Give us more info on the elog:
What PD are you using? How much power the beams on the recombining BS are? What kind of BS is it?
How are you looking for the beat note? (on the scope? or spectrum analyzer?)
What was the scanned temp range?
Three points to be checked:
I'm using a 1611 New Focus PD (1 GHZ, with maximum input power 1mW), and the total power hitting on the PD is of about 0.650 mW.
The current of the NPRO laser is set to 1.38 A, so that the input power is 19 mW. The beam is initially damped by a 10% reflection BS and then it hits a 33% reflection BS (where it recombines with the PSL pick-off beam) with 2 mW power.
After this second BS the power is reduced to 0.592 mW.
The PSL pick-off hits on the 33% reflection BS with 65.5 uW power, and it exit with a 47 uW power.
I connected a power supply to apply a Voltage to the slow frequency BNC, in way to tune the laser frequency.
I'm using the AGILENT 4395A Spectrum analyzer to make the measurement. I tried to use the HEWELETT PACKARD 8591E spectrum analyzer, but the monitor didn't turn on.
The temperature spanned until now in only of about 10 deg C, because I realized that I needed a better alignment, so I added a lens in front of the PD and I did a better alignment.
Moreover, the current of the laser is too low, so I need to increase it and add more beam splitters in the beam path to dump the beam, in way to don't reach the PD threshold.
I knew that both the beams are s-polarized, but maybe I can check it again.
A rotation stage has been ordered.
Specification: Newport 481-A
This looks cool, we should have something similar, can be really useful.
Is it better than Luxor? https://labcit.ligo.caltech.edu/~jharms/luxor.html
I tried to compare the cavity scan data we get from the Finesse simulation and that we expect from the Analytical solution. The diagram of the cavity I defined in Finesse is given below along with the values of different quantities I used. For the analytical solution I have used two different equations and they are listed below.
Analytical 1 - Blue Graph
Analytical 2 - Red Graph
The graph obtained from both these solutions completely matches with each other.
The cavity which I defined in Finesse is shown below. The solution from Finesse and the Analytical solution also matches with each other. Another plot is made by taking the difference between Finesse solution and Analytical solution. The difference seems to be of the order of .
The Difference plot is also attached below.
The cavity scan data obtained from the Finesse simulation is attached here. Fig1 indicates the cavity scan data in the absence of induced misalignment. In that case only the fundemental mode is resonating. But when a misalignment is induced, higher order modes are also present as seen in Fig2. This is in the absence of surface figure error in the mirrors. Now I am trying to provide perturbations to the mirror surface in the form of zernike polynomials and get the scan data fom the simulation. These cavity scan data can be used to develop fitting models. Once we have a model, we can use it to analyse the data from the experimental cavity scan.
The unit mentioned in the x-axis was wrong. So I have remade the graphs. The point where frequency equals to zero is actually the frequency corresponding to the laser, which is in the range of 1014 Hz and it caliberated as zero.
I've been working on putting together a Finesse model for the current 40m configuration. The idea was to see if I could reproduce a model that is in agreement with what we have been seeing during the recent DRFPMI locks. With Antonio and EricQs help, I've been making slow progress in my forays into Finesse and pyKat. Here is a summary of what I have so far.
Having put together the .kat file (code attached, but this is probably useless, the new model with RC folding mirrors the right way will be what is relevant), I was able to recover a power recycling gain of ~7.5. The arm transmission at full lock also matches the expected value (125*80uW ~ 10mW) based on a recent measurement I did while putting the X endtable together. I also tuned the arm losses to see (qualitatively) that the power recycling gain tracked this curve by Yutaro. EricQ suggested I do a few more checks:
Conclusion: It doesn't look like I've done anything crazy. So unless anyone thinks there are any further checks I should do on this "toy" model, I will start putting together the "correct" model - using RC folding mirrors that are oriented the right way, and using the "ideal" RC cavity lengths as detailed on this wiki page. The plan of action then is
Sidenote to self: It would be nice to consolidate the most recent cavity length measurements in one place sometime...
