Turbulent flow at high Reynolds number  Convergence problem
Re: Turbulent flow at high Reynolds number  Convergence problem
@Matthias : Yes, that's what I did, of course...
I'm just reporting this "bug"...
Anyway, the main objective of this thread is to discuss about the solver settings and configuration in view to obtain convergence of turbulent, relatively high Reynolds number flows.
I've found these examples (see previous post) and I am still looking for some advices about the solver settings.
I am quite perplexed with the choices of the users. I progress slowly and, even I am pretty sure that it is possible to reach convergence of quite high Reynolds number cases with Elmer, I miss information (even in the manuals) on how to set the solver configuration.
In the 2D example upper (lost exhaust), I was able to increase the outlet velocity without any problems after suppressing the Newton method (of course).
But the treatment of a 3D case is still a problem... and I don't understand the choices of the solver settings.
It is also the case for the diffuser_sst model and I am really looking for help to understand the choices made for the solver settings.
If someone has already tested some cases at higher air inlet velocities and obtain a good convergence, I will really appreciate the help.
I think it is very interesting to be able to use Elmer (as stand alone sotware) in multiphysics with a good simulation of turbulent flow and I am sure that several users have reach a good convergence with this software.
If there is a good explanation of solver settings for turbulent flows somewhere, I am also very interested.
Thanks again to all for your help,
Best Regards
I'm just reporting this "bug"...
Anyway, the main objective of this thread is to discuss about the solver settings and configuration in view to obtain convergence of turbulent, relatively high Reynolds number flows.
I've found these examples (see previous post) and I am still looking for some advices about the solver settings.
I am quite perplexed with the choices of the users. I progress slowly and, even I am pretty sure that it is possible to reach convergence of quite high Reynolds number cases with Elmer, I miss information (even in the manuals) on how to set the solver configuration.
In the 2D example upper (lost exhaust), I was able to increase the outlet velocity without any problems after suppressing the Newton method (of course).
But the treatment of a 3D case is still a problem... and I don't understand the choices of the solver settings.
It is also the case for the diffuser_sst model and I am really looking for help to understand the choices made for the solver settings.
If someone has already tested some cases at higher air inlet velocities and obtain a good convergence, I will really appreciate the help.
I think it is very interesting to be able to use Elmer (as stand alone sotware) in multiphysics with a good simulation of turbulent flow and I am sure that several users have reach a good convergence with this software.
If there is a good explanation of solver settings for turbulent flows somewhere, I am also very interested.
Thanks again to all for your help,
Best Regards

 Posts: 31
 Joined: 17 Jun 2015, 10:04
 Antispam: Yes
Re: Turbulent flow at high Reynolds number  Convergence problem
Dear All,
I have been experimenting with Elmer for some time to model a few different turbulent flows and have gathered a few observations.
(1) In my experience, Elmer is quite good at modeling 2D turbulent flows using Navierstokes and kepsilon solvers, even in the case of moderate to high turbulent behavior in complex geometries. Convergence problems arise, but they can be mitigated by the standard techniques viz., using 1 nonlinear iteration per steady state iteration cycle, using low values (< 0.3) of nonlinear relaxation factors, employing BiCGStabl solver with appropriate preconditioning etc. An example is discussed below.
(2) It is almost always the 3D problems that lead to convergence issues or crash, especially in complex geometries. In my test cases, it is usually the kepsilon model that blows up first, followed by NS. I am not sure if this has to do with the discretization and stabilization techniques in Elmer or if it is an inherent FE limitation.
Attached is an example of a 2D test case of a recirculatory flow induced by a narrow jet in a rectangular tank. This is a reasonably complex turbulent flow with jet mixing and expansion, shear layers, recirculation etc. Water enters the rectangular tank (2.5 m x 2 m) through a central slot nozzle of 0.1 m width at the right and exits through an identical nozzle at the left. Due to symmetry, only one half of the domain is simulated.
Kumar
I have been experimenting with Elmer for some time to model a few different turbulent flows and have gathered a few observations.
(1) In my experience, Elmer is quite good at modeling 2D turbulent flows using Navierstokes and kepsilon solvers, even in the case of moderate to high turbulent behavior in complex geometries. Convergence problems arise, but they can be mitigated by the standard techniques viz., using 1 nonlinear iteration per steady state iteration cycle, using low values (< 0.3) of nonlinear relaxation factors, employing BiCGStabl solver with appropriate preconditioning etc. An example is discussed below.
(2) It is almost always the 3D problems that lead to convergence issues or crash, especially in complex geometries. In my test cases, it is usually the kepsilon model that blows up first, followed by NS. I am not sure if this has to do with the discretization and stabilization techniques in Elmer or if it is an inherent FE limitation.
Attached is an example of a 2D test case of a recirculatory flow induced by a narrow jet in a rectangular tank. This is a reasonably complex turbulent flow with jet mixing and expansion, shear layers, recirculation etc. Water enters the rectangular tank (2.5 m x 2 m) through a central slot nozzle of 0.1 m width at the right and exits through an identical nozzle at the left. Due to symmetry, only one half of the domain is simulated.
Kumar
 Attachments

