stripeline and frequency
The following is my 'qestions' as appeared under another topic "Difference between 2.5D and 3D simulation tools":
---------------
Among other things, Dr. Rautio must be referring to the stripline benchmark problem for which we can find interesting and impressive details on Sonnet's website. Anyone using EM solvers shouldn't miss that.
In the reference list (on Sonnet's website), there are two papers (letters in a column about MIC simulation in IJMMW) whose titles suggest some exchange with Zeland on the problem. I think that must be very interesting, but unfortunately I don't have access to those volumes of the Journal. I would deeply appreciate it if Dr. Rautio or some one else can post the relevant materials here.
I notice that EM3DS offers an example on simulation of thick stripline, which I don't see being discussed on Sonnet's website. Obviously the purpose of that example is to show the accuracy of EM3DS' 3D model. The conductor is however perfect, thus the currents stay on the surface and the reuslt of that example doesn't say too much about the 3D model. It would be very interesting to compare the results for a thick lossy stripeline between Sonnet and EM3DS.
Following Jim Rautio's suggestion we are moving to a nice new topic I deem very useful to anyone involved in the EM reality. I follow the last suggestion of Dr. Rautio and post my latest post in here. Not sure how relevant is that last post but I did not want to bloat the topic with my longer postings, yet, should there be a need I will repost my earlier comments on 2.5D versus 3D discussion as per our previous discussion.
My latest post follows:
__________________________________________________ ________________
Loucy, I agree with you that a good comparison would be to compare the thick lossy stripline and other structures with Em3DS, Sonnet and possibly others. I can run the Em3DS sim and present results here if someone is interested. but we have no access to Sonnet, hence, someone else should run that part.
Please let me know.
Best regards,
Cheng
P.S I can also run in 2.5D mode (em3ds) and compare too should that be of interest to anyone.
Let's start with Loucy's suggestion of the standard stripline. Attached is an Excel spreadsheet that has both single and coupled (on the second worksheet) stripline. There is an exact solution for lossless infinitely thin stripline which is calculated to about 8 decimal places. Zo (even and odd for the coupled case), S-parameters, and Z-parameters are calculated.
The idea is to analyze a stripline of known dimensions with a known exact result. Start with a coarse, fast meshing and then successively refine the meshing. Preferably cut the cell size in half each time. This can be done for both surface and volume meshing codes. The basic idea is described in
J. C. Rautio, "An Ultra-High Precision Benchmark For Validation Of Planar Electromagnetic Analyses," IEEE Tran. Microwave Theory Tech., Vol. 42, No. 11, Nov. 1994, pp. 2046-2050.
You can get it at IEEE Xplore, or I can email you copies of this or any papers I have authored, including those that Loucy mentioned. The above paper pretty much covers it all.
One of the things that is important to know is that most of the error in Zo is due to cell width. Cell length mostly affects error in velocity of propagation. However, there is some coupling between the two. Unless the cell length is kept really tiny, it can cancel error due to cell width. When this happens accidentally, the error convergence is not monotonic as you shrink cell width, it oscillates about the correct value. If an unethical researcher realizes this is happening, then they can selectively choose cell width/length combinations that take advantage of this. They then repoirt only the good ones and throw out the bad ones. I've actually seen this happen. If you see very large cell sizes reported as having very small error, this is probably what is happening.
To keep this from happening, always do a small cell length, then start with a large cell width and plot Zo as you keep cutting cell width in half. Alternatively, for a 50 Ohm line, you can plot S11 mag for a line exactly 1/4 wavelength long. In this case, S11 mag is about equal to the percent error in Zo.
As noted in the above paper, the velocity of propagation error tends to be proportional to the square of the cell length in terms of wavelengths, while Zo error tends to be proportional to the cell width in terms of the line width. This second item shows why the common statement that "cell size must be small with respect to wavelength" is insufficient.
Some ground rules if you want to present comparative data:
1) If you have a vested interest in the outcome (i.e., you work for a vendor, as I do), you are on your honor to tell us. Let's not play games on this one, folks.
2) You should state your relative skill level with each tool for which you present results. This can have a big effect on the outcome. Let's use three levels if its OK with everyone: 1) Beginner or occasional user, 2) Regular user, 3) Expert user.
3) State the processor (e.g., Pentium IV, AMD, etc.) and clock speed, assuming you post times.
4) Post all files needed to duplicate your results, at least for the coarsest meshing. Irreproducable results are not allowed!
5) Let's recognize that we might make mistakes and be open to feedback.
6) If someone starts getting angry or upset, then we should all recognize that we have found a weak spot that deserves serious investigation. I.e., let's keep the discussions rational and fact based. After all, we are engineers.
If you want to go beyond the infinitely thin stripline, I have attached a pdf on a thick lossless stripline test case. I have a spread sheet for that too, but it is not prettied up. Will hand it out on special request to very capable users.
There are many many other test cases. If there is interest, I can bring them up as we get each case work out.
The zero-thickness and thick stripeline geometries both support TEM mode. The accuracy of S11 for the fixed length lines bascially reflects the accuracy of Zo and propagation constant Bo. Some 3D solvers use 2D modal analysis to obtain these two values. The results are accurate with relatively coarse mesh (in terms total number of cells, some cells are quite small). By constrast, Sonnet and a number of other 2.5D (3Dplanar) tools solve 3D problems to obtain Zo and Bo. These latter class of tools obviously require more computer resource to obtain similar accuracy because the effect of higher order modes need to be captured and removed. If 2D modal analysis is used in a 2.5D solver, this solver will be faster.