Having played around with a toy finesse model, I went about setting up a model in which the RC folding mirrors are not flipped. I then repeated the low-level tests detailed in the earlier elog, after which I ran a few spatial mode overlap analyses, the results of which are presented here. It remains to do a stability analysis.
Overview of model parameters (more details to follow):
Results (general note: positive RoC in these plots mean a concave surface as seen by the beam):
Next step is to carry out a stability analysis...
I think you should use the current actual PRC & SRC cavity lengths as measured, as it would be simplest to simply replace the folding mirror optics without changing the macroscopic lengths / optic positions. (EDIT: Gautam rightly points out that we have to move things around regardless, since our current lengths include propagation through the folding mirror subtrates)
Moreover, the recycling cavity lengths you posted are not the right "ideal" lengths to use, as they do not account for the complex reflectivities of the sidebands off of the arm cavities (I have made this mistake myself). See this 40m wiki page for details.
In short, given our current modulation frequency, the ideal lengths to use would be:
These are the lengths that the recycling cavity optics were positioned for (though we did not achieve them perfectly). If you do a finer PRC/SRC length scan around the DRFPMI resonance of your model, you would presumably see some undesired sideband splitting.
So I finished writing a script which takes an .ipc file (the one which defines channel names and numbers for use with the RCG code generator), parses it, checks for duplicate channel names and ipcNums, and then parses and .mdl file looking for channel names, and outputs a new .ipc file with all the new channels added (without modifying existing channels).
The script is written in python, and for the moment can be found in /home/controls/advLigoRTS/src/epics/simLink/parse_mdl.py
I still need to add all the nice command line interface stuff, but the basic core works. And already found an error in my previous .ipc file, where I used the channel number 21 twice, apparently.
Right now its hard coded to read in C1.ipc and spy.mdl, and outputs to H1.ipc, but I should have that fixed tonight.
This IPC stuff looks really a nice improvement of CDS.
Please just maintain the wiki updated so that we can keep the latest procedures and scripts to build the models.
We have completed modifications and testing of the HAM Coil driver D1100687 units with serial numbers listed below. The DCC tree reflects these changes and tests (Run/Acq modes transfer functions).
** A fix had to be done on the DC power supply for these. The units' regulated power boards were not connected to the raw DC power, so the cabling had to be modified accordingly (see Attachment #1)
** A fix had to be done on the DC power supply for these. The units' regulated power boards were not connected to the raw DC power, so the cabling had to be modified accordingly.
Further, Paco fixed the two even serial number units (S2101648, S211650) that failed the test. The issues were minor soldering mistakes that were easily resolved.
S2100619 was fixed by Koji and it passed the test after that.
The MEDM screen capture tab is now working and up on the summary pages: https://nodus.ligo.caltech.edu:30889/detcharsummary/day/20160725/medm/
Please let me know if you have any suggestions or notice any issues.
Looks pretty great. However, there's two problems:
1) Some of the MEDM screens don't show the time. You can fix this by editing the screens and copy/paste from screens which have working screens.
2) The snapshot script seems to not grab the full MEDM screen sometimes.
These are not a very big deal, so you can get the microphones working first and we can take care of this afterwards.
Gautam helped me drill holes in a metal box and I set up my circuit inside. Everything seems to be working so far. Tomorrow I'll be suspending the box near the PSL and setting up a data channel. Attached are some pictures of the box- sorry some of the angles turned out weird.
[Jenne, Kiwamu, Steve, Sharmila, Katherine, Joe]
We finished bolting the door on the new ITMX (old ITMY) and putting the access connector section back into place. We finished with torquing all the bolts to 40 foot-pounds.
The GPIB interfaces have been updated to the new 192.168.113.xxx addresses, with Alberto's help.
Spare ethernet cables have been moved into a cabinet halfway down the x-arm.
The illuminators have a white V error on the alarm handler, but I'm not sure why. I can turn them on and off using the video screen controls (except for the x arm, which has no computer control, just walk out and turn it on).
There's a laptop or two I haven't tracked down yet, but that should be it on IPs.
