Convergence in steady state simulations vs transient ones

 Register Blogs Members List Search Today's Posts Mark Forums Read

January 20, 2018, 18:21
#2
Super Moderator

Glenn Horrocks
Join Date: Mar 2009
Location: Sydney, Australia
Posts: 17,728
Rep Power: 143
Quote:
 So it seems that transient simulation does not work for this system
That means you are missing some important physics in the transient simulation. But you appear to be saying that if you use a steady state simulation then the time history (as modelled with steady state pseudo time presumably) is a better match for the expected results. Is this correct? If so this is concerning as steady state simulations do not guarantee time accurate simulations, several simplifications have been made to make convergence easier at the cost of inaccurate time history. The only thing which matters in a steady state simulation is the final converged result.

Quote:
 RMS is also not a good candidate to judge convergence in the steady state in my problem
Yes, that can happen. You might be sensitive to the small regions of high residual which RMS tends to average out. That is why the MAX option is there, to handle these cases.

 January 20, 2018, 18:38 #3 Member   Join Date: Jun 2017 Posts: 40 Rep Power: 8 In a nutshell, I have tried a SS solution to my problem and achieved convergence (max RMS < 5e-4) after ~14K iterations, increasing timescale by a factor of 1000 after iteration number 10K. Monitoring concentrations at some points, I realised that they monotonically increase and reach a SS after ~500 iterations, then they start to slightly decrease after iteration 5-6K to reach a new SS at iteration 14 K. The convergence criteria of max RMS < 0.0005 just reached after 14 K although my monitor points were flat at the first SS. I am actually interested in the first SS (the one after 500 iterations), so I used a stop criteria (Peclet

January 20, 2018, 18:46
#4
Member

Join Date: Jun 2017
Posts: 40
Rep Power: 8
Quote:
 Originally Posted by ghorrocks That means you are missing some important physics in the transient simulation. But you appear to be saying that if you use a steady state simulation then the time history (as modelled with steady state pseudo time presumably) is a better match for the expected results. Is this correct? If so this is concerning as steady state simulations do not guarantee time accurate simulations, several simplifications have been made to make convergence easier at the cost of inaccurate time history. The only thing which matters in a steady state simulation is the final converged result. Yes, that can happen. You might be sensitive to the small regions of high residual which RMS tends to average out. That is why the MAX option is there, to handle these cases.
Hi Glenn,

Thanks a lot. My system seems to have 2 steady states. I monitored the concentrations at some monitor points, and I reach a flat curve after 500 iterations and they stay at the same level for almost 5-6K more iterations, then they start slight depletion to a new SS, where the max RMS criteria will be met (after 14K iterations)

I am actually interested in the first SS, but I am a bit concerned about the max RMS not being reached although the monitor points concentration curves are flat.

I tried to switch to transient and to ensure that I get a converged solution at every timestep, but I got completely different results more similar to the final SS after 14 K after simulating my system for 70 sec and non of the transient solutions at smaller times are similar to the first SS.

January 21, 2018, 05:00
#5
Super Moderator

Glenn Horrocks
Join Date: Mar 2009
Location: Sydney, Australia
Posts: 17,728
Rep Power: 143
Quote:
 My system seems to have 2 steady states.
What you have described does not sound like 2 steady states. Rather, it sounds like you have a simulation with two time scales and the first time scale converges in about 500 iterations and the second time scale converges in about 14k iterations. Many complex simulations have multiple time scales so this is not unusual.

Often when convergence initial proceeds well, then levels off then starts converging again and goes to convergence; this is the sign of something from the inlet making its way to the outlet. It could be when the flow from the inlet finally reaches the outlet, or it could be that the heat from one side of a block finally reaches the other side.

Do not be fooled into thinking that the leveling off of the convergence at 500 iterations is the sign of anything meaningful. It is still just an unconverged simulation and should not be used for anything until your convergence criteria is achieved for ALL variables.

The normal way of dealing with large differences in time scales like this in a steady state simulation is to advance the different equations at different rates. For instance in fluid flow + solid heat transfer simulations usually the fluid domain converges quickly but the heat in the solid transfer is much slower. To help in these cases we use "Solid Timescale Factor" which is an acceleration factor for the heat equation. It can greatly accelerate convergence in the solid domain.

You need to work out which equation is converging fast and which is slow, and accelerate the slow one.

A final point: if you are interested in the transient response, then you need to run a transient simulation. You cannot use any acceleration then, you just have to run it until all the transients you wish to resolve (both the fast and slow ones) are completed.

 January 21, 2018, 10:59 #6 Member   Join Date: Jun 2017 Posts: 40 Rep Power: 8 Thanks a lot Glenn, I will work on that and update this post.