1 Introduction↩︎

The LIGO [1] and Virgo [2] detectors started the era of GW astronomy when the first GW signal from the merger of two black holes was detected in 2015 [3]. Since then, the LIGO, Virgo and KAGRA collaborations have published 90 probable detections of signals involving black holes and neutron stars, including the spectacular multi-messenger discovery in August 2017 [4]–[7]. The data used for these analyses is publicly available in the Gravitational Wave Open Science Center (GWOSC) ¹. The LIGO and KAGRA detectors started taking data in the fourth Observing run (O4) on May 24, 2023. The first part of the run, O4a, finished on January 16, 2024. The LIGO and Virgo detectors started taking data again on April 10, 2024 (O4b). In O4a, 92 significant detection candidates were shared as public alerts, including 11 retracted for data-quality problems. The 81 significant detections are almost as many as the 90 detections published in the GWTC-3 catalog [6] from the previous three Observing runs.

As shown in , the broadband sensitivity of the LIGO detectors in O4a was significantly better than in O3b. Some key upgrades commissioned between O3 and O4 include frequency-dependent squeezing [8], replacement of some optics that had small defects (point absorbers) allowing for increased laser power, and replacing squeezing subsystem Faraday isolators with more efficient ones to increase the level of squeezing achieved [9]. As a measure of sensitivity, we can calculate the distance to which the noise would allow a detection of a binary neutron star system (1.4 M\(_\odot\) each) with a signal-to-noise ratio of 8; this is often called the “binary neutron star inspiral range" or”BNS range". The spectra shown in correspond to dates in O3b (H1 March 19, 2020, and L1 January 4, 2020) with BNS ranges 112 Mpc and 134 Mpc respectively; and dates in O4a (H1 December 12 2023, and L1 December 31 2023) with BNS ranges 160 Mpc and 158 Mpc respectively.

At the junction between instrument scientists and data analysis groups, the DetChar group works to understand instrumental noise and improve data-quality [10]. The group carries out instrumental investigations in collaboration with instrument scientists at the observatory sites, striving to understand and mitigate sources of noise that are not inherent to the detector design and thereby to improve detector performance. The DetChar group also provides data-quality information for GW searches, which is crucial for searches to avoid or remove noise artifacts, and to evaluate and validate gravitational-wave signals.

The DetChar group carries out a variety of regular activities to monitor noise and provide data-quality information to searches. The LIGO DetChar summary pages [11] provide detector-site and off-site scientists, technicians, and staff an overview of the performance of the LIGO detectors and environmental monitors through a series of regularly updated webpages. The summary pages centralize plots, figures of merit, and links to additional resources for analyses of the GW strain data and various detector subsystems. During Observing runs, members of the DetChar group participate in data-quality shifts, which utilize the summary pages to closely monitor the performance of the detector. Members also engage in event validation shifts in order to assess any data-quality issues near and during the time of GW signals [12]. Finally, the DetChar group works together with GW analysis efforts to construct data-quality products in order to avoid noise contamination in GW analyses (see .)

Data recorded by the LIGO detectors can be characterized as generally Gaussian and stationary with non-Gaussian noise appearing in several forms: 1) short-duration artifacts, often referred to as “glitches”, which are typically broadband in nature; 2) persistent or slowly time-varying, broadband artifacts; 3) persistent or slowly time-varying, narrowband artifacts, often referred to as “lines” [13]. Typically, glitches impact transient GW searches and parameter estimation of candidate transient GW events but do not strongly impact searches for persistent GW. Similarly, lines impact persistent GW searches but do not strongly impact searches and parameter estimation for transient GW. Persistent broadband artifacts impact all types of searches because the artifacts elevate the noise background.

Glitches and lines can be further categorized by the morphology of the artifact. Many common glitch classes have been named (e.g., Tomte, Fast Scattering, Low Frequency Burst) based on a combination of their shape in spectrogram plots and what is known of their origin, see  [14], [15]. Some glitch classes are more detrimental to transient gravitational-wave searches and CBC parameter estimation than others. Lines, by contrast, have fewer morphological distinctions, see  [16]. The primary distinction is between individual line artifacts and combs of lines. Combs arise when narrowband noise that couples into the gravitational-wave data is not purely sinusoidal, creating a series of peaks with common frequency spacing. Combs are particularly problematic because one source can impact multiple narrow frequency bands at the same time.

This paper describes the activities performed to characterize the strain data measured by the LIGO detectors, and investigations of instrumental and environmental noise between the end of O3 and end of O4a. In , we describe the instrumental investigations. In we describe the activities to promptly validate gravitational-wave candidates, using tools to analyze the data-quality surrounding the event. In , we describe the use of data-quality products in searches of gravitational-waves of different kinds (compact binary coalescences, un-modelled transients, continuous waves and stochastic background). In , we summarize the results, present conclusions and present prospects for the near future.

2 Instrumental Investigations↩︎

Instrumental investigations carried out by the DetChar group are crucial for understanding the impact of various noise sources on detector data quality [19]. The PEM investigations are often carried out at the sites and require a strong co-ordination between the DetChar group and the instrument scientists at the site [20]. These investigations rely on vibration, acoustic, and magnetic injections, which enable us to estimate environmental couplings between different parts of the detector and the GW strain data. As explained in detail later in this section, some noise couplings can be reduced or eliminated thereby reducing the amount of noise in GW strain channel. In addition to artificially inducing environmental noise, we also routinely induce differential-arm displacements via the photon calibrator [21] to study coupling between the GW channel and interferometer auxiliary channels (see ) and add simulated gravitational waveforms for end-to-end testing of analysis pipelines [22].

In addition to artificially induced environmental noise and signals studies, the DetChar group also analyzes instrument data in other ways to enhance our understanding of the detector. These studies usually use several DetChar tools for determining the coupling between the environment, auxiliary sensors, and the detector noise characteristics. For example, through these investigations we may find that ground motion at a specific location in the detector is more correlated to the noise in the GW strain channel than at other locations. Such hypothesis can then be tested using the PEM tests.

These investigations are central to solving problems that adversely impact the detector data-quality and uptime, consequently reducing the number of gravitational-wave observations. The transient noise impacts the parameter estimation process and could generate false alerts, which have to be retracted later. In this section, we first give an overview of transient noise in O4a at both the detectors, and then we discuss several DetChar investigations carried out between the end of O3 and the end of O4a.

2.1 Transient noise investigations at both sites↩︎

2.1.1 Transient Noise↩︎

Glitches are short-duration bursts of excess power with their origins in environmental and/or instrumental couplings. Omicron is an event trigger generator that is used to search for this excess power in \(h(t)\) and auxiliary channels [23]. The time, frequency, and SNR of short-duration transient noise can be visualized using “glitchgrams”. shows an example of a glitchgram which is generated by plotting Omicron triggers.

Different tools are used to investigate glitches, and one of them is Gravity Spy, which classifies noise transients by their morphologies in spectrograms [14], [15], [24], [25]. shows time-frequency spectrograms for six common glitch categories at the two sites  [26], [27]. Certain environmental or instrumental conditions can generate similar signals. Glitch classification enables the identification of patterns in the data for similar noise transients.