At some point, a find and replace on 131.215.xxx.yyy addresses to 192.168.xxx.yyy should be done on the wiki. I also need to generate an up to date ethernet IP spreadsheet and post it to the wiki.
v: Edit on Dec 15 10PM
v: Edit on Dec 16 10PM
JD: We should check OSEMs for all optics *after* table leveling. Some of them (esp. BS and ITMX) are currently close to their limits right now.
KA: Check green alignment.
Take photos of the tables.
Fix the leveling weights
Location Optics Action
@ITMX - v POX alignment
v POP1/POP2 alignment
v Table Leveling
@ITMY - POY mirror replacement (45deg->0deg) / alignment
v SR2-TT alignment
v SRM Tower alignment / EQ-stop release
v SRM alignment
v SRM OSEM
vvSRM OPLEV (X2) install (VIS)/ alignment
v ITMY OPLEV (X2) install (VIS)/ alignment
v OM1/OM2 install (DLC 45deg)/ alignment
v Table Leveling
@BSC - v OM3 install (DLC 45deg/ alignment)
v OM4(PZT) neutralize, adjustment
IPPOS steering alignment
v BS OPLEV alignment
v PRM OPLEV(x2) alignment
@BSC - v OM3 install (DLC 45deg/ alignment)
OM4(PZT) neutralize, adjustment
IPPOS steering alignment
v BS OPLEV alignment
PRM OPLEV(x2) alignment
@IMC - v REFL mirror replacement (45deg->0deg)
@ETMX - Al foil removal
@ETMY - ETMY damping
Al foil removal
@OMC - v OM5(PZT) neutralize, adjustment
@ITM/ETM - Mirror Wiping
The (masked) tech accessed all areas in the lab (office area, control room, VEA) between ~230pm-3pm. The laser safety goggles he used have been kept aside for appropriate sanitaiton.
A technician came to the lab today at ~4pm. He entered the VEA (with booties and googles), and also the clean and bake lab. The whole procedure lasted ~10 minutes. I did not follow him around, but was available in the control room throughout the process. I think the whole episode went without incident.
BTW, this guy didn't ring the doorbell, I just happened to be here when he came by. I don't know if this is usual practise - are we happy with the technicians entering the VEA and/or clean and bake labs without supervision? AFAIK, this wasn't scheduled.
Had to restart the elog many times. For some reason firefox 8 on Win 7 kills the elog pretty consistently when trying to make a new entry. IE9 works fine ....
[Jon, Keerthana, Sandrine]
Thu.-Fri. we continued with PRC scans using the AUX laser, but now the "scanned" parameter is the frequency of AM sidebands, rather than the frequency of the AUX carrier itself. The switch to AM (or PM) allows us to coherently measure the cavity transfer as a function of modulation frequency.
In order to make a sentinel measurement, I installed a broadband PDA255 at an unused pickoff behind the first AUX steering mirror on the AS table. The sentinel PD measures the AM actually imprinted on the light going into the IFO, making our measurement independent of the AOM response. This technique removes not only the (non-flat) AOM transfer function, but also any non-linearities from, e.g., overdriving the AOM. The below photo shows the new PD (center) on the AS table.
With the sentinel PD installed, we proceeded as follows.
The below photo shows the measured transfer function [AUX Reflection / AUX Injection]. The measurement coherence is high to ~55 MHz (the AOM bandwidth is 60 MHz). We clearly resolve two FSRs, visible as Lorentzian dips at which more AUX power couples into the cavity. The SURFs have these data and will be separately posting figures for the measurements.
With the basic system working, we attempted to produce HOMs, first by partially occluding the injected AUX beam with a razor blade, then by placing a thin two-prong fork in the beam path. We also experimented with using a razor blade on the output to partially occlude the reflection beam just before the sensor. We were able to observe an apparent secondary dip indicative of an HOM a few times, as shown below, but could not repeat this deterministically. Besides not having fine control over the occlusion of the beams, there is also large few-Hz angular noise shaking the AS beam position. I suspect from moment to moment the HOM content is varying considerably due to the movement of the AS beam relative to the occluding object. I'm now thinking about more systematic ways to approach this.
