I spent the afternoon/evening finishing building the temporary breadboard version of the PDH2 and testing it. As a reminder, this is necessary because we want to get a new, low gyro noise spectrum before the LV meeting for the poster/talk and we won't have time to get the PCB done before then.
Here she is in all her glory:
I spared no expense; she's equipped with all four (DIP-switchable) boosts, as well as an invert and a switch-engaged SWEEP input.
I originally built it with all AD829s as we are planning (?) on doing for the final servo, but I had problems with oscillations. I followed the instructions for shunt compensation capacitors in the datasheet, but Frank and I found that there were unavoidable problems from stray capacitances and inductances in the push-in components and the board. We systematically replaced each stage with an OP27 from the input to the output until we had no more issues. We don't expect to see the electronics noise floor at the moment, so the difference between 1.7 nV/rHz and 3 nV/rHz isn't a big one in our case, though we would like to keep as few slower parts in there as possible to retain phase. In any case, we were able to keep two of the AD829s without a problem.
Taking a transfer function was a big pain in the ass because of the 105-dB DC gain (i.e. without the boost). I have a trimpot to adjust the offset of the input stage, but the way the Agilent makes swept-sine measurements causes a small change in offset over the course of the sweep that is enough to rail the servo. I ended up avoiding the problem by taking the transfer function with random-noise excitation. I used an SR560 immediately after the source with a 2nd-order HPF so that I had enough at high frequencies where there is much attenuation. Here is the experimental setup and the SR560 settings (for one of the span settings):
I got a good measurement in the end; you can see that it agrees very well with the model:
Note: this says "LISO estimate", but that's a mistake; it's actually an ideal model of a TF with the same poles and zeros. This shows that the phase lag introduced by the OP27s is negligible in our region of (current) interest.
As for the noise level, the agreement isn't quite as great. Here is the input-referred spectrum along with the LISO prediction:
It converges in the middle frequencies but doesn't at either end. I was suspicious that the analyzer noise was the culprit at higher frequencies---you can see that the input range is different over the higher three measurements---so I looked at how the output noise compared:
I would be willing to bet that the high-frequency divergence is indeed caused by the analyzer noise, and I will retake these measurements tomorrow. As for the low-frequency part, I am not quite so sure. It could be that one or more of the stages has a higher-than-quoted noise corner frequency, but this doesn't seem all that likely.
Since the transfer function looked right, I tried very briefly to lock the cavity with it. It didn't work right away (I got some audio-band oscillations), but I had to take off so I didn't mess with it too much. I think that if I adjust the gain by adjusting the input power to the experiment I can get it to lock. This is first on the docket tomorrow. In the meantime, I am taking a low-frequency measurement of the servo noise overnight for inclusion in the noise model. Hopefully this will help to identify the low-frequency noise culprit.