From 23-12-27

Yuhang, Michael

To summarize, we verified the capability of the CC2 phase scan for Taiwan's Machine Learning project and also took squeezing measurements from 4.1 to 4.6 mV MZ offset (20 - 45 mW injection). We saw a total of 28% optical loss and 24 mrad phase noise, however, this measurement is dubious due to continued issues with CC2 glitch noise and rushed alignment of tabletop mode cleaners.

The PLL was left on from the previous day and remained locked when we started. To prepare for the data taking via the DGS system, we took measurements of the background noises. The baseline is Analogue to Digital Converter noise, which is at a set level, measured by putting a 50 Ohm terminator on the ADC port and taking a spectrum with diaggui. For the Taiwan machine leaning data, we send homodyne to SR560 and then to the spare ADC port K1:FDS-ADCspare_1_. The electronic noise of the SR560 is taken by terminating it with 50 Ohms at the input. The gain of the SR560 is set to put the SR560 electronic noise level just above the ADC noise level. This ensures that the optical signal from the homodyne is amplified enough to clear the ADC noise baseline. We take homodyne dark noise by measuring homodyne with no light input, which is only valid at one gain selection of SR560 (we end up having to change it later). Then we take homodyne shot noise by measuring the spectrum injecting just the local oscillator (after checking homodyne is balanced). Still pictures cannot be posted to elog, but the homodyne power spectrum shot noise is about 20x above homodyne dark, 40x above SR560 electronic noise and 50x above ADC noise (approximately).

For the Taiwan ML project, we want to somehow scan the squeezing and antisqueezing and take a long time record in the homodyne signal sent to DGS. The high voltage drivers in TAMA for whatever reason cannot scan over a range large than corresponding to 120 degrees phase shift of the local oscillator, and then the loop unlocks. Anyway, we sent a triangle wave to CC2 Perturb In to check. We want to take data with a scan of about 0.01 Hz, ~ 16 mVpk (any more than this and CC2 unlocks), for about 10 cycles. So we allocate about half an hour for each level of green power injection. We were a bit confused about how to exactly extract long term data from individual channels considering that we can only pull up quite recent data on dataviewer. In the end we just took the basic option to let Taiwan have the full sized GW frame files (.gwf) and they can extract the data using the usual python extraction tools. The size of the files are quite large so we should find a better option - upgrade DGS, set up a remote connection accessible to Taiwan, or both.

We proceeded to characterize losses using the method of measuring squeezing vs antisqueezing for different green pump power to OPO. After changing the pump power, we use CC1 error signal to characterize roughly the amount of squeezing obtained. Normally when the servo is set to "scan" mode, there is a ramp signal sent to some actuator that allows us to inspect the error signal over an FSR or some other selectable range, however, the CC1 scan mode amplitude should be set to zero for normal operation (I personally don't know the exact reason but I was told the response time of the CC1 servo to ramp signals affects lock acquisition). If we want to see the error signal we should turn the amplitude knob on the CC1 servo - after some value we can see a sinusoidal wave whose amplitude corresponds directly to the amount of squeezing applied (with the scan amplitude adjustment just turning the sine wave "on/off"), - CC1 actuates at the green phase shifter, so a scan of GRPS covers amplification and deamplification of infrared (i.e. CC) inside the OPO. Before we would use some more convoluted method like looking at BAB transmission amplification/deamplification caused by modulating the phase of the green input at the green phase shifter high voltage driver. But now we just check quickly without rearranging anything. We don't have any reference levels for the CC1 error signal characterization and it turned out to be quite erratic though. Still, we used the error signal to optimize temperature and ppol frequency for each pump power. 

At 50 mW injection the CC1 error signal is very unstable. In the past we have been able to reach 70 mW so this should be fixed. There is a lot of noise coming from the CC PLL control loop at this level of green injection, for whatever reason. 

Looking at the data afterwards, it seems that while we can perhaps fool ourselves into thinking that the squeeze level is about 7 dB on the spectrum analyzer, comparing the average levels of shot noise (-132.5 dBVrms/rtHz homodyne power spectrum) vs optimal squeezing (-137.5 dBm) gives only 5 dB squeezing at 35 mW injection. The squeezing versus antisqueezing follows:

Actually the 45 mW point was recorded wrong, and accidentally wrote the same spectrum as 40 mW squeezing. So it was removed to leave 6 points. The resultant curve fitting gives 28.6% optical loss and 24.1 mrad phase noise. Even the 40 mW injection point was noted to have quite a lot of CC2 glitch noise, so we tried removing it. The result is 29.1 % optical loss and 17.5 mrad phase noise. This is quite suspect, since phase noise is mostly determined from the high injection power part of the curve, but we still expect that the real phase noise level is lower than what the current CC PLL can provide.