How much was the osc freq of the marconi? And then how much was the resulting freq offset between PSL and AUX?
Are we supposed to see two dips with the separation of an FSR? Or four dips (you have two sidebands)?
And the distance between the dips (28MHz-ish?) seems too large to be the FSR (22MHz-ish).
(Jon, Keerthana, Sandrine)
I am attaching the plots of the Reflected and transmitted AUX beam. In the transmission graph, we are getting peak corresponding to the resonance frequencies, as at that frequency maximum power goes to the cavity. But in the Reflection graph, we are obtaining dips corresponding to the resonance frequency because maximum power goes to the cavity and the reflected beam intensity becomes very less at those points.
First Contact Training with Margot
Made a dry run of the in-situ cleaning for a 3inch optic.
Attachment 1: The Al dummy mass is clamped in the suspension cage.
Attachment 2: The front surface was painted. The nominal brush with the FC bottle was used.
Attachment 3: Zoom in of the front surface.
Attachment 4: The back surface was painted.
Attachment 5: The back surface was peeled.
Attachment 6: The front surface was peeled too.
Attachment 7: The peeled layers.
1. To paint a thick layer (particlarly on the rim) is the key to peel it nicely.
2. It was helpful for easier peeling to have mutiple peek tabs. Two tabs were sufficient for ~1" circle.
3. The nominal brush with the bottle was OK although one has to apply the liquid many times to cover such a large area. A larger brush may cause dripping.
4. The nominal brush was sufficiently long once the OSEMs are removed. In any case it is better to remove the OSEMs.
After much fussing, we got a picture of MMT1 with the beam.
Using the iris doesn't seem feasible. Since it has to be significantly separated from the optic, it is hard to judge whether it is centered, especially in yaw.
It took ~30 min to get this picture. Comments on whether this kind of picture is good enough are welcomed, since there are many more to be taken.
Looks good. Any way that you can tell in an unambiguous way, where the beam is, is very good. Ideally we want to have1-2 mm accuracy.
[Jenne, Kiwamu, Steve]
Round 1 of measuring the MC mode is pretty much done. Yay.
Earlier today, Steve and I launched the MC beam off the flat mirror just after the Faraday, and sent it down toward ETMY(new convention). We ended up not being able to see it all the way at the ETM because we were hitting the beam tube, but at the ITM chamber we could see that the beam looked nice and circle-y, so wasn't being clipped in the Faraday or anywhere else. To do this we removed 2 1inch oplev optics. One was removed from the BS table, and wrapped in foil and put in a plastic box. The other was just layed on its' side on the BS table.
I then took the beam out of the BS chamber, in order to begin measuring the mode. I left the flat fixed mirror in the place of what will be PZT SM1, and instead used the PZT mirror to turn the beam and get it out the BS chamber door. (Thoughts of getting the beam to the BS oplev table were abandoned since this was way easier, since Kiwamu and Steve had made the nifty table leg things.) Kiwamu and I borrowed an 2inch 45P Y1 optic from the collection on Koji's desk (since we have ZERO 2inch optics on the random-optic-shelf....no good), to shoot the beam down the hallway of the Yarm (new convention). We used the beam scan on a rolling cart to measure the beam at various distances. I made some sweet impromptu plum bobs to help make our distance measurements a bit more accurate.
We stopped at ~25 feet from the BS chamber, since the spot was getting too big for the beam scanner. If it turns out that I can't get a good fit with the points I have, I'll keep everything in-chamber the same, and do the farther distances using the good ol' razor blade technique.
I have measurements for the distances between the beam scan head and the opening of the BS chamber. Tomorrow, or very soon after, I need to measure the distances in-chamber between the MC and the BS chamber opening. Plots etc will come after I have those distances.
