I reran the probe dimension sensitivity analysis again with a much finer attention to detail and the results look good. The first thing I did was to bump up the number of tetrahedrals to around 400K and also narrowed the range simulated to just around the TE_{0,1} mode, 2.435Ghz to 2.45Ghz (15Mhz spread).

Quick summary – A probe at 28mm in diameter is ideal

The results I sent around before with the 4dB insertion loss had a well meshed model but did not have enough samples in a tight enough frequency spread. One realization over the past few days was that there are two types of precision in EM simulations, the first is the number tetrahedrons and the second is the number of samples. I will describe the difference in the next two paragraphs.

When CST builds a model, the first type of precision is how many tetrahedral are necessary to accurately simulate the model at a single given frequency. The larger the space or the higher the frequency, requires more tetrahedrals. The primary method I use to control the number of tetrahedrals is by specifying the number of steps per wavelength, currently seven (default is 4). I also increase the “curvature refinement ratio” to 0.25 (from 0.5) and the “Max number of steps from curvature ref:” to 150 (from 100) because our cavity is entirely made out of round parts. The end result is that for the cavity which is 280mm in diameter and 390mm long simulated at 2.45Ghz, CST requires about 400K of tetrahedrals to meet an accuracy of 1% (it usually stops mesh refinement at 0.5%). The trick is to get as much accuracy as possible without running out of memory or making the simulations take days. That many tetradedrals uses about 4GB of RAM when simulating.

The second type of precision is that, depending on the width of the frequency range, CST runs the simulation at as many separate frequencies as necessary to make the interpolation of points on the S-parameter graph smooth. In the S-Parameter plots below, the frequency width is only 15Mhz and CST sampled 15 different frequencies, each running with about 440K tetrahedrals. In the previous S-parameter graph (showing 4dB insertion loss), the frequency range was 300Mhz and I stopped the number of samples at 60.

In short, to get an accurate S-parameter graph, it is necessary to both maximize the number of tetrahedrals and the number of samples in a tight a frequency range as possible.

Below are S-parameter graphs where I varied the probe diameter from 30mm (on the left, as they are currently constructed in the cavity) to 26mm (on the right).

It is clear from the graphs that the sweet spot is at 28mm with an S_{1,1} of -40dB and an insertion loss of -0.5. It is also clear that having a probe a millimetre on either side will mean a drop of about 20dB for S_{1,1}.

While the simulations were running, I also went back and have started putting all status updates going back three years (to 2009) into a blog you can get to here. There are currently about sixty status reports up, but I should have them all up by the end of the week.

My next few tasks will be to:

- Find a way to built mm accurate probes out of copper. The good news is they are small and should be fairly inexpensive to get made.
- Add the rest of the status updates to the blog.
- I will also run a monster simulation with the probe at 28mm, across a large frequency range that will probably take a day or so (likely running into hundreds of sample frequencies) . With the resulting S-parameter chart, I can then compare it against the measured S-parameter chart and quickly find the TE
_{0,1}mode.