Examining the relationship between noise in the primary gravitational-wave channel and multiple auxiliary channels can potentially lead us to the source of the noise. Hveto is one of the main tools used to study the correlation between transient noise in the primary strain and auxiliary data [28]. Auxiliary channels that also witness the same noise are called witness channels. The glitchgram shown in is from a day when seismometers recorded elevated ground motion, which caused the excess transient noise at low frequencies around \(10\)–\(20~\mathrm{Hz}\). Sometimes, this motion is so strong that the interferometer cannot maintain the servo-controlled resonance condition and stops observing [29]. This is referred to as losing lock of the interferometer.

Multiple variables affect the glitch rate in the detectors, including instrumental upgrades, addition of new subsystems, and environmental conditions such as wind, elevated ground motion, or the passing of trucks or trains near the site. shows the comparison of glitch rates between O3 and O4 for two different SNR thresholds. In O3, transient noise at both detectors was dominated by stray light. These couplings were greatly reduced during O3b and after O3. Further details are discussed in [17], [18] and in .

The O4a transient noise at LHO was dominated by low SNR glitches mostly in \(10\)–\(50~\mathrm{Hz}\). Most of the LHO transient noise was low SNR as can be seen in and in , which shows the daily glitch rate at LHO and LLO during O4a. These broadband transients had a common source and were mitigated during O4a. provides more details on this.

The transient noise at LLO was dominated by low-frequency ground motion that induce laser light scattering. Most of this ground motion can be attributed to the impact of atmosphere driven ocean waves on the ocean surface, also known as microseism [30]. The microseismic motion is seasonal and is caused by intense ocean wave activity from winter storms. This is why we see an increase in the glitch rate during the latter half of O4a as shown in . provides more detail on this noise.

2.1.2 Overview of scattering noise↩︎

Noise due to light scattering is a common problem at both detectors. Scattering happens when a small fraction of laser light gets reflected off a test mass, hits another moving surface in the vicinity, and then rejoins the main laser beam. This rejoining leads to the introduction of a time dependent phase modulation to the main laser beam [31]. The additional phase noise shows up as \(h_{\rm{ph}}(f)\),

\[h_{\rm{ph}}(f) = \frac{K}{2}\frac{\lambda}{4\pi L}\mathcal{F}[\sin{\delta \phi}] \label{eq1}\tag{1}\] where

\[\phi(t) = \phi_{0} + \delta \phi_{\rm{sc}}(t) = \frac{4 \pi}{\lambda}[x_{0} + \delta x_{\rm{sc}}(t)]. \label{eq2}\tag{2}\] Here, \(K\) is the ratio of stray light amplitude to the amplitude of light in the main beam (usually unknown but small, of the order of 1e-9 in O4a microseismic scatter), \(\mathcal{F}\) indicates a Fourier transform, \(\lambda\) is laser wavelength (1064 nm) and \(L\) is the length of interferometer arms (4 km) [17], [32]–[34]. When the relative motion between the scattering surfaces is not small compared to the laser wavelength, we get fringe wrapping and the noise shows up as arches in time-frequency spectrograms (see ). Transient scattering noise can be classified into two categories, which depends on the frequency of the ground motion that produces it: Slow Scattering and Fast Scattering [17]. Scattering is discussed in the following section, and again in .

2.1.3 Cryo-manifold baffle noise↩︎

Most of the optics in LIGO are housed in corner and end stations (X-arm and Y-arm). During O3, vibrating CBs at the LIGO end stations were identified as a cause of light scattering noise [18], [35]. These baffles prevent most light reflecting from the beamtube reduction flange, where the beamtube narrows at the corner and end stations, from reentering the detector arms [36]. Some light reflected from these baffles, however, still interferes with the circulating light in the main beam. When a mechanical resonance of the CB with a high quality factor was rung up from heightened ground motion, scattered light noise at about 4 Hz and harmonics were visible in \(h(t)\).

At both detectors, rubber viton dampers were installed on three of the four CBs to decrease the quality factor of the mechanical resonance prior to the start of O4a. Damping reduced the velocity of the reflecting surface and thus lowered the cutoff frequency of the scattering noise to low frequencies where the interferometer is not sensitive. The vibration coupling at the remaining undamped cryobaffles increased between O3 and O4 as the power in the arms increased. The difference in coupling between 60 W and 75 W at the undamped CBs was about a factor of about 3 at LHO [37].

In the O4a-O4b commissioning break an incursion was made into each detector’s X-arm end station vacuum enclosure to damp the final cryo-manifold baffle [38]. After the viton damper installation, the quality factor of the \(\sim\) 4 Hz mechanical mode in the CB dropped by a factor of \(\sim\)​20, reducing the maximum frequency of the scattering noise by roughly the same amount [39].

2.1.4 Safety studies↩︎

A crucial aspect of verifying gravitational-wave candidates is ensuring that they are not introduced by environmental noise sources observed in auxiliary channels. In general, however, the transfer functions between \(h(t)\) (GW strain) channel and the thousands of auxiliary channels is unknown, so it is possible for real gravitational-wave events to produce signals in some of the auxiliary channels. If we were to use such a channel (an “unsafe” channel) to veto a gravitational-wave event, without knowledge of this coupling, we would in effect be using an astrophysical signal to veto itself. To probe the “safety” of auxiliary channels for vetoing candidate events, we use the photon calibrator to inject sine-Gaussian waveforms into \(h(t)\) at each detector, mimicking a gravitational-wave signal. A safe auxiliary channel should not respond to these injected waveforms in \(h(t)\). Then, we perform a statistical analysis of the \(\mathcal{O}(5000)\) auxiliary channels sampled above 16 Hz using the pointy-poisson tool [40] to classify auxiliary channels as either “safe” (i.e., acceptable to use to veto potential gravitational-wave events) or “unsafe.” The resulting list of safe channels is then passed on to downstream data-quality analyses.

The injection and safety analysis process was repeated every few months in each detector during Engineering Run 15 (ER15) (April 26, 2023 - May 24, 2023) and O4a (May 24, 2023 - January 16, 2024) to track possible changes in safety and to ensure that any new channels were correctly classified as either safe or unsafe. During the first set of analyses in ER15, a handful of LHO channels were found to be unsafe compared to the existing list from O3: the ESD voltage monitors at X-arm end station, and a handful of suspension rack magnetometers. We found no substantial changes in channel safety during the subsequent duration of ER15 and O4a.

2.1.5 Narrow spectral artifacts↩︎

A typical daily-averaged, high-resolution spectrum at either LHO or LLO reveals hundreds or thousands of narrow spectral artifacts (lines) within 10-2000 Hz, the band of particular interest to persistent gravitational-wave searches. These lines display a variety of amplitudes, widths, and shapes. Some are stable over long periods of time (weeks, months), while others are variable. Identifying and mitigating the most problematic lines requires both routine monitoring and focused investigations.

Since noise lines impact CW and stochastic gravitational-wave searches in slightly different ways, specific tools are used to evaluate these artifacts to aid the analyses. CW-focused line studies typically use a tool called Fscan [41], while stochastic-focused studies use tools called STAMP-PEM [42] and StochMon [43]. These tools provide complementary information about lines, and their results are often used together to inform line investigations, mitigation efforts, and data-quality products.

