I need help to know if converged...

L3munoz · June 24, 2009, 04:24

Hi everybody!

I was running a simulation and i got the following residual and Cd monitors. According to them, and from my point of view, I guess the simulation isnt yet converged, but after 10000 iterations I think it's already enough, no?

I know that number of iterations of course depends on the number of cells, the viscosity model and so on, but it's a huge number, no? I give you here my simulation data and I'd appreciate a lot if you could give your advice about the convergence, the residuals and the CD value...Should I still wait??

http://i723.photobucket.com/albums/ww240/L3munoz/Residuales.jpg
http://i723.photobucket.com/albums/ww240/L3munoz/CD.jpg

# of cells: around 6 millions.
Viscous model: k-e realizable
Discretization order: 2nd (first simulation on 1st order)
Under-relaxation factors: I decreased them in order to get a smoother convergence (0.3 0.8 0.8 0.6 0.4 0.4 0.8)

I also include the data of the last iteration:
iter: 10255
continuity: 2.0676e-03
x-velocity: 9.8457e-05
y-velocity: 9.4748e-05
z-velocity: 4.5349e-05
k: 1.8427e-04
epsilon: 4.3215e-04
cd: 6.7836e-02
time/iter: 64:46:01
turbulent viscosity limited to viscosity ratio of 1.000000e+05 in 82139 cells

Thanks in advance,

Jorge.

nstar · June 24, 2009, 09:50

I don't think it's converged.
In terms of residuales, sometimes there's no much thing you can do if it does not drop to 10^-3 (or -4, -5). Smaller relaxation factor gives you a smoother residuales but slows down the convergence.
However, look at the Cd you are monitoring, it's certainly not converged. I'd say when the Cd looks pretty stable it's converged.

L3munoz · June 25, 2009, 07:09

Hi nstar,

thanks a lot for replying my post. I was wondering the same, so, as you say and as I was thinking, I'll keep the simulation running...But, from ur point of view, how long should it take to converge?? Dont u think 10000 iterations are already a lot even for a # of cells of 6 millions?? which is the reason?? For me the problem is the relaxation factors, but...

and about still having "turbulent viscosity limited to viscosity ratio of 1.000000e+05 in 82139 cells" is it really a problem??

Jorge.

nstar · June 25, 2009, 10:01

Hi,
I really can't say how long it's gonna take to converge, but looks at your Cd monitor it seems to be close. It sometimes depends on your geometry and mesh quality as I understand. for example, in region with mixing flow you may want to have fine dense hex volume mesh. It also depends on what kind of model, i.e., turbulence model, etc., you are using.
6M mesh is quite a large model to simulate. For the turbulence viscosity limit, well, I have been seeing this too many times, normally as long as the results look o.k, I just ignore it since I don't know what's the exact reason and how to deal with that. LOL.

Quote:

Originally Posted by L3munoz

Hi nstar,

thanks a lot for replying my post. I was wondering the same, so, as you say and as I was thinking, I'll keep the simulation running...But, from ur point of view, how long should it take to converge?? Dont u think 10000 iterations are already a lot even for a # of cells of 6 millions?? which is the reason?? For me the problem is the relaxation factors, but...

and about still having "turbulent viscosity limited to viscosity ratio of 1.000000e+05 in 82139 cells" is it really a problem??

Jorge.

mystic_cfd · June 25, 2009, 15:53

the oscillations on your residuals plot indicate that a part or parts of your simulation domain has periodic (transient) flow. the solver is trying to find an equilibrium (steady state) which does not exist. this is often the case when you discretize a domain with a fine mesh as you can easily start to resolve phenomena like vortex shedding off some corner or edge.

an option is to increase underrelaxation factors again - low under ralaxation factors (URFs) will result in very slow convergence.

always remember to check normalized convergence of the key parameter. i usually take the latest value of the parameter (e.g. Cd for you) and divide the series by that value to to see what the percentage change is from one iteration to the next.

L3munoz · June 28, 2009, 09:00

Nstar: Hi! Yes, it seems CD is going close to the end...I checked the last 900 iterations and it is oscillating among an almost constant intervale...However I'll wait a little more...About what u say about geometry and mesh quality I tried to prepare a mesh as hex-volume meshed as possible, but obviously close to the train and in certain parts of it was impossible...Do u think changes from meshing types affect the convergence? I mean, trying to keep it with as many hexahedra as possible I did several changes, passing from tetrahedra to hexahedra and so. And also I have some interfaces... About the turbulent viscosity, jajaj, good answer!

I hope someone could give us a better answer, although I guess looking for it in the forum we'll find it. Thanks a lot!

mystic_cfd: thank u very much for replying! I decreased them, yes, in order to smooth the converge as much as possible, but convergence maybe is too slow...So I guess I'll have to get a compromise between them. Do u have any idea of how define them? Good idea also this of checking convergence by the main parameter, but what do u mean dividing the series?