 2DJetMixingTank.tar.gz
 (248.88 KiB) Downloaded 33 times

 velElmer.png
 (33.4 KiB) Not downloaded yet
Re: Turbulent flow at high Reynolds number  Convergence problem
Dear Kumar,
Thanks a lot to your very detailed post. I really appreciate...
Indeed, my last tests showed me that it was possible, under some conditions (especially what you specify about the solver (I based my work on the linked test case in one of my previous posts)) to reach relatively high Reynolds number with Elmer's turbulent models (there komega SST). But all cases were in 2D, indeed...
I will try to evaluate a simple 3D case, in the same solver conditions, and I will come back to you (and other Elmer users) to discuss about the results.
Thanks again for your very relevant post.
Paul
Thanks a lot to your very detailed post. I really appreciate...
Indeed, my last tests showed me that it was possible, under some conditions (especially what you specify about the solver (I based my work on the linked test case in one of my previous posts)) to reach relatively high Reynolds number with Elmer's turbulent models (there komega SST). But all cases were in 2D, indeed...
I will try to evaluate a simple 3D case, in the same solver conditions, and I will come back to you (and other Elmer users) to discuss about the results.
Thanks again for your very relevant post.
Paul

 Posts: 31
 Joined: 17 Jun 2015, 10:04
 Antispam: Yes
Re: Turbulent flow at high Reynolds number  Convergence problem
You are most welcome Paul! Perhaps we can learn something together.
As I said before, I never had any serious issues with 2D problems. The solver settings in the above example almost always worked well with NS and kepsilon (also with sstkomega).
It would be very nice if we could identify a couple of benchmark turbulent 3D cases, solve them with Elmer, compare it with OpenFOAM and try to arrive at the optimal solution strategies with Elmer Solver. I believe this might also be very useful for a lot of Elmer users.
Thanks and have a nice day.
Kumar
As I said before, I never had any serious issues with 2D problems. The solver settings in the above example almost always worked well with NS and kepsilon (also with sstkomega).
It would be very nice if we could identify a couple of benchmark turbulent 3D cases, solve them with Elmer, compare it with OpenFOAM and try to arrive at the optimal solution strategies with Elmer Solver. I believe this might also be very useful for a lot of Elmer users.
Thanks and have a nice day.
Kumar
Re: Turbulent flow at high Reynolds number  Convergence problem
Dear Kumar, dear all,
I will try a simple case : an air flow in an elbow with free outlet.
I have already tried a 2D case and I had no problem to converge with an air velocity inlet of 20 m/s. I will now test higher velocities.
The solver parameter were about those you mentionned upper in your example.
I have also tried a 3D case (the 2D case has the same geometry) and I wasn't able, with the same solver parameter, to reach higher air inlet velocity than 0.3 m/s, even by refining the mesh.
For higher velocities, the simulation didn't converge, even by changing some solver parameters.
I have obtained slight improvements by adjusting the initial conditions but this solution will not be possible for more complex cases... Anyway, there's no convergence.
I will now try to test various solver parameters combination in view to converge for higher velocities...
May be a scanning simulation, or a restart from a converged solution may be also the solution ?
I will post the case files later, when I will complete these tests.
I will try a simple case : an air flow in an elbow with free outlet.
I have already tried a 2D case and I had no problem to converge with an air velocity inlet of 20 m/s. I will now test higher velocities.
The solver parameter were about those you mentionned upper in your example.
I have also tried a 3D case (the 2D case has the same geometry) and I wasn't able, with the same solver parameter, to reach higher air inlet velocity than 0.3 m/s, even by refining the mesh.
For higher velocities, the simulation didn't converge, even by changing some solver parameters.
I have obtained slight improvements by adjusting the initial conditions but this solution will not be possible for more complex cases... Anyway, there's no convergence.
I will now try to test various solver parameters combination in view to converge for higher velocities...
May be a scanning simulation, or a restart from a converged solution may be also the solution ?
I will post the case files later, when I will complete these tests.
Re: Turbulent flow at high Reynolds number  Convergence problem
Hi guys!
Here is my experiments on Kumar's case with triangular 2D mesh and 3D tetrahedral mesh.
2D mesh:
https://drive.google.com/open?id=1t1cWZ ... wyYMrz7wG9
All is OK.
3D mesh:
https://drive.google.com/open?id=1bm07Z ... J0rQlEZKmO
God damn hard... One have to decrease Relaxation factor to 0.05 (!). Convergence od KE stops on RELC = 0.08.
Conclusion: High dependence of KE solver on mesh quality, geometry and initial state.
Here is my experiments on Kumar's case with triangular 2D mesh and 3D tetrahedral mesh.
2D mesh:
https://drive.google.com/open?id=1t1cWZ ... wyYMrz7wG9
All is OK.
3D mesh:
https://drive.google.com/open?id=1bm07Z ... J0rQlEZKmO
God damn hard... One have to decrease Relaxation factor to 0.05 (!). Convergence od KE stops on RELC = 0.08.
Conclusion: High dependence of KE solver on mesh quality, geometry and initial state.
 Attachments

