BCGSTAB
LINEAR_SOLVER_PREC= ILU
LINEAR_SOLVER_ERROR= 1E-3
LINEAR_SOLVER_ITER= 1000 (just to make sure the linear solver converges)
DISCADJ_LIN_SOLVER= BCGSTAB
DISCADJ_LIN_PREC= ILU

I suggest the switch to BiCGSTAB because that way you can increase the number of iterations to make sure the linear solver converges during the first iterations without increasing memory usage (that would happen with FGMRES). If BiCGSTAB fails (it should not with the CFL you have) switch back to FGMRES, keep ILU and reduce the number of iterations to the 10-100 range.

You mentioned that Euler works fine, keep in mind that SU2 does not have wall functions, therefore you need y+ around 1. If your wall cell size is already that small be careful with very high edge ratios (1000:1 is as far as I would go for testing).

If all this fails there is always the option to freeze the turbulence in the adjoint solver (I do not recall the name but you can check the template config file).

Cheers,
Pedro

June 18, 2018, 05:30	Problem with Discrete adjoint based on RANS	#1
SumanVajjala New Member Suman Vajjala Join Date: Jun 2014 Posts: 4 Rep Power: 12	Hi. I am using SU v 6.1 to extract adjoints for a simple onera M6 wing (mesh provided in the SU2 Testcases suite) using discrete adjoint for RANS and the simulations are blowing-up when the solver starts the AD part after I get the direct solution. However, the same simulation does not blow when running Euler simulation instead of RANS. I have tried many options- reducing CFL no, changing the flux calculation (Least squares and Green-Gauss) but none worked. Can you please suggest what to do in this regard. Best Suman

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
the gradients calculated through discrete adjoint are near zero	Haowang	SU2	9	October 17, 2019 23:53
SU2 LES, discrete adjoint	testcase	SU2	1	July 27, 2017 05:46
Problem with Application based on a faceZone in parallel	psilkeit	OpenFOAM Programming & Development	2	April 28, 2016 10:47
SQP method for adjoint based aerodynamic shape optimization	cfdbooks	Main CFD Forum	2	May 26, 2009 11:40
steady RANS for flow problem with unsteady inlet?	kat	Main CFD Forum	0	February 16, 2009 11:24

July 4, 2018, 04:36		#2
Zen Member Zeno Join Date: Sep 2013 Location: Delft, The Netherlands Posts: 63 Rep Power: 13	Hi Suman, Can you attach your config file? Zen

July 31, 2018, 06:36		#4
pcg Senior Member Pedro Gomes Join Date: Dec 2017 Posts: 466 Rep Power: 13	Hi Suman, The universal advise with discrete adjoint is that you should converge the direct solver very well. Do note that the turbulence residuals usually take longer to drop, I usually use a much larger CFL for the turbulence equations (by setting CFL_REDUCTION_TURB > 1). With regards to solver settings you can try: MGLEVEL= 2 or 3 LINEAR_SOLVER= BCGSTAB LINEAR_SOLVER_PREC= ILU LINEAR_SOLVER_ERROR= 1E-3 LINEAR_SOLVER_ITER= 1000 (just to make sure the linear solver converges) DISCADJ_LIN_SOLVER= BCGSTAB DISCADJ_LIN_PREC= ILU I suggest the switch to BiCGSTAB because that way you can increase the number of iterations to make sure the linear solver converges during the first iterations without increasing memory usage (that would happen with FGMRES). If BiCGSTAB fails (it should not with the CFL you have) switch back to FGMRES, keep ILU and reduce the number of iterations to the 10-100 range. You mentioned that Euler works fine, keep in mind that SU2 does not have wall functions, therefore you need y+ around 1. If your wall cell size is already that small be careful with very high edge ratios (1000:1 is as far as I would go for testing). If all this fails there is always the option to freeze the turbulence in the adjoint solver (I do not recall the name but you can check the template config file). Cheers, Pedro