Adjoint Solver Convergence

devesh.baghel · August 7, 2017, 12:14

Hi All,

I need to do an optimization study on Auto component.
I have few question on adjoint solver.
1. Can we run Adjoint with Segregated solver or it only works with coupled solver ?
2. I run a air flow model on segregated and coupled solver i.e. I used same case with two solvers. Case with segregated solver converged in 2500 iterations while coupled solver converged with 15000 iterations. I didn't understand why it is happening ?
3. I tried to run by change in parameter of coupled solver i.e. Grid Sequencing, Expert Driver and Continuity Convergence Accelerator but no result in terms of convergence. It takes almost same number of iterations.

Can anyone please throw some light on above mentioned points.

Thanks in advance.

Reagrds
Devesh

me3840 · August 8, 2017, 11:16

1. No, coupled only.
2. What did you use as the CFL? Star's coupled solver is density-based, so if your flow is incompressible you have to do a lot of tuning to get the solver to work well.

What is the component?

devesh.baghel · August 8, 2017, 12:10

Hi,

Thank you for prompt reply on query.

Yes, It is incompressible flow (constant density) model.

1. CFD I used default value of 5 for steady state flow.
Even when I observed divergence, I started run with 1st oder (coupled) along with CFD-10. Afterwards I reduced CFD 10 to 5, even sometimes I used 2 as well.
2. I re-meshed the model with different settings of prism layer (2-5) layers and observed in few cases mass flow monitor and convergence were good upto 100-200 iterations. Afterwards suddenly recirculation observed at outlet boundary and simulation got diverged.
3. When started with lower mass flow (50%) of actual and then gradually increased to actual mass flow rate, case is running smoothly but taking long run to get mass flow balance.

Please put opinion on that, so would get direction to proceed further.

Thanks to all

me3840 · August 8, 2017, 12:44

What are you simulating? Those CFLs are far too low to beat the segregated solver's convergence rate.

devesh.baghel · August 13, 2017, 22:55

Hi,
@me3840... you were absolutely true on convergence rate of coupled solver....
I tried with higher CFL(20-50), but case diverged every time within 500 iterations even though started with 1st order...
I used typically conventional method of start with low flow rate with gradual increment to have control on solution.... and finally it's done....

Sent from my Redmi 3S using CFD Online Forum mobile app

devesh.baghel · August 28, 2017, 11:32

Hi guys,

Need a quick opinion on adjoint solver parameters.
Can we run through multiple adjoint cost function in single simulation?
For example, If I am running an air flow simulation in cylinder; can we apply pressure, mass flow, force etc cost function simultaneously in one model ?

How about the reliability of solution, if used multiple cost function ?

Thanks for help...

urosgrivc · September 26, 2018, 05:32

Yes, you can combine multiple Observables into one and run the solver with combined result.
You can do that with operation types under settings

urosgrivc · October 9, 2018, 08:05

has anybody managed to converge the adjoint while
energy term being activated?
are there any turbulence model limitations while doing this?
I would like to optimize temperature variance at the outlet of water mixer with
two inlets but cant converge it

I observed that using enhanced wall treatment causes
the divergence of adjoint energy
And that the use of inflation layers causes slower convergence

It seems that adjoint solver is very sensitive to model settings for k-epsilon and near wall treatment.. are there any rules on what to use?

August 7, 2017, 12:14	Adjoint Solver Convergence	#1
devesh.baghel Member Devesh Baghel Join Date: Mar 2009 Posts: 84 Rep Power: 17	Hi All, I need to do an optimization study on Auto component. I have few question on adjoint solver. 1. Can we run Adjoint with Segregated solver or it only works with coupled solver ? 2. I run a air flow model on segregated and coupled solver i.e. I used same case with two solvers. Case with segregated solver converged in 2500 iterations while coupled solver converged with 15000 iterations. I didn't understand why it is happening ? 3. I tried to run by change in parameter of coupled solver i.e. Grid Sequencing, Expert Driver and Continuity Convergence Accelerator but no result in terms of convergence. It takes almost same number of iterations. Can anyone please throw some light on above mentioned points. Thanks in advance. Reagrds Devesh