 20180725_163102.png (29.69 KiB) Viewed 282 times

 20180725_163031.png (22.42 KiB) Viewed 282 times

 20180725_162853.png (104.47 KiB) Viewed 282 times

 Posts: 31
 Joined: 17 Jun 2015, 10:04
 Antispam: Yes
Re: Turbulent flow at high Reynolds number  Convergence problem
Dear All,
What Dmitry remarked is what exactly what I had observed too. In 3D flows, the relaxation factors of both solvers have to be reduced so low (< 0.1) that they are no longer usable.
I noticed that in 3D cases, the k and epsilon solutions have a tendency to oscillate more wildly than in 2D domains. Sometimes k and epsilon values reach so large values that makes no sense and thereafter the linear system starts to blow up. In the solver code, there are lower limits (I think 1.0e10 or so) on k and eps values to prevent them going negative.
Perhaps putting some upper limits on k and eps would be useful? The other thing I noticed is that only Bubble stabilization is implemented in the solver code. This almost doubles the solver time. It would be interesting to see if the standard stabilization does a better job. Perhaps some developer can comment on this.
If just one of the RANS turbulence solvers (say KESolver) is made more robust, Elmer would be suitable for a much wider range of industrial applications, esp. those involve CFD and FEA.
What Dmitry remarked is what exactly what I had observed too. In 3D flows, the relaxation factors of both solvers have to be reduced so low (< 0.1) that they are no longer usable.
I noticed that in 3D cases, the k and epsilon solutions have a tendency to oscillate more wildly than in 2D domains. Sometimes k and epsilon values reach so large values that makes no sense and thereafter the linear system starts to blow up. In the solver code, there are lower limits (I think 1.0e10 or so) on k and eps values to prevent them going negative.
Perhaps putting some upper limits on k and eps would be useful? The other thing I noticed is that only Bubble stabilization is implemented in the solver code. This almost doubles the solver time. It would be interesting to see if the standard stabilization does a better job. Perhaps some developer can comment on this.
If just one of the RANS turbulence solvers (say KESolver) is made more robust, Elmer would be suitable for a much wider range of industrial applications, esp. those involve CFD and FEA.

 Posts: 31
 Joined: 17 Jun 2015, 10:04
 Antispam: Yes
Re: Turbulent flow at high Reynolds number  Convergence problem
Hi All, Here is the result from a hexahedral mesh.