2.1.5.1 Line investigations with Fscan

Fscan produces high-resolution spectra averaged over long periods of time. It was largely rewritten for O4: modernizing the code, improving stability of data generation, making new data products and visualization tools available for analysis, and enabling production of custom spectra. In O4a, Fscan was used to generate daily, weekly, and monthly spectra for about 80 channels at each observatory site, using FFTs of 1800-s-long data segments. Additional analyses were performed to track lines of interest (determining witness channels and times at which the artifacts changed) using Fscan data.

In , we highlight examples of successful investigation and mitigation efforts in O4a at LHO. Because LLO has generally cleaner data for persistent gravitational-wave searches, there have been fewer notable examples of mitigation. A number of high-priority narrow spectral noise artifacts, however, have not yet been mitigated, including artifacts present at both detector sites. Additional work is ongoing in this area. The highest-priority artifacts are those that contaminate a broad spectral region (i.e., combs, especially those with many visible peaks) and artifacts that are present at both sites (e.g., 60 Hz power mains) because these have a disproportionate impact on persistent gravitational-wave searches.

The highest line artifacts in have known cause, typically due to design choices inherent to the detector design (mirror suspension resonances at various frequencies, strongest at 300 Hz and 500 Hz and harmonics) or calibration and dither lines to monitor or control interferometric cavities, respectively.

2.1.5.2 Monitoring strain-strain narrow-band coherence.

Stochastic searches rely on cross-correlating \(h(t)\) data from detector pairs, so understanding and monitoring potential noise sources that could detrimentally impact the cross-correlated data is crucial. A stochastic monitoring tool called StochMon is specifically designed to calculate strain–strain coherence in medium-latency and flag problematic frequency bins that have excess coherence. These are flagged when they pass a Gaussian coherence threshold of \(1 - \left(1/N_{f}\right)^{1/(N_{\rm eff} -1)}\), where \(N_f\) is the number of frequency bins, and \(N_{\rm eff}\) is the effective number of segments used in the coherence calculation. An example of the calculated coherence and its outliers is shown in . The outlier bins shown have all been traced back to specific detector noise sources. The \(505.75~\mathrm{Hz}\) and \(1496.2~\mathrm{Hz}\) outliers lies within to the first and third violin modes of the detectors, which are resonances of the detector’s mirror suspension fibers and are caused by changes in the fiber tension. The \(24.5 ~\mathrm{Hz}\) outlier corresponds to a calibration line that was turned on between July 25, 2023 to August 9, 2023  [44]. The \(960 ~\mathrm{Hz}\) outlier corresponds to one of the Duotone signals in the timing system  [45].

In practice, the investigation to determine the instrumental causes of the coherence outliers is crucial. They are carried out by spectral monitoring tools including Fscan, and STAMP-PEM. STAMP-PEM is another stochastic monitoring tool that keeps track of daily, weekly coherence between \(h(t)\) and physical environmental channels at 0.1 Hz resolution. In terms of computing resources, a moderately low resolution of 0.1 Hz allows it to monitor a wide range of channels (about 1000 channels per observatory site). The high-resolution spectral information from Fscan and auxiliary channel information from STAMP-PEM provide complementary information to support the strain-strain coherence outliers investigation.

2.1.6 Broadband persistent artifacts↩︎

2.1.6.1 Investigations of coherence noise

Correlated magnetic noise was investigated as a potential noise source for stochastic searches. During O4a, two sets of coordinated magnetic injections were performed to study the coupling between the coherence of \(h(t)\) in the presence of an increased correlated magnetic field, thereby determining the magnetic noise budget. The first set of injections lasted 5 minutes and was composed of broadband white noise to test the coordinated injection capability between the sites. The second set lasted \(\sim\)​45 minutes and consisted of a realistic Schumann-like spectrum (see e.g., [46] for a description of Schumann noise relevant for ground-based gravitational-wave detectors). shows the coherence spectrums between the magnetometers as well as the coherence between \(h(t)\) at LHO and LLO during the second set of injections compared to a reference background time. Details on the level of correlated magnetic noise estimate will be provided in the O4 stochastic analyses release.

2.2 LIGO Hanford noise investigations↩︎

2.2.1 Electronics ground noise↩︎

During the commissioning period before O4, spectra of the variation in current flowing from building electronics ground to neutral earth were observed to be correlated with noise in \(h(t)\). A newly developed electronics ground injection system showed that noise in \(h(t)\) could be produced by injecting \(\sim\)​100 mA currents onto the building electronics ground. The coupling is thought to be produced by fluctuations in the potential of the electronics ground system due to the variations in current flows across the finite resistance between electronics ground and true neutral earth, measured to be about 2 \(\Omega\) at LHO [47]. Forces on the charged test mass may fluctuate with the potentials of nearby electronic systems that are referenced to the fluctuating electronics ground, such as the electrostatic drive and ring heaters.

The noise from electronics ground potential fluctuations was reduced in two ways. First, the resistance between certain electronics chassis and the building electronics ground were reduced in order to reduce the total resistance to neutral earth for those electronics. Second, the biases of the electrostatic drives were swept and set to values that minimized the coupling to \(h(t)\) of injections onto the electronics ground. It is thought that, at the coupling minimum, the forces on the charged test mass due to ground potential fluctuations are partially canceled out by an opposite dipole force associated with the bias-polarization of the test mass [47]. These mitigations resulted in range improvements of a few megaparsecs. Further mitigation could be obtained by shielding electronics inside the chamber from the test mass with shields connected to the chamber.

The LIGO test masses are held in place by a quadruple pendulum suspension [48]. The top three stages of the pendulum suspension employ magnetic actuators to hold the suspension fixed, while the test mass itself is held in place by four low-noise ESDs [49]. The test masses located at ends of the interferometer arms require more actuation than the input test masses [48]. The electric force that the ESDs apply to the end test masses depends on the electric potential between the test mass and its surroundings as well as the voltage supplied to the ESD electrodes.

A changing electrical environment local to the end test masses modulates the force applied by the the ESDs to the test masses, introducing noise into the GW data. At LHO it was found that lowering the resistance on the grounding wires for electronics chassis used to control test mass motion made the electronics less sensitive to ground potential fluctuations [50], [51]. Changing the grounding of controls chassis located at the end stations lowered the noise in \(h(t)\) overall and also reduced coherence between test mass motion and current to ground below 100 Hz [47]. Sensitivity to ground potential fluctuations was further reduced by selecting a bias voltage for the DC component of the ESDs which minimized coupling between \(h(t)\) and currents injected onto an electronics chassis at each end station [47]. The results of these two changes is illustrated in . The bias voltage that minimizes currents coupling to \(h(t)\) changes over time and continues to be tracked [52].

A minimum noise setting for Y-end ESD bias was identified at LLO before O4 [53] and midrun changes in the X-arm end station-end ESD bias were found to reduce noise at \(\sim\)​11 Hz and \(\sim\)​60 Hz and harmonics of these frequencies [54].

