Hmm, convergence might be an issue. I didn’t check the command line output due to the relatively long computation time (size of screen buffer was too small to scroll back). Maybe I should redirect the output to a file in future.
My case makes use of a Fortran subroutine; I’ve attached the libraries for Linux and Windows, respectively. To calculate and plot the return loss you can use the script “plot_return_loss.r”.
Thanks a lot for looking into it!
Tom
Accuracy of results / microwave waveguide
Re: Accuracy of results / microwave waveguide
- Attachments
-
- wr28_waveguide_bandpass_filter.zip
- (784.59 KiB) Downloaded 352 times
Re: Accuracy of results / microwave waveguide
Dear Peter,
Have you had a chance to run the case?
Tom
Have you had a chance to run the case?
Tom
-
- Site Admin
- Posts: 4825
- Joined: 22 Aug 2009, 11:57
- Antispam: Yes
- Location: Espoo, Finland
- Contact:
Re: Accuracy of results / microwave waveguide
Sorry, not yet. Maybe first week of 2019. -Peter
-
- Site Admin
- Posts: 4825
- Joined: 22 Aug 2009, 11:57
- Antispam: Yes
- Location: Espoo, Finland
- Contact:
Re: Accuracy of results / microwave waveguide
Hi Tom,
I had a quick look.
It seems that the system (with lowest order edge elements) converges well without diagonal stabilization so I removed. Also removed nonlinear iteration. It is basically not harmful but in case your nonlinear tolerances are stricter than the linear ones, you may end up using resources in vain.
For the 2nd order edges I can confirm challenges in convergence. That would require some further work. So this could explain why the results with quadratic elements where not an improvement.
Didn't look at results but I think that the time spent is reduced quite a bit. Commented out some lines from f90 too, since they are not needed. Probably effect of that is very minimal.
-Peter
I had a quick look.
It seems that the system (with lowest order edge elements) converges well without diagonal stabilization so I removed. Also removed nonlinear iteration. It is basically not harmful but in case your nonlinear tolerances are stricter than the linear ones, you may end up using resources in vain.
For the 2nd order edges I can confirm challenges in convergence. That would require some further work. So this could explain why the results with quadratic elements where not an improvement.
Didn't look at results but I think that the time spent is reduced quite a bit. Commented out some lines from f90 too, since they are not needed. Probably effect of that is very minimal.
-Peter
- Attachments
-
- EMParam_WR28.f90
- modified f90 file
- (2.17 KiB) Downloaded 264 times
-
- Filter_Zhai_pr.sif
- modified sif file
- (3.54 KiB) Downloaded 312 times
Re: Accuracy of results / microwave waveguide
Hi Peter,
Happy new year and thanks you for looking into it! With your modified sif file computation time is indeed reduced a lot. For the second order edges, what in particular can I do to improve the results?
Tom
Happy new year and thanks you for looking into it! With your modified sif file computation time is indeed reduced a lot. For the second order edges, what in particular can I do to improve the results?
Tom
-
- Site Admin
- Posts: 4825
- Joined: 22 Aug 2009, 11:57
- Antispam: Yes
- Location: Espoo, Finland
- Contact:
Re: Accuracy of results / microwave waveguide
Hi Tom,
We should play around with the preconditioners. There are some ideas but currently no funding to try them out...
-Peter
We should play around with the preconditioners. There are some ideas but currently no funding to try them out...
-Peter
Re: Accuracy of results / microwave waveguide
With ILUT as preconditioner and relatively small values for the ILUT tolerance the system converges even for a more complex case than the given one. This comes at a cost of longer computation time. For the waveguide filter, Elmer's second order results are now on a similar level as the results from the commercial solver.
Tom
Tom
-
- Site Admin
- Posts: 4825
- Joined: 22 Aug 2009, 11:57
- Antispam: Yes
- Location: Espoo, Finland
- Contact:
Re: Accuracy of results / microwave waveguide
Hi Tom,
Nice results! Maybe you can share the updated comparison.
Can you confirm that the second order elements very not properly converging and with better convergence you can now compare favorably to the commercial code?
-Peter
Nice results! Maybe you can share the updated comparison.
Can you confirm that the second order elements very not properly converging and with better convergence you can now compare favorably to the commercial code?
-Peter
Re: Accuracy of results / microwave waveguide
Sure, here it is:
You can clearly see the difference by comparing the dashed line in figure 3 with the light orange line in figure 4. For the results in figure 4 I used your modified sif/f90 files, ILUT as preconditioner and a threshold level of 1.0e-4. Also, your modification enabled me to run the case in parallel -- that didn't work properly before.
Tom
You can clearly see the difference by comparing the dashed line in figure 3 with the light orange line in figure 4. For the results in figure 4 I used your modified sif/f90 files, ILUT as preconditioner and a threshold level of 1.0e-4. Also, your modification enabled me to run the case in parallel -- that didn't work properly before.
Tom
-
- Site Admin
- Posts: 4825
- Joined: 22 Aug 2009, 11:57
- Antispam: Yes
- Location: Espoo, Finland
- Contact:
Re: Accuracy of results / microwave waveguide
Hi Tom
That looks really good! I think this is the first time the 2nd order elements for the wave solver really showed its potential. Initially only the linear basis worked and the quadratic operation was fixed only rather recently.
-Peter
That looks really good! I think this is the first time the 2nd order elements for the wave solver really showed its potential. Initially only the linear basis worked and the quadratic operation was fixed only rather recently.
-Peter