Next on the to-do list:
1. Measure distances in-chamber for first mode scan.
2. Plot spot size vs. distance, see if we need more points. Take more points if needed.
3. Put in MMT1, repeat measurements.
4. Put in MMT2, rinse and repeat.
5. Move the PZT mirror to its new place as SM1, and figure out how to connect it. Right now the little wires are hooked up on the BS table, but we're going to need to make / find a connector to the outside world from the IOO table. This is potentially a pretty big pain, if we don't by happenstance have open connectors on the IOO table.
[Jenne, Jamie, Mirko]
We got the first version of the oaf code based on Matt"s code running!! :-)
Produces already data for e.g. MICH DOF. But don"t trust that. It's only 10 taps long and delay is not adjusted.
Today, we finally crossed the last hurdle and got a successful converging coil balancing run.
This week I wrote Matlab code, most of which can be found in /users/masha
First, I wrote a simulation seismicFilter.m which filters noisy seismic noise with a desired signal of non-seismic noise. The signals are purely simulated, so I played around with zero-pole-gain generation of transfer functions to obtain them. The function takes the number of taps, the filter type (Wiener or adaptive nlms) as well as an iteration step size and number of iterations, and generates PSD plots of the witness signal, the desired signal, the estimated (filtered) signal, and the error. I'm not sure that I am properly implementing the Wiener part of the code, and I assume the line "[W, R, P] = miso_firlev(TAPS, noisySeismicSignal1, seismicSignal2); " generates W, a filter with TAPS number of weights, but then "[y, error] = filter(W, 1, noisySeismicSignal1);" generates an error signal of size TAPS rather than N, the size of the original signal. Perhaps I should calculate error using e(t) = d(t + a) - w(t)*x(t), where "a" is the delay.
I have various screenshots in my directory of what seismicFilter.m generates, and I will take a larger screenshot, as well as generate a learning curve (for error vs. number of taps) when I can use Sasha's computer for a bit, since it both has more computing power and a larger screen.
The funciton filterConvergence.m, meanwhile, is similar, except it takes two file names as real data, and uses realDataFilter.m to run the filtering. Currently, I am working with data from C1:IOO-MC_F_DQ-Online and C1:PEM-SEIS_GUR1_X_IN1_DQ-Online, and I will include screenshots of these once I get on Sasha's computer.
In order to generate the data, meanwhile, I had to modify the python script, and thus wrote mashaImportingData.py for myself. Likewise, plotSignalFromFile.m visualizes this data, both in the time domain and in the frequency domain.
On the side, I wrote an NLMS filter in adaptiveFilterSimulationNLMS.m, and compared is to Matlab's NLMS filter in NLMStest.m (using generated data) and adaptiveFilterSimulation.m using twn input signals. Right now, it's faster on smaller inputs and smaller tap sizes, but then begins to choke and run slower than the Matlab one when these get to big. In order to improve it, I have to develop a better method of generating the initial weights.
As far as machine learning goes, once I find the number of taps for the convergence of both the Wiener filter and the NLMS filter, I will email Denis for further instructions. At some point, however, I should generate learning sets from the seismometers and the MCL (or the DARM), and thus have to find adequate times at which I can take data (probably not from the DARM, however, because it was rarely on).
Thanks for reading!
This week, the other SURF students and I got acquainted with the caltech campus, LIGO 40m lab and the expectations of the SURF program. We went to a lot of safety meetings and lectures that established a framework for the jobs we will be doing over the course of the summer. I went on several tours of the 40m interferometer (one each with Jenne, Jamie and Steve) to get an overview of the layout and specifics of the setup. I read parts of R. Ward and A. Parameswaran's theses and Saulson's book in order to prepare myself and gain a broader understanding of the purpose of LIGO.
I also began working in Python this week, primarily graphing PSDs of data from the C1:SUS-ETMY_SENSOR_LR, C1:SUS-ETMY_SENSOR_LL, C1:SUS-ETMY_SENSOR_UR, and C1:SUS-ETMY_SENSOR_UL channels. I will eventually be using Python to generate the plots for the summary pages, so this is good practice. The code that I have been working on can be found in /users/elizabeth.davison/script5.py. Additionally, I have been going through the G1 summary pages and attempting to understand the plots available on them and the code that is available.