September 26, 2018, 05:32		#7
urosgrivc Senior Member urosgrivc Join Date: Dec 2015 Location: Slovenija Posts: 365 Rep Power: 12	Yes, you can combine multiple Observables into one and run the solver with combined result. You can do that with operation types under settings Last edited by urosgrivc; October 10, 2018 at 08:25.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Adjoint solver for flow problems	jelmer	OpenFOAM Running, Solving & CFD	11	March 12, 2017 07:52
Creating Helmholtz Solver: no convergence	serles	OpenFOAM Programming & Development	14	September 28, 2016 10:22
fluent divergence for no reason	sufjanst	FLUENT	2	March 23, 2016 17:08
convergence of density-based solver for unsteady flow	zhengjg	Main CFD Forum	0	June 16, 2014 12:37
force convergence problems in CFX 6DOF rigid body solver	ajay_ks	CFX	8	March 25, 2013 05:02

August 8, 2017, 11:16		#2
me3840 Senior Member Join Date: Nov 2010 Location: USA Posts: 1,232 Rep Power: 25	1. No, coupled only. 2. What did you use as the CFL? Star's coupled solver is density-based, so if your flow is incompressible you have to do a lot of tuning to get the solver to work well. What is the component?

August 8, 2017, 12:10		#3
devesh.baghel Member Devesh Baghel Join Date: Mar 2009 Posts: 84 Rep Power: 17	Hi, Thank you for prompt reply on query. Yes, It is incompressible flow (constant density) model. 1. CFD I used default value of 5 for steady state flow. Even when I observed divergence, I started run with 1st oder (coupled) along with CFD-10. Afterwards I reduced CFD 10 to 5, even sometimes I used 2 as well. 2. I re-meshed the model with different settings of prism layer (2-5) layers and observed in few cases mass flow monitor and convergence were good upto 100-200 iterations. Afterwards suddenly recirculation observed at outlet boundary and simulation got diverged. 3. When started with lower mass flow (50%) of actual and then gradually increased to actual mass flow rate, case is running smoothly but taking long run to get mass flow balance. Please put opinion on that, so would get direction to proceed further. Thanks to all

August 8, 2017, 12:44		#4
me3840 Senior Member Join Date: Nov 2010 Location: USA Posts: 1,232 Rep Power: 25	What are you simulating? Those CFLs are far too low to beat the segregated solver's convergence rate.

August 13, 2017, 22:55		#5
devesh.baghel Member Devesh Baghel Join Date: Mar 2009 Posts: 84 Rep Power: 17	Hi, @me3840... you were absolutely true on convergence rate of coupled solver.... I tried with higher CFL(20-50), but case diverged every time within 500 iterations even though started with 1st order... I used typically conventional method of start with low flow rate with gradual increment to have control on solution.... and finally it's done.... Sent from my Redmi 3S using CFD Online Forum mobile app

August 28, 2017, 11:32		#6
devesh.baghel Member Devesh Baghel Join Date: Mar 2009 Posts: 84 Rep Power: 17	Hi guys, Need a quick opinion on adjoint solver parameters. Can we run through multiple adjoint cost function in single simulation? For example, If I am running an air flow simulation in cylinder; can we apply pressure, mass flow, force etc cost function simultaneously in one model ? How about the reliability of solution, if used multiple cost function ? Thanks for help...

October 9, 2018, 08:05		#8
urosgrivc Senior Member urosgrivc Join Date: Dec 2015 Location: Slovenija Posts: 365 Rep Power: 12	has anybody managed to converge the adjoint while energy term being activated? are there any turbulence model limitations while doing this? I would like to optimize temperature variance at the outlet of water mixer with two inlets but cant converge it I observed that using enhanced wall treatment causes the divergence of adjoint energy And that the use of inflation layers causes slower convergence It seems that adjoint solver is very sensitive to model settings for k-epsilon and near wall treatment.. are there any rules on what to use?