June 24, 2009, 04:24	I need help to know if converged...	#1
L3munoz Member Jorge Muñoz Join Date: Apr 2009 Posts: 56 Rep Power: 17	Hi everybody! I was running a simulation and i got the following residual and Cd monitors. According to them, and from my point of view, I guess the simulation isnt yet converged, but after 10000 iterations I think it's already enough, no? I know that number of iterations of course depends on the number of cells, the viscosity model and so on, but it's a huge number, no? I give you here my simulation data and I'd appreciate a lot if you could give your advice about the convergence, the residuals and the CD value...Should I still wait?? http://i723.photobucket.com/albums/ww240/L3munoz/Residuales.jpg http://i723.photobucket.com/albums/ww240/L3munoz/CD.jpg # of cells: around 6 millions. Viscous model: k-e realizable Discretization order: 2nd (first simulation on 1st order) Under-relaxation factors: I decreased them in order to get a smoother convergence (0.3 0.8 0.8 0.6 0.4 0.4 0.8) I also include the data of the last iteration: iter: 10255 continuity: 2.0676e-03 x-velocity: 9.8457e-05 y-velocity: 9.4748e-05 z-velocity: 4.5349e-05 k: 1.8427e-04 epsilon: 4.3215e-04 cd: 6.7836e-02 time/iter: 64:46:01 turbulent viscosity limited to viscosity ratio of 1.000000e+05 in 82139 cells Thanks in advance, Jorge.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Solution converged, and then diverge again.	jrg	FLUENT	2	April 1, 2013 23:10
How to import an existing converged soultion into a similar refined mesh?	Sophie	FLUENT	1	March 22, 2009 22:31
how to identify transient simulation converged	littlelz	CFX	5	January 27, 2009 18:13
When is it converged?	Pascal	Main CFD Forum	4	February 8, 2008 04:58
Calculate a converged flow solution	kumar	FLUENT	2	May 24, 2007 01:33

June 24, 2009, 09:50		#2
nstar New Member Ling Join Date: Jun 2009 Posts: 25 Rep Power: 16	I don't think it's converged. In terms of residuales, sometimes there's no much thing you can do if it does not drop to 10^-3 (or -4, -5). Smaller relaxation factor gives you a smoother residuales but slows down the convergence. However, look at the Cd you are monitoring, it's certainly not converged. I'd say when the Cd looks pretty stable it's converged.

June 25, 2009, 07:09		#3
L3munoz Member Jorge Muñoz Join Date: Apr 2009 Posts: 56 Rep Power: 17	Hi nstar, thanks a lot for replying my post. I was wondering the same, so, as you say and as I was thinking, I'll keep the simulation running...But, from ur point of view, how long should it take to converge?? Dont u think 10000 iterations are already a lot even for a # of cells of 6 millions?? which is the reason?? For me the problem is the relaxation factors, but... and about still having "turbulent viscosity limited to viscosity ratio of 1.000000e+05 in 82139 cells" is it really a problem?? Jorge.

June 25, 2009, 15:53		#5
mystic_cfd New Member HP Kritzinger Join Date: May 2009 Location: South Africa Posts: 25 Rep Power: 16	the oscillations on your residuals plot indicate that a part or parts of your simulation domain has periodic (transient) flow. the solver is trying to find an equilibrium (steady state) which does not exist. this is often the case when you discretize a domain with a fine mesh as you can easily start to resolve phenomena like vortex shedding off some corner or edge. an option is to increase underrelaxation factors again - low under ralaxation factors (URFs) will result in very slow convergence. always remember to check normalized convergence of the key parameter. i usually take the latest value of the parameter (e.g. Cd for you) and divide the series by that value to to see what the percentage change is from one iteration to the next.

June 28, 2009, 09:00		#6
L3munoz Member Jorge Muñoz Join Date: Apr 2009 Posts: 56 Rep Power: 17	Nstar: Hi! Yes, it seems CD is going close to the end...I checked the last 900 iterations and it is oscillating among an almost constant intervale...However I'll wait a little more...About what u say about geometry and mesh quality I tried to prepare a mesh as hex-volume meshed as possible, but obviously close to the train and in certain parts of it was impossible...Do u think changes from meshing types affect the convergence? I mean, trying to keep it with as many hexahedra as possible I did several changes, passing from tetrahedra to hexahedra and so. And also I have some interfaces... About the turbulent viscosity, jajaj, good answer! I hope someone could give us a better answer, although I guess looking for it in the forum we'll find it. Thanks a lot! mystic_cfd: thank u very much for replying! I decreased them, yes, in order to smooth the converge as much as possible, but convergence maybe is too slow...So I guess I'll have to get a compromise between them. Do u have any idea of how define them? Good idea also this of checking convergence by the main parameter, but what do u mean dividing the series?