My plans for the upcoming week begin with modifying my code and potentially calibrating the channel data so that it is in units of length instead of counts. I will also access the code from the G1 pages and go over it in depth, hopefully gaining insight into the structure of the website.
[Jenne, Steve, Nancy, Gopal]
We made an attempt at hanging some of the Tip Tilt eddy current dampers today.
Photo 1 shows the 2 ECDs suspended.
(1) Loosen the #4-40 screws on the side of the ECDs, so the wire can be threaded through the clamps.
(2) Place the ECDs in the locator jigs (not shown), and the locator jigs in the backplane (removed from main TT structure), all laying flat on the table.
(3) Get a length of Tungsten wire (0.007 inch OD = 180um OD), wipe it with acetone, and cut it into 4 ~8cm long segments (long enough to go from the top of the backplane to the bottom).
(4) Thread a length of wire through the clamps on the ECDs, one length going through both ECDs' clamps.
(5) One person hold the wire taught, and straight, and as horizontal as possible, the other person tightened the clamping screws on the ECDs.
(6) Again holding the wire in place, one person put the clamps onto the backplane (the horizontal 'sticks' with 3 screws in them).
(7) The end. In the future, we'll also clip off extra pieces of wire.
When we held up the backplane to check out our handy work, it was clear that the bottom ECD was a much softer pendulum than the top one, since the top one has the wire held above and below, while the bottom one only has the wire held on the top. I assume we'll trim the wire so that the upper ECD is only held on the top as well?
* This may be a 3 person job, or a 2 people who are good at multitasking job. The wire needs to be held, the ECDs need to be held in place so they don't move during the screwing/clamping process, and the screws need to be tightened.
* Make sure to actually hold the wire taught. This didn't end up happening successfully for the leftmost wire in the photo, and the wire is a bit loose between the 2 ECDs. This will need to be redone.
* We aren't sure that we have the correct screws for the clamps holding the wire to the backplane. We only have 3/16" screws, and we aren't getting very many threads into the aluminum of the backplane. Rana is ordering some 316 Stainless Steel (low magnetism) 1/4" #4-40 screws. We're going for Stainless because Brass (the screws in the photo), while they passed their RGA scan, aren't really good for the vacuum. And titanium is very expensive.
The 2nd photo is of the magnet sticking out of the optic holder. The hole that the magnet is sitting in has an aluminum piece ~2/3 of the way through. A steel disk has been placed on one side, and the magnet on the other. By doing this, we don't need to do any press-fitting (which was a concern whether or not the magnets could withstand that procedure), and we don't need to do any epoxying. We'll have to wait until the ECDs are hung, and the optic holder suspended, to see whether or not the magnet is sticking out far enough to get to the ECDs.
The efficiency of the mode matching (MM) to PRC (Power-Recycling Cavity) has been estimated by using the interferometer.
The estimated MM efficiency is about 74 % when losses in the cavity are assumed to be zero.
Here are the measured values in REFLDC
Anyways the estimated MM efficiency with the sidebands effect included and without loss effect is
MM efficiency = 73.7 +/- 1.7 % (1 sigma error) or +/- 8.7 % (5 sigma error)
"^2"s are missing in the second equation, but the calculation results seem correct.
PRX and PRY have different mode matching because of the Michelson asymmetry.
Are individually estimated mode matching indicates any sign of reasonable mode mismatch?
(The difference can be very small because the asymmetry is not so big.)
On the topic of high AS55_Q RFPD offset, it seems it stems from a small residual offset on top of the 42 dB whitening filter gain (previously 3 dB). We verified this by looking in the past using dtt and seeing an offset of ~ 100 counts, which are consistent with the hotfix. We reverted the whitening filter gain to +24 dB, in order to accomodate the 10% power difference from AS2. We decided to move forward, and try locking MICH using AS55_Q_ERR. The IQ mixing angle was changed to -167 deg from -122 deg to minimize the signal in AS55_I_ERR. We have also added comb60 filters for AS55. The LSC_MICH filter gain was adjusted to -6 (used to be -13 in the configuration script) to get a MICH_OLTF UGF of 90 Hz (which is the previously measured value as of 2021 July), see Attachment #1 for the MICH OLTF estimate.
