Weird divergence behaviour

Numerical methods and mathematical models of Elmer
Post Reply
Andrea_P
Posts: 176
Joined: 22 Mar 2021, 18:39
Antispam: Yes
Location: POLYTHECNIC UNIVERSITY OF TURIN

Weird divergence behaviour

Post by Andrea_P »

Hi,
My simulation diverges after 900 iteration and I can't explain this weird behaviour. Until that moment, the result are good and physically correct. Have you got any idea of this behaviour? I have tried also to reduce the domain and to refine the mesh, nevertheless the divergence always occurs. I attached an image that shows the evolution of the profile and the convergence history with the case.sif in a zip folder. The actual mesh I am using is: https://we.tl/t-drNNRZxPlc
Regards,
Andrea
Attachments
DOMINIO_B=1_RIDOTTO.zip
(51.98 KiB) Downloaded 127 times
Immagine1.png
(57.37 KiB) Not downloaded yet
POLYTECHNIC UNIVERSITY OF TURIN-DIMEAS
Andrea_P
Posts: 176
Joined: 22 Mar 2021, 18:39
Antispam: Yes
Location: POLYTHECNIC UNIVERSITY OF TURIN

Re: Weird divergence behaviour

Post by Andrea_P »

I don't understand why the convergence occurs in this problem viewtopic.php?f=3&t=7507 and in the simulation, I am doing it doesn't.
-Andrea
POLYTECHNIC UNIVERSITY OF TURIN-DIMEAS
raback
Site Admin
Posts: 4801
Joined: 22 Aug 2009, 11:57
Antispam: Yes
Location: Espoo, Finland
Contact:

Re: Weird divergence behaviour

Post by raback »

Hi Andrea,

You've been working hard with a very complex case that would deserve some closer attention! Unfortunately I haven't had time to look it that closely. It is easier for some cases where you get an idea just by reading through the sif file. Convergence issues are more difficult.

Could you maybe classify a little to a post what all you have tried, what has been succesfull, where are your struggling and need the most help. There are quite a few cases on the forum now to choose from ;-)

-Peter
Andrea_P
Posts: 176
Joined: 22 Mar 2021, 18:39
Antispam: Yes
Location: POLYTHECNIC UNIVERSITY OF TURIN

Re: Weird divergence behaviour

Post by Andrea_P »

raback wrote: 20 Jul 2021, 01:46 Hi Andrea,

You've been working hard with a very complex case that would deserve some closer attention! Unfortunately I haven't had time to look it that closely. It is easier for some cases where you get an idea just by reading through the sif file. Convergence issues are more difficult.

Could you maybe classify a little to a post what all you have tried, what has been succesfull, where are your struggling and need the most help. There are quite a few cases on the forum now to choose from ;-)

-Peter
First of all, I tried to reduce the domain and refine the mesh, but these measures didn't work.
Secondly, I used the Mumps direct method, indeed an iterative method doesn't work (maybe due to the level set, I don't know honestly). I launched a paralleling simulation with Mumps. It worked but it diverged. So, I decided to keep on with Umfpack.
For a couple of days, I have been trying to reduce the relaxation factor(<=0.1) and using the Picard method, instead of Newton one. Apparently, the convergence behavior is pretty good, nevertheless, the obtained velocity profile is far from the exact one. Perhaps, a very low relaxation factor reduces the convergence velocity even if improving it.
Instead, with the relaxation factor of 0.1, the differences between the velocity computed by Elmer and the reference one were negligible. However, the divergence in this case occurred. I attach some images and case.sif just to show us the results. My uniform magnetic fields are B=0.5 T and B=1 T. I adopted a Gaussian magnetic field as the third example too. Thanks in advance for your help.
Regards,
Andrea
Attachments
casenewton_b_1_relax_0.1.sif
(6.27 KiB) Downloaded 126 times
casi_elmer.zip
(169.63 KiB) Downloaded 124 times
Immagine1.png
(60.83 KiB) Not downloaded yet
POLYTECHNIC UNIVERSITY OF TURIN-DIMEAS
Andrea_P
Posts: 176
Joined: 22 Mar 2021, 18:39
Antispam: Yes
Location: POLYTHECNIC UNIVERSITY OF TURIN

Re: Weird divergence behaviour

Post by Andrea_P »

I have noted that if I use either a relaxation factor <0.1 or the Picard method, the convergence of the problem gets really slow. Indeed the velocity profiles are pretty different from the ones computed in the reference paper.
-Andrea
POLYTECHNIC UNIVERSITY OF TURIN-DIMEAS
Post Reply