2.2.2 Broadband transient noise↩︎

We noticed an increased noise at low frequencies which showed non-stationary behavior in the frequency band \(10-50~\mathrm{Hz}\). A bicoherence analysis of \(h(t)\) noise with itself found that this noise was modulated by a low frequency \(h(t)\) signal mostly around 2.6 Hz [55], [56]. Most of the longitudinal drive control sent to the ESD was being sent in the band \(1-3~\mathrm{Hz}\) which could have contributed to this increased noise. A new longitudinal control scheme that reduced the amount of control sent at these low frequencies was developed and implemented . This led to significant improvement in \(h(t)\) noise, thereby reducing the non-stationary or glitch behavior as well [57].

2.2.3 Modifying input power↩︎

The amount of laser power input into the LIGO interferometers has increased in each Observing run. Increased power circulating in the arms improves the high-frequency sensitivity of the LIGO detectors by reducing the effect of quantum shot noise. In O4a, both LIGO detectors were slated to operate with \(75~\mathrm{W}\) of laser power sent into the IMC.

As input laser power was increased at both detectors prior to the fourth Observing run, vibration coupling also increased. This includes vibration coupling through both scattered light noise and input beam jitter noise. The dramatic increases in coupling have prompted a warning that vibration coupling may become increasingly problematic as power increases [37]. A possible explanation is that increased thermal distortion of the test mass surfaces around coating defects may increase scattered light and also reduces the symmetry of the arms, decreasing common mode rejection of input noise.

Due to duty cycle and control scheme concerns associated with high-power operation the laser power sent to the IMC was reduced from \(75~\mathrm{W}\) to \(60~\mathrm{W}\) at LHO during O4a [58]. shows the reduction in vibrational coupling between PEM sensors placed around the in-air optics table where the input laser light is produced and the apparent differential arm length.

2.2.4 Weekly magnetic monitoring↩︎

Regular measurements were taken before and during O4a to understand the potential for coupling between local magnetic fields and the interferometer [53], [59]. Local magnetic fields were generated by running a current through large coils of wire placed near electronics racks used for interferometer controls as well as in the experiment hall. The response of each interferometer to the resulting magnetic fields was quantified using the network of PEM magnetometers set up around each observatory [20], [60]. In order to vet GW candidates at kilohertz frequencies, such as transients from neutron star f-modes [61]–[63], more computing space was allocated prior to O4a to store accelerometer and magnetometer data up to \(8192\) Hz. Nearly every week of O4a, a broadband magnetic field was injected at \(1000-4096\) Hz at \(7\) locations around LHO for \(36\) s at each coil to quantify magnetic coupling in the newly-monitored part of the kilohertz regime. These injections provoked a response in \(h(t)\) at the LIGO LHO corner station. shows \(h(t)\) response to several of these weekly injections compared to a reference background time. Weekly probes of the high frequency magnetic coupling could be used to more accurately estimate the total environmental contribution to a high-frequency GW candidate [64]. Broadband magnetic injections were also made over \(10-100\) Hz and \(100-1000\) Hz from these \(7\) coils as part of the weekly injection campaign.

During the course of O4a, the magnetic coupling of the detector at kilohertz frequencies fluctuated from week to week. At the beginning of O4a, the LHO detector was somewhat sensitive to large external magnetic fields applied at the Corner Station. Towards the middle of O4a, the detector’s response to these applied magnetic fields increased, before dropping to being only weakly coupled towards the end of O4a. The most likely mechanism for the observed magnetic coupling in the kilohertz regime is magnetic interference with cables that control the LIGO suspensions and optics. Specific mid-run electronics configuration changes which affected the degree of magnetic coupling, such as cables being rerouted, have not yet been identified.

2.2.5 Cosmic ray glitch correlations↩︎

A high energy cosmic ray shower is a potential mechanism for creating glitches via momentum transfer to, heating of, or changing the electric potential near the test masses [65], [66]. Cosmic rays are monitored by four photomultiplier tubes placed beneath the vacuum chamber which houses the X-arm input test mass at LHO. During O2 and O3, we compared the time difference between cosmic ray arrival times and blip glitches at LHO and found no evidence of a correlation [10]. We expanded this search in O4a to include blips, low-frequency blips, repeating blips, and tomtes as identified by Gravity Spy [14]. A description of the cosmic ray sensor systematics in O4a can be found at [67]. No temporal correlation was found between cosmic rays and any of these glitch classes. Additionally, we found the amplitude of cosmic rays which struck LHO within a second of a glitch were consistent with the overall amplitude distribution of cosmic rays witnessed by the cosmic ray detectors installed at LHO.

2.2.6 Scattering noise from the input arm↩︎

In O4a, short shutdowns of the LHO building HVAC system produced several percent increases in astrophysical range [68]. Localized vibration injections indicated that the coupling of the HVAC vibrations at the corner station was in the input arm of the interferometer [69]. The coupling was further localized, using laser vibrometry, to particular baffles in the input arm [70]. During the O4a-O4b break, the baffles were moved and damped, and new baffles added, greatly reducing the vibration coupling in the input arm [71], [72].

2.2.7 Comb investigations↩︎

During O4a, two sources of comb artifacts were identified and mitigated at LHO. The first source was in the electronics driving a mirror heating element. This created a comb of approximately 1.6611 Hz (though different mitigation efforts caused changes in the spacing) centered around 280 Hz. A time-correlation was found between changing electronic settings for the mirror heating element and variations in the 1.6611 Hz comb amplitude, which suggested a possible source for the comb [73]. This insight motivated subsequent mitigation efforts that more clearly identified the problem. Electrical connections were changed to stop the 1.6611 Hz comb from being created [74].

The second source created a near-1 Hz comb, as well as a near-5 Hz and near-7 Hz comb at various times. All three combs were traced to Hartmann wavefront wensors (HWS), which are part of the interferometer ASC subsystem [75]. It was determined that the comb frequency spacing changed when the HWS camera shutter frequency setting was changed [76], [77]. The low amplitude of the comb makes observing this artifact in short stretches of data (less than \(\sim\)​1 day) challenging, and thus mitigation efforts more time-consuming. Once the connection between changes in the comb spacing to changes in HWS hardware settings was established, efforts to change the hardware configuration while in observing mode helped to mitigate this comb [78].

2.3 LIGO Livingston noise investigations↩︎

2.3.1 Slow scattering↩︎

Noise due to high ground motion in the band \(0.1\)–\(0.5~\mathrm{Hz}\) was the most dominant source of glitches in the LLO data during O4a. These glitches, also known as Slow Scattering, adversely impacted the strain sensitivity mostly in \(10\)–\(50~\mathrm{Hz}\) band. shows glitch rate and ground motion for three days. For two of these days, ground motion in the band \(0.1\)–\(1~\mathrm{Hz}\) was high, which led to a high rate of Slow Scattering glitches in the data.

The additional phase noise as given by shows up as arches in the time-frequency spectrogram as shown in the middle plots in figure 2. The time separation between subsequent scattering arches gives a direct measure of the frequency with which the scattering surface is moving. During O4a, we have seen that the frequency of the scattering surface is not constant because it moves at whatever frequency is dominant in the ground motion [34]. We have not yet found any optics which have enough velocity to create noise above \(10~\mathrm{Hz}\) in \(h(t)\).

