ansoft sample circuit

Here's the question: let's say you have an arbitrarily shaped pcb trace on .030 rogers 4003. You want to characterize the trace (is it capacitive, inductive, etc and of what value).

In Ansoft Designer you have two options; you assign an edge port directly on the trace with "gap source", "port solver" and "post process port" checked (normalized to 50ohms) and run the simulation. Or, you can add a 50ohm track of indeterminate length to the unknown trace, run the same simulation again but this time "deembed" back the length of the 50ohm track. Technically speaking, you should get the same answer, but of course you don't (they're close). Ansoft recommends using a track that matches the port impedance to minimize modes and other artifacts, yet it seems the more correct answer is without the track. At low frequencies this doesn't matter, but at 5GHz and above, the differences become significant.

So:
(1) anyone have experience with this and can comment?
(2) does anyone know of calibrated 2D objects such as a precisely measured and dimensioned parallel plate capacitor that can be used to measure the software package against (dimensions, material, and test results posted)?

Hi smt -- There are very few planar structures that have exact solutions. However, there is one well known and understood one: lossless, infinitely thin stripline. Stripline is microstrip with an identical substrate and second ground plane above, as well as below. I posted a description it on this forum, just search for my name, rautio, and go to the posting titled EM Accuracy.

At that posting you will find a pdf describing use of a stripline to test any EM analysis. Using such you can determine which is the most accurate way to proceed with your work. This test will not catch error due to inaccurate loss calculation or error due to having multiple dielectric constants (since above and below the microstrip are the same dielectric) or error due to metal thickness. However, it will catch all other errors including error due to de-embedding accuracy. In otherwords, this test is a necessary but not sufficient condition for accuracy.

There is also an Excel spreadsheet at that posting that gives the exact solution for both single and coupled stripline.


Another test is described in:

Erik H. Lenzing and James C. Rautio, "A Model for Discretization Error in Electromagnetic Analysis of Capacitors," IEEE Transactions on Microwave Theory and Techniques, Vol. 46, No. 2, February 1998, pp. 162-166.

You can get it from IEEE Xplore, or I will email you a pdf on request.

This paper describes some work Erik and I did with a very detailed error analysis of a capacitor. You can use this test also with any EM tool you want. Volume meshers have trouble with this one because of the thin dielectric. It should be no problem for any planar tool.

More tests are available at http://www.sonnetsoftware.com/produc...nchmarking.asp.

I do want to frankly express concern that you are analyzing a planar circuit with a volume meshing tool. Such tools (any volume meshing tool) tend to be much much slower than any planar tool for planar circuits. Also, judging by looking at current distributions, the accuracy you can hope for is moderate at best. And finally, it has been a real awakening looking at threads on this forum of all the multitude of problems people are going through just to get simple planar circuits to work in volume meshers. It is amazing what people will tolerate!

If there is even a portion of your circuit that has a 3-D aspect to it (not including thickness), by all means, use a volume mesher. But for a planar circuit, you should almost always be using a planar tool. A list of vendors includes Sonnet (my company), Agilent, AWR, Zeland, Eagleware, and Ansoft. Perhaps you have the Ansoft Ensemble tool, if so use it, not HFSS for planar circuits. If you have no budget for another EM tool, just get our free SonnetLite (no time-out), www.sonnetsoftware.com. That will be plenty to handle simple circuits. As for time you need to invest, it takes about 5 minutes to install, and 45 minutes to go through the tutorial and start doing circuits. Most circuits analyze in a few seconds. Perhaps worth the effort, don't you think?

Dr. Rautio:
Ensemble has become part of Ansoft Designer, so smt is talking about a planar tool.

SMT:
can you tell us why you think "yet it seems the more correct answer is without the track"?

Rautio,

Thank you for the comments. I have access to IEEE trans; will take a look at the article and benchmarks.