We then calibrate MICH using the fringe amplitude, so that , where is the amplitude of the error point (C1:LSC-AS55_Q_ERR_DQ) in our case ~ 110 +- 2 counts. The calibrated error point spectral density is shown in Attachment #2. Calibration is done into meters in terms of difference between BS to ITMX length and BS to ITMY length.
In brief, I trained a deep neural network (DNN) to recosntuct the cavity length, using as input only the transmitted power and the reflection PDH signals. The training was performed with simulated data, computed along 0.25s long trajectories sampled at 8kHz, with random ending point in the [-lambda/4, lambda/4] unique region and with random velocity.
The goal of thsi work is to validate the whole approach of length reconstruction witn DNN in the Fabry-Perot case, by comparing the DNN reconstruction with the ALS caivity lenght measurement. The final target is to deploy a system to lock PRMI and DRMI. Actually, the Fabry-Perot cavity problem is harder for a DNN: the cavity linewidth is quite narrow, forcing me to use very high sampling frequency (8kHz) to be able to capture a few samples at each resonance crossing. I'm using a recurrent neural network (RNN), in the input layers of the DNN, and this is traine using truncated backpropagation in time (TBPT): during training each layer of RNN is unrolled into as many copies as there are input time samples (8192 * 0.25 = 2048). So in practice I'm training a DNN with >2000 layers! The limit here is computational, mostly the GPU memory. That's why I'm not able to use longer data stretches.
But in brief, the DNN reconstruction is performing well for the first attempt.
In the results shown below, I'm using a pre-trained network with parameters that do not match very well the actual data, in particular for the distribution of mirror velocity and the sensing noises. I'm working on improving the training.
I used the following parameters for the Fabry-Perot cavity:
The uncertaint is assumed to be the 90% confidence level of a gaussian distribution. The DNN is trained on 100000 examples, each one a 0.25/8kHz long trajectory with random velocity between 0.1 and 5 um/s, and ending point distributed as follow: 33% uniform on the [-lambda/4, lambda/4] region, plus 33% gaussian distribution peaked at the center with 5 nm width. In addition there are 33% more static examples, distributed near the center.
For each point along the trajectory, the signals TRA, POX11_I and POX11_Q are computed and used as input to the DNN.
Gautam collected about 10 minutes of data with the free swinging cavity, with ALS locked on the arm. Some more data were collected with the cavity driven, to increase the motion. I used the driven dataset in the analysis below.
The ALS signal is calibrated in green Hz. After converting it to meters, I checked the calibration by measuring the distance between carrier peaks. It turned out that the ALS signal is undercalibrated by about 26%. After correcting for this, I found that there is a small non-linearity in the ALS response over multiple FSR. So I binned the ALS signal over the entire range and averaged the TRA power in each bin, to get the transmission signals as a function of ALS (in nm) below:
I used a peak detection algorithm to extract the carrier and 11 MHz sideband peaks, and compared them with the nominal positions. The difference between the expected and measured peak positions as a function of the ALS signal is shown below, with a quadratic fit that I used to improve the ALS calibration
The result is
The ALS calibrated z error from the peak position is of the order of 3 nm (one sigma)
Using the calibrated ALS signal, I computed the cavity length velocity. The histogram below shows that this is well described by a gaussian with width of about 3 um/s. In my DNN training I used a different velocity distribution, but this shouldn't have a big impact. I'm retraining with a different distirbution.
The plot below shows a stretch of time domain DNN reconstruction, compared with the ALS calibrated signal. The DNN output is limited in the [-lambda/4, lambda/4] region, so the ALS signal is also wrapped in the same region. In general the DNN reconstruction follows reasonably well the real motion, mostly failing when the velocity is small and the cavity is simultanously out of resonance. This is a limitation that i see also in simulation, and it is due to the short training time of 0.25s.