2.3.2 Fast scattering↩︎

During O3, Fast Scattering glitches were the most common glitch source at LLO, making up about 27% of all glitches with a confidence of 90%, according to Gravity Spy [14], [15], [24]. Fast scattering is typically found to be correlated with ground motion in the microseismic band \(0.1\)–\(0.3~\mathrm{Hz}\), and the anthropogenic band \(1\)–\(6~\mathrm{Hz}\). Fast scattering arches are short in duration, shown in Figure 2, and impact the detector sensitivity in the \(10\)–\(100~\mathrm{Hz}\) frequency range. Trains, logging, construction, and other human activity were the main sources of Fast Scattering, as the anthropogenic motion upconverts to higher frequency  [15], [18].

In O3, Trains near the LLO Y-arm end station produced low-frequency seismic noise that would upconvert into the gravitational-wave sensitive frequency band. For this reason, they provided an avenue to study how periods of large ground motion impacted the detector. Spectrograms of the ground motion revealed many harmonic lines with changing frequency, and short bursts of increased amplitude in the strain data. The suspicion was that each burst was produced by the low-frequency ground motion exciting mechanical resonances of some scattering surface. Two methods, Lasso [79] regression and Spearman correlation [80], were employed to identify which narrow band seismic frequencies contributed the most to increased detector noise. Both methods consistently pointed to ground motion in the \(1.8\)–\(2.2~\mathrm{Hz}\) range as the primary factor correlating with heightened strain noise at the corner station [81]. The subsequent mitigation of noise from these frequencies for O4 is discussed in .

From roughly June 2023 through August 2023, there was a significant amount of logging occurring near LLO [82], [83]. The anthropogenic ground motion in the vertical direction near the corner station consistently reached amplitudes greater than 1000 nm/s, as shown in Figure 13. These high ground motion levels caused the detector to lose lock multiple times for many hours during the daytime. After August 2023, the logging activity ceased, significantly reducing the disruptive ground motion near the corner station.

2.3.3 Arm cavity baffle resonances↩︎

ACBs [84] are located at each of the test masses, attached to the first stage of the active seismic isolation system (HEPI [85]), and are used to catch the light from wide angle scattering. ACB resonances are sensitive to the physical state of the system and changes to it can led to the shift in resonant frequencies [18]. After O3, but before O4, at the corner and Y-arm end station, the ACBs had a high-quality-factor resonance at around 1.6 Hz [86]. When rung up, noise appears in the gravitational-wave data from around \(20\)–\(100~\mathrm{Hz}\). In the absence of high microseismic ground motion, 1.6 Hz motion would create scattering noise at 3.2 Hz. After O3 ended, 3.3 Hz scattering noise was observed that had not been seen before; this can be explained by ACB resonances at 1.6 Hz. We suspect that during O3, the ACB resonance was around 2 Hz, which would produce the common 4 Hz Fast Scattering observed. In late 2022, the ACB resonances at the corner and Y-arm end station were mechanically damped. As a result, the rate of Fast Scattering decreased dramatically and subsequently it was found that this noise coupling was no longer present [18]. This effect of this remediation can be seen in the sweep injections performed in July 2022 (see ) and again in February 2023 [87].

Comparisons of the impact of logging during O3 and O4 showed that for similar ground motion amplitudes, the rate of transients was significantly reduced by a factor of about 50 [18]. In O3, anthropogenic ground motion at such high levels would have produced many Omicron glitches (assuming lock is not lost). Due to the damping of the arm cavity baffles, we did not observe significant strain noise due to the logging activities in O4 [89].

2.3.4 Binary neutron star range ocillations↩︎

During O4a, from time to time, the observed BNS range exhibited oscillations with a period of about 30 minutes and a range variation of about 5-15 Mpc, lasting for all or part of a day. These variations can be seen in . The range variations are the result of broadband excess noise in \(h(t)\).

Searches for the cause of these oscillations identified accelerometers that seemed to witness motion that aligned with the oscillations [90]–[92]. There is a line at around 30 Hz produced by the HVAC system, and initial investigations hypothesized that this line changing amplitude could be responsible for the observed BNS range oscillations. To check, a shaker injection at 30.5 Hz was performed on the vacuum enclosure of the chamber containing the X-arm end test mass. The 30.5 Hz shaker injection could not re-create the broadband effect we observe in \(h(t)\) [93].

The hunt for what may be causing these oscillations was continued by analyzing the output of the summary page tool Lasso. As described in [94], Lasso can produce overlays of the BNS range with auxiliary channels, allowing identification of channels with similar periodicity to the range variations. However, range variations often include secular trends combined with the oscillations, such that the channels found by Lasso often fail to align with the period of the oscillations and are thus unlikely to yield information about their cause. The channels that seemed to correlate the most, according to the Lasso algorithm, were primarily located at the X-arm end station and were related to temperature sensors. Additional days with similar leading channels were found – along with channels not explicitly measuring temperature but sensitive to it [95]. Additional investigations have mentioned issues with temperature control [96]. While the evidence seems to implicate temperature effects at the X-arm end station, a deeper study during all of O4a could yield the actual cause, as the oscillations are still present in O4b. The coupling mechanism between the thermal variations and \(h(t)\) is not currently known.

2.3.5 84 Hz \(h(t)\) noise↩︎

During ER15, which immediately preceded O4a, excess noise around 84 Hz in \(h(t)\) data was present. The noise would appear and disappear, suggesting that its cause might be related to dehumidifiers and fans which turn off and on.

This noise was present at both end stations, but was louder at the Y-arm end station. Analysis of spectrograms and outputs of the channels that monitor the dehumidifiers and fans pointed towards the 84 Hz source being related to two exhaust fans located at the Y-arm end station [97]. To further confirm that this excess noise is coming from the Y-arm end station, broadband acoustic injections were done at both the X-arm end station and the Y-arm end station. These injections revealed a sharp mechanical resonance at around 84 Hz at the Y-arm end station [98]. The exact coupling mechanism of the fans into \(h(t)\) is still unknown, but the source was removed by moving the fans off the beam tube enclosure doors [99].

3 Event Validation in O4↩︎

Validation of GW candidates is a crucial step that enhances our confidence in the astrophysical origin of the candidate events and the reliability of source parameter-estimation results. Event Validation refers to the process of checking for the presence of any data-quality issues surrounding the time of an event and conveying this information to the relevant data analysis groups in the collaboration. These assessments build upon the initial vetting conducted by the Rapid Response Team (RRT), a joint LVK working group tasked with promptly responding to event candidate alerts. The RRT conducts a series of prescribed data-quality checks, informed by the DQR ; see , and verifies the overall functioning of the low-latency pipeline infrastructure. This team provides round the clock coverage, comprising rotating on-shift scientists and experts from the various areas of detector operations, including DetChar. The prompt assessment, made after the alert for a significant event is produced by online searches, is then reconsidered more thoroughly and in more detail by the Event Validation task force to determine a final evaluation of the data-quality for every event candidate. Multiple checks are performed on the data, e.g., making sure the detector is operating in a nominal data-taking configuration, identifying any noise artifacts that could bias source property estimation, and checking for any inconsistencies in the output of the PEM sensors as that may suggest noise coupling between the environment and strain data. If the data-quality around the candidate event is found to be unsatisfactory, further data processing techniques such as bayesian noise inference and transient noise removal may be applied [100]. For example, BayesWave data cleaning and linear noise subtraction were applied to a total of 17 events during O3. The catalog papers discuss these events and techniques in more detail [101]–[103].