Ansoft's Designer has a built in planar solver, yet regardless of the solver used (zeland, momentum, etc) they all have "issues" which if known beforehand, can be worked around. The problem is having a suitable reference to work against that is similar to the problem you wish to solve. What's not clear in Designer is the accuracy (if that's the correct term) of the deembeding function. Regardless of the length used for the ideal t-line (and of course there's absolutely no guidance as to what that length should be, other than long enough to attenuate higher order modes) the result should always be the same.

That is simply not the case.

Varying the length of the tline can vary the impedance of the unknown artifact by several pF or more - not trivial by any means. And measuring the capacitance of the artifact (in this case a component pad) whose dimensions are very small when compared to the wavelength and harmonics - well I just can't see how higher order modes even come into play. So when I see a difference between measurements using no transmission line and a line of arbitrary length that is deembedded, well I'm curious as to which is the more accurate result.

The goal here is to accurately measure component pads on various substrates and include them in our designs. Our frequency range is up to 10GHz. Any insights greatly appreciated.

loucy, regarding your question on which I think is the correct answer, I've noticed (and this is strictly and empirical observation) that the deembed function increases the capacitance of the pad with increasing transmission line length (recall that an ideal 50 ohm tline is added to the pad I wish to measure, and then an edge port is added to the 50ohm line. The measurement is made and the line deembedded back to the source). Not quite sure why this occurrs, but feel that the actual capacitance is lower.

Loucy -- Thanks for the info on Ensemble being merged into HFSS, I had heard that but did not realize people would refer to Ensemble as HFSS, in addition to the finite element solver. Just a bit confusing when the same name is used for two different things.

smt -- Checking convergence as a function of de-embedding length is an important test many people miss. Your effort and concern is well placed. In our case (Sonnet) we find on the short end, things fail if the de-embedding length is less than 1 substrate thickness, for really high accuracy it should be longer than 2 substrate thicknesses, in most typical geometries.

On the long side, it fails when it is so long that there are box resonances. For unshielded analysis, the corresponding failure would presumably be when radiation kicks up. The de-embedding works to full accuracy even when the length is a multiple of a half wavelength long, however, the Zo and Eeff values printed out are not valid.

In between these two lengths, things are rock solid.

If you use a length of stripline as your exact standard, you will at least know precisely what the error is when you get a number, you won't have to guess which result is better. Should be able to do the same thing with the capacitors described in my paper, they are basically just the parallel plate capacitance plus a tiny bit of fringing. Never tried using those as standards, however.

Just a caution I'm sure you will think of: When you settle on an "accurate" length for one known structure, make sure that that same length gives you just as accurate an answer for a second, different, known structure.

Also, be sure to check convergence as cell size gets smaller. If it is working properly, you can successively cut the cell/mesh size in half and plot the result. Each time the error will cut in half (in most cases) and it is easy to extrapolate to almost the exact answer. This is called Richardson extrapolation. You can also use the differences to get a good solid estimate of the remaining error.

For accurate S-parameters of pads (presumably for SMT?), you probably should consider using the GLG de-embdedding that I just published. GLG stands for General Local Ground. GLG can de-embed a local ground hanging anywhere in your circuit and, to within the precision of the underlying EM analysis, it is exact. It was designed for SMT on top of LTCC (Low Temp Cofired Ceramic), but should work just as well for FR4. See the Feb. 2005 issue of the MTT Transactions, third to last paper.

Good luck!

rautio,

good info on the dembedding function. Will definitely check out your paper in the MTT Trans.

All this brings me to the next logical question, and feel free to comment on the sonnet package.

Must every pad we wish to characterize first be terminated in a 50ohm tline which then is deembedded back to the pad? Seems like a lot of work when the layout has already been imported into the EM solver. I should just (and this is I'm sure an over-simplification) be able to pick any pad I want, add the appropriate edge port and run the solver. Does Sonnet require the additional t-line, or can I just import the layout and "Pick a Pad" to measure.

Again, greatly appreciate the help.