So the stripeline benchmark problem mainly test the excitation and corresponding de-embedding scheme in an EM solver. It appears that Sonnet does a better job than most other MoM solvers in the thick stripeline case. But Sonnet's speed and accuracy for this problem are not so impressive when compared with 3D solvers with the idea of wave-port implemented. In addition, EM3DS seems to achieve similar results for the thick stripeline with much less number of cells (but not necessarily less memory) mainly because it uses a non-uniform mesh. In my opinion, this points out one weakness of Sonnet's thick conductor model-the sheets are uniformly spaced in the vertical direction, which is not necessary. Let us welcome Dr. Rautio's comments on why it has done so.
I know of no commercial MoM tool using the wave-port (2D modal analysis to determine the excitation). Please point it out if there is actually something out there. It would be intereting to run it on the stripeline problems.
Zeland's IE3D offers an option for port setting called "extension for wave", but doesn't appear to be using a 2D modal analysis. If anyone has exact information on this option, we would like to hear more about it.
The above is based on my experience using a number of EM solvers. I made no attempt to record the time, but I did plot the |S11| to observe the accurarcy. In some case (solver A vs solver B), the time difference is quite obvious and the reasons (as I perceived) have been described above. The story might be different for lossy thick stripeline.
Thanks for your comments Loucy. If what you want is Zo, then a modal 2-D analysis, such as used in the volume meshers is faster. However, you also have to ask which definition of Zo is being used, voltage current, voltage power, or power current. At low frequencies all give the same answer. For lossless stripline (homogeneous dielectric) all give the same answer. For microstrip at high frequency, all three give different answers. Sonnet uses a TEM equivalent Zo, i.e., it looks at the S-parameters of a length of line and reports what Zo corresponds to those S-parameters. Seems a lot more reasonable to me. It also gives a Zo(f) that corresponds to a causal system (no output before you give it an input), as will be published in a paper I presently have in the review cycle.
To test the volume meshers for their performance, rather than testing the incorporated 2-D Zo solver, compare S11 for a quarter wavelength line. Make sure it is exactly 1/4 wavelength long. An exact 50 Ohm line will have S11 exactly equal to zero. The calculated value will of course not be exactly zero. The voltage magnitude (not dB) will be very closely equal to the percentage error in Zo calculation. This assumes that the length of the cells or mesh is so small that the error in velocity of propagation is not important. For example, if mag S11 = 0.01, you have 1% error in Zo. So to test volume meshers, see how mag S11 converges. Things to look for are how much analysis time does it take to converge to a given level, and whether or not the convergence is smooth and uniform.
As for Sonnet's uniform meshing of thick metal (i.e., equal distribution of sheets through the thickness of the metal), that does require a lot of unnecessary work when you need to use a lot of sheets. Let's consider three cases:
1) Only one sheet is needed. This is the most common case, and should be used if at all possible. To test, get S-parameters with one sheet, then with two. See if the difference is large or small compared to requirements. For most cases, one sheet is enough as long as you have no gaps that are less than twice the thickness and no critical lines with width less than twice the thickness. There are exceptions, so when there is doubt, test your specific case. This advice is true for all EM analyses.
2) Two sheet model is the most common when thickness is needed. Again, if in doubt, test with one sheet, or four sheets. Usually needed when pushing past the limits described for one sheet above.
3) Multi-sheet is needed when gaps are << thickness. When multi-sheet is needed, you can get the accuracy of multi-sheet using two sheet, see
James C. Rautio, "A Space-Mapped Model of Thick, Tightly Coupled Conductors for Planar Electromagnetic Analysis," IEEE Microwave Magazine, Vol. 5, No. 3, September 2004, pp. 62 - 72.
and
David I. Sanderson, James C. Rautio, Robert A. Groves, and Sanjay Raman, "Accurate Modeling of Monolithic Inductors Using Conformal Meshing for Reduced Computation," IEEE Microwave Magazine, Vol. 4, No. 4 December 2003, pp. 87 - 96.
Available at IEEE Xplore, or I will email copies on request.
So, the multi-sheet model actually needs to be used in only releatively few cases and due to analysis time requirements should be avoided if at all possible, again this is true for all EM analyses.
I usually hate to go around all the time saying. "The problem will be solved in the next release," but in this case it is true. Sonnet's next release will have non-uniform distribution of sheets for thick metal. I have been helping with the solution myself. But again, I advise, for Sonnet and for all other EM analyses, do whatever you can to use the simplest model you can to get the results to the level of accuracy you need. Multi-sheet should only be an absolute last resort.
As for loss, you can get a good idea of the accuracy an analysis will have for loss by just looking at the current distribution. If you want to get the I2R loss right, you have to get the I right. The high edge current (where there is high loss) is most important to have accurately represented.
Also, just saying something is "accurate" really does not say very much. Everyone has a different, sometimes extremely different, idea of what accurate means. Something much more useful would be analysis time or mesh size required to converge to 1% error and 0.1% error.