We have many peaks below 40 Hz that couple, at least partly, through input beam jitter. Last summer Sam and Genevieve determined that a wide variety of site equipment produced these peaks, including the office area air handler, mini-splits in the CER, and chiller compressors for the main HVAC (LIGO-G2402140). The figure shows that there is coherence between DARM and the IMC WFS that might be used to clean these low frequency peaks, but that they are currently not being cleaned.
Eventually, we would rather not have peaks that need cleaning, but instead, reduce the source vibration and/or the vibration coupling to DARM. I think that the best plan is to reduce the source vibration of the largest peaks, but to mainly focus on reducing the coupling, because many of these peaks are just 2-5 times the vibration background at the coupling sites, so even eliminating the vibration of the sources will not be enough to get us to our design sensitivity.
The coupling of relatively low amplitude vibrations at low frequencies seems to be associated with coupling resonances. For example, when one of the frequencies of the office area air handler drifted into the 35Hz peak frequency of one of these coupling resonances, the peak in DARM was huge, but was greatly reduced by changing the operation frequency of the air handler (82986). Ill try to map out these low frequency coupling resonances during commissioning periods as a step in understanding their cause. But for now, it would be nice to see how much we can reduce the peaks with cleaning.
The nonsens training for the cleaning is set to clean over the band from 20 Hz to 8 kHz. However, the most appreciable cleaning occurs above 100 Hz. I have attached two plots from the recent training Matt and I ran. The first compares the strain before and after the code runs an offline cleaning of the data. Even in the offline cleaning, it does not perform any subtraction below 60 Hz. The contributions plot shows that the code measures a contribution from IMC WFS A pitch and yaw that is approximately 2 orders of magnitude below the strain.
Similarly, the noise budget injections usually indicate a very low jitter coupling below 60 Hz. This plot is the jitter subbudget showing pitch and yaw contributions. I removed the "total H1" line, since it's currently incorrect. However, this plot only shows contributions from IMC WFS A, and jitter is measured using the IMC PZT, which may only allow us to capture one gouy phase.
All of this is to say, despite this coherence, the nonsens algorithm doesn't find anything to subtract at low frequency. Our noise budget also doesn't show significant coupling here.
Adding: Robert and I think it may be a resolution issue. The noise budget resolution is quite broad at 0.3 Hz, so that may be why those peaks are not captured in the injection. I'm not sure how to address or test the nonsens cleaning resolution.
I have iterated through many different parameters in the nonsens algorithm, including length of time, frequency resolution, number of second order sections, maximum permitted Q value, training method, and frequency band. I am unable to achieve subtraction that is comparable to the measured coherence of these lines. At best, I have achieved 40% reduction of two of the many lines. At best I can achieve 10% reduction of some of the broadband noise. Since I am training offline, I don't expect this to be the result of some funny phase delay between the models. I'm not sure why cleaning these features isn't possible.