7 Replies Latest reply on Jan 22, 2013 10:39 AM by Chris Michalski

    Steady state v. transient

    Chris Michalski

      Okay, got a question regarding "steady state" versus time dependent simulations.  My understanding of flow is that it solves every problem as time dependent.  If you set it as a time dependent it simply means that it will save data along the way.

       

      This is a simulation of a gas exiting a slot and impining a plate - colors represent concentrations of various gases on the impinged surface.

       

      In the attached, I have a video.  Frame 1 shows the result of a steady state simulation that was taken as the starting point (0seconds) of a time dependent simulation.  The entire video represents a time dependent 10 second simulation.  This suggests that my "steady state" may have been steady as far as the overall model (95% unshown in these images), but obviously not stabilized for this region.  This 10 seconds took about 2200 iterations.  And as it stabilizes about 2 seconds in I would expect under 1000 iterations to stabilize.

       

      Based on this, I added more goals locally to ensure this was accounted for in my steady state simulation.  I have captured screen shots of the steady state along the way recently.  There are oscillations at the ends of the slot which result in the wild goal graphs (as the data shows, the oscillations are in mm/sec).

       

      Assuming my opening statement (steady state is solved the same but only saves the end result), then why does continuing the steady state simulation for an extended period not produce the same result?  Why doesn't adding 2000 iterations to the end of a simulation produce the same end result as a transient simulation that consumes 2000 iterations?

       

      My problem is, actual testing verifies that the steady state solution presented is accurate.  So why might the time dependent simulation converge at a different condition?  And how do I know in the future which answer is accurate? 

       

      Has anybody else run into a discrepancy like this in any models?

        • Re: Steady state v. transient
          Mike Pogue

          Chris,

           

          The steady-state solver does not advance through the time-steps until the system comes to a steady state; this would be very wasteful. Instead, it sets the time dependent terms in the Navier Stokes equations to zero before setting up the systems of equations.

          The transient study, on the other hand, iterates until it converges in space, then steps in time, then iterates until it converges in space...

           

          The reason you get different answers, I'm guessing, is that there are fluctuations in the actual solution. These cannot be captured by the steady state solution, because they depend on time, and you have artificially imposed time-independence.

           

          Think of vortex shedding, or a flickering lighter. It is impossible to capture unsteady phenomena with a steady-state solution.

            • Re: Steady state v. transient
              Chris Michalski

              According to SW Flow documentation:

              "It is interesting to note that the Flow Simulation solver assumes that all analyses are transient.  For a "steady-state" analysis, the solver runs the transient analysis and looks for convergence in the flow field which would mean that the analysis has reached a steady-state."

               

              But if you watch a vortex shedding simulation as it iterates in a "steady-state" simulation you will see that oscillation and it will never converge.  That is more like what is happening at the ends of the slot where the curvatures exist, these are from oscillations.

               

              My time dependent animation shows that while these minor oscillations at the ends of the jet continually vary, the center location moving away from the slot stabilizes quite quickly.  This is my area of concern.  Why does the stable portion of a time dependent vary from the stable portion of a steady-state?

                • Re: Steady state v. transient
                  Mike Pogue

                  Well, then I'm curious too. Have you looked at the convergence?

                    • Re: Steady state v. transient
                      Chris Michalski

                      I didn't think to look at the convergence of the time dependent while it was running.  When I hit the 9000 iteration mark that I told it to continue up to now I'll open that configuration back up and take a look at that.

                       

                      In my screen captures of the steady state progression, you can see there is a bunch of oscillation in local goals (bottom of the images).  The picture below shows the local mesh body and highlights where those goals are.  The 2 highlighted ends are the left and right ends of the screen in the other images.  You are looking down from above - the thin line is at the middle of that tall rectangle. 

                      So I chose those 2 faces and added flow, concentration, velocity, (you name it) goals to verify those conditions stabilized.  They oscillate slightly, but the overall values are in mm/sec compared to the bulk flow in m/sec and they are not progressing, simply oscillating.  I'm wondering if the auto time steps chosen in the time dependent are too wide and cause the deviation?

                       

                      I pinged my former VAR (I had to drop our maintenance contract) to see how to get some feedback from them regarding what to change to improve my confidence in one answer or the other.

                      • Re: Steady state v. transient
                        Chris Michalski

                        Mike (or anyone else) -

                        how often do you do time dependent studies?  Most of my designs don't end up having fluctuations so I'm not very experienced at manual tuning of time steps.

                        Because it appears that about 1second in the time dependent study takes off like a rocket from the steady state, could it be a time step issue?  That Flow is adjusting the time step between iterations and the step gets so large that it loses the convergence that the steady state had developped?

                        I'm wondering if setting a defined time step instead of letting Flow determine it could bring the steady and transient in line.

                        Thoughts?

                          • Re: Steady state v. transient
                            Mike Pogue

                            Fluctuations are a sign of too long a time step in certain implicit schemes. But you would probably see fluctuations anyway if the underlying solution fluctuates, so I'm not sure that is the problem. If I remember the rule of thumb, you don't want material to travel more than 1 cell per time step.

                             

                            There is also spacial convergence, though. As you reduce the grid size, the difference between subsequent solutions should tend to zero. I sent a link for a straightforward method of accelerating the process. Sometimes there are model problems that cause the solution to diverge. If you had such a problem, your results would be bogus in both simulations, and you would be chasing a ghost trying to make them agree.

                              • Re: Steady state v. transient
                                Chris Michalski

                                I think I've found the culprit.  Joe Galliera pointed out that I should reduce the model size so I took advantage of symmetry and eliminated some needless complexity.  That change alone let the steady state simulation converge with respect to all goals (in under 1000 iterations).

                                 

                                Using that as the initial conditions I am running a time dependent, but I manually specified the time step of 2ms.  For the first 3 iterations it said it was too big, then it accepted 2ms.  Then later, it actually stepped back down to 1ms at times.  It is up to 4.2sec currently with the same visual appearance as the steady state solution. 

                                 

                                It would seem that given completely authority, Flow incremented the time steps to the point of losing convergence on the flow characteristics (I know I saw it jump more than 20ms at times before).  I am on my way to 6sec to be sure, but I believe that too large of a time step was to blame for the inconsistency.

                                 

                                Thanks for the feedback Mike.  I don't do much time dependent, but your comment about needing time steps small enough to make sure flow is registered in each and doesn't skip over cells makes sense.