Hi SMT -- To demonstrate, I went ahead and did an example circuit, attached. Took about 20 minutes end-to-end. Analysis times ranged from 1 second / freq to 2 seconds / freq. I used a Pentium M 1.3 GHz Notebook.

SonnetLite (free at www.sonnetsoftware.com, no timeout on the software) allows up to four ports. So I put in four pads.

You most certainly do NOT want to use 50 Ohm lines to connect the pads to edge ports, unless you are using the same 50 Ohm lines in the actual circuit. You should use lines that are the same width as what you are going to use in the actual circuit. The reason is that part of the pad capacitance is the capacitance from the pad to the line, this step discontinuity must be included in the analysis. You can have the same pad size, but if the connecting lines are of different width or connect at different places, you will get different capacitances. Whether this is important or not depends on your specific geometry and requirements.

In this case, I made the pads 1 mm square and the lines 0.2 mm wide, and the connecting lines 3.5 mm long. I set the reference planes right at the pads (see image below). The lines up to the reference plane are automatically removed by SonnetLite. De-embedding removes only the transmission line mode, the fringing fields to the pad are left in the data.

I did three analyses. The first de-embedded 3.5 mm connecting lines, the second de-embedded 7 mm lines, and the third de-embedded 14 mm lines.

Then I also enabled the Sonnet SPICE PI model synthesis (Analysis->Output Files). Very easy in this case , but for some circuits it can be tricky, be sure to read the manual. Only thing special I did in this case was set the lower limit on capacitors from 0.01 pF to 0.001 pF. Turns out the capacitance from pad 1 to pad 4 is about 0.004 pF.

The SPICE synthesis takes data at two frequencies and synthesizes an equivalent PI network with ports acting as nodes. I analyzed .5 to 5 GHz step 0.5 GHz, it made a model at each pair of frequencies.

Full data is in the zip file in *.lib.txt files. Some sample data is below.

Reference plane length 3.5mm:

* Analysis frequencies: 500.0, 1000.0 MHz
.subckt FourPads_0 1 2 3 4 GND
C_C1 1 GND 0.080566pf
C_C2 1 2 0.013731pf
C_C3 1 3 0.013749pf
C_C4 1 4 0.004496pf
C_C5 2 GND 0.080566pf
C_C6 2 3 0.004496pf
C_C7 2 4 0.013749pf
C_C8 3 GND 0.082386pf
C_C9 3 4 0.013951pf
C_C10 4 GND 0.082386pf
.ends FourPads_0

Reference plane length 7.0 mm:

* Analysis frequencies: 500.0, 1000.0 MHz
.subckt FourPads_2xLen_0 1 2 3 4 GND
C_C1 1 GND 0.079625pf
C_C2 1 2 0.013846pf
C_C3 1 3 0.014074pf
C_C4 1 4 0.00474pf
C_C5 2 GND 0.079625pf
C_C6 2 3 0.00474pf
C_C7 2 4 0.014074pf
C_C8 3 GND 0.08131pf
C_C9 3 4 0.014173pf
C_C10 4 GND 0.08131pf
.ends FourPads_2xLen_0

The substrate thickness is 2 mm, so the 3.5 mm reference plane is a little less than 2x substrate thickness which introduces some error. Thus there is a small difference between the two models. I also did 14 mm reference planes (included in the zip file), and the model is nearly identical to the 7 mm reference plane.

When I look at how the model changes with frequency, the above model error is < about 2% up to 2 GHz, and < 10% up to 5 GHz.

The GLG de-embedding I mentioned prevously is available in the full Sonnet on special request. It is not in SonnetLite.

One thing to watch out for is de-embedding closely coupled ports, as we have here. It is no problem in Sonnet. To check other EM analyses, just de-embed a closely coupled 4-port coupled line to zero length and look at the result. Some work, some don't. I don't know what Ensemble does for this case.

OK, I definitely need to digest what you just did and try it in sonnet lite.

thanks