There have been a number of changes and improvements in the event validation procedure since the last observation run. These changes include:

• An LVK event validation roster for all active detectors: O4a event validation infrastructure is designed to take information from LIGO, Virgo and KAGRA interferometers. This centralization has reduced the person power and time required to perform event validation compared to past Observing runs, during which the LIGO and Virgo Collaborations conducted validation of their detector data separately [12], [104]. Additionally, the unified framework has ensured more standardization and uniformity in the procedures for evaluating data-quality and the tools utilized for the assessment.

• More automated event validation software infrastructure: The event validation infrastructure in O4 is centralized and is maintained using git on the event validation website, accessible to all LVK members. This centralized infrastructure allows easier information flow from DQR, to noise mitigation teams and other data analysis groups. The event validation website acts as a repository of all the details related to event validation including a list of events, the contact information of the volunteers and RRT experts, links to event’s DQR report and the event validation. For each event, a Gitlab issue page is created where any additional details regarding the event’s data-quality can be discussed. This page is also linked from the Event validation website.

• More automated DQR infrastructure: The DQR is a DetChar tool used to assess the data-quality surrounding an event time, as detailed in . The O4 version of the DQR has undergone significant upgrades compared to O3, incorporating automated checks that offer insights into the interferometer’s state and data-quality around the event time. The results of these checks are displayed in the form of labels (Pass, DQ issue, or Task fail), indicating whether a particular data-quality check has been successful or not. The event validation volunteers in O4 have made extensive use of these results for validation purposes.

3.1 Data Quality Report↩︎

As mentioned, the primary tool used in event validation is the DQR [105]. In O3, similar DQR toolkits were separately used by the LSC [12] and Virgo Collaboration [106] to evaluate candidates from the GWTC catalog. For O4, we used the experience gained from O3 to improve the DQR, with a focus on improving the speed and robustness of analyses, increasing the fraction of analyses that were automated, and generalizing the software to support analysis of data from all ground-based gravitational-wave detectors. One key upgrade was the use of a p-value to automatically identify data-quality issues that could impact the detection or analysis of gravitational-wave candidates. Additional details about the DQR architecture used in O4a can be found in [105].

A wide variety of different analyses were used as part of the DQR framework during O4a. Tests that were used to analyze LHO and LLO data include: estimates of noise contributions from the observatory environment [64], statistical correlations between strain data and auxiliary-channel information [28], predictions of the presence of glitches using only auxiliary information [107], analytic identification of excess power in spectrograms of the strain data [108], [109], machine-learning image classification of spectrograms of the candidate [110], quantitative estimates of the data stationarity [111], estimates of the Rayleigh statistic of the data, measurements of the local glitch rate [112], and monitors of the detector range at the time of the candidate. These tasks were completed on two different timescales. Most tasks were completed within 5 minutes of a DQR being launched, allowing these tasks to be used as part of the initial rapid response to identified gravitational-wave signals. Additional tasks were available within a few hours to help with additional offline event validation of each candidate.

A key feature of this updated DQR was the ability to automatically flag DQ issues in the candidate events identified by tasks based on the reported p-value. Candidates with DQ issues reported by the DQR underwent additional scrutiny as part of the rapid response to gravitational-wave candidates in O4a. No additional human follow-up of the candidate data-quality was completed in low latency when no DQ issue was identified. All candidates were further analyzed offline, however, regardless of the conclusion reached in low latency.

We found that the choice of p-value threshold strongly impacted the rate of false alarms from the DQR. The p-value threshold chosen to identify a data-quality issue was changed partway through the run for this reason; at the start of O4a, a threshold of 0.1 was used, but was eventually changed to 0.05. This lower threshold reduced the rate of false alarms with minimal reduction in the true alarm rate. We also changed the set of tasks used throughout the run to improve the true alarm rate and reduce the false alarm rate. Using a single p-value threshold was also suboptimal, as the exact definition of the reported p-value varied between tasks. This led to tasks either overreporting or underreporting the presence of DQ issues. This limitation has been addressed for O4b by introducing task-specific thresholds that are informed by our O4a experience.

3.2 Event Validation procedure↩︎

The Event Validation workflow is shown in fig 17. For each event, the volunteer assigned to the week-long validation shift is immediately notified. They receive all necessary information about the event, as well as instructions on how to validate it. This includes links to the links to the Event Validation form, the DQR and the GraceDB (Gravitational-Wave Candidate Event Database) ² page for the event. In the event validation form, the validator can fill the event details for each detector. These details include the “validation conclusion for the detector”; the options are “Not Observing”, “No Data Quality Issues”, and “Data Quality Issues”. In case the validator finds DQ (Data Quality) issues, they can fill the noise duration and frequency bandwidth in the “Noise Box” for time and frequency. This information is then used as input for glitch subtraction before parameter-estimation analysis, as detailed in the next paragraphs. The validation form also has an option to write any important details about the event in the “Notes” tab. The Gitlab page for each event allows for any further data-quality related discussion. Once the validator is satisfied with their findings, they can submit the validation form. This validation conclusion is then passed to the noise mitigation review team.

The noise mitigation team is responsible for assessing whether any excess power within the target time-frequency analysis window of any candidate is sufficiently non-Gaussian to require further action [111], [113]. We do this by comparing the PSD noise variance in each identified time-frequency region and check it is consistent with Gaussian noise. For regions which are not consistent with Gaussian noise (p \(<\) 0.01) there are two options available. If the noise is sufficiently isolated in time and frequency the noise transient can be subtracted from the data. All noise-subtracted data in O4a were produced by the BayesWave algorithm [100], [114]. The procedure of how this is done is described in the Appendix of Ref [6].

= [rectangle, minimum width=2cm, minimum height=8mm, text centered, text width=3cm, draw=black] = [diamond, minimum width=2cm, minimum height=2cm, text centered, text width=15mm, draw=black] = [thick,->,>=stealth]

To assess the efficacy of the noise subtracted data we compare the Gaussianity of the noise-subtracted data with the targeted time–frequency window to Gaussian noise [108]. Noise-subtracted data consistent with Gaussian noise were deemed sufficiently stationary for parameter estimation. If the noise is extended in time and frequency such that noise subtraction is not appropriate, or the noise subtracted data were not sufficiently stationary, the noise mitigation team recommended restricting the time-frequency analysis window, so the parameter estimation analysis does not take into account any of the noise. The final recommended time-frequency analysis window, along with the recommended data frame name, are then sent through CBCFlow for parameter estimation to automatically retrieve the information and start their analyses. CBCFlow is a Python library that facilitates storage and transfer of event metadata [115].