I did not hand-pick a good period, this is representative of the average performance. To get a better understanding of the performance, here's a histogram of the error for 100 seconds of data:
The central peak was fitted with a gaussian, just to give a rough idea of its width, although the tails are much wider. A more interesting plot is the hisrogram below of the reconstructed position as a function of the ALS position, Ideally one would expect a perfect diagonal. The result isn't too far from the expectation:
The largest off diagonal peak is at (-27, 125) and marked with the red cross. Its origin is more clear in the plot below, which shows the mean, RMS and maximum error as a function of the cavity length. The second peak corresponds to where the 55 MHz sideband resonate. In my training model, there were no 55 MHz sidebands nor higher order modes.
The DNN reconstruction performance is already quite good, considering that the DNN couldn't be trained optimally because of computation power limitations. This is a validation of the whole idea of training the DNN offline on a simulation and then deploy the system online.
I'm working to improve the results by
However I won't spend too much time on this, since I think the idea has been already validated.
I included the 55 MHz sideband and higher order modes in my training examples. To keep things simple, I just assumed there are higher order modes up to n+m=4 in the input beam. The power in each HOM is randomly chosen from a random gaussian distribution with width determined from experimental cavity scans. I used a value of 0.913+-0.01 rad for the Gouy phase (again estimated from cavity scans, but in reasonable agreement with the nominal radius of curvature of ETMX)
Results are improved. The plot belows show the performance of the neural network on 100s of experimental data
For reference, the plots below show the performance of the same network on simulated data (that includes sensing noise but no higher order modes)
I've begun cleaning the optics that will eventually go back onto the newly installed X-endtable. We decided that First Contact was the way to go (as opposed to methanol drag wiping). Koji demonstrated the application of the (red) First Contact solution onto a 2" mirror - I then proceeded to work on the rest of the optics. We are broadly following the procedure in E1000079 - first one coat of First Contact solution is applied, then a small piece of PEEK is embedded by applying a second layer of solution over it (this will enable us to pull off the First Contact once we are ready - the plan is to do this after roughly placing the optic on the table. As of now, I've finished coating most of the optics that are part of the IR Transmon path - I will continue later in the evening.
The new endtable is almost ready for re-population. Steve just needs to shim the enclosure which will be done tomorrow morning. The game-plan as discussed at the meeting today is to first try and set up the IR Transmon path. This will allow us to verify that the endtable height is such that we can maintain a beam height of 4" everywhere on the table (I suspect we may have to compromise at some poing and do some fine adjustment of 1/4 to 1/2" somewhere though). It will also allow me to define the cavity axis relative to the table, which will be useful to place the green steering optics eventually. Doing this will be challenging though as right now, I can't see any of the arm flashes on the endtable using an IR card. Ideally, we want to somehow lock the X arm and then do the checks mentioned at the endtable, before beginning to put the endtable back together.
I used a pair of tweezers to remove the stray fiber of first contact. As Koji predicted, this was rather dry and so it didn't have the usual elasticity, so while I was able to pull most of it off, there is a small spot remaining on the HR surface of the ETM. We will remove this with a fresh application of a small patch of FC.
I the meantime, I'm curious if this has actually fixed the suspension woes, so yet another round of freeswinging data collection is ongoing. From the first 5 mins, looks positive, I see 4 peaks around 1Hz !
Update 730pm: There are now four well-defined peaks around 1 Hz. Together with the Bounce and Roll modes, that makes six. The peak at 0.92 Hz, which I believe corresponds to the Yaw eigenmode, is significantly lower than the other three. I want to get some info about the input matrix but there was some NDS dropout and large segments of data aren't available using the python nds fetch method, so I am trying again, kicked ETMY at 1828 PDT. It may be that we could benefit from some adjustment of the OSEM positions, the coupling of bounce mode to LL is high. Also the SIDE/POS resonances aren't obviously deconvolved. The stray first contact has to be removed too. But overall I think it was a successful removal, and the suspension characteristics are more in line with what is "expected".