
[Sponsors] 
August 22, 2012, 05:44 
Adaptive Timestep and Residuals

#1 
Senior Member
Stuart
Join Date: Jul 2009
Location: Portsmouth, England
Posts: 488
Rep Power: 15 
Hi,
I'm running a transient external aero simulation around some bluff bodies (ICEM hybrid mesh ~20 million elements  on my 8 core PC takes ages to run). Re = 6 million. I did the usuall steady state simulations (SST komega) to determine mesh and physcial time scale sensitivity in order to get a flowfield to initialise the transient run. I expect a turbulent flowfield behind the bluff bodies with a range of scales  not a periodic shedding of vortices. So I don't know what transient time setup to use. Therefore, following Glenn's reply ([CFX] Simulation of Flow Separation in a rectangular diffuser  Convergence Problem) I'm using the adaptive timestepping to CFX can work it out it self. But I've found the RMS residuals have increased greatly (see image of the run in progress). In using the SASSST with Bounded CD advection scheme following the ANSYS SRS Best Practice document. I could not get the residuals any lower in the steady state run as the mesh size was getting far too large (however, the steady state monitor points became constant and the imbalances were all 0%). Should I let this simulation to continue (it'll take ages) or does anyone known how to improve things. I'm also watching the monitor points which are starting to oscillate (which I expect was they are points in the turbulent wake). Thanks 

August 22, 2012, 06:42 

#2 
Super Moderator
Glenn Horrocks
Join Date: Mar 2009
Location: Sydney, Australia
Posts: 12,832
Rep Power: 100 
You cannot do time scale sensitivity on a steady state run. It has to be done on a transient run. Doing mesh sensitivity on a steady state run sounds OK. But if you use adaptive time stepping that avoid the problem  but you still need to define a convergence criteria and that requires a sensitivity analysis.
You cannot directly compare residuals from steady state and transient runs. They are normalised differently. So do a sensitivity analysis on your convergence criteria on the transient run and if that shows it is OK then you are fine. 

August 22, 2012, 07:05 

#3 
Senior Member
Stuart
Join Date: Jul 2009
Location: Portsmouth, England
Posts: 488
Rep Power: 15 
Thanks Glenn,
Since a timescale (in my case a physical timescale) is needed for a steady state simulation I always thought it a good idea to do a sensitivty test on that parameter. Do you consider that pointless? With regards to a sensitivity analysis on the convergence criteria. The images shows from the CFXSolver Manager Out File (the run is still going) the Timestep and Solver Controls I set. BTW the 1 second limit was arbitary and I'll end up stopping the run before then. So which input parameter would I need to change? I guess it must be the Coeff Loops because since the RMS residuals (mass and momemtum) are not even getting close to 1e06 then any change to that would, surely, not make any difference. Regards 

August 22, 2012, 07:14 

#4 
Super Moderator
Glenn Horrocks
Join Date: Mar 2009
Location: Sydney, Australia
Posts: 12,832
Rep Power: 100 
Sensitivity analysis on physical time step size for a steady state run is pointless. The only thing which matters is how tight the convergence is. The time step is just a means to achieve convergence.
You have specified a very tight convergence. Are you sure you need to be that tight? Why are you using a central difference advection scheme (very accurate), second order time differencing (very accurate) but only first order turbulence numerics? This is not necessarily wrong, just that it is unusual. 

August 22, 2012, 07:27 

#5 
Senior Member
Stuart
Join Date: Jul 2009
Location: Portsmouth, England
Posts: 488
Rep Power: 15 
Hi Glenn,
My reason for setting a target residual = 1e06 was because I thought the solver would never get to that level and so the solver would not stop running until it got the the specified number of iterations  which I set to 200 to ensure all the monitor points became flat for a good handful of iterations. I'm looking through the CFX guides about the transient residuals are normalized differently but I'm not seeing anything to say it's a different method to steady state residuals. I was under the impression from reading through the CFX guides that because I'm not using LES (or a RANSLES) model I should leave the turbulence numerics to 1storder as with a RANS simulation. Do you recommend changing it for SASSST? Again the ANSYS SRS Best Practice gude says nothing with that respect. So for a sensitivty test on the convergence, what parameter(s) do you suggest I change? Regards 

August 22, 2012, 07:37 

#6 
Super Moderator
Glenn Horrocks
Join Date: Mar 2009
Location: Sydney, Australia
Posts: 12,832
Rep Power: 100 
OK, you are just running it until it converges as far as it can.
The transient residual contains a transient term which obviously the steady state one does not. The SASSST model is a transient model. You cannot run it steady state. For sensitivity on convergence, the main parameter is the residual tolerance, but for some models the imbalance is important. 

August 23, 2012, 02:21 

#7 
Senior Member
Stuart
Join Date: Jul 2009
Location: Portsmouth, England
Posts: 488
Rep Power: 15 
I know SASSST is transient only so cannot be run in steady state  I'm not sure where in the topic I gave the impression that I was trying to run it as steady state.
Anyway, the simulation crashed last night. Not 100% why but it did have a high Mach number notice (M ~ 8) which is completely none physical for this simulation. So I've set a fixed timestep and try again. If that fails I'll try SSTURANS instead of SASSST although that's not the preferred model. 

Thread Tools  
Display Modes  


Similar Threads  
Thread  Thread Starter  Forum  Replies  Last Post 
Adaptive Timestepping: Possible Glitch?  Josh  CFX  9  August 25, 2010 21:52 
How does openFoam calculate residuals?  sahm  OpenFOAM  0  March 1, 2010 15:10 
Monitor residuals  Juan CatelĂ©n  CFX  4  March 26, 2007 08:25 
Use of Timestep in obtaining solution.  hagupta  CFX  7  February 28, 2006 14:14 
Convergence  scaled vs unscaled residuals  HS  FLUENT  1  November 7, 2005 06:45 