3.3 Validation of O4a events found by online search pipelines↩︎

In O4a, online searches generated alerts for 92 significant event candidates. Out of these, 11 were retracted by the RRT due to evident contamination from noise artifacts or other issues that resulted in inaccurately estimated event significance, rendering their astrophysical origin improbable [116]–[126]. Some of the events from the remaining 81 candidates required noise mitigation through glitch subtraction. The procedures for assessing the necessity of glitch subtraction and its execution, along with the evaluation of the result, are detailed in [113].

For the remaining events showing data-quality issues but deemed not to require glitch subtraction, restrictions were implemented on the analyzed times and frequency bands surrounding the events. A common problem was low-frequency non-stationary noise, particularly in the lowest part of the detector sensitivity range, between 10 and 40 Hz, often caused by ground motion. Examples of this noise can be seen in the middle and lower panels of Fig. 2. While GW searches and parameter estimation typically start at 20 Hz [6] to avoid the noise wall below 15 Hz (as shown in Fig. 1), when non-stationary noise is present, the lowest frequency may be increased to exclude the affected frequency range and avoid biases in the analysis results.

4 Data Quality for Astrophysical Searches↩︎

DetChar seeks to help mitigate or eliminate identified noise sources as a top priority. Since this is not always feasible and excess non-Gaussian noise remains in archival data even in cases where the issue was corrected, the DetChar group prepares various data-quality products applied to astrophysical searches to reduce the impact of non-Gaussianity of the data on these searches.

4.1 Data quality products for all searches↩︎

As in previous observation runs [12], the LIGO DetChar group recommended that specific periods when the data is unusable due to severe data-quality issues be removed from analyzed data prior to performing astrophysical searches. This is handled through defining segments (time periods specified by start and stop times) to be vetoed by data-quality flags. Category 1 flags define times to be removed prior to running an analysis. It is generally recommended that a consistent set of these flags be applied across all searches, making them relevant for all analyses described in later subsections. Searches for gravitational waves have continued to become more sophisticated in their handling of suboptimal detector data, including utilization of noise subtraction techniques to extract gravitational-wave signals in data with noise transients present. For this reason, LIGO DetChar was less aggressive about recommending periods of data for removal through defining data-quality flags in O4a than in previous runs.

Periods of non-stationarity with significant salvageable data in some frequency bands were left in place, but there were still stretches of data when the detector was nominally supposed to be operating as an astronomical observatory during which the data were in practice un-analyzable due to severe issues sufficient to bias the PSD or the detector status otherwise being inconsistent with the detection of gravitational-waves. The total deadtime percentage, defined as percentage of observation time removed by data-quality flags, was less than one tenth of one percent for each interferometer during O4a. Category 1 flags in this category covered issues including:

• Incorrect line subtraction for the LHO or LLO, generally at the start of lock stretches.

• Parametric instability mode [127] rung up and severely impacting data shortly before causing lockloss. This issue occurred infrequently in both LIGO interferometers in O4a and also occurred in the third Observing run [12].

• A servo causing severe issues with squeezing [8] at LHO.

• Violin modes [128], i.e. resonances of the mirror suspension fibers, rung up severely at LLO, in one case leading directly to distorted strain data followed by lockloss, and in another case causing issues with data calibration.

• Observing mode was defined incorrectly in either LIGO interferometer early in the ER15 Engineering run prior to O4a.

• Observing mode was defined but \(h(t)\) data was not stored permanently due to technical issues.

4.2 Data quality for transient searches↩︎

Searches for transient gravitational-waves cover gravitational-wave emission that will be in the detectable LIGO frequency band for relatively short duration (sub-second to minutes). These searches include matched-filter analyses detecting compact binary coalescences (CBCs) as well as searches for less well-modelled phenomena referred to as GW bursts. In previous runs, both types of transient searches used Category 2 flags, which are typically shorter in duration than Category 1 and targeted the needs of specific analyses. Due to improved confidence in gravitational-wave detection in the presence of noise, CBC searches have eliminated the use of traditional Category 2 flags, although some of these searches use supplementary data-quality products in their place as described below. Unmodelled transient searches, which cannot rely on the characteristic chirp structure of the gravitational-wave signal for confirmation, continue to use these additional data-quality flags, but fewer kinds of flag resulting in reduced deadtime compared to previous runs.

4.2.1 Data quality for compact binary coalescence transient searches↩︎

The primary data-quality products used in CBC searches were the iDQ  [107] timeseries. The iDQ pipeline produces statistical data-quality information based on auxiliary channel activity. In O4a, the Ordered Veto List (OVL)  [129] algorithm was implemented within iDQ to create and rank an ensemble of vetoes for strain data triggered on auxiliary channels. The internal rank of OVL is then calibrated to probabilistic statements on the presence of a glitch by iDQ. To produce timeseries, the generated vetoes are applied to the strain data, and the probability that any time in the strain data contains a glitch is given by the highest ranked veto active at that time. These output timeseries are available for each detector in real-time and contain a number of statistics calculated by OVL for each time sample: the ranking statistic; the false-alarm probability (FAP), which is the probability that a random time with no glitch would be ranked at least as high as the current sample; the natural logarithm of the likelihood that transient noise is present (\(\log L\)); and a state vector for iDQ indicating the quality of the iDQ data.

CBC searches in O4a were performed in two different operating modes, online and offline. In online operations, CBC detection pipelines search for gravitational waves in near real-time with initial low-latency alerts sent to the public for significant detections on a timescale of seconds to minutes. In offline operations, CBC detection pipelines analyze archival data in high latency on a timescale of days to weeks. These offline searches have data from the entire Observing period available to them as well as additional data-quality products such as the Category 1 vetoes described previously. This makes the offline configuration of detection pipelines typically more sensitive than their low-latency counterparts, but comes at an additional computational cost and by definition cannot provide real-time alerts for astronomers.

The iDQ pipeline was also run in online and offline modes. As with CBC searches, the online configuration produced data available in near real-time to detection pipelines and data-quality experts. One detection pipeline, PyCBC Live [130], integrated the iDQ FAP timeseries into their search to reject candidate gravitational-wave signals caused by glitches. This implementation discarded all gravitational-wave candidates with coalescence times within one second of any time satisfying \(\mathrm{FAP}(t) < 10^{-4}\).

Offline iDQ differs from the low-latency version by having access to larger amounts of data when ranking and calibrating vetoes, allowing for more accurate estimation of their statistical properties. The log-likelihood timeseries produced by offline iDQ were used to construct data-quality flags. All of the times satisfying \(\log L(t) \geq 5\) were identified, and segments were constructed covering times from 0.25 s before to 0.25 s after each identified time. Data-quality flags were made by taking the union of all such segments. Different CBC searches may use these flags as they see fit. For example, the flags may be used as vetoes in the style of the Category 2 flags used in previous Observing runs, or they may be incorporated into a ranking statistic as in [131].

4.2.2 Data quality for unmodelled transient searches↩︎

The coherent WaveBurst (cWB) pipeline [132] was the primary online search algorithm for unmodelled transients, or bursts, used in O4a with other algorithms (oLIB [133] and MLy [134]) planned to be used later in O4. cWB has several modes of operation which allow it to be applied to both short duration and long duration searches for GW signals from the entire sky, and searches for signals from binary black holes, Galactic core collapse supernovae or magnetar bursts or flares.

Times of poor data-quality were removed from the burst searches through Category 1 and 2 flags as described above. All searches applied the same Category 1 flags as applied to the CBC searches, while Category 2 flags were developed by determining auxiliary data channels (those that monitor environmental or instrumental changes that are not sensitive to the effects of gravitational-waves) with high correlation to glitches that affect the burst searches. Category 2 flags are applied primarily to the offline burst searches. Similar to issues related to light intensity dips in O3, two Category 2 flags (one for the LHO and another for the LLO) were developed to exclude very loud glitches (usually with SNRs \(>100\)) [12]. These flags had deadtime percentages of 0.122% for LHO and 0.019% for LLO during O4a. There were 60 Hz glitches at LLO witnessed by ESD monitors which were used to develop an effective Category 2 flag specific to this observatory, with deadtime percentage of 0.069% during O4a.

4.3 Data quality for persistent searches↩︎

Persistent gravitational-wave signals are predicted to take a variety of forms. For example, rapidly-rotating non-axisymmetric neutron stars can emit nearly monochromatic gravitational-wave signals [135], a superposition of many gravitational-wave emitters can create a broadband stationary stochastic background [136], conditions in the early Universe may have generated a stochastic background signal [137], etc. The wide variety of possible signal models and our knowledge (or lack thereof) of waveform parameters motivates a similarly wide variety of analysis efforts [135], [138]. Nevertheless, these different search techniques may be served well by a relatively small number of data-quality products. Common data-quality products include information on which frequency bands are contaminated by narrowband noise lines, and cleaned strain data sets in which loud non-Gaussian transient noise has been removed. Different searches take differing approaches towards handling non-stationary noise; CW searches tend to de-weight those time-frequency intervals of higher noise while stochastic searches typically remove those periods from the analysis. Below we describe the data-quality products used for persistent searches in greater detail.

4.3.1 Data quality for continuous wave searches↩︎

4.3.1.1 Self-gated \(h(t)\)

Although noise transients typically do not impact CW analyses, the cumulative effect of frequent and loud noise transients can degrade search sensitivity by effectively increasing the noise background, especially below \(\sim\)​500 Hz [12]. Removing these loud transient artifacts has proven useful in improving the sensitivity of broadband CW searches. In O4, we have employed a more sophisticated algorithm to identify and remove such artifacts, creating a new calibrated \(h(t)\) dataset useful for CW searches [139].

4.3.1.2 Lists of narrow-band instrumental artifacts

Most CW searches depend on a catalog of known instrumental lines to veto spurious candidates or remove contaminated spectral bands from analysis. To produce the catalog, all lines visible in a high-resolution O4a-averaged spectrum (using Hann-windowed, 50% overlapping FFTs of 7200-second-long data segments) are listed and evaluated. Artifacts that are confirmed to be non-astrophysical are added to a curated list and made available for searches to use, as in [140]. A corresponding O4a list is currently being produced. Combs are always considered non-astrophysical because they do not align with a CW signal model. Other lines are considered non-astrophysical when their instrumental/environmental source is known. Artifacts that do not have identified non-astrophysical causes are placed in a separate curated list, as in [141]. Only the confirmed non-astrophysical list is used to veto outliers, whereas the unvetted list is used for investigation purposes.

4.3.2 Data quality for stochastic searches↩︎

4.3.2.1 Non-stationarity cuts in stochastic searches.

Stochastic searches in LVK data are optimal under the assumption that the noise is stationary and Gaussian [136]. This is not the case in general, however, as introduced in and described throughout . To mitigate these effects on stochastic analyses, we split our data into smaller segments, historically 192 s [142]–[144]. The data stationarity is enforced by removing \(h(t)\) segments with a standard deviation that varies by more than a chosen threshold between adjacent segments  [145].

4.3.2.2 Auto-gating and stochastic DQ vetoes.

To mitigate the effect of glitches on frequency-domain non-stationarity cuts, a gating procedure is implemented to pre-process the data. In O4a, gating in stochastic searches is handled by the pygwb workflow [145] as described in [146] by multiplying the data with an inverse Planck-taper window. Periods around samples in the whitened data with an absolute value above a chosen threshold are marked for gating to remove the entirety of the glitch present in the data segment. Occasionally, the gating method implemented in pygwb produces extended gates (\(\geq\)​20 s) that cover periods marked with Category 1 flags, as explained in . This tool helps monitoring the emergence of new Category 1 flagged periods.

In O4a, there are time segments marked with Category 1 flags unique to the stochastic searches collected in a stochastic veto definer file (VF). These involve periods when a violin mode was rung up and interfering with the stationarity of the data and a specific instance when calibration lines were not being properly subtracted. The stochastic isotropic search [146] was run with and without VF vetoed times to verify that the VF correctly excludes data segments that trigger long gates. In , we compare gate distributions between detectors. VF effectively removed long gates in pygwb while maintaining overall bulk of the distribution. Of these gates, over 80% match the minimum gate duration (2 s); others cluster between 2.5-4.5 s.

4.3.2.3 Notch lists.

As described in , we monitor the coherence between \(h(t)\) channels at different sites, and between \(h(t)\) channel at one site and auxiliary channels at the same site. The coherence data between \(h(t)\) channels indicates the frequency bins that pass the coherence threshold as shown in . They are further examined for possible instrumental causes by spectral monitor tools that also keep track of auxiliary channels (Fscan, STAMP-PEM). Frequency bins containing lines known to have an instrumental origin are documented and removed from the analysis.

5 Summary and Future Prospects↩︎

Detector characterization is the study of the detectors and how various noise sources couple and impact GW strain data. In this paper, we summarize the work of the DetChar group between the end of the third Observing run and the end of the first half of the fourth Observing run. These efforts led to a factor of \(\sim\)​50 reduction in the rate of daytime laser light scattering at LLO as detailed in 2.3.3, identification of the high-frequency magnetic noise coupling at LHO, identification of several persistent narrow-band noise features in LLO and LHO, a more comprehensive Event Validation and DQRframework and an overall better understanding of the noise characteristics for astrophysical searches.

As the LIGO detectors become more sensitive, detector characterization will become more challenging. Increased sensitivity translates to higher rate of events but could also lead to an increase in glitch rate. Our work entails not just glitch characterization and reduction, but also validation of the data quality surrounding an event. Lower detector noise across the band also implies higher sensitivity to persistent signals as well as narrow-band, broad-band, and/or correlated terrestrial noise sources. Continuous monitoring of the data quality, identification of potential noise couplings in the detector, and improvement in our software tools are some of the prerequisites for time timely dissemination of robust, and accurate astrophysical results. This would, among other things, require more person power, increased automation of the tools such as acDQR and Event Validation, and a stronger collaboration between the instrument science and DetChar group. These efforts will lead to more robust identification of weak astrophysical GW signals in noisy LIGO data, and thus a deeper probe of the GW sky.

Data-quality products described in this paper from previous Observing Runs have been released publicly on the Gravitational-Wave Open Science Center webpage (GWOSC) ³, and when the O4a data is publicly released, data quality products will be released alongside [147].

References↩︎

---

1. www.gwosc.org↩︎

2. www.gracedb.ligo.org↩︎

3. www.